2005-04-16 16:20:36 -06:00
|
|
|
/*
|
|
|
|
* Routines having to do with the 'struct sk_buff' memory handlers.
|
|
|
|
*
|
2008-10-13 20:01:08 -06:00
|
|
|
* Authors: Alan Cox <alan@lxorguk.ukuu.org.uk>
|
2005-04-16 16:20:36 -06:00
|
|
|
* Florian La Roche <rzsfl@rz.uni-sb.de>
|
|
|
|
*
|
|
|
|
* Fixes:
|
|
|
|
* Alan Cox : Fixed the worst of the load
|
|
|
|
* balancer bugs.
|
|
|
|
* Dave Platt : Interrupt stacking fix.
|
|
|
|
* Richard Kooijman : Timestamp fixes.
|
|
|
|
* Alan Cox : Changed buffer format.
|
|
|
|
* Alan Cox : destructor hook for AF_UNIX etc.
|
|
|
|
* Linus Torvalds : Better skb_clone.
|
|
|
|
* Alan Cox : Added skb_copy.
|
|
|
|
* Alan Cox : Added all the changed routines Linus
|
|
|
|
* only put in the headers
|
|
|
|
* Ray VanTassle : Fixed --skb->lock in free
|
|
|
|
* Alan Cox : skb_copy copy arp field
|
|
|
|
* Andi Kleen : slabified it.
|
|
|
|
* Robert Olsson : Removed skb_head_pool
|
|
|
|
*
|
|
|
|
* NOTE:
|
|
|
|
* The __skb_ routines should be called with interrupts
|
|
|
|
* disabled, or you better be *real* sure that the operation is atomic
|
|
|
|
* with respect to whatever list is being frobbed (e.g. via lock_sock()
|
|
|
|
* or via disabling bottom half handlers, etc).
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public License
|
|
|
|
* as published by the Free Software Foundation; either version
|
|
|
|
* 2 of the License, or (at your option) any later version.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The functions in this file will not compile correctly with gcc 2.4.x
|
|
|
|
*/
|
|
|
|
|
2012-05-16 13:58:40 -06:00
|
|
|
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
#include <linux/module.h>
|
|
|
|
#include <linux/types.h>
|
|
|
|
#include <linux/kernel.h>
|
2008-08-30 04:16:35 -06:00
|
|
|
#include <linux/kmemcheck.h>
|
2005-04-16 16:20:36 -06:00
|
|
|
#include <linux/mm.h>
|
|
|
|
#include <linux/interrupt.h>
|
|
|
|
#include <linux/in.h>
|
|
|
|
#include <linux/inet.h>
|
|
|
|
#include <linux/slab.h>
|
2014-01-26 02:58:16 -07:00
|
|
|
#include <linux/tcp.h>
|
|
|
|
#include <linux/udp.h>
|
2016-06-02 12:05:43 -06:00
|
|
|
#include <linux/sctp.h>
|
2005-04-16 16:20:36 -06:00
|
|
|
#include <linux/netdevice.h>
|
|
|
|
#ifdef CONFIG_NET_CLS_ACT
|
|
|
|
#include <net/pkt_sched.h>
|
|
|
|
#endif
|
|
|
|
#include <linux/string.h>
|
|
|
|
#include <linux/skbuff.h>
|
2007-11-07 00:30:13 -07:00
|
|
|
#include <linux/splice.h>
|
2005-04-16 16:20:36 -06:00
|
|
|
#include <linux/cache.h>
|
|
|
|
#include <linux/rtnetlink.h>
|
|
|
|
#include <linux/init.h>
|
2007-04-02 21:19:53 -06:00
|
|
|
#include <linux/scatterlist.h>
|
2009-02-11 22:03:37 -07:00
|
|
|
#include <linux/errqueue.h>
|
2011-05-20 13:50:29 -06:00
|
|
|
#include <linux/prefetch.h>
|
net: Always untag vlan-tagged traffic on input.
Currently the functionality to untag traffic on input resides
as part of the vlan module and is build only when VLAN support
is enabled in the kernel. When VLAN is disabled, the function
vlan_untag() turns into a stub and doesn't really untag the
packets. This seems to create an interesting interaction
between VMs supporting checksum offloading and some network drivers.
There are some drivers that do not allow the user to change
tx-vlan-offload feature of the driver. These drivers also seem
to assume that any VLAN-tagged traffic they transmit will
have the vlan information in the vlan_tci and not in the vlan
header already in the skb. When transmitting skbs that already
have tagged data with partial checksum set, the checksum doesn't
appear to be updated correctly by the card thus resulting in a
failure to establish TCP connections.
The following is a packet trace taken on the receiver where a
sender is a VM with a VLAN configued. The host VM is running on
doest not have VLAN support and the outging interface on the
host is tg3:
10:12:43.503055 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27243,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x48d9), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294837885 ecr 0,nop,wscale 7], length 0
10:12:44.505556 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27244,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x44ee), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294838888 ecr 0,nop,wscale 7], length 0
This connection finally times out.
I've only access to the TG3 hardware in this configuration thus have
only tested this with TG3 driver. There are a lot of other drivers
that do not permit user changes to vlan acceleration features, and
I don't know if they all suffere from a similar issue.
The patch attempt to fix this another way. It moves the vlan header
stipping code out of the vlan module and always builds it into the
kernel network core. This way, even if vlan is not supported on
a virtualizatoin host, the virtual machines running on top of such
host will still work with VLANs enabled.
CC: Patrick McHardy <kaber@trash.net>
CC: Nithin Nayak Sujir <nsujir@broadcom.com>
CC: Michael Chan <mchan@broadcom.com>
CC: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Acked-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-08 12:42:13 -06:00
|
|
|
#include <linux/if_vlan.h>
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
#include <net/protocol.h>
|
|
|
|
#include <net/dst.h>
|
|
|
|
#include <net/sock.h>
|
|
|
|
#include <net/checksum.h>
|
2014-01-09 03:02:46 -07:00
|
|
|
#include <net/ip6_checksum.h>
|
2005-04-16 16:20:36 -06:00
|
|
|
#include <net/xfrm.h>
|
|
|
|
|
2016-12-24 12:46:01 -07:00
|
|
|
#include <linux/uaccess.h>
|
2009-04-14 17:39:12 -06:00
|
|
|
#include <trace/events/skb.h>
|
2012-04-05 03:35:15 -06:00
|
|
|
#include <linux/highmem.h>
|
2015-01-30 11:29:32 -07:00
|
|
|
#include <linux/capability.h>
|
|
|
|
#include <linux/user_namespace.h>
|
2006-10-19 14:08:53 -06:00
|
|
|
|
2012-04-30 02:10:34 -06:00
|
|
|
struct kmem_cache *skbuff_head_cache __read_mostly;
|
2006-12-06 21:33:20 -07:00
|
|
|
static struct kmem_cache *skbuff_fclone_cache __read_mostly;
|
2016-02-03 01:26:57 -07:00
|
|
|
int sysctl_max_skb_frags __read_mostly = MAX_SKB_FRAGS;
|
|
|
|
EXPORT_SYMBOL(sysctl_max_skb_frags);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
2013-02-11 06:30:38 -07:00
|
|
|
* skb_panic - private function for out-of-line support
|
|
|
|
* @skb: buffer
|
|
|
|
* @sz: size
|
|
|
|
* @addr: address
|
2013-02-13 04:20:27 -07:00
|
|
|
* @msg: skb_over_panic or skb_under_panic
|
2005-04-16 16:20:36 -06:00
|
|
|
*
|
2013-02-11 06:30:38 -07:00
|
|
|
* Out-of-line support for skb_put() and skb_push().
|
|
|
|
* Called via the wrapper skb_over_panic() or skb_under_panic().
|
|
|
|
* Keep out of line to prevent kernel bloat.
|
|
|
|
* __builtin_return_address is not used because it is not always reliable.
|
2005-04-16 16:20:36 -06:00
|
|
|
*/
|
2013-02-11 06:30:38 -07:00
|
|
|
static void skb_panic(struct sk_buff *skb, unsigned int sz, void *addr,
|
2013-02-13 04:20:27 -07:00
|
|
|
const char msg[])
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2012-05-16 13:58:40 -06:00
|
|
|
pr_emerg("%s: text:%p len:%d put:%d head:%p data:%p tail:%#lx end:%#lx dev:%s\n",
|
2013-02-13 04:20:27 -07:00
|
|
|
msg, addr, skb->len, sz, skb->head, skb->data,
|
2012-05-16 13:58:40 -06:00
|
|
|
(unsigned long)skb->tail, (unsigned long)skb->end,
|
|
|
|
skb->dev ? skb->dev->name : "<NULL>");
|
2005-04-16 16:20:36 -06:00
|
|
|
BUG();
|
|
|
|
}
|
|
|
|
|
2013-02-11 06:30:38 -07:00
|
|
|
static void skb_over_panic(struct sk_buff *skb, unsigned int sz, void *addr)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2013-02-11 06:30:38 -07:00
|
|
|
skb_panic(skb, sz, addr, __func__);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2013-02-11 06:30:38 -07:00
|
|
|
static void skb_under_panic(struct sk_buff *skb, unsigned int sz, void *addr)
|
|
|
|
{
|
|
|
|
skb_panic(skb, sz, addr, __func__);
|
|
|
|
}
|
2012-07-31 17:44:19 -06:00
|
|
|
|
|
|
|
/*
|
|
|
|
* kmalloc_reserve is a wrapper around kmalloc_node_track_caller that tells
|
|
|
|
* the caller if emergency pfmemalloc reserves are being used. If it is and
|
|
|
|
* the socket is later found to be SOCK_MEMALLOC then PFMEMALLOC reserves
|
|
|
|
* may be used. Otherwise, the packet data may be discarded until enough
|
|
|
|
* memory is free
|
|
|
|
*/
|
|
|
|
#define kmalloc_reserve(size, gfp, node, pfmemalloc) \
|
|
|
|
__kmalloc_reserve(size, gfp, node, _RET_IP_, pfmemalloc)
|
2012-12-28 11:24:28 -07:00
|
|
|
|
|
|
|
static void *__kmalloc_reserve(size_t size, gfp_t flags, int node,
|
|
|
|
unsigned long ip, bool *pfmemalloc)
|
2012-07-31 17:44:19 -06:00
|
|
|
{
|
|
|
|
void *obj;
|
|
|
|
bool ret_pfmemalloc = false;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Try a regular allocation, when that fails and we're not entitled
|
|
|
|
* to the reserves, fail.
|
|
|
|
*/
|
|
|
|
obj = kmalloc_node_track_caller(size,
|
|
|
|
flags | __GFP_NOMEMALLOC | __GFP_NOWARN,
|
|
|
|
node);
|
|
|
|
if (obj || !(gfp_pfmemalloc_allowed(flags)))
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
/* Try again but now we are using pfmemalloc reserves */
|
|
|
|
ret_pfmemalloc = true;
|
|
|
|
obj = kmalloc_node_track_caller(size, flags, node);
|
|
|
|
|
|
|
|
out:
|
|
|
|
if (pfmemalloc)
|
|
|
|
*pfmemalloc = ret_pfmemalloc;
|
|
|
|
|
|
|
|
return obj;
|
|
|
|
}
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
/* Allocate a new skbuff. We do this ourselves so we can fill in a few
|
|
|
|
* 'private' fields and also do memory statistics to find all the
|
|
|
|
* [BEEP] leaks.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
2013-04-17 00:46:58 -06:00
|
|
|
struct sk_buff *__alloc_skb_head(gfp_t gfp_mask, int node)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb;
|
|
|
|
|
|
|
|
/* Get the HEAD */
|
|
|
|
skb = kmem_cache_alloc_node(skbuff_head_cache,
|
|
|
|
gfp_mask & ~__GFP_DMA, node);
|
|
|
|
if (!skb)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Only clear those fields we need to clear, not those that we will
|
|
|
|
* actually initialise below. Hence, don't put any more fields after
|
|
|
|
* the tail pointer in struct sk_buff!
|
|
|
|
*/
|
|
|
|
memset(skb, 0, offsetof(struct sk_buff, tail));
|
2013-06-03 03:28:43 -06:00
|
|
|
skb->head = NULL;
|
2013-04-17 00:46:58 -06:00
|
|
|
skb->truesize = sizeof(struct sk_buff);
|
|
|
|
atomic_set(&skb->users, 1);
|
|
|
|
|
2013-05-29 01:16:05 -06:00
|
|
|
skb->mac_header = (typeof(skb->mac_header))~0U;
|
2013-04-17 00:46:58 -06:00
|
|
|
out:
|
|
|
|
return skb;
|
|
|
|
}
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
/**
|
2005-08-17 15:57:30 -06:00
|
|
|
* __alloc_skb - allocate a network buffer
|
2005-04-16 16:20:36 -06:00
|
|
|
* @size: size to allocate
|
|
|
|
* @gfp_mask: allocation mask
|
2012-07-31 17:44:19 -06:00
|
|
|
* @flags: If SKB_ALLOC_FCLONE is set, allocate from fclone cache
|
|
|
|
* instead of head cache and allocate a cloned (child) skb.
|
|
|
|
* If SKB_ALLOC_RX is set, __GFP_MEMALLOC will be used for
|
|
|
|
* allocations in case the data is required for writeback
|
2006-12-06 21:32:36 -07:00
|
|
|
* @node: numa node to allocate memory on
|
2005-04-16 16:20:36 -06:00
|
|
|
*
|
|
|
|
* Allocate a new &sk_buff. The returned buffer has no headroom and a
|
2012-06-06 09:23:37 -06:00
|
|
|
* tail room of at least size bytes. The object has a reference count
|
|
|
|
* of one. The return is the buffer. On a failure the return is %NULL.
|
2005-04-16 16:20:36 -06:00
|
|
|
*
|
|
|
|
* Buffers may only be allocated from interrupts using a @gfp_mask of
|
|
|
|
* %GFP_ATOMIC.
|
|
|
|
*/
|
2005-10-07 00:46:04 -06:00
|
|
|
struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
|
2012-07-31 17:44:19 -06:00
|
|
|
int flags, int node)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2006-12-06 21:33:20 -07:00
|
|
|
struct kmem_cache *cache;
|
2006-01-03 15:06:50 -07:00
|
|
|
struct skb_shared_info *shinfo;
|
2005-04-16 16:20:36 -06:00
|
|
|
struct sk_buff *skb;
|
|
|
|
u8 *data;
|
2012-07-31 17:44:19 -06:00
|
|
|
bool pfmemalloc;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2012-07-31 17:44:19 -06:00
|
|
|
cache = (flags & SKB_ALLOC_FCLONE)
|
|
|
|
? skbuff_fclone_cache : skbuff_head_cache;
|
|
|
|
|
|
|
|
if (sk_memalloc_socks() && (flags & SKB_ALLOC_RX))
|
|
|
|
gfp_mask |= __GFP_MEMALLOC;
|
2006-01-23 17:32:45 -07:00
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
/* Get the HEAD */
|
2006-12-06 21:32:36 -07:00
|
|
|
skb = kmem_cache_alloc_node(cache, gfp_mask & ~__GFP_DMA, node);
|
2005-04-16 16:20:36 -06:00
|
|
|
if (!skb)
|
|
|
|
goto out;
|
2010-05-05 02:07:37 -06:00
|
|
|
prefetchw(skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2011-10-13 01:28:54 -06:00
|
|
|
/* We do our best to align skb_shared_info on a separate cache
|
|
|
|
* line. It usually works because kmalloc(X > SMP_CACHE_BYTES) gives
|
|
|
|
* aligned memory blocks, unless SLUB/SLAB debug is enabled.
|
|
|
|
* Both skb->head and skb_shared_info are cache line aligned.
|
|
|
|
*/
|
2011-11-02 07:40:28 -06:00
|
|
|
size = SKB_DATA_ALIGN(size);
|
2011-10-13 01:28:54 -06:00
|
|
|
size += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
|
2012-07-31 17:44:19 -06:00
|
|
|
data = kmalloc_reserve(size, gfp_mask, node, &pfmemalloc);
|
2005-04-16 16:20:36 -06:00
|
|
|
if (!data)
|
|
|
|
goto nodata;
|
2011-10-13 01:28:54 -06:00
|
|
|
/* kmalloc(size) might give us more room than requested.
|
|
|
|
* Put skb_shared_info exactly at the end of allocated zone,
|
|
|
|
* to allow max possible filling before reallocation.
|
|
|
|
*/
|
|
|
|
size = SKB_WITH_OVERHEAD(ksize(data));
|
2010-05-05 02:07:37 -06:00
|
|
|
prefetchw(data + size);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2007-03-19 07:48:59 -06:00
|
|
|
/*
|
2008-05-03 21:56:42 -06:00
|
|
|
* Only clear those fields we need to clear, not those that we will
|
|
|
|
* actually initialise below. Hence, don't put any more fields after
|
|
|
|
* the tail pointer in struct sk_buff!
|
2007-03-19 07:48:59 -06:00
|
|
|
*/
|
|
|
|
memset(skb, 0, offsetof(struct sk_buff, tail));
|
2011-10-13 01:28:54 -06:00
|
|
|
/* Account for allocated memory : skb + skb->head */
|
|
|
|
skb->truesize = SKB_TRUESIZE(size);
|
2012-07-31 17:44:19 -06:00
|
|
|
skb->pfmemalloc = pfmemalloc;
|
2005-04-16 16:20:36 -06:00
|
|
|
atomic_set(&skb->users, 1);
|
|
|
|
skb->head = data;
|
|
|
|
skb->data = data;
|
2007-04-19 21:29:13 -06:00
|
|
|
skb_reset_tail_pointer(skb);
|
2007-04-19 21:43:29 -06:00
|
|
|
skb->end = skb->tail + size;
|
2013-05-29 01:16:05 -06:00
|
|
|
skb->mac_header = (typeof(skb->mac_header))~0U;
|
|
|
|
skb->transport_header = (typeof(skb->transport_header))~0U;
|
2009-06-16 23:23:27 -06:00
|
|
|
|
2006-01-03 15:06:50 -07:00
|
|
|
/* make sure we initialize shinfo sequentially */
|
|
|
|
shinfo = skb_shinfo(skb);
|
2010-05-05 02:07:37 -06:00
|
|
|
memset(shinfo, 0, offsetof(struct skb_shared_info, dataref));
|
2006-01-03 15:06:50 -07:00
|
|
|
atomic_set(&shinfo->dataref, 1);
|
2011-01-25 16:18:38 -07:00
|
|
|
kmemcheck_annotate_variable(shinfo->destructor_arg);
|
2006-01-03 15:06:50 -07:00
|
|
|
|
2012-07-31 17:44:19 -06:00
|
|
|
if (flags & SKB_ALLOC_FCLONE) {
|
2014-09-29 14:29:15 -06:00
|
|
|
struct sk_buff_fclones *fclones;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2014-09-29 14:29:15 -06:00
|
|
|
fclones = container_of(skb, struct sk_buff_fclones, skb1);
|
|
|
|
|
|
|
|
kmemcheck_annotate_bitfield(&fclones->skb2, flags1);
|
2005-08-17 15:57:30 -06:00
|
|
|
skb->fclone = SKB_FCLONE_ORIG;
|
2014-09-29 14:29:15 -06:00
|
|
|
atomic_set(&fclones->fclone_ref, 1);
|
2005-08-17 15:57:30 -06:00
|
|
|
|
2014-12-03 18:04:39 -07:00
|
|
|
fclones->skb2.fclone = SKB_FCLONE_CLONE;
|
2005-08-17 15:57:30 -06:00
|
|
|
}
|
2005-04-16 16:20:36 -06:00
|
|
|
out:
|
|
|
|
return skb;
|
|
|
|
nodata:
|
2006-01-23 17:32:45 -07:00
|
|
|
kmem_cache_free(cache, skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
skb = NULL;
|
|
|
|
goto out;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(__alloc_skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
/**
|
2015-04-24 17:05:01 -06:00
|
|
|
* __build_skb - build a network buffer
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
* @data: data buffer provided by caller
|
2015-04-24 17:05:01 -06:00
|
|
|
* @frag_size: size of data, or 0 if head was kmalloced
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
*
|
|
|
|
* Allocate a new &sk_buff. Caller provides space holding head and
|
2013-07-23 13:22:39 -06:00
|
|
|
* skb_shared_info. @data must have been allocated by kmalloc() only if
|
2015-04-24 17:05:01 -06:00
|
|
|
* @frag_size is 0, otherwise data should come from the page allocator
|
|
|
|
* or vmalloc()
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
* The return is the new skb buffer.
|
|
|
|
* On a failure the return is %NULL, and @data is not freed.
|
|
|
|
* Notes :
|
|
|
|
* Before IO, driver allocates only data buffer where NIC put incoming frame
|
|
|
|
* Driver should add room at head (NET_SKB_PAD) and
|
|
|
|
* MUST add room at tail (SKB_DATA_ALIGN(skb_shared_info))
|
|
|
|
* After IO, driver calls build_skb(), to allocate sk_buff and populate it
|
|
|
|
* before giving packet to stack.
|
|
|
|
* RX rings only contains data buffers, not full skbs.
|
|
|
|
*/
|
2015-04-24 17:05:01 -06:00
|
|
|
struct sk_buff *__build_skb(void *data, unsigned int frag_size)
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
{
|
|
|
|
struct skb_shared_info *shinfo;
|
|
|
|
struct sk_buff *skb;
|
net: allow skb->head to be a page fragment
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-26 18:33:38 -06:00
|
|
|
unsigned int size = frag_size ? : ksize(data);
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
|
|
|
|
skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
|
|
|
|
if (!skb)
|
|
|
|
return NULL;
|
|
|
|
|
net: allow skb->head to be a page fragment
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-26 18:33:38 -06:00
|
|
|
size -= SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
|
|
|
|
memset(skb, 0, offsetof(struct sk_buff, tail));
|
|
|
|
skb->truesize = SKB_TRUESIZE(size);
|
|
|
|
atomic_set(&skb->users, 1);
|
|
|
|
skb->head = data;
|
|
|
|
skb->data = data;
|
|
|
|
skb_reset_tail_pointer(skb);
|
|
|
|
skb->end = skb->tail + size;
|
2013-05-29 01:16:05 -06:00
|
|
|
skb->mac_header = (typeof(skb->mac_header))~0U;
|
|
|
|
skb->transport_header = (typeof(skb->transport_header))~0U;
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
|
|
|
|
/* make sure we initialize shinfo sequentially */
|
|
|
|
shinfo = skb_shinfo(skb);
|
|
|
|
memset(shinfo, 0, offsetof(struct skb_shared_info, dataref));
|
|
|
|
atomic_set(&shinfo->dataref, 1);
|
|
|
|
kmemcheck_annotate_variable(shinfo->destructor_arg);
|
|
|
|
|
|
|
|
return skb;
|
|
|
|
}
|
2015-04-24 17:05:01 -06:00
|
|
|
|
|
|
|
/* build_skb() is wrapper over __build_skb(), that specifically
|
|
|
|
* takes care of skb->head and skb->pfmemalloc
|
|
|
|
* This means that if @frag_size is not zero, then @data must be backed
|
|
|
|
* by a page fragment, not kmalloc() or vmalloc()
|
|
|
|
*/
|
|
|
|
struct sk_buff *build_skb(void *data, unsigned int frag_size)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb = __build_skb(data, frag_size);
|
|
|
|
|
|
|
|
if (skb && frag_size) {
|
|
|
|
skb->head_frag = 1;
|
2015-08-21 15:11:51 -06:00
|
|
|
if (page_is_pfmemalloc(virt_to_head_page(data)))
|
2015-04-24 17:05:01 -06:00
|
|
|
skb->pfmemalloc = 1;
|
|
|
|
}
|
|
|
|
return skb;
|
|
|
|
}
|
net: introduce build_skb()
One of the thing we discussed during netdev 2011 conference was the idea
to change some network drivers to allocate/populate their skb at RX
completion time, right before feeding the skb to network stack.
In old days, we allocated skbs when populating the RX ring.
This means bringing into cpu cache sk_buff and skb_shared_info cache
lines (since we clear/initialize them), then 'queue' skb->data to NIC.
By the time NIC fills a frame in skb->data buffer and host can process
it, cpu probably threw away the cache lines from its caches, because lot
of things happened between the allocation and final use.
So the deal would be to allocate only the data buffer for the NIC to
populate its RX ring buffer. And use build_skb() at RX completion to
attach a data buffer (now filled with an ethernet frame) to a new skb,
initialize the skb_shared_info portion, and give the hot skb to network
stack.
build_skb() is the function to allocate an skb, caller providing the
data buffer that should be attached to it. Drivers are expected to call
skb_reserve() right after build_skb() to adjust skb->data to the
Ethernet frame (usually skipping NET_SKB_PAD and NET_IP_ALIGN, but some
drivers might add a hardware provided alignment)
Data provided to build_skb() MUST have been allocated by a prior
kmalloc() call, with enough room to add SKB_DATA_ALIGN(sizeof(struct
skb_shared_info)) bytes at the end of the data without corrupting
incoming frame.
data = kmalloc(NET_SKB_PAD + NET_IP_ALIGN + 1536 +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
GFP_ATOMIC);
...
skb = build_skb(data);
if (!skb) {
recycle_data(data);
} else {
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
...
}
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-13 23:03:34 -07:00
|
|
|
EXPORT_SYMBOL(build_skb);
|
|
|
|
|
2016-02-08 05:14:59 -07:00
|
|
|
#define NAPI_SKB_CACHE_SIZE 64
|
|
|
|
|
|
|
|
struct napi_alloc_cache {
|
|
|
|
struct page_frag_cache page;
|
2016-11-18 17:47:56 -07:00
|
|
|
unsigned int skb_count;
|
2016-02-08 05:14:59 -07:00
|
|
|
void *skb_cache[NAPI_SKB_CACHE_SIZE];
|
|
|
|
};
|
|
|
|
|
2015-05-06 22:11:57 -06:00
|
|
|
static DEFINE_PER_CPU(struct page_frag_cache, netdev_alloc_cache);
|
2016-02-08 05:14:59 -07:00
|
|
|
static DEFINE_PER_CPU(struct napi_alloc_cache, napi_alloc_cache);
|
2014-12-09 20:40:42 -07:00
|
|
|
|
|
|
|
static void *__netdev_alloc_frag(unsigned int fragsz, gfp_t gfp_mask)
|
|
|
|
{
|
2015-05-06 22:11:57 -06:00
|
|
|
struct page_frag_cache *nc;
|
2014-12-09 20:40:42 -07:00
|
|
|
unsigned long flags;
|
|
|
|
void *data;
|
|
|
|
|
|
|
|
local_irq_save(flags);
|
2015-05-06 22:11:40 -06:00
|
|
|
nc = this_cpu_ptr(&netdev_alloc_cache);
|
2017-01-10 17:58:06 -07:00
|
|
|
data = page_frag_alloc(nc, fragsz, gfp_mask);
|
2012-05-17 23:12:12 -06:00
|
|
|
local_irq_restore(flags);
|
|
|
|
return data;
|
|
|
|
}
|
2012-07-31 17:44:19 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* netdev_alloc_frag - allocate a page fragment
|
|
|
|
* @fragsz: fragment size
|
|
|
|
*
|
|
|
|
* Allocates a frag from a page for receive buffer.
|
|
|
|
* Uses GFP_ATOMIC allocations.
|
|
|
|
*/
|
|
|
|
void *netdev_alloc_frag(unsigned int fragsz)
|
|
|
|
{
|
|
|
|
return __netdev_alloc_frag(fragsz, GFP_ATOMIC | __GFP_COLD);
|
|
|
|
}
|
2012-05-17 23:12:12 -06:00
|
|
|
EXPORT_SYMBOL(netdev_alloc_frag);
|
|
|
|
|
2014-12-09 20:40:42 -07:00
|
|
|
static void *__napi_alloc_frag(unsigned int fragsz, gfp_t gfp_mask)
|
|
|
|
{
|
2016-02-08 05:14:59 -07:00
|
|
|
struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
|
2015-05-06 22:11:40 -06:00
|
|
|
|
2017-01-10 17:58:06 -07:00
|
|
|
return page_frag_alloc(&nc->page, fragsz, gfp_mask);
|
2014-12-09 20:40:42 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
void *napi_alloc_frag(unsigned int fragsz)
|
|
|
|
{
|
|
|
|
return __napi_alloc_frag(fragsz, GFP_ATOMIC | __GFP_COLD);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(napi_alloc_frag);
|
|
|
|
|
2014-12-09 20:40:49 -07:00
|
|
|
/**
|
|
|
|
* __netdev_alloc_skb - allocate an skbuff for rx on a specific device
|
|
|
|
* @dev: network device to receive on
|
2015-08-24 07:56:54 -06:00
|
|
|
* @len: length to allocate
|
2014-12-09 20:40:49 -07:00
|
|
|
* @gfp_mask: get_free_pages mask, passed to alloc_skb
|
|
|
|
*
|
|
|
|
* Allocate a new &sk_buff and assign it a usage count of one. The
|
|
|
|
* buffer has NET_SKB_PAD headroom built in. Users should allocate
|
|
|
|
* the headroom they think they need without accounting for the
|
|
|
|
* built in space. The built in space is used for optimisations.
|
|
|
|
*
|
|
|
|
* %NULL is returned if there is no free memory.
|
|
|
|
*/
|
2015-05-06 22:11:40 -06:00
|
|
|
struct sk_buff *__netdev_alloc_skb(struct net_device *dev, unsigned int len,
|
|
|
|
gfp_t gfp_mask)
|
2014-12-09 20:40:49 -07:00
|
|
|
{
|
2015-05-06 22:11:57 -06:00
|
|
|
struct page_frag_cache *nc;
|
2015-05-06 22:11:40 -06:00
|
|
|
unsigned long flags;
|
2014-12-09 20:40:49 -07:00
|
|
|
struct sk_buff *skb;
|
2015-05-06 22:11:40 -06:00
|
|
|
bool pfmemalloc;
|
|
|
|
void *data;
|
|
|
|
|
|
|
|
len += NET_SKB_PAD;
|
2014-12-09 20:40:49 -07:00
|
|
|
|
2015-05-06 22:11:40 -06:00
|
|
|
if ((len > SKB_WITH_OVERHEAD(PAGE_SIZE)) ||
|
2015-11-06 17:28:21 -07:00
|
|
|
(gfp_mask & (__GFP_DIRECT_RECLAIM | GFP_DMA))) {
|
2015-05-13 14:34:13 -06:00
|
|
|
skb = __alloc_skb(len, gfp_mask, SKB_ALLOC_RX, NUMA_NO_NODE);
|
|
|
|
if (!skb)
|
|
|
|
goto skb_fail;
|
|
|
|
goto skb_success;
|
|
|
|
}
|
2014-12-09 20:40:49 -07:00
|
|
|
|
2015-05-06 22:11:40 -06:00
|
|
|
len += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
|
|
|
|
len = SKB_DATA_ALIGN(len);
|
|
|
|
|
|
|
|
if (sk_memalloc_socks())
|
|
|
|
gfp_mask |= __GFP_MEMALLOC;
|
|
|
|
|
|
|
|
local_irq_save(flags);
|
|
|
|
|
|
|
|
nc = this_cpu_ptr(&netdev_alloc_cache);
|
2017-01-10 17:58:06 -07:00
|
|
|
data = page_frag_alloc(nc, len, gfp_mask);
|
2015-05-06 22:11:40 -06:00
|
|
|
pfmemalloc = nc->pfmemalloc;
|
|
|
|
|
|
|
|
local_irq_restore(flags);
|
|
|
|
|
|
|
|
if (unlikely(!data))
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
skb = __build_skb(data, len);
|
|
|
|
if (unlikely(!skb)) {
|
2015-05-06 22:12:03 -06:00
|
|
|
skb_free_frag(data);
|
2015-05-06 22:11:40 -06:00
|
|
|
return NULL;
|
2006-08-07 17:09:04 -06:00
|
|
|
}
|
2014-12-09 20:40:49 -07:00
|
|
|
|
2015-05-06 22:11:40 -06:00
|
|
|
/* use OR instead of assignment to avoid clearing of bits in mask */
|
|
|
|
if (pfmemalloc)
|
|
|
|
skb->pfmemalloc = 1;
|
|
|
|
skb->head_frag = 1;
|
|
|
|
|
2015-05-13 14:34:13 -06:00
|
|
|
skb_success:
|
2015-05-06 22:11:40 -06:00
|
|
|
skb_reserve(skb, NET_SKB_PAD);
|
|
|
|
skb->dev = dev;
|
|
|
|
|
2015-05-13 14:34:13 -06:00
|
|
|
skb_fail:
|
2006-07-31 23:35:23 -06:00
|
|
|
return skb;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(__netdev_alloc_skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2014-12-09 20:40:49 -07:00
|
|
|
/**
|
|
|
|
* __napi_alloc_skb - allocate skbuff for rx in a specific NAPI instance
|
|
|
|
* @napi: napi instance this buffer was allocated for
|
2015-08-24 07:56:54 -06:00
|
|
|
* @len: length to allocate
|
2014-12-09 20:40:49 -07:00
|
|
|
* @gfp_mask: get_free_pages mask, passed to alloc_skb and alloc_pages
|
|
|
|
*
|
|
|
|
* Allocate a new sk_buff for use in NAPI receive. This buffer will
|
|
|
|
* attempt to allocate the head from a special reserved region used
|
|
|
|
* only for NAPI Rx allocation. By doing this we can save several
|
|
|
|
* CPU cycles by avoiding having to disable and re-enable IRQs.
|
|
|
|
*
|
|
|
|
* %NULL is returned if there is no free memory.
|
|
|
|
*/
|
2015-05-06 22:11:40 -06:00
|
|
|
struct sk_buff *__napi_alloc_skb(struct napi_struct *napi, unsigned int len,
|
|
|
|
gfp_t gfp_mask)
|
2014-12-09 20:40:49 -07:00
|
|
|
{
|
2016-02-08 05:14:59 -07:00
|
|
|
struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
|
2014-12-09 20:40:49 -07:00
|
|
|
struct sk_buff *skb;
|
2015-05-06 22:11:40 -06:00
|
|
|
void *data;
|
|
|
|
|
|
|
|
len += NET_SKB_PAD + NET_IP_ALIGN;
|
2014-12-09 20:40:49 -07:00
|
|
|
|
2015-05-06 22:11:40 -06:00
|
|
|
if ((len > SKB_WITH_OVERHEAD(PAGE_SIZE)) ||
|
2015-11-06 17:28:21 -07:00
|
|
|
(gfp_mask & (__GFP_DIRECT_RECLAIM | GFP_DMA))) {
|
2015-05-13 14:34:13 -06:00
|
|
|
skb = __alloc_skb(len, gfp_mask, SKB_ALLOC_RX, NUMA_NO_NODE);
|
|
|
|
if (!skb)
|
|
|
|
goto skb_fail;
|
|
|
|
goto skb_success;
|
|
|
|
}
|
2015-05-06 22:11:40 -06:00
|
|
|
|
|
|
|
len += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
|
|
|
|
len = SKB_DATA_ALIGN(len);
|
|
|
|
|
|
|
|
if (sk_memalloc_socks())
|
|
|
|
gfp_mask |= __GFP_MEMALLOC;
|
2014-12-09 20:40:49 -07:00
|
|
|
|
2017-01-10 17:58:06 -07:00
|
|
|
data = page_frag_alloc(&nc->page, len, gfp_mask);
|
2015-05-06 22:11:40 -06:00
|
|
|
if (unlikely(!data))
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
skb = __build_skb(data, len);
|
|
|
|
if (unlikely(!skb)) {
|
2015-05-06 22:12:03 -06:00
|
|
|
skb_free_frag(data);
|
2015-05-06 22:11:40 -06:00
|
|
|
return NULL;
|
2014-12-09 20:40:49 -07:00
|
|
|
}
|
|
|
|
|
2015-05-06 22:11:40 -06:00
|
|
|
/* use OR instead of assignment to avoid clearing of bits in mask */
|
2016-02-08 05:14:59 -07:00
|
|
|
if (nc->page.pfmemalloc)
|
2015-05-06 22:11:40 -06:00
|
|
|
skb->pfmemalloc = 1;
|
|
|
|
skb->head_frag = 1;
|
|
|
|
|
2015-05-13 14:34:13 -06:00
|
|
|
skb_success:
|
2015-05-06 22:11:40 -06:00
|
|
|
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
|
|
|
|
skb->dev = napi->dev;
|
|
|
|
|
2015-05-13 14:34:13 -06:00
|
|
|
skb_fail:
|
2014-12-09 20:40:49 -07:00
|
|
|
return skb;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(__napi_alloc_skb);
|
|
|
|
|
2008-10-07 15:22:33 -06:00
|
|
|
void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off,
|
2012-03-23 17:59:33 -06:00
|
|
|
int size, unsigned int truesize)
|
2008-10-07 15:22:33 -06:00
|
|
|
{
|
|
|
|
skb_fill_page_desc(skb, i, page, off, size);
|
|
|
|
skb->len += size;
|
|
|
|
skb->data_len += size;
|
2012-03-23 17:59:33 -06:00
|
|
|
skb->truesize += truesize;
|
2008-10-07 15:22:33 -06:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_add_rx_frag);
|
|
|
|
|
2013-11-01 00:07:47 -06:00
|
|
|
void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,
|
|
|
|
unsigned int truesize)
|
|
|
|
{
|
|
|
|
skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
|
|
|
|
|
|
|
|
skb_frag_size_add(frag, size);
|
|
|
|
skb->len += size;
|
|
|
|
skb->data_len += size;
|
|
|
|
skb->truesize += truesize;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_coalesce_rx_frag);
|
|
|
|
|
2006-07-13 20:26:39 -06:00
|
|
|
static void skb_drop_list(struct sk_buff **listp)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2013-06-24 07:26:00 -06:00
|
|
|
kfree_skb_list(*listp);
|
2006-07-13 20:26:39 -06:00
|
|
|
*listp = NULL;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2006-07-13 20:26:39 -06:00
|
|
|
static inline void skb_drop_fraglist(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
skb_drop_list(&skb_shinfo(skb)->frag_list);
|
|
|
|
}
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
static void skb_clone_fraglist(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
struct sk_buff *list;
|
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
skb_walk_frags(skb, list)
|
2005-04-16 16:20:36 -06:00
|
|
|
skb_get(list);
|
|
|
|
}
|
|
|
|
|
net: allow skb->head to be a page fragment
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-26 18:33:38 -06:00
|
|
|
static void skb_free_head(struct sk_buff *skb)
|
|
|
|
{
|
2015-05-06 22:12:03 -06:00
|
|
|
unsigned char *head = skb->head;
|
|
|
|
|
net: allow skb->head to be a page fragment
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-26 18:33:38 -06:00
|
|
|
if (skb->head_frag)
|
2015-05-06 22:12:03 -06:00
|
|
|
skb_free_frag(head);
|
net: allow skb->head to be a page fragment
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-26 18:33:38 -06:00
|
|
|
else
|
2015-05-06 22:12:03 -06:00
|
|
|
kfree(head);
|
net: allow skb->head to be a page fragment
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-26 18:33:38 -06:00
|
|
|
}
|
|
|
|
|
2006-06-29 14:02:35 -06:00
|
|
|
static void skb_release_data(struct sk_buff *skb)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2014-09-23 19:39:30 -06:00
|
|
|
struct skb_shared_info *shinfo = skb_shinfo(skb);
|
|
|
|
int i;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2014-09-23 19:39:30 -06:00
|
|
|
if (skb->cloned &&
|
|
|
|
atomic_sub_return(skb->nohdr ? (1 << SKB_DATAREF_SHIFT) + 1 : 1,
|
|
|
|
&shinfo->dataref))
|
|
|
|
return;
|
2011-07-06 06:22:12 -06:00
|
|
|
|
2014-09-23 19:39:30 -06:00
|
|
|
for (i = 0; i < shinfo->nr_frags; i++)
|
|
|
|
__skb_frag_unref(&shinfo->frags[i]);
|
2011-07-06 06:22:12 -06:00
|
|
|
|
2014-09-23 19:39:30 -06:00
|
|
|
/*
|
|
|
|
* If skb buf is from userspace, we need to notify the caller
|
|
|
|
* the lower device DMA has done;
|
|
|
|
*/
|
|
|
|
if (shinfo->tx_flags & SKBTX_DEV_ZEROCOPY) {
|
|
|
|
struct ubuf_info *uarg;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2014-09-23 19:39:30 -06:00
|
|
|
uarg = shinfo->destructor_arg;
|
|
|
|
if (uarg->callback)
|
|
|
|
uarg->callback(uarg, true);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2014-09-23 19:39:30 -06:00
|
|
|
|
|
|
|
if (shinfo->frag_list)
|
|
|
|
kfree_skb_list(shinfo->frag_list);
|
|
|
|
|
|
|
|
skb_free_head(skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Free an skbuff by memory without cleaning the state.
|
|
|
|
*/
|
2007-11-26 08:11:19 -07:00
|
|
|
static void kfree_skbmem(struct sk_buff *skb)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2014-09-29 14:29:15 -06:00
|
|
|
struct sk_buff_fclones *fclones;
|
2005-08-17 15:57:30 -06:00
|
|
|
|
|
|
|
switch (skb->fclone) {
|
|
|
|
case SKB_FCLONE_UNAVAILABLE:
|
|
|
|
kmem_cache_free(skbuff_head_cache, skb);
|
2014-12-03 18:04:39 -07:00
|
|
|
return;
|
2005-08-17 15:57:30 -06:00
|
|
|
|
|
|
|
case SKB_FCLONE_ORIG:
|
2014-09-29 14:29:15 -06:00
|
|
|
fclones = container_of(skb, struct sk_buff_fclones, skb1);
|
2005-08-17 15:57:30 -06:00
|
|
|
|
2014-12-03 18:04:39 -07:00
|
|
|
/* We usually free the clone (TX completion) before original skb
|
|
|
|
* This test would have no chance to be true for the clone,
|
|
|
|
* while here, branch prediction will be good.
|
2005-08-17 15:57:30 -06:00
|
|
|
*/
|
2014-12-03 18:04:39 -07:00
|
|
|
if (atomic_read(&fclones->fclone_ref) == 1)
|
|
|
|
goto fastpath;
|
|
|
|
break;
|
2014-11-21 12:47:16 -07:00
|
|
|
|
2014-12-03 18:04:39 -07:00
|
|
|
default: /* SKB_FCLONE_CLONE */
|
|
|
|
fclones = container_of(skb, struct sk_buff_fclones, skb2);
|
2005-08-17 15:57:30 -06:00
|
|
|
break;
|
2007-04-20 18:09:22 -06:00
|
|
|
}
|
2014-12-03 18:04:39 -07:00
|
|
|
if (!atomic_dec_and_test(&fclones->fclone_ref))
|
|
|
|
return;
|
|
|
|
fastpath:
|
|
|
|
kmem_cache_free(skbuff_fclone_cache, fclones);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2008-10-01 03:33:12 -06:00
|
|
|
static void skb_release_head_state(struct sk_buff *skb)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2009-06-01 23:19:30 -06:00
|
|
|
skb_dst_drop(skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
#ifdef CONFIG_XFRM
|
|
|
|
secpath_put(skb->sp);
|
|
|
|
#endif
|
2005-04-19 23:39:42 -06:00
|
|
|
if (skb->destructor) {
|
|
|
|
WARN_ON(in_irq());
|
2005-04-16 16:20:36 -06:00
|
|
|
skb->destructor(skb);
|
|
|
|
}
|
2011-12-11 19:58:22 -07:00
|
|
|
#if IS_ENABLED(CONFIG_NF_CONNTRACK)
|
2017-01-23 10:21:56 -07:00
|
|
|
nf_conntrack_put(skb_nfct(skb));
|
2011-01-12 12:25:08 -07:00
|
|
|
#endif
|
2014-10-01 03:19:17 -06:00
|
|
|
#if IS_ENABLED(CONFIG_BRIDGE_NETFILTER)
|
2005-04-16 16:20:36 -06:00
|
|
|
nf_bridge_put(skb->nf_bridge);
|
|
|
|
#endif
|
2008-10-01 03:33:12 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Free everything but the sk_buff shell. */
|
|
|
|
static void skb_release_all(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
skb_release_head_state(skb);
|
2013-06-03 03:28:43 -06:00
|
|
|
if (likely(skb->head))
|
2013-04-17 00:46:58 -06:00
|
|
|
skb_release_data(skb);
|
2007-11-26 08:11:19 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* __kfree_skb - private function
|
|
|
|
* @skb: buffer
|
|
|
|
*
|
|
|
|
* Free an sk_buff. Release anything attached to the buffer.
|
|
|
|
* Clean the state. This is an internal helper function. Users should
|
|
|
|
* always call kfree_skb
|
|
|
|
*/
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2007-11-26 08:11:19 -07:00
|
|
|
void __kfree_skb(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
skb_release_all(skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
kfree_skbmem(skb);
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(__kfree_skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2006-03-20 22:28:35 -07:00
|
|
|
/**
|
|
|
|
* kfree_skb - free an sk_buff
|
|
|
|
* @skb: buffer to free
|
|
|
|
*
|
|
|
|
* Drop a reference to the buffer and free it if the usage count has
|
|
|
|
* hit zero.
|
|
|
|
*/
|
|
|
|
void kfree_skb(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
if (unlikely(!skb))
|
|
|
|
return;
|
|
|
|
if (likely(atomic_read(&skb->users) == 1))
|
|
|
|
smp_rmb();
|
|
|
|
else if (likely(!atomic_dec_and_test(&skb->users)))
|
|
|
|
return;
|
2009-03-11 03:49:55 -06:00
|
|
|
trace_kfree_skb(skb, __builtin_return_address(0));
|
2006-03-20 22:28:35 -07:00
|
|
|
__kfree_skb(skb);
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(kfree_skb);
|
2006-03-20 22:28:35 -07:00
|
|
|
|
2013-06-24 07:26:00 -06:00
|
|
|
void kfree_skb_list(struct sk_buff *segs)
|
|
|
|
{
|
|
|
|
while (segs) {
|
|
|
|
struct sk_buff *next = segs->next;
|
|
|
|
|
|
|
|
kfree_skb(segs);
|
|
|
|
segs = next;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(kfree_skb_list);
|
|
|
|
|
2012-11-01 03:16:28 -06:00
|
|
|
/**
|
|
|
|
* skb_tx_error - report an sk_buff xmit error
|
|
|
|
* @skb: buffer that triggered an error
|
|
|
|
*
|
|
|
|
* Report xmit error if a device callback is tracking this skb.
|
|
|
|
* skb must be freed afterwards.
|
|
|
|
*/
|
|
|
|
void skb_tx_error(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
if (skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) {
|
|
|
|
struct ubuf_info *uarg;
|
|
|
|
|
|
|
|
uarg = skb_shinfo(skb)->destructor_arg;
|
|
|
|
if (uarg->callback)
|
|
|
|
uarg->callback(uarg, false);
|
|
|
|
skb_shinfo(skb)->tx_flags &= ~SKBTX_DEV_ZEROCOPY;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_tx_error);
|
|
|
|
|
2009-03-11 03:49:55 -06:00
|
|
|
/**
|
|
|
|
* consume_skb - free an skbuff
|
|
|
|
* @skb: buffer to free
|
|
|
|
*
|
|
|
|
* Drop a ref to the buffer and free it if the usage count has hit zero
|
|
|
|
* Functions identically to kfree_skb, but kfree_skb assumes that the frame
|
|
|
|
* is being dropped after a failure and notes that
|
|
|
|
*/
|
|
|
|
void consume_skb(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
if (unlikely(!skb))
|
|
|
|
return;
|
|
|
|
if (likely(atomic_read(&skb->users) == 1))
|
|
|
|
smp_rmb();
|
|
|
|
else if (likely(!atomic_dec_and_test(&skb->users)))
|
|
|
|
return;
|
2010-08-23 03:46:12 -06:00
|
|
|
trace_consume_skb(skb);
|
2009-03-11 03:49:55 -06:00
|
|
|
__kfree_skb(skb);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(consume_skb);
|
|
|
|
|
2016-02-08 05:14:59 -07:00
|
|
|
void __kfree_skb_flush(void)
|
|
|
|
{
|
|
|
|
struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
|
|
|
|
|
|
|
|
/* flush skb_cache if containing objects */
|
|
|
|
if (nc->skb_count) {
|
|
|
|
kmem_cache_free_bulk(skbuff_head_cache, nc->skb_count,
|
|
|
|
nc->skb_cache);
|
|
|
|
nc->skb_count = 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-02-08 05:15:04 -07:00
|
|
|
static inline void _kfree_skb_defer(struct sk_buff *skb)
|
2016-02-08 05:14:59 -07:00
|
|
|
{
|
|
|
|
struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
|
|
|
|
|
|
|
|
/* drop skb->head and call any destructors for packet */
|
|
|
|
skb_release_all(skb);
|
|
|
|
|
|
|
|
/* record skb to CPU local list */
|
|
|
|
nc->skb_cache[nc->skb_count++] = skb;
|
|
|
|
|
|
|
|
#ifdef CONFIG_SLUB
|
|
|
|
/* SLUB writes into objects when freeing */
|
|
|
|
prefetchw(skb);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* flush skb_cache if it is filled */
|
|
|
|
if (unlikely(nc->skb_count == NAPI_SKB_CACHE_SIZE)) {
|
|
|
|
kmem_cache_free_bulk(skbuff_head_cache, NAPI_SKB_CACHE_SIZE,
|
|
|
|
nc->skb_cache);
|
|
|
|
nc->skb_count = 0;
|
|
|
|
}
|
|
|
|
}
|
2016-02-08 05:15:04 -07:00
|
|
|
void __kfree_skb_defer(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
_kfree_skb_defer(skb);
|
|
|
|
}
|
2016-02-08 05:14:59 -07:00
|
|
|
|
|
|
|
void napi_consume_skb(struct sk_buff *skb, int budget)
|
|
|
|
{
|
|
|
|
if (unlikely(!skb))
|
|
|
|
return;
|
|
|
|
|
2016-03-11 01:43:58 -07:00
|
|
|
/* Zero budget indicate non-NAPI context called us, like netpoll */
|
2016-02-08 05:14:59 -07:00
|
|
|
if (unlikely(!budget)) {
|
2016-03-11 01:43:58 -07:00
|
|
|
dev_consume_skb_any(skb);
|
2016-02-08 05:14:59 -07:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (likely(atomic_read(&skb->users) == 1))
|
|
|
|
smp_rmb();
|
|
|
|
else if (likely(!atomic_dec_and_test(&skb->users)))
|
|
|
|
return;
|
|
|
|
/* if reaching here SKB is ready to free */
|
|
|
|
trace_consume_skb(skb);
|
|
|
|
|
|
|
|
/* if SKB is a clone, don't handle this case */
|
2016-03-20 12:27:47 -06:00
|
|
|
if (skb->fclone != SKB_FCLONE_UNAVAILABLE) {
|
2016-02-08 05:14:59 -07:00
|
|
|
__kfree_skb(skb);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2016-02-08 05:15:04 -07:00
|
|
|
_kfree_skb_defer(skb);
|
2016-02-08 05:14:59 -07:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(napi_consume_skb);
|
|
|
|
|
net: reorganize sk_buff for faster __copy_skb_header()
With proliferation of bit fields in sk_buff, __copy_skb_header() became
quite expensive, showing as the most expensive function in a GSO
workload.
__copy_skb_header() performance is also critical for non GSO TCP
operations, as it is used from skb_clone()
This patch carefully moves all the fields that were not copied in a
separate zone : cloned, nohdr, fclone, peeked, head_frag, xmit_more
Then I moved all other fields and all other copied fields in a section
delimited by headers_start[0]/headers_end[0] section so that we
can use a single memcpy() call, inlined by compiler using long
word load/stores.
I also tried to make all copies in the natural orders of sk_buff,
to help hardware prefetching.
I made sure sk_buff size did not change.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-28 23:18:47 -06:00
|
|
|
/* Make sure a field is enclosed inside headers_start/headers_end section */
|
|
|
|
#define CHECK_SKB_FIELD(field) \
|
|
|
|
BUILD_BUG_ON(offsetof(struct sk_buff, field) < \
|
|
|
|
offsetof(struct sk_buff, headers_start)); \
|
|
|
|
BUILD_BUG_ON(offsetof(struct sk_buff, field) > \
|
|
|
|
offsetof(struct sk_buff, headers_end)); \
|
|
|
|
|
2007-10-14 01:37:30 -06:00
|
|
|
static void __copy_skb_header(struct sk_buff *new, const struct sk_buff *old)
|
|
|
|
{
|
|
|
|
new->tstamp = old->tstamp;
|
net: reorganize sk_buff for faster __copy_skb_header()
With proliferation of bit fields in sk_buff, __copy_skb_header() became
quite expensive, showing as the most expensive function in a GSO
workload.
__copy_skb_header() performance is also critical for non GSO TCP
operations, as it is used from skb_clone()
This patch carefully moves all the fields that were not copied in a
separate zone : cloned, nohdr, fclone, peeked, head_frag, xmit_more
Then I moved all other fields and all other copied fields in a section
delimited by headers_start[0]/headers_end[0] section so that we
can use a single memcpy() call, inlined by compiler using long
word load/stores.
I also tried to make all copies in the natural orders of sk_buff,
to help hardware prefetching.
I made sure sk_buff size did not change.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-28 23:18:47 -06:00
|
|
|
/* We do not copy old->sk */
|
2007-10-14 01:37:30 -06:00
|
|
|
new->dev = old->dev;
|
net: reorganize sk_buff for faster __copy_skb_header()
With proliferation of bit fields in sk_buff, __copy_skb_header() became
quite expensive, showing as the most expensive function in a GSO
workload.
__copy_skb_header() performance is also critical for non GSO TCP
operations, as it is used from skb_clone()
This patch carefully moves all the fields that were not copied in a
separate zone : cloned, nohdr, fclone, peeked, head_frag, xmit_more
Then I moved all other fields and all other copied fields in a section
delimited by headers_start[0]/headers_end[0] section so that we
can use a single memcpy() call, inlined by compiler using long
word load/stores.
I also tried to make all copies in the natural orders of sk_buff,
to help hardware prefetching.
I made sure sk_buff size did not change.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-28 23:18:47 -06:00
|
|
|
memcpy(new->cb, old->cb, sizeof(old->cb));
|
2010-05-11 17:19:48 -06:00
|
|
|
skb_dst_copy(new, old);
|
2008-10-28 14:24:06 -06:00
|
|
|
#ifdef CONFIG_XFRM
|
2007-10-14 01:37:30 -06:00
|
|
|
new->sp = secpath_get(old->sp);
|
|
|
|
#endif
|
net: reorganize sk_buff for faster __copy_skb_header()
With proliferation of bit fields in sk_buff, __copy_skb_header() became
quite expensive, showing as the most expensive function in a GSO
workload.
__copy_skb_header() performance is also critical for non GSO TCP
operations, as it is used from skb_clone()
This patch carefully moves all the fields that were not copied in a
separate zone : cloned, nohdr, fclone, peeked, head_frag, xmit_more
Then I moved all other fields and all other copied fields in a section
delimited by headers_start[0]/headers_end[0] section so that we
can use a single memcpy() call, inlined by compiler using long
word load/stores.
I also tried to make all copies in the natural orders of sk_buff,
to help hardware prefetching.
I made sure sk_buff size did not change.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-28 23:18:47 -06:00
|
|
|
__nf_copy(new, old, false);
|
|
|
|
|
|
|
|
/* Note : this field could be in headers_start/headers_end section
|
|
|
|
* It is not yet because we do not want to have a 16 bit hole
|
|
|
|
*/
|
|
|
|
new->queue_mapping = old->queue_mapping;
|
|
|
|
|
|
|
|
memcpy(&new->headers_start, &old->headers_start,
|
|
|
|
offsetof(struct sk_buff, headers_end) -
|
|
|
|
offsetof(struct sk_buff, headers_start));
|
|
|
|
CHECK_SKB_FIELD(protocol);
|
|
|
|
CHECK_SKB_FIELD(csum);
|
|
|
|
CHECK_SKB_FIELD(hash);
|
|
|
|
CHECK_SKB_FIELD(priority);
|
|
|
|
CHECK_SKB_FIELD(skb_iif);
|
|
|
|
CHECK_SKB_FIELD(vlan_proto);
|
|
|
|
CHECK_SKB_FIELD(vlan_tci);
|
|
|
|
CHECK_SKB_FIELD(transport_header);
|
|
|
|
CHECK_SKB_FIELD(network_header);
|
|
|
|
CHECK_SKB_FIELD(mac_header);
|
|
|
|
CHECK_SKB_FIELD(inner_protocol);
|
|
|
|
CHECK_SKB_FIELD(inner_transport_header);
|
|
|
|
CHECK_SKB_FIELD(inner_network_header);
|
|
|
|
CHECK_SKB_FIELD(inner_mac_header);
|
|
|
|
CHECK_SKB_FIELD(mark);
|
|
|
|
#ifdef CONFIG_NETWORK_SECMARK
|
|
|
|
CHECK_SKB_FIELD(secmark);
|
|
|
|
#endif
|
|
|
|
#ifdef CONFIG_NET_RX_BUSY_POLL
|
|
|
|
CHECK_SKB_FIELD(napi_id);
|
2007-10-14 01:37:30 -06:00
|
|
|
#endif
|
xps: fix xps for stacked devices
A typical qdisc setup is the following :
bond0 : bonding device, using HTB hierarchy
eth1/eth2 : slaves, multiqueue NIC, using MQ + FQ qdisc
XPS allows to spread packets on specific tx queues, based on the cpu
doing the send.
Problem is that dequeues from bond0 qdisc can happen on random cpus,
due to the fact that qdisc_run() can dequeue a batch of packets.
CPUA -> queue packet P1 on bond0 qdisc, P1->ooo_okay=1
CPUA -> queue packet P2 on bond0 qdisc, P2->ooo_okay=0
CPUB -> dequeue packet P1 from bond0
enqueue packet on eth1/eth2
CPUC -> dequeue packet P2 from bond0
enqueue packet on eth1/eth2 using sk cache (ooo_okay is 0)
get_xps_queue() then might select wrong queue for P1, since current cpu
might be different than CPUA.
P2 might be sent on the old queue (stored in sk->sk_tx_queue_mapping),
if CPUC runs a bit faster (or CPUB spins a bit on qdisc lock)
Effect of this bug is TCP reorders, and more generally not optimal
TX queue placement. (A victim bulk flow can be migrated to the wrong TX
queue for a while)
To fix this, we have to record sender cpu number the first time
dev_queue_xmit() is called for one tx skb.
We can union napi_id (used on receive path) and sender_cpu,
granted we clear sender_cpu in skb_scrub_packet() (credit to Willem for
this union idea)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Nandita Dukkipati <nanditad@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-04 00:48:24 -07:00
|
|
|
#ifdef CONFIG_XPS
|
|
|
|
CHECK_SKB_FIELD(sender_cpu);
|
|
|
|
#endif
|
2007-10-14 01:37:30 -06:00
|
|
|
#ifdef CONFIG_NET_SCHED
|
net: reorganize sk_buff for faster __copy_skb_header()
With proliferation of bit fields in sk_buff, __copy_skb_header() became
quite expensive, showing as the most expensive function in a GSO
workload.
__copy_skb_header() performance is also critical for non GSO TCP
operations, as it is used from skb_clone()
This patch carefully moves all the fields that were not copied in a
separate zone : cloned, nohdr, fclone, peeked, head_frag, xmit_more
Then I moved all other fields and all other copied fields in a section
delimited by headers_start[0]/headers_end[0] section so that we
can use a single memcpy() call, inlined by compiler using long
word load/stores.
I also tried to make all copies in the natural orders of sk_buff,
to help hardware prefetching.
I made sure sk_buff size did not change.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-28 23:18:47 -06:00
|
|
|
CHECK_SKB_FIELD(tc_index);
|
2007-10-14 01:37:30 -06:00
|
|
|
#endif
|
2013-06-10 02:39:50 -06:00
|
|
|
|
2007-10-14 01:37:30 -06:00
|
|
|
}
|
|
|
|
|
2009-05-22 16:11:37 -06:00
|
|
|
/*
|
|
|
|
* You should not add any new code to this function. Add it to
|
|
|
|
* __copy_skb_header above instead.
|
|
|
|
*/
|
2007-10-14 01:37:52 -06:00
|
|
|
static struct sk_buff *__skb_clone(struct sk_buff *n, struct sk_buff *skb)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
|
|
|
#define C(x) n->x = skb->x
|
|
|
|
|
|
|
|
n->next = n->prev = NULL;
|
|
|
|
n->sk = NULL;
|
2007-10-14 01:37:30 -06:00
|
|
|
__copy_skb_header(n, skb);
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
C(len);
|
|
|
|
C(data_len);
|
2007-03-16 16:00:46 -06:00
|
|
|
C(mac_len);
|
[SKBUFF]: Keep track of writable header len of headerless clones
Currently NAT (and others) that want to modify cloned skbs copy them,
even if in the vast majority of cases its not necessary because the
skb is a clone made by TCP and the portion NAT wants to modify is
actually writable because TCP release the header reference before
cloning.
The problem is that there is no clean way for NAT to find out how
long the writable header area is, so this patch introduces skb->hdr_len
to hold this length. When a headerless skb is cloned skb->hdr_len
is set to the current headroom, for regular clones it is copied from
the original. A new function skb_clone_writable(skb, len) returns
whether the skb is writable up to len bytes from skb->data. To avoid
enlarging the skb the mac_len field is reduced to 16 bit and the
new hdr_len field is put in the remaining 16 bit.
I've done a few rough benchmarks of NAT (not with this exact patch,
but a very similar one). As expected it saves huge amounts of system
time in case of sendfile, bringing it down to basically the same
amount as without NAT, with sendmsg it only helps on loopback,
probably because of the large MTU.
Transmit a 1GB file using sendfile/sendmsg over eth0/lo with and
without NAT:
- sendfile eth0, no NAT: sys 0m0.388s
- sendfile eth0, NAT: sys 0m1.835s
- sendfile eth0: NAT + path: sys 0m0.370s (~ -80%)
- sendfile lo, no NAT: sys 0m0.258s
- sendfile lo, NAT: sys 0m2.609s
- sendfile lo, NAT + patch: sys 0m0.260s (~ -90%)
- sendmsg eth0, no NAT: sys 0m2.508s
- sendmsg eth0, NAT: sys 0m2.539s
- sendmsg eth0, NAT + patch: sys 0m2.445s (no change)
- sendmsg lo, no NAT: sys 0m2.151s
- sendmsg lo, NAT: sys 0m3.557s
- sendmsg lo, NAT + patch: sys 0m2.159s (~ -40%)
I expect other users can see a similar performance improvement,
packet mangling iptables targets, ipip and ip_gre come to mind ..
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2007-06-25 05:35:20 -06:00
|
|
|
n->hdr_len = skb->nohdr ? skb_headroom(skb) : skb->hdr_len;
|
2008-01-07 22:56:41 -07:00
|
|
|
n->cloned = 1;
|
2005-04-16 16:20:36 -06:00
|
|
|
n->nohdr = 0;
|
|
|
|
n->destructor = NULL;
|
|
|
|
C(tail);
|
|
|
|
C(end);
|
2008-01-07 22:56:41 -07:00
|
|
|
C(head);
|
net: allow skb->head to be a page fragment
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-26 18:33:38 -06:00
|
|
|
C(head_frag);
|
2008-01-07 22:56:41 -07:00
|
|
|
C(data);
|
|
|
|
C(truesize);
|
|
|
|
atomic_set(&n->users, 1);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
atomic_inc(&(skb_shinfo(skb)->dataref));
|
|
|
|
skb->cloned = 1;
|
|
|
|
|
|
|
|
return n;
|
2007-10-14 01:37:52 -06:00
|
|
|
#undef C
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_morph - morph one skb into another
|
|
|
|
* @dst: the skb to receive the contents
|
|
|
|
* @src: the skb to supply the contents
|
|
|
|
*
|
|
|
|
* This is identical to skb_clone except that the target skb is
|
|
|
|
* supplied by the user.
|
|
|
|
*
|
|
|
|
* The target skb is returned upon exit.
|
|
|
|
*/
|
|
|
|
struct sk_buff *skb_morph(struct sk_buff *dst, struct sk_buff *src)
|
|
|
|
{
|
2007-11-26 08:11:19 -07:00
|
|
|
skb_release_all(dst);
|
2007-10-14 01:37:52 -06:00
|
|
|
return __skb_clone(dst, src);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_morph);
|
|
|
|
|
2012-07-10 04:55:09 -06:00
|
|
|
/**
|
|
|
|
* skb_copy_ubufs - copy userspace skb frags buffers to kernel
|
2011-08-31 02:03:29 -06:00
|
|
|
* @skb: the skb to modify
|
|
|
|
* @gfp_mask: allocation priority
|
|
|
|
*
|
|
|
|
* This must be called on SKBTX_DEV_ZEROCOPY skb.
|
|
|
|
* It will copy all frags into kernel and drop the reference
|
|
|
|
* to userspace pages.
|
|
|
|
*
|
|
|
|
* If this function is called from an interrupt gfp_mask() must be
|
|
|
|
* %GFP_ATOMIC.
|
|
|
|
*
|
|
|
|
* Returns 0 on success or a negative error code on failure
|
|
|
|
* to allocate kernel memory to copy to.
|
|
|
|
*/
|
|
|
|
int skb_copy_ubufs(struct sk_buff *skb, gfp_t gfp_mask)
|
2011-07-06 06:22:12 -06:00
|
|
|
{
|
|
|
|
int i;
|
|
|
|
int num_frags = skb_shinfo(skb)->nr_frags;
|
|
|
|
struct page *page, *head = NULL;
|
|
|
|
struct ubuf_info *uarg = skb_shinfo(skb)->destructor_arg;
|
|
|
|
|
|
|
|
for (i = 0; i < num_frags; i++) {
|
|
|
|
u8 *vaddr;
|
|
|
|
skb_frag_t *f = &skb_shinfo(skb)->frags[i];
|
|
|
|
|
2012-07-16 20:05:29 -06:00
|
|
|
page = alloc_page(gfp_mask);
|
2011-07-06 06:22:12 -06:00
|
|
|
if (!page) {
|
|
|
|
while (head) {
|
2013-07-12 01:17:23 -06:00
|
|
|
struct page *next = (struct page *)page_private(head);
|
2011-07-06 06:22:12 -06:00
|
|
|
put_page(head);
|
|
|
|
head = next;
|
|
|
|
}
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
2012-04-05 03:35:15 -06:00
|
|
|
vaddr = kmap_atomic(skb_frag_page(f));
|
2011-07-06 06:22:12 -06:00
|
|
|
memcpy(page_address(page),
|
2011-10-18 15:00:24 -06:00
|
|
|
vaddr + f->page_offset, skb_frag_size(f));
|
2012-04-05 03:35:15 -06:00
|
|
|
kunmap_atomic(vaddr);
|
2013-07-12 01:17:23 -06:00
|
|
|
set_page_private(page, (unsigned long)head);
|
2011-07-06 06:22:12 -06:00
|
|
|
head = page;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* skb frags release userspace buffers */
|
2012-07-16 20:05:29 -06:00
|
|
|
for (i = 0; i < num_frags; i++)
|
2011-10-19 17:01:49 -06:00
|
|
|
skb_frag_unref(skb, i);
|
2011-07-06 06:22:12 -06:00
|
|
|
|
2012-11-01 03:16:22 -06:00
|
|
|
uarg->callback(uarg, false);
|
2011-07-06 06:22:12 -06:00
|
|
|
|
|
|
|
/* skb frags point to kernel buffers */
|
2012-07-16 20:05:29 -06:00
|
|
|
for (i = num_frags - 1; i >= 0; i--) {
|
|
|
|
__skb_fill_page_desc(skb, i, head, 0,
|
|
|
|
skb_shinfo(skb)->frags[i].size);
|
2013-07-12 01:17:23 -06:00
|
|
|
head = (struct page *)page_private(head);
|
2011-07-06 06:22:12 -06:00
|
|
|
}
|
2011-08-31 02:03:29 -06:00
|
|
|
|
|
|
|
skb_shinfo(skb)->tx_flags &= ~SKBTX_DEV_ZEROCOPY;
|
2011-07-06 06:22:12 -06:00
|
|
|
return 0;
|
|
|
|
}
|
2012-07-20 03:23:20 -06:00
|
|
|
EXPORT_SYMBOL_GPL(skb_copy_ubufs);
|
2011-07-06 06:22:12 -06:00
|
|
|
|
2007-10-14 01:37:52 -06:00
|
|
|
/**
|
|
|
|
* skb_clone - duplicate an sk_buff
|
|
|
|
* @skb: buffer to clone
|
|
|
|
* @gfp_mask: allocation priority
|
|
|
|
*
|
|
|
|
* Duplicate an &sk_buff. The new one is not owned by a socket. Both
|
|
|
|
* copies share the same packet data but not structure. The new
|
|
|
|
* buffer has a reference count of 1. If the allocation fails the
|
|
|
|
* function returns %NULL otherwise the new buffer is returned.
|
|
|
|
*
|
|
|
|
* If this function is called from an interrupt gfp_mask() must be
|
|
|
|
* %GFP_ATOMIC.
|
|
|
|
*/
|
|
|
|
|
|
|
|
struct sk_buff *skb_clone(struct sk_buff *skb, gfp_t gfp_mask)
|
|
|
|
{
|
2014-09-29 14:29:15 -06:00
|
|
|
struct sk_buff_fclones *fclones = container_of(skb,
|
|
|
|
struct sk_buff_fclones,
|
|
|
|
skb1);
|
2014-12-03 18:04:39 -07:00
|
|
|
struct sk_buff *n;
|
2007-10-14 01:37:52 -06:00
|
|
|
|
2012-07-20 03:23:10 -06:00
|
|
|
if (skb_orphan_frags(skb, gfp_mask))
|
|
|
|
return NULL;
|
2011-07-06 06:22:12 -06:00
|
|
|
|
2007-10-14 01:37:52 -06:00
|
|
|
if (skb->fclone == SKB_FCLONE_ORIG &&
|
2014-12-03 18:04:39 -07:00
|
|
|
atomic_read(&fclones->fclone_ref) == 1) {
|
|
|
|
n = &fclones->skb2;
|
|
|
|
atomic_set(&fclones->fclone_ref, 2);
|
2007-10-14 01:37:52 -06:00
|
|
|
} else {
|
2012-07-31 17:44:19 -06:00
|
|
|
if (skb_pfmemalloc(skb))
|
|
|
|
gfp_mask |= __GFP_MEMALLOC;
|
|
|
|
|
2007-10-14 01:37:52 -06:00
|
|
|
n = kmem_cache_alloc(skbuff_head_cache, gfp_mask);
|
|
|
|
if (!n)
|
|
|
|
return NULL;
|
2008-08-30 04:16:35 -06:00
|
|
|
|
|
|
|
kmemcheck_annotate_bitfield(n, flags1);
|
2007-10-14 01:37:52 -06:00
|
|
|
n->fclone = SKB_FCLONE_UNAVAILABLE;
|
|
|
|
}
|
|
|
|
|
|
|
|
return __skb_clone(n, skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_clone);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2013-03-07 06:21:40 -07:00
|
|
|
static void skb_headers_offset_update(struct sk_buff *skb, int off)
|
|
|
|
{
|
2013-10-19 12:42:54 -06:00
|
|
|
/* Only adjust this if it actually is csum_start rather than csum */
|
|
|
|
if (skb->ip_summed == CHECKSUM_PARTIAL)
|
|
|
|
skb->csum_start += off;
|
2013-03-07 06:21:40 -07:00
|
|
|
/* {transport,network,mac}_header and tail are relative to skb->head */
|
|
|
|
skb->transport_header += off;
|
|
|
|
skb->network_header += off;
|
|
|
|
if (skb_mac_header_was_set(skb))
|
|
|
|
skb->mac_header += off;
|
|
|
|
skb->inner_transport_header += off;
|
|
|
|
skb->inner_network_header += off;
|
2013-03-07 06:21:46 -07:00
|
|
|
skb->inner_mac_header += off;
|
2013-03-07 06:21:40 -07:00
|
|
|
}
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
static void copy_skb_header(struct sk_buff *new, const struct sk_buff *old)
|
|
|
|
{
|
2007-10-14 01:37:30 -06:00
|
|
|
__copy_skb_header(new, old);
|
|
|
|
|
2006-06-22 03:40:14 -06:00
|
|
|
skb_shinfo(new)->gso_size = skb_shinfo(old)->gso_size;
|
|
|
|
skb_shinfo(new)->gso_segs = skb_shinfo(old)->gso_segs;
|
|
|
|
skb_shinfo(new)->gso_type = skb_shinfo(old)->gso_type;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2012-07-31 17:44:19 -06:00
|
|
|
static inline int skb_alloc_rx_flag(const struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
if (skb_pfmemalloc(skb))
|
|
|
|
return SKB_ALLOC_RX;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
/**
|
|
|
|
* skb_copy - create private copy of an sk_buff
|
|
|
|
* @skb: buffer to copy
|
|
|
|
* @gfp_mask: allocation priority
|
|
|
|
*
|
|
|
|
* Make a copy of both an &sk_buff and its data. This is used when the
|
|
|
|
* caller wishes to modify the data and needs a private copy of the
|
|
|
|
* data to alter. Returns %NULL on failure or the pointer to the buffer
|
|
|
|
* on success. The returned buffer has a reference count of 1.
|
|
|
|
*
|
|
|
|
* As by-product this function converts non-linear &sk_buff to linear
|
|
|
|
* one, so that &sk_buff becomes completely private and caller is allowed
|
|
|
|
* to modify all the data of returned buffer. This means that this
|
|
|
|
* function is not recommended for use in circumstances when only
|
|
|
|
* header is going to be modified. Use pskb_copy() instead.
|
|
|
|
*/
|
|
|
|
|
2005-10-07 00:46:04 -06:00
|
|
|
struct sk_buff *skb_copy(const struct sk_buff *skb, gfp_t gfp_mask)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2010-08-31 23:25:10 -06:00
|
|
|
int headerlen = skb_headroom(skb);
|
2012-05-04 08:26:56 -06:00
|
|
|
unsigned int size = skb_end_offset(skb) + skb->data_len;
|
2012-07-31 17:44:19 -06:00
|
|
|
struct sk_buff *n = __alloc_skb(size, gfp_mask,
|
|
|
|
skb_alloc_rx_flag(skb), NUMA_NO_NODE);
|
2010-08-31 23:25:10 -06:00
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
if (!n)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
/* Set the data pointer */
|
|
|
|
skb_reserve(n, headerlen);
|
|
|
|
/* Set the tail pointer and length */
|
|
|
|
skb_put(n, skb->len);
|
|
|
|
|
|
|
|
if (skb_copy_bits(skb, -headerlen, n->head, headerlen + skb->len))
|
|
|
|
BUG();
|
|
|
|
|
|
|
|
copy_skb_header(n, skb);
|
|
|
|
return n;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_copy);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
2014-06-11 16:36:26 -06:00
|
|
|
* __pskb_copy_fclone - create copy of an sk_buff with private head.
|
2005-04-16 16:20:36 -06:00
|
|
|
* @skb: buffer to copy
|
2011-12-03 14:39:53 -07:00
|
|
|
* @headroom: headroom of new skb
|
2005-04-16 16:20:36 -06:00
|
|
|
* @gfp_mask: allocation priority
|
2014-06-11 16:36:26 -06:00
|
|
|
* @fclone: if true allocate the copy of the skb from the fclone
|
|
|
|
* cache instead of the head cache; it is recommended to set this
|
|
|
|
* to true for the cases where the copy will likely be cloned
|
2005-04-16 16:20:36 -06:00
|
|
|
*
|
|
|
|
* Make a copy of both an &sk_buff and part of its data, located
|
|
|
|
* in header. Fragmented data remain shared. This is used when
|
|
|
|
* the caller wishes to modify only header of &sk_buff and needs
|
|
|
|
* private copy of the header to alter. Returns %NULL on failure
|
|
|
|
* or the pointer to the buffer on success.
|
|
|
|
* The returned buffer has a reference count of 1.
|
|
|
|
*/
|
|
|
|
|
2014-06-11 16:36:26 -06:00
|
|
|
struct sk_buff *__pskb_copy_fclone(struct sk_buff *skb, int headroom,
|
|
|
|
gfp_t gfp_mask, bool fclone)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2011-12-03 14:39:53 -07:00
|
|
|
unsigned int size = skb_headlen(skb) + headroom;
|
2014-06-11 16:36:26 -06:00
|
|
|
int flags = skb_alloc_rx_flag(skb) | (fclone ? SKB_ALLOC_FCLONE : 0);
|
|
|
|
struct sk_buff *n = __alloc_skb(size, gfp_mask, flags, NUMA_NO_NODE);
|
2010-08-31 23:25:10 -06:00
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
if (!n)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
/* Set the data pointer */
|
2011-12-03 14:39:53 -07:00
|
|
|
skb_reserve(n, headroom);
|
2005-04-16 16:20:36 -06:00
|
|
|
/* Set the tail pointer and length */
|
|
|
|
skb_put(n, skb_headlen(skb));
|
|
|
|
/* Copy the bytes */
|
2007-03-27 15:55:52 -06:00
|
|
|
skb_copy_from_linear_data(skb, n->data, n->len);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2006-11-07 15:57:15 -07:00
|
|
|
n->truesize += skb->data_len;
|
2005-04-16 16:20:36 -06:00
|
|
|
n->data_len = skb->data_len;
|
|
|
|
n->len = skb->len;
|
|
|
|
|
|
|
|
if (skb_shinfo(skb)->nr_frags) {
|
|
|
|
int i;
|
|
|
|
|
2012-07-20 03:23:10 -06:00
|
|
|
if (skb_orphan_frags(skb, gfp_mask)) {
|
|
|
|
kfree_skb(n);
|
|
|
|
n = NULL;
|
|
|
|
goto out;
|
2011-07-06 06:22:12 -06:00
|
|
|
}
|
2005-04-16 16:20:36 -06:00
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
|
|
|
|
skb_shinfo(n)->frags[i] = skb_shinfo(skb)->frags[i];
|
2011-08-22 17:44:58 -06:00
|
|
|
skb_frag_ref(skb, i);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
skb_shinfo(n)->nr_frags = i;
|
|
|
|
}
|
|
|
|
|
2010-08-23 01:13:46 -06:00
|
|
|
if (skb_has_frag_list(skb)) {
|
2005-04-16 16:20:36 -06:00
|
|
|
skb_shinfo(n)->frag_list = skb_shinfo(skb)->frag_list;
|
|
|
|
skb_clone_fraglist(n);
|
|
|
|
}
|
|
|
|
|
|
|
|
copy_skb_header(n, skb);
|
|
|
|
out:
|
|
|
|
return n;
|
|
|
|
}
|
2014-06-11 16:36:26 -06:00
|
|
|
EXPORT_SYMBOL(__pskb_copy_fclone);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* pskb_expand_head - reallocate header of &sk_buff
|
|
|
|
* @skb: buffer to reallocate
|
|
|
|
* @nhead: room to add at head
|
|
|
|
* @ntail: room to add at tail
|
|
|
|
* @gfp_mask: allocation priority
|
|
|
|
*
|
2013-11-07 06:18:26 -07:00
|
|
|
* Expands (or creates identical copy, if @nhead and @ntail are zero)
|
|
|
|
* header of @skb. &sk_buff itself is not changed. &sk_buff MUST have
|
2005-04-16 16:20:36 -06:00
|
|
|
* reference count of 1. Returns zero in the case of success or error,
|
|
|
|
* if expansion failed. In the last case, &sk_buff is not changed.
|
|
|
|
*
|
|
|
|
* All the pointers pointing into skb header may change and must be
|
|
|
|
* reloaded after call to this function.
|
|
|
|
*/
|
|
|
|
|
2005-07-08 15:57:47 -06:00
|
|
|
int pskb_expand_head(struct sk_buff *skb, int nhead, int ntail,
|
2005-10-07 00:46:04 -06:00
|
|
|
gfp_t gfp_mask)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2017-01-27 08:11:27 -07:00
|
|
|
int i, osize = skb_end_offset(skb);
|
|
|
|
int size = osize + nhead + ntail;
|
2005-04-16 16:20:36 -06:00
|
|
|
long off;
|
2017-01-27 08:11:27 -07:00
|
|
|
u8 *data;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2008-10-01 08:09:38 -06:00
|
|
|
BUG_ON(nhead < 0);
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
if (skb_shared(skb))
|
|
|
|
BUG();
|
|
|
|
|
|
|
|
size = SKB_DATA_ALIGN(size);
|
|
|
|
|
2012-07-31 17:44:19 -06:00
|
|
|
if (skb_pfmemalloc(skb))
|
|
|
|
gfp_mask |= __GFP_MEMALLOC;
|
|
|
|
data = kmalloc_reserve(size + SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
|
|
|
|
gfp_mask, NUMA_NO_NODE, NULL);
|
2005-04-16 16:20:36 -06:00
|
|
|
if (!data)
|
|
|
|
goto nodata;
|
2012-04-10 14:08:39 -06:00
|
|
|
size = SKB_WITH_OVERHEAD(ksize(data));
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/* Copy only real data... and, alas, header. This should be
|
2010-08-31 23:25:10 -06:00
|
|
|
* optimized for the cases when header is void.
|
|
|
|
*/
|
|
|
|
memcpy(data + nhead, skb->head, skb_tail_pointer(skb) - skb->head);
|
|
|
|
|
|
|
|
memcpy((struct skb_shared_info *)(data + size),
|
|
|
|
skb_shinfo(skb),
|
2010-07-22 13:09:08 -06:00
|
|
|
offsetof(struct skb_shared_info, frags[skb_shinfo(skb)->nr_frags]));
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2012-05-04 08:26:51 -06:00
|
|
|
/*
|
|
|
|
* if shinfo is shared we must drop the old head gracefully, but if it
|
|
|
|
* is not we can just drop the old head and let the existing refcount
|
|
|
|
* be since all we did is relocate the values
|
|
|
|
*/
|
|
|
|
if (skb_cloned(skb)) {
|
2011-07-06 06:22:12 -06:00
|
|
|
/* copy this zero copy skb frags */
|
2012-07-20 03:23:10 -06:00
|
|
|
if (skb_orphan_frags(skb, gfp_mask))
|
|
|
|
goto nofrags;
|
2010-09-02 17:09:32 -06:00
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
|
2011-08-22 17:44:58 -06:00
|
|
|
skb_frag_ref(skb, i);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2010-09-02 17:09:32 -06:00
|
|
|
if (skb_has_frag_list(skb))
|
|
|
|
skb_clone_fraglist(skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2010-09-02 17:09:32 -06:00
|
|
|
skb_release_data(skb);
|
2012-05-04 08:26:51 -06:00
|
|
|
} else {
|
|
|
|
skb_free_head(skb);
|
2010-09-02 17:09:32 -06:00
|
|
|
}
|
2005-04-16 16:20:36 -06:00
|
|
|
off = (data + nhead) - skb->head;
|
|
|
|
|
|
|
|
skb->head = data;
|
net: allow skb->head to be a page fragment
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-26 18:33:38 -06:00
|
|
|
skb->head_frag = 0;
|
2005-04-16 16:20:36 -06:00
|
|
|
skb->data += off;
|
2007-04-19 21:43:29 -06:00
|
|
|
#ifdef NET_SKBUFF_DATA_USES_OFFSET
|
|
|
|
skb->end = size;
|
2007-04-09 12:45:04 -06:00
|
|
|
off = nhead;
|
2007-04-19 21:43:29 -06:00
|
|
|
#else
|
|
|
|
skb->end = skb->head + size;
|
2007-04-09 12:45:04 -06:00
|
|
|
#endif
|
2007-04-19 21:29:13 -06:00
|
|
|
skb->tail += off;
|
2013-06-06 07:27:21 -06:00
|
|
|
skb_headers_offset_update(skb, nhead);
|
2005-04-16 16:20:36 -06:00
|
|
|
skb->cloned = 0;
|
[SKBUFF]: Keep track of writable header len of headerless clones
Currently NAT (and others) that want to modify cloned skbs copy them,
even if in the vast majority of cases its not necessary because the
skb is a clone made by TCP and the portion NAT wants to modify is
actually writable because TCP release the header reference before
cloning.
The problem is that there is no clean way for NAT to find out how
long the writable header area is, so this patch introduces skb->hdr_len
to hold this length. When a headerless skb is cloned skb->hdr_len
is set to the current headroom, for regular clones it is copied from
the original. A new function skb_clone_writable(skb, len) returns
whether the skb is writable up to len bytes from skb->data. To avoid
enlarging the skb the mac_len field is reduced to 16 bit and the
new hdr_len field is put in the remaining 16 bit.
I've done a few rough benchmarks of NAT (not with this exact patch,
but a very similar one). As expected it saves huge amounts of system
time in case of sendfile, bringing it down to basically the same
amount as without NAT, with sendmsg it only helps on loopback,
probably because of the large MTU.
Transmit a 1GB file using sendfile/sendmsg over eth0/lo with and
without NAT:
- sendfile eth0, no NAT: sys 0m0.388s
- sendfile eth0, NAT: sys 0m1.835s
- sendfile eth0: NAT + path: sys 0m0.370s (~ -80%)
- sendfile lo, no NAT: sys 0m0.258s
- sendfile lo, NAT: sys 0m2.609s
- sendfile lo, NAT + patch: sys 0m0.260s (~ -90%)
- sendmsg eth0, no NAT: sys 0m2.508s
- sendmsg eth0, NAT: sys 0m2.539s
- sendmsg eth0, NAT + patch: sys 0m2.445s (no change)
- sendmsg lo, no NAT: sys 0m2.151s
- sendmsg lo, NAT: sys 0m3.557s
- sendmsg lo, NAT + patch: sys 0m2.159s (~ -40%)
I expect other users can see a similar performance improvement,
packet mangling iptables targets, ipip and ip_gre come to mind ..
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2007-06-25 05:35:20 -06:00
|
|
|
skb->hdr_len = 0;
|
2005-04-16 16:20:36 -06:00
|
|
|
skb->nohdr = 0;
|
|
|
|
atomic_set(&skb_shinfo(skb)->dataref, 1);
|
2017-01-27 08:11:27 -07:00
|
|
|
|
|
|
|
/* It is not generally safe to change skb->truesize.
|
|
|
|
* For the moment, we really care of rx path, or
|
|
|
|
* when skb is orphaned (not attached to a socket).
|
|
|
|
*/
|
|
|
|
if (!skb->sk || skb->destructor == sock_edemux)
|
|
|
|
skb->truesize += size - osize;
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
return 0;
|
|
|
|
|
2011-07-06 06:22:12 -06:00
|
|
|
nofrags:
|
|
|
|
kfree(data);
|
2005-04-16 16:20:36 -06:00
|
|
|
nodata:
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(pskb_expand_head);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/* Make private copy of skb with writable head and some headroom */
|
|
|
|
|
|
|
|
struct sk_buff *skb_realloc_headroom(struct sk_buff *skb, unsigned int headroom)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb2;
|
|
|
|
int delta = headroom - skb_headroom(skb);
|
|
|
|
|
|
|
|
if (delta <= 0)
|
|
|
|
skb2 = pskb_copy(skb, GFP_ATOMIC);
|
|
|
|
else {
|
|
|
|
skb2 = skb_clone(skb, GFP_ATOMIC);
|
|
|
|
if (skb2 && pskb_expand_head(skb2, SKB_DATA_ALIGN(delta), 0,
|
|
|
|
GFP_ATOMIC)) {
|
|
|
|
kfree_skb(skb2);
|
|
|
|
skb2 = NULL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return skb2;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_realloc_headroom);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_copy_expand - copy and expand sk_buff
|
|
|
|
* @skb: buffer to copy
|
|
|
|
* @newheadroom: new free bytes at head
|
|
|
|
* @newtailroom: new free bytes at tail
|
|
|
|
* @gfp_mask: allocation priority
|
|
|
|
*
|
|
|
|
* Make a copy of both an &sk_buff and its data and while doing so
|
|
|
|
* allocate additional space.
|
|
|
|
*
|
|
|
|
* This is used when the caller wishes to modify the data and needs a
|
|
|
|
* private copy of the data to alter as well as more space for new fields.
|
|
|
|
* Returns %NULL on failure or the pointer to the buffer
|
|
|
|
* on success. The returned buffer has a reference count of 1.
|
|
|
|
*
|
|
|
|
* You must pass %GFP_ATOMIC as the allocation priority if this function
|
|
|
|
* is called from an interrupt.
|
|
|
|
*/
|
|
|
|
struct sk_buff *skb_copy_expand(const struct sk_buff *skb,
|
2005-07-08 15:57:47 -06:00
|
|
|
int newheadroom, int newtailroom,
|
2005-10-07 00:46:04 -06:00
|
|
|
gfp_t gfp_mask)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
|
|
|
/*
|
|
|
|
* Allocate the copy buffer
|
|
|
|
*/
|
2012-07-31 17:44:19 -06:00
|
|
|
struct sk_buff *n = __alloc_skb(newheadroom + skb->len + newtailroom,
|
|
|
|
gfp_mask, skb_alloc_rx_flag(skb),
|
|
|
|
NUMA_NO_NODE);
|
2007-04-10 19:30:09 -06:00
|
|
|
int oldheadroom = skb_headroom(skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
int head_copy_len, head_copy_off;
|
|
|
|
|
|
|
|
if (!n)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
skb_reserve(n, newheadroom);
|
|
|
|
|
|
|
|
/* Set the tail pointer and length */
|
|
|
|
skb_put(n, skb->len);
|
|
|
|
|
2007-04-10 19:30:09 -06:00
|
|
|
head_copy_len = oldheadroom;
|
2005-04-16 16:20:36 -06:00
|
|
|
head_copy_off = 0;
|
|
|
|
if (newheadroom <= head_copy_len)
|
|
|
|
head_copy_len = newheadroom;
|
|
|
|
else
|
|
|
|
head_copy_off = newheadroom - head_copy_len;
|
|
|
|
|
|
|
|
/* Copy the linear header and data. */
|
|
|
|
if (skb_copy_bits(skb, -head_copy_len, n->head + head_copy_off,
|
|
|
|
skb->len + head_copy_len))
|
|
|
|
BUG();
|
|
|
|
|
|
|
|
copy_skb_header(n, skb);
|
|
|
|
|
2013-10-19 12:42:54 -06:00
|
|
|
skb_headers_offset_update(n, newheadroom - oldheadroom);
|
2007-04-10 19:30:09 -06:00
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
return n;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_copy_expand);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_pad - zero pad the tail of an skb
|
|
|
|
* @skb: buffer to pad
|
|
|
|
* @pad: space to pad
|
|
|
|
*
|
|
|
|
* Ensure that a buffer is followed by a padding area that is zero
|
|
|
|
* filled. Used by network drivers which may DMA or transfer data
|
|
|
|
* beyond the buffer end onto the wire.
|
|
|
|
*
|
2006-06-23 03:06:41 -06:00
|
|
|
* May return error in out of memory cases. The skb is freed on error.
|
2005-04-16 16:20:36 -06:00
|
|
|
*/
|
2007-02-09 07:24:36 -07:00
|
|
|
|
2006-06-23 03:06:41 -06:00
|
|
|
int skb_pad(struct sk_buff *skb, int pad)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2006-06-23 03:06:41 -06:00
|
|
|
int err;
|
|
|
|
int ntail;
|
2007-02-09 07:24:36 -07:00
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
/* If the skbuff is non linear tailroom is always zero.. */
|
2006-06-23 03:06:41 -06:00
|
|
|
if (!skb_cloned(skb) && skb_tailroom(skb) >= pad) {
|
2005-04-16 16:20:36 -06:00
|
|
|
memset(skb->data+skb->len, 0, pad);
|
2006-06-23 03:06:41 -06:00
|
|
|
return 0;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2006-06-23 03:06:41 -06:00
|
|
|
|
2007-04-19 21:43:29 -06:00
|
|
|
ntail = skb->data_len + pad - (skb->end - skb->tail);
|
2006-06-23 03:06:41 -06:00
|
|
|
if (likely(skb_cloned(skb) || ntail > 0)) {
|
|
|
|
err = pskb_expand_head(skb, 0, ntail, GFP_ATOMIC);
|
|
|
|
if (unlikely(err))
|
|
|
|
goto free_skb;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* FIXME: The use of this function with non-linear skb's really needs
|
|
|
|
* to be audited.
|
|
|
|
*/
|
|
|
|
err = skb_linearize(skb);
|
|
|
|
if (unlikely(err))
|
|
|
|
goto free_skb;
|
|
|
|
|
|
|
|
memset(skb->data + skb->len, 0, pad);
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
free_skb:
|
2005-04-16 16:20:36 -06:00
|
|
|
kfree_skb(skb);
|
2006-06-23 03:06:41 -06:00
|
|
|
return err;
|
2007-02-09 07:24:36 -07:00
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_pad);
|
2007-02-09 07:24:36 -07:00
|
|
|
|
2013-11-07 06:18:24 -07:00
|
|
|
/**
|
|
|
|
* pskb_put - add data to the tail of a potentially fragmented buffer
|
|
|
|
* @skb: start of the buffer to use
|
|
|
|
* @tail: tail fragment of the buffer to use
|
|
|
|
* @len: amount of data to add
|
|
|
|
*
|
|
|
|
* This function extends the used data area of the potentially
|
|
|
|
* fragmented buffer. @tail must be the last fragment of @skb -- or
|
|
|
|
* @skb itself. If this would exceed the total buffer size the kernel
|
|
|
|
* will panic. A pointer to the first byte of the extra data is
|
|
|
|
* returned.
|
|
|
|
*/
|
|
|
|
|
|
|
|
unsigned char *pskb_put(struct sk_buff *skb, struct sk_buff *tail, int len)
|
|
|
|
{
|
|
|
|
if (tail != skb) {
|
|
|
|
skb->data_len += len;
|
|
|
|
skb->len += len;
|
|
|
|
}
|
|
|
|
return skb_put(tail, len);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(pskb_put);
|
|
|
|
|
2008-03-27 18:43:41 -06:00
|
|
|
/**
|
|
|
|
* skb_put - add data to a buffer
|
|
|
|
* @skb: buffer to use
|
|
|
|
* @len: amount of data to add
|
|
|
|
*
|
|
|
|
* This function extends the used data area of the buffer. If this would
|
|
|
|
* exceed the total buffer size the kernel will panic. A pointer to the
|
|
|
|
* first byte of the extra data is returned.
|
|
|
|
*/
|
|
|
|
unsigned char *skb_put(struct sk_buff *skb, unsigned int len)
|
|
|
|
{
|
|
|
|
unsigned char *tmp = skb_tail_pointer(skb);
|
|
|
|
SKB_LINEAR_ASSERT(skb);
|
|
|
|
skb->tail += len;
|
|
|
|
skb->len += len;
|
|
|
|
if (unlikely(skb->tail > skb->end))
|
|
|
|
skb_over_panic(skb, len, __builtin_return_address(0));
|
|
|
|
return tmp;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_put);
|
|
|
|
|
2008-03-27 18:52:40 -06:00
|
|
|
/**
|
|
|
|
* skb_push - add data to the start of a buffer
|
|
|
|
* @skb: buffer to use
|
|
|
|
* @len: amount of data to add
|
|
|
|
*
|
|
|
|
* This function extends the used data area of the buffer at the buffer
|
|
|
|
* start. If this would exceed the total buffer headroom the kernel will
|
|
|
|
* panic. A pointer to the first byte of the extra data is returned.
|
|
|
|
*/
|
|
|
|
unsigned char *skb_push(struct sk_buff *skb, unsigned int len)
|
|
|
|
{
|
|
|
|
skb->data -= len;
|
|
|
|
skb->len += len;
|
|
|
|
if (unlikely(skb->data<skb->head))
|
|
|
|
skb_under_panic(skb, len, __builtin_return_address(0));
|
|
|
|
return skb->data;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_push);
|
|
|
|
|
2008-03-27 18:47:24 -06:00
|
|
|
/**
|
|
|
|
* skb_pull - remove data from the start of a buffer
|
|
|
|
* @skb: buffer to use
|
|
|
|
* @len: amount of data to remove
|
|
|
|
*
|
|
|
|
* This function removes data from the start of a buffer, returning
|
|
|
|
* the memory to the headroom. A pointer to the next data in the buffer
|
|
|
|
* is returned. Once the data has been pulled future pushes will overwrite
|
|
|
|
* the old data.
|
|
|
|
*/
|
|
|
|
unsigned char *skb_pull(struct sk_buff *skb, unsigned int len)
|
|
|
|
{
|
2010-05-02 03:21:44 -06:00
|
|
|
return skb_pull_inline(skb, len);
|
2008-03-27 18:47:24 -06:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_pull);
|
|
|
|
|
2008-03-27 18:54:01 -06:00
|
|
|
/**
|
|
|
|
* skb_trim - remove end from a buffer
|
|
|
|
* @skb: buffer to alter
|
|
|
|
* @len: new length
|
|
|
|
*
|
|
|
|
* Cut the length of a buffer down by removing data from the tail. If
|
|
|
|
* the buffer is already under the length specified it is not modified.
|
|
|
|
* The skb must be linear.
|
|
|
|
*/
|
|
|
|
void skb_trim(struct sk_buff *skb, unsigned int len)
|
|
|
|
{
|
|
|
|
if (skb->len > len)
|
|
|
|
__skb_trim(skb, len);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_trim);
|
|
|
|
|
2006-06-09 17:13:38 -06:00
|
|
|
/* Trims skb to length len. It can change skb pointers.
|
2005-04-16 16:20:36 -06:00
|
|
|
*/
|
|
|
|
|
2006-06-09 17:13:38 -06:00
|
|
|
int ___pskb_trim(struct sk_buff *skb, unsigned int len)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2006-07-13 20:26:39 -06:00
|
|
|
struct sk_buff **fragp;
|
|
|
|
struct sk_buff *frag;
|
2005-04-16 16:20:36 -06:00
|
|
|
int offset = skb_headlen(skb);
|
|
|
|
int nfrags = skb_shinfo(skb)->nr_frags;
|
|
|
|
int i;
|
2006-07-13 20:26:39 -06:00
|
|
|
int err;
|
|
|
|
|
|
|
|
if (skb_cloned(skb) &&
|
|
|
|
unlikely((err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC))))
|
|
|
|
return err;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2006-07-30 21:20:28 -06:00
|
|
|
i = 0;
|
|
|
|
if (offset >= len)
|
|
|
|
goto drop_pages;
|
|
|
|
|
|
|
|
for (; i < nfrags; i++) {
|
2011-10-18 15:00:24 -06:00
|
|
|
int end = offset + skb_frag_size(&skb_shinfo(skb)->frags[i]);
|
2006-07-13 20:26:39 -06:00
|
|
|
|
|
|
|
if (end < len) {
|
|
|
|
offset = end;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_set(&skb_shinfo(skb)->frags[i++], len - offset);
|
2006-07-13 20:26:39 -06:00
|
|
|
|
2006-07-30 21:20:28 -06:00
|
|
|
drop_pages:
|
2006-07-13 20:26:39 -06:00
|
|
|
skb_shinfo(skb)->nr_frags = i;
|
|
|
|
|
|
|
|
for (; i < nfrags; i++)
|
2011-08-22 17:44:58 -06:00
|
|
|
skb_frag_unref(skb, i);
|
2006-07-13 20:26:39 -06:00
|
|
|
|
2010-08-23 01:13:46 -06:00
|
|
|
if (skb_has_frag_list(skb))
|
2006-07-13 20:26:39 -06:00
|
|
|
skb_drop_fraglist(skb);
|
2006-07-30 21:20:28 -06:00
|
|
|
goto done;
|
2006-07-13 20:26:39 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
for (fragp = &skb_shinfo(skb)->frag_list; (frag = *fragp);
|
|
|
|
fragp = &frag->next) {
|
|
|
|
int end = offset + frag->len;
|
|
|
|
|
|
|
|
if (skb_shared(frag)) {
|
|
|
|
struct sk_buff *nfrag;
|
|
|
|
|
|
|
|
nfrag = skb_clone(frag, GFP_ATOMIC);
|
|
|
|
if (unlikely(!nfrag))
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
nfrag->next = frag->next;
|
2012-04-18 20:24:53 -06:00
|
|
|
consume_skb(frag);
|
2006-07-13 20:26:39 -06:00
|
|
|
frag = nfrag;
|
|
|
|
*fragp = frag;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2006-07-13 20:26:39 -06:00
|
|
|
|
|
|
|
if (end < len) {
|
|
|
|
offset = end;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (end > len &&
|
|
|
|
unlikely((err = pskb_trim(frag, len - offset))))
|
|
|
|
return err;
|
|
|
|
|
|
|
|
if (frag->next)
|
|
|
|
skb_drop_list(&frag->next);
|
|
|
|
break;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2006-07-30 21:20:28 -06:00
|
|
|
done:
|
2006-07-13 20:26:39 -06:00
|
|
|
if (len > skb_headlen(skb)) {
|
2005-04-16 16:20:36 -06:00
|
|
|
skb->data_len -= skb->len - len;
|
|
|
|
skb->len = len;
|
|
|
|
} else {
|
2006-07-13 20:26:39 -06:00
|
|
|
skb->len = len;
|
|
|
|
skb->data_len = 0;
|
2007-04-19 21:29:13 -06:00
|
|
|
skb_set_tail_pointer(skb, len);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(___pskb_trim);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* __pskb_pull_tail - advance tail of skb header
|
|
|
|
* @skb: buffer to reallocate
|
|
|
|
* @delta: number of bytes to advance tail
|
|
|
|
*
|
|
|
|
* The function makes a sense only on a fragmented &sk_buff,
|
|
|
|
* it expands header moving its tail forward and copying necessary
|
|
|
|
* data from fragmented part.
|
|
|
|
*
|
|
|
|
* &sk_buff MUST have reference count of 1.
|
|
|
|
*
|
|
|
|
* Returns %NULL (and &sk_buff does not change) if pull failed
|
|
|
|
* or value of new tail of skb in the case of success.
|
|
|
|
*
|
|
|
|
* All the pointers pointing into skb header may change and must be
|
|
|
|
* reloaded after call to this function.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/* Moves tail of skb head forward, copying data from fragmented part,
|
|
|
|
* when it is necessary.
|
|
|
|
* 1. It may fail due to malloc failure.
|
|
|
|
* 2. It may change skb pointers.
|
|
|
|
*
|
|
|
|
* It is pretty complicated. Luckily, it is called only in exceptional cases.
|
|
|
|
*/
|
|
|
|
unsigned char *__pskb_pull_tail(struct sk_buff *skb, int delta)
|
|
|
|
{
|
|
|
|
/* If skb has not enough free space at tail, get new one
|
|
|
|
* plus 128 bytes for future expansions. If we have enough
|
|
|
|
* room at tail, reallocate without expansion only if skb is cloned.
|
|
|
|
*/
|
2007-04-19 21:43:29 -06:00
|
|
|
int i, k, eat = (skb->tail + delta) - skb->end;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
if (eat > 0 || skb_cloned(skb)) {
|
|
|
|
if (pskb_expand_head(skb, 0, eat > 0 ? eat + 128 : 0,
|
|
|
|
GFP_ATOMIC))
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2007-04-19 21:29:13 -06:00
|
|
|
if (skb_copy_bits(skb, skb_headlen(skb), skb_tail_pointer(skb), delta))
|
2005-04-16 16:20:36 -06:00
|
|
|
BUG();
|
|
|
|
|
|
|
|
/* Optimization: no fragments, no reasons to preestimate
|
|
|
|
* size of pulled pages. Superb.
|
|
|
|
*/
|
2010-08-23 01:13:46 -06:00
|
|
|
if (!skb_has_frag_list(skb))
|
2005-04-16 16:20:36 -06:00
|
|
|
goto pull_pages;
|
|
|
|
|
|
|
|
/* Estimate size of pulled pages. */
|
|
|
|
eat = delta;
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
|
2011-10-18 15:00:24 -06:00
|
|
|
int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
|
|
|
|
|
|
|
|
if (size >= eat)
|
2005-04-16 16:20:36 -06:00
|
|
|
goto pull_pages;
|
2011-10-18 15:00:24 -06:00
|
|
|
eat -= size;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
/* If we need update frag list, we are in troubles.
|
|
|
|
* Certainly, it possible to add an offset to skb data,
|
|
|
|
* but taking into account that pulling is expected to
|
|
|
|
* be very rare operation, it is worth to fight against
|
|
|
|
* further bloating skb head and crucify ourselves here instead.
|
|
|
|
* Pure masohism, indeed. 8)8)
|
|
|
|
*/
|
|
|
|
if (eat) {
|
|
|
|
struct sk_buff *list = skb_shinfo(skb)->frag_list;
|
|
|
|
struct sk_buff *clone = NULL;
|
|
|
|
struct sk_buff *insp = NULL;
|
|
|
|
|
|
|
|
do {
|
2006-01-08 23:24:28 -07:00
|
|
|
BUG_ON(!list);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
if (list->len <= eat) {
|
|
|
|
/* Eaten as whole. */
|
|
|
|
eat -= list->len;
|
|
|
|
list = list->next;
|
|
|
|
insp = list;
|
|
|
|
} else {
|
|
|
|
/* Eaten partially. */
|
|
|
|
|
|
|
|
if (skb_shared(list)) {
|
|
|
|
/* Sucks! We need to fork list. :-( */
|
|
|
|
clone = skb_clone(list, GFP_ATOMIC);
|
|
|
|
if (!clone)
|
|
|
|
return NULL;
|
|
|
|
insp = list->next;
|
|
|
|
list = clone;
|
|
|
|
} else {
|
|
|
|
/* This may be pulled without
|
|
|
|
* problems. */
|
|
|
|
insp = list;
|
|
|
|
}
|
|
|
|
if (!pskb_pull(list, eat)) {
|
2009-02-24 17:37:32 -07:00
|
|
|
kfree_skb(clone);
|
2005-04-16 16:20:36 -06:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
} while (eat);
|
|
|
|
|
|
|
|
/* Free pulled out fragments. */
|
|
|
|
while ((list = skb_shinfo(skb)->frag_list) != insp) {
|
|
|
|
skb_shinfo(skb)->frag_list = list->next;
|
|
|
|
kfree_skb(list);
|
|
|
|
}
|
|
|
|
/* And insert new clone at head. */
|
|
|
|
if (clone) {
|
|
|
|
clone->next = list;
|
|
|
|
skb_shinfo(skb)->frag_list = clone;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
/* Success! Now we may commit changes to skb data. */
|
|
|
|
|
|
|
|
pull_pages:
|
|
|
|
eat = delta;
|
|
|
|
k = 0;
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
|
2011-10-18 15:00:24 -06:00
|
|
|
int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
|
|
|
|
|
|
|
|
if (size <= eat) {
|
2011-08-22 17:44:58 -06:00
|
|
|
skb_frag_unref(skb, i);
|
2011-10-18 15:00:24 -06:00
|
|
|
eat -= size;
|
2005-04-16 16:20:36 -06:00
|
|
|
} else {
|
|
|
|
skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
|
|
|
|
if (eat) {
|
|
|
|
skb_shinfo(skb)->frags[k].page_offset += eat;
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
|
2005-04-16 16:20:36 -06:00
|
|
|
eat = 0;
|
|
|
|
}
|
|
|
|
k++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
skb_shinfo(skb)->nr_frags = k;
|
|
|
|
|
|
|
|
skb->tail += delta;
|
|
|
|
skb->data_len -= delta;
|
|
|
|
|
2007-04-19 21:29:13 -06:00
|
|
|
return skb_tail_pointer(skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(__pskb_pull_tail);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2011-07-29 12:37:31 -06:00
|
|
|
/**
|
|
|
|
* skb_copy_bits - copy bits from skb to kernel buffer
|
|
|
|
* @skb: source skb
|
|
|
|
* @offset: offset in source
|
|
|
|
* @to: destination buffer
|
|
|
|
* @len: number of bytes to copy
|
|
|
|
*
|
|
|
|
* Copy the specified number of bytes from the source skb to the
|
|
|
|
* destination buffer.
|
|
|
|
*
|
|
|
|
* CAUTION ! :
|
|
|
|
* If its prototype is ever changed,
|
|
|
|
* check arch/{*}/net/{*}.S files,
|
|
|
|
* since it is called from BPF assembly code.
|
|
|
|
*/
|
2005-04-16 16:20:36 -06:00
|
|
|
int skb_copy_bits(const struct sk_buff *skb, int offset, void *to, int len)
|
|
|
|
{
|
2007-04-27 16:21:23 -06:00
|
|
|
int start = skb_headlen(skb);
|
2009-06-09 01:18:59 -06:00
|
|
|
struct sk_buff *frag_iter;
|
|
|
|
int i, copy;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
if (offset > (int)skb->len - len)
|
|
|
|
goto fault;
|
|
|
|
|
|
|
|
/* Copy header. */
|
2007-04-27 16:21:23 -06:00
|
|
|
if ((copy = start - offset) > 0) {
|
2005-04-16 16:20:36 -06:00
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
2007-03-27 15:55:52 -06:00
|
|
|
skb_copy_from_linear_data_offset(skb, offset, to, copy);
|
2005-04-16 16:20:36 -06:00
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return 0;
|
|
|
|
offset += copy;
|
|
|
|
to += copy;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
|
2007-04-27 16:21:23 -06:00
|
|
|
int end;
|
2012-04-05 03:35:15 -06:00
|
|
|
skb_frag_t *f = &skb_shinfo(skb)->frags[i];
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2008-07-25 22:43:18 -06:00
|
|
|
WARN_ON(start > offset + len);
|
2007-04-27 16:21:23 -06:00
|
|
|
|
2012-04-05 03:35:15 -06:00
|
|
|
end = start + skb_frag_size(f);
|
2005-04-16 16:20:36 -06:00
|
|
|
if ((copy = end - offset) > 0) {
|
|
|
|
u8 *vaddr;
|
|
|
|
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
|
|
|
|
2012-04-05 03:35:15 -06:00
|
|
|
vaddr = kmap_atomic(skb_frag_page(f));
|
2005-04-16 16:20:36 -06:00
|
|
|
memcpy(to,
|
2012-04-05 03:35:15 -06:00
|
|
|
vaddr + f->page_offset + offset - start,
|
|
|
|
copy);
|
|
|
|
kunmap_atomic(vaddr);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return 0;
|
|
|
|
offset += copy;
|
|
|
|
to += copy;
|
|
|
|
}
|
2007-04-27 16:21:23 -06:00
|
|
|
start = end;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
skb_walk_frags(skb, frag_iter) {
|
|
|
|
int end;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
WARN_ON(start > offset + len);
|
|
|
|
|
|
|
|
end = start + frag_iter->len;
|
|
|
|
if ((copy = end - offset) > 0) {
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
|
|
|
if (skb_copy_bits(frag_iter, offset - start, to, copy))
|
|
|
|
goto fault;
|
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return 0;
|
|
|
|
offset += copy;
|
|
|
|
to += copy;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2009-06-09 01:18:59 -06:00
|
|
|
start = end;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2011-07-06 06:22:12 -06:00
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
if (!len)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
fault:
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_copy_bits);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2007-11-07 00:30:13 -07:00
|
|
|
/*
|
|
|
|
* Callback from splice_to_pipe(), if we need to release some pages
|
|
|
|
* at the end of the spd in case we error'ed out in filling the pipe.
|
|
|
|
*/
|
|
|
|
static void sock_spd_release(struct splice_pipe_desc *spd, unsigned int i)
|
|
|
|
{
|
2009-01-19 18:03:56 -07:00
|
|
|
put_page(spd->pages[i]);
|
|
|
|
}
|
2007-11-07 00:30:13 -07:00
|
|
|
|
2012-04-23 21:06:11 -06:00
|
|
|
static struct page *linear_to_page(struct page *page, unsigned int *len,
|
|
|
|
unsigned int *offset,
|
2013-01-11 07:46:37 -07:00
|
|
|
struct sock *sk)
|
2009-01-19 18:03:56 -07:00
|
|
|
{
|
net: use a per task frag allocator
We currently use a per socket order-0 page cache for tcp_sendmsg()
operations.
This page is used to build fragments for skbs.
Its done to increase probability of coalescing small write() into
single segments in skbs still in write queue (not yet sent)
But it wastes a lot of memory for applications handling many mostly
idle sockets, since each socket holds one page in sk->sk_sndmsg_page
Its also quite inefficient to build TSO 64KB packets, because we need
about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit
page allocator more than wanted.
This patch adds a per task frag allocator and uses bigger pages,
if available. An automatic fallback is done in case of memory pressure.
(up to 32768 bytes per frag, thats order-3 pages on x86)
This increases TCP stream performance by 20% on loopback device,
but also benefits on other network devices, since 8x less frags are
mapped on transmit and unmapped on tx completion. Alexander Duyck
mentioned a probable performance win on systems with IOMMU enabled.
Its possible some SG enabled hardware cant cope with bigger fragments,
but their ndo_start_xmit() should already handle this, splitting a
fragment in sub fragments, since some arches have PAGE_SIZE=65536
Successfully tested on various ethernet devices.
(ixgbe, igb, bnx2x, tg3, mellanox mlx4)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Vijay Subramanian <subramanian.vijay@gmail.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-09-23 17:04:42 -06:00
|
|
|
struct page_frag *pfrag = sk_page_frag(sk);
|
2009-02-01 01:41:42 -07:00
|
|
|
|
net: use a per task frag allocator
We currently use a per socket order-0 page cache for tcp_sendmsg()
operations.
This page is used to build fragments for skbs.
Its done to increase probability of coalescing small write() into
single segments in skbs still in write queue (not yet sent)
But it wastes a lot of memory for applications handling many mostly
idle sockets, since each socket holds one page in sk->sk_sndmsg_page
Its also quite inefficient to build TSO 64KB packets, because we need
about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit
page allocator more than wanted.
This patch adds a per task frag allocator and uses bigger pages,
if available. An automatic fallback is done in case of memory pressure.
(up to 32768 bytes per frag, thats order-3 pages on x86)
This increases TCP stream performance by 20% on loopback device,
but also benefits on other network devices, since 8x less frags are
mapped on transmit and unmapped on tx completion. Alexander Duyck
mentioned a probable performance win on systems with IOMMU enabled.
Its possible some SG enabled hardware cant cope with bigger fragments,
but their ndo_start_xmit() should already handle this, splitting a
fragment in sub fragments, since some arches have PAGE_SIZE=65536
Successfully tested on various ethernet devices.
(ixgbe, igb, bnx2x, tg3, mellanox mlx4)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Vijay Subramanian <subramanian.vijay@gmail.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-09-23 17:04:42 -06:00
|
|
|
if (!sk_page_frag_refill(sk, pfrag))
|
|
|
|
return NULL;
|
2009-02-01 01:41:42 -07:00
|
|
|
|
net: use a per task frag allocator
We currently use a per socket order-0 page cache for tcp_sendmsg()
operations.
This page is used to build fragments for skbs.
Its done to increase probability of coalescing small write() into
single segments in skbs still in write queue (not yet sent)
But it wastes a lot of memory for applications handling many mostly
idle sockets, since each socket holds one page in sk->sk_sndmsg_page
Its also quite inefficient to build TSO 64KB packets, because we need
about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit
page allocator more than wanted.
This patch adds a per task frag allocator and uses bigger pages,
if available. An automatic fallback is done in case of memory pressure.
(up to 32768 bytes per frag, thats order-3 pages on x86)
This increases TCP stream performance by 20% on loopback device,
but also benefits on other network devices, since 8x less frags are
mapped on transmit and unmapped on tx completion. Alexander Duyck
mentioned a probable performance win on systems with IOMMU enabled.
Its possible some SG enabled hardware cant cope with bigger fragments,
but their ndo_start_xmit() should already handle this, splitting a
fragment in sub fragments, since some arches have PAGE_SIZE=65536
Successfully tested on various ethernet devices.
(ixgbe, igb, bnx2x, tg3, mellanox mlx4)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Vijay Subramanian <subramanian.vijay@gmail.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-09-23 17:04:42 -06:00
|
|
|
*len = min_t(unsigned int, *len, pfrag->size - pfrag->offset);
|
2009-02-01 01:41:42 -07:00
|
|
|
|
net: use a per task frag allocator
We currently use a per socket order-0 page cache for tcp_sendmsg()
operations.
This page is used to build fragments for skbs.
Its done to increase probability of coalescing small write() into
single segments in skbs still in write queue (not yet sent)
But it wastes a lot of memory for applications handling many mostly
idle sockets, since each socket holds one page in sk->sk_sndmsg_page
Its also quite inefficient to build TSO 64KB packets, because we need
about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit
page allocator more than wanted.
This patch adds a per task frag allocator and uses bigger pages,
if available. An automatic fallback is done in case of memory pressure.
(up to 32768 bytes per frag, thats order-3 pages on x86)
This increases TCP stream performance by 20% on loopback device,
but also benefits on other network devices, since 8x less frags are
mapped on transmit and unmapped on tx completion. Alexander Duyck
mentioned a probable performance win on systems with IOMMU enabled.
Its possible some SG enabled hardware cant cope with bigger fragments,
but their ndo_start_xmit() should already handle this, splitting a
fragment in sub fragments, since some arches have PAGE_SIZE=65536
Successfully tested on various ethernet devices.
(ixgbe, igb, bnx2x, tg3, mellanox mlx4)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Vijay Subramanian <subramanian.vijay@gmail.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-09-23 17:04:42 -06:00
|
|
|
memcpy(page_address(pfrag->page) + pfrag->offset,
|
|
|
|
page_address(page) + *offset, *len);
|
|
|
|
*offset = pfrag->offset;
|
|
|
|
pfrag->offset += *len;
|
2009-01-19 18:03:56 -07:00
|
|
|
|
net: use a per task frag allocator
We currently use a per socket order-0 page cache for tcp_sendmsg()
operations.
This page is used to build fragments for skbs.
Its done to increase probability of coalescing small write() into
single segments in skbs still in write queue (not yet sent)
But it wastes a lot of memory for applications handling many mostly
idle sockets, since each socket holds one page in sk->sk_sndmsg_page
Its also quite inefficient to build TSO 64KB packets, because we need
about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit
page allocator more than wanted.
This patch adds a per task frag allocator and uses bigger pages,
if available. An automatic fallback is done in case of memory pressure.
(up to 32768 bytes per frag, thats order-3 pages on x86)
This increases TCP stream performance by 20% on loopback device,
but also benefits on other network devices, since 8x less frags are
mapped on transmit and unmapped on tx completion. Alexander Duyck
mentioned a probable performance win on systems with IOMMU enabled.
Its possible some SG enabled hardware cant cope with bigger fragments,
but their ndo_start_xmit() should already handle this, splitting a
fragment in sub fragments, since some arches have PAGE_SIZE=65536
Successfully tested on various ethernet devices.
(ixgbe, igb, bnx2x, tg3, mellanox mlx4)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Vijay Subramanian <subramanian.vijay@gmail.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-09-23 17:04:42 -06:00
|
|
|
return pfrag->page;
|
2007-11-07 00:30:13 -07:00
|
|
|
}
|
|
|
|
|
2012-04-22 06:26:16 -06:00
|
|
|
static bool spd_can_coalesce(const struct splice_pipe_desc *spd,
|
|
|
|
struct page *page,
|
|
|
|
unsigned int offset)
|
|
|
|
{
|
|
|
|
return spd->nr_pages &&
|
|
|
|
spd->pages[spd->nr_pages - 1] == page &&
|
|
|
|
(spd->partial[spd->nr_pages - 1].offset +
|
|
|
|
spd->partial[spd->nr_pages - 1].len == offset);
|
|
|
|
}
|
|
|
|
|
2007-11-07 00:30:13 -07:00
|
|
|
/*
|
|
|
|
* Fill page/offset/length into spd, if it can hold more pages.
|
|
|
|
*/
|
2012-04-23 21:06:11 -06:00
|
|
|
static bool spd_fill_page(struct splice_pipe_desc *spd,
|
|
|
|
struct pipe_inode_info *pipe, struct page *page,
|
|
|
|
unsigned int *len, unsigned int offset,
|
2013-01-11 07:46:37 -07:00
|
|
|
bool linear,
|
2012-04-23 21:06:11 -06:00
|
|
|
struct sock *sk)
|
2007-11-07 00:30:13 -07:00
|
|
|
{
|
2012-04-22 06:26:16 -06:00
|
|
|
if (unlikely(spd->nr_pages == MAX_SKB_FRAGS))
|
2012-04-23 21:06:11 -06:00
|
|
|
return true;
|
2007-11-07 00:30:13 -07:00
|
|
|
|
2009-01-19 18:03:56 -07:00
|
|
|
if (linear) {
|
2013-01-11 07:46:37 -07:00
|
|
|
page = linear_to_page(page, len, &offset, sk);
|
2009-01-19 18:03:56 -07:00
|
|
|
if (!page)
|
2012-04-23 21:06:11 -06:00
|
|
|
return true;
|
2012-04-22 06:26:16 -06:00
|
|
|
}
|
|
|
|
if (spd_can_coalesce(spd, page, offset)) {
|
|
|
|
spd->partial[spd->nr_pages - 1].len += *len;
|
2012-04-23 21:06:11 -06:00
|
|
|
return false;
|
2012-04-22 06:26:16 -06:00
|
|
|
}
|
|
|
|
get_page(page);
|
2007-11-07 00:30:13 -07:00
|
|
|
spd->pages[spd->nr_pages] = page;
|
2009-02-01 01:41:42 -07:00
|
|
|
spd->partial[spd->nr_pages].len = *len;
|
2007-11-07 00:30:13 -07:00
|
|
|
spd->partial[spd->nr_pages].offset = offset;
|
|
|
|
spd->nr_pages++;
|
2009-01-19 18:03:56 -07:00
|
|
|
|
2012-04-23 21:06:11 -06:00
|
|
|
return false;
|
2007-11-07 00:30:13 -07:00
|
|
|
}
|
|
|
|
|
2012-04-23 21:06:11 -06:00
|
|
|
static bool __splice_segment(struct page *page, unsigned int poff,
|
|
|
|
unsigned int plen, unsigned int *off,
|
2013-01-11 07:46:37 -07:00
|
|
|
unsigned int *len,
|
2012-04-23 21:35:04 -06:00
|
|
|
struct splice_pipe_desc *spd, bool linear,
|
2012-04-23 21:06:11 -06:00
|
|
|
struct sock *sk,
|
|
|
|
struct pipe_inode_info *pipe)
|
2007-11-07 00:30:13 -07:00
|
|
|
{
|
2008-07-15 01:49:11 -06:00
|
|
|
if (!*len)
|
2012-04-23 21:06:11 -06:00
|
|
|
return true;
|
2008-07-15 01:49:11 -06:00
|
|
|
|
|
|
|
/* skip this segment if already processed */
|
|
|
|
if (*off >= plen) {
|
|
|
|
*off -= plen;
|
2012-04-23 21:06:11 -06:00
|
|
|
return false;
|
2008-06-27 18:27:21 -06:00
|
|
|
}
|
2007-11-07 00:30:13 -07:00
|
|
|
|
2008-07-15 01:49:11 -06:00
|
|
|
/* ignore any bits we already processed */
|
2013-01-05 14:31:18 -07:00
|
|
|
poff += *off;
|
|
|
|
plen -= *off;
|
|
|
|
*off = 0;
|
2007-11-07 00:30:13 -07:00
|
|
|
|
2013-01-11 07:46:37 -07:00
|
|
|
do {
|
|
|
|
unsigned int flen = min(*len, plen);
|
2008-07-15 01:49:11 -06:00
|
|
|
|
2013-01-11 07:46:37 -07:00
|
|
|
if (spd_fill_page(spd, pipe, page, &flen, poff,
|
|
|
|
linear, sk))
|
|
|
|
return true;
|
|
|
|
poff += flen;
|
|
|
|
plen -= flen;
|
|
|
|
*len -= flen;
|
|
|
|
} while (*len && plen);
|
2008-07-15 01:49:11 -06:00
|
|
|
|
2012-04-23 21:06:11 -06:00
|
|
|
return false;
|
2008-07-15 01:49:11 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2012-04-23 21:06:11 -06:00
|
|
|
* Map linear and fragment data from the skb to spd. It reports true if the
|
2008-07-15 01:49:11 -06:00
|
|
|
* pipe is full or if we already spliced the requested length.
|
|
|
|
*/
|
2012-04-23 21:06:11 -06:00
|
|
|
static bool __skb_splice_bits(struct sk_buff *skb, struct pipe_inode_info *pipe,
|
|
|
|
unsigned int *offset, unsigned int *len,
|
|
|
|
struct splice_pipe_desc *spd, struct sock *sk)
|
2008-07-15 01:49:11 -06:00
|
|
|
{
|
|
|
|
int seg;
|
2016-03-07 15:11:04 -07:00
|
|
|
struct sk_buff *iter;
|
2008-07-15 01:49:11 -06:00
|
|
|
|
2012-04-26 20:10:03 -06:00
|
|
|
/* map the linear part :
|
2012-05-02 12:18:42 -06:00
|
|
|
* If skb->head_frag is set, this 'linear' part is backed by a
|
|
|
|
* fragment, and if the head is not shared with any clones then
|
|
|
|
* we can avoid a copy since we own the head portion of this page.
|
2008-07-15 01:49:11 -06:00
|
|
|
*/
|
|
|
|
if (__splice_segment(virt_to_page(skb->data),
|
|
|
|
(unsigned long) skb->data & (PAGE_SIZE - 1),
|
|
|
|
skb_headlen(skb),
|
2013-01-11 07:46:37 -07:00
|
|
|
offset, len, spd,
|
2012-05-02 19:09:42 -06:00
|
|
|
skb_head_is_locked(skb),
|
2012-04-26 20:10:03 -06:00
|
|
|
sk, pipe))
|
2012-04-23 21:06:11 -06:00
|
|
|
return true;
|
2007-11-07 00:30:13 -07:00
|
|
|
|
|
|
|
/*
|
|
|
|
* then map the fragments
|
|
|
|
*/
|
|
|
|
for (seg = 0; seg < skb_shinfo(skb)->nr_frags; seg++) {
|
|
|
|
const skb_frag_t *f = &skb_shinfo(skb)->frags[seg];
|
|
|
|
|
2011-08-22 17:44:58 -06:00
|
|
|
if (__splice_segment(skb_frag_page(f),
|
2011-10-18 15:00:24 -06:00
|
|
|
f->page_offset, skb_frag_size(f),
|
2013-01-11 07:46:37 -07:00
|
|
|
offset, len, spd, false, sk, pipe))
|
2012-04-23 21:06:11 -06:00
|
|
|
return true;
|
2007-11-07 00:30:13 -07:00
|
|
|
}
|
|
|
|
|
2016-03-07 15:11:04 -07:00
|
|
|
skb_walk_frags(skb, iter) {
|
|
|
|
if (*offset >= iter->len) {
|
|
|
|
*offset -= iter->len;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
/* __skb_splice_bits() only fails if the output has no room
|
|
|
|
* left, so no point in going over the frag_list for the error
|
|
|
|
* case.
|
|
|
|
*/
|
|
|
|
if (__skb_splice_bits(iter, pipe, offset, len, spd, sk))
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2012-04-23 21:06:11 -06:00
|
|
|
return false;
|
2007-11-07 00:30:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Map data from the skb to a pipe. Should handle both the linear part,
|
2016-03-07 15:11:04 -07:00
|
|
|
* the fragments, and the frag list.
|
2007-11-07 00:30:13 -07:00
|
|
|
*/
|
2015-05-21 09:00:00 -06:00
|
|
|
int skb_splice_bits(struct sk_buff *skb, struct sock *sk, unsigned int offset,
|
2007-11-07 00:30:13 -07:00
|
|
|
struct pipe_inode_info *pipe, unsigned int tlen,
|
2016-09-17 19:02:10 -06:00
|
|
|
unsigned int flags)
|
2007-11-07 00:30:13 -07:00
|
|
|
{
|
2012-04-22 06:26:16 -06:00
|
|
|
struct partial_page partial[MAX_SKB_FRAGS];
|
|
|
|
struct page *pages[MAX_SKB_FRAGS];
|
2007-11-07 00:30:13 -07:00
|
|
|
struct splice_pipe_desc spd = {
|
|
|
|
.pages = pages,
|
|
|
|
.partial = partial,
|
2012-06-12 07:24:40 -06:00
|
|
|
.nr_pages_max = MAX_SKB_FRAGS,
|
2007-11-07 00:30:13 -07:00
|
|
|
.flags = flags,
|
2014-01-22 11:36:57 -07:00
|
|
|
.ops = &nosteal_pipe_buf_ops,
|
2007-11-07 00:30:13 -07:00
|
|
|
.spd_release = sock_spd_release,
|
|
|
|
};
|
2010-05-20 02:43:18 -06:00
|
|
|
int ret = 0;
|
|
|
|
|
2016-03-07 15:11:04 -07:00
|
|
|
__skb_splice_bits(skb, pipe, &offset, &tlen, &spd, sk);
|
2007-11-07 00:30:13 -07:00
|
|
|
|
2015-05-21 09:00:00 -06:00
|
|
|
if (spd.nr_pages)
|
2016-09-17 19:02:10 -06:00
|
|
|
ret = splice_to_pipe(pipe, &spd);
|
2007-11-07 00:30:13 -07:00
|
|
|
|
2010-05-20 02:43:18 -06:00
|
|
|
return ret;
|
2007-11-07 00:30:13 -07:00
|
|
|
}
|
2015-05-21 09:00:01 -06:00
|
|
|
EXPORT_SYMBOL_GPL(skb_splice_bits);
|
2007-11-07 00:30:13 -07:00
|
|
|
|
2005-04-19 23:30:14 -06:00
|
|
|
/**
|
|
|
|
* skb_store_bits - store bits from kernel buffer to skb
|
|
|
|
* @skb: destination buffer
|
|
|
|
* @offset: offset in destination
|
|
|
|
* @from: source buffer
|
|
|
|
* @len: number of bytes to copy
|
|
|
|
*
|
|
|
|
* Copy the specified number of bytes from the source buffer to the
|
|
|
|
* destination skb. This function handles all the messy bits of
|
|
|
|
* traversing fragment lists and such.
|
|
|
|
*/
|
|
|
|
|
2007-04-20 17:40:01 -06:00
|
|
|
int skb_store_bits(struct sk_buff *skb, int offset, const void *from, int len)
|
2005-04-19 23:30:14 -06:00
|
|
|
{
|
2007-04-27 16:21:23 -06:00
|
|
|
int start = skb_headlen(skb);
|
2009-06-09 01:18:59 -06:00
|
|
|
struct sk_buff *frag_iter;
|
|
|
|
int i, copy;
|
2005-04-19 23:30:14 -06:00
|
|
|
|
|
|
|
if (offset > (int)skb->len - len)
|
|
|
|
goto fault;
|
|
|
|
|
2007-04-27 16:21:23 -06:00
|
|
|
if ((copy = start - offset) > 0) {
|
2005-04-19 23:30:14 -06:00
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
2007-03-31 08:55:19 -06:00
|
|
|
skb_copy_to_linear_data_offset(skb, offset, from, copy);
|
2005-04-19 23:30:14 -06:00
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return 0;
|
|
|
|
offset += copy;
|
|
|
|
from += copy;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
|
|
|
|
skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
|
2007-04-27 16:21:23 -06:00
|
|
|
int end;
|
|
|
|
|
2008-07-25 22:43:18 -06:00
|
|
|
WARN_ON(start > offset + len);
|
2005-04-19 23:30:14 -06:00
|
|
|
|
2011-10-18 15:00:24 -06:00
|
|
|
end = start + skb_frag_size(frag);
|
2005-04-19 23:30:14 -06:00
|
|
|
if ((copy = end - offset) > 0) {
|
|
|
|
u8 *vaddr;
|
|
|
|
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
|
|
|
|
2012-04-05 03:35:15 -06:00
|
|
|
vaddr = kmap_atomic(skb_frag_page(frag));
|
2007-04-27 16:21:23 -06:00
|
|
|
memcpy(vaddr + frag->page_offset + offset - start,
|
|
|
|
from, copy);
|
2012-04-05 03:35:15 -06:00
|
|
|
kunmap_atomic(vaddr);
|
2005-04-19 23:30:14 -06:00
|
|
|
|
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return 0;
|
|
|
|
offset += copy;
|
|
|
|
from += copy;
|
|
|
|
}
|
2007-04-27 16:21:23 -06:00
|
|
|
start = end;
|
2005-04-19 23:30:14 -06:00
|
|
|
}
|
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
skb_walk_frags(skb, frag_iter) {
|
|
|
|
int end;
|
2005-04-19 23:30:14 -06:00
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
WARN_ON(start > offset + len);
|
|
|
|
|
|
|
|
end = start + frag_iter->len;
|
|
|
|
if ((copy = end - offset) > 0) {
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
|
|
|
if (skb_store_bits(frag_iter, offset - start,
|
|
|
|
from, copy))
|
|
|
|
goto fault;
|
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return 0;
|
|
|
|
offset += copy;
|
|
|
|
from += copy;
|
2005-04-19 23:30:14 -06:00
|
|
|
}
|
2009-06-09 01:18:59 -06:00
|
|
|
start = end;
|
2005-04-19 23:30:14 -06:00
|
|
|
}
|
|
|
|
if (!len)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
fault:
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_store_bits);
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
/* Checksum skb data. */
|
2013-10-30 04:50:51 -06:00
|
|
|
__wsum __skb_checksum(const struct sk_buff *skb, int offset, int len,
|
|
|
|
__wsum csum, const struct skb_checksum_ops *ops)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2007-04-27 16:21:23 -06:00
|
|
|
int start = skb_headlen(skb);
|
|
|
|
int i, copy = start - offset;
|
2009-06-09 01:18:59 -06:00
|
|
|
struct sk_buff *frag_iter;
|
2005-04-16 16:20:36 -06:00
|
|
|
int pos = 0;
|
|
|
|
|
|
|
|
/* Checksum header. */
|
|
|
|
if (copy > 0) {
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
2013-10-30 04:50:51 -06:00
|
|
|
csum = ops->update(skb->data + offset, copy, csum);
|
2005-04-16 16:20:36 -06:00
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return csum;
|
|
|
|
offset += copy;
|
|
|
|
pos = copy;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
|
2007-04-27 16:21:23 -06:00
|
|
|
int end;
|
2012-04-05 03:35:15 -06:00
|
|
|
skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
|
2007-04-27 16:21:23 -06:00
|
|
|
|
2008-07-25 22:43:18 -06:00
|
|
|
WARN_ON(start > offset + len);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2012-04-05 03:35:15 -06:00
|
|
|
end = start + skb_frag_size(frag);
|
2005-04-16 16:20:36 -06:00
|
|
|
if ((copy = end - offset) > 0) {
|
2006-11-14 22:36:14 -07:00
|
|
|
__wsum csum2;
|
2005-04-16 16:20:36 -06:00
|
|
|
u8 *vaddr;
|
|
|
|
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
2012-04-05 03:35:15 -06:00
|
|
|
vaddr = kmap_atomic(skb_frag_page(frag));
|
2013-10-30 04:50:51 -06:00
|
|
|
csum2 = ops->update(vaddr + frag->page_offset +
|
|
|
|
offset - start, copy, 0);
|
2012-04-05 03:35:15 -06:00
|
|
|
kunmap_atomic(vaddr);
|
2013-10-30 04:50:51 -06:00
|
|
|
csum = ops->combine(csum, csum2, pos, copy);
|
2005-04-16 16:20:36 -06:00
|
|
|
if (!(len -= copy))
|
|
|
|
return csum;
|
|
|
|
offset += copy;
|
|
|
|
pos += copy;
|
|
|
|
}
|
2007-04-27 16:21:23 -06:00
|
|
|
start = end;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
skb_walk_frags(skb, frag_iter) {
|
|
|
|
int end;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
WARN_ON(start > offset + len);
|
|
|
|
|
|
|
|
end = start + frag_iter->len;
|
|
|
|
if ((copy = end - offset) > 0) {
|
|
|
|
__wsum csum2;
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
2013-10-30 04:50:51 -06:00
|
|
|
csum2 = __skb_checksum(frag_iter, offset - start,
|
|
|
|
copy, 0, ops);
|
|
|
|
csum = ops->combine(csum, csum2, pos, copy);
|
2009-06-09 01:18:59 -06:00
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return csum;
|
|
|
|
offset += copy;
|
|
|
|
pos += copy;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2009-06-09 01:18:59 -06:00
|
|
|
start = end;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2006-01-08 23:24:28 -07:00
|
|
|
BUG_ON(len);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
return csum;
|
|
|
|
}
|
2013-10-30 04:50:51 -06:00
|
|
|
EXPORT_SYMBOL(__skb_checksum);
|
|
|
|
|
|
|
|
__wsum skb_checksum(const struct sk_buff *skb, int offset,
|
|
|
|
int len, __wsum csum)
|
|
|
|
{
|
|
|
|
const struct skb_checksum_ops ops = {
|
2013-11-04 09:10:25 -07:00
|
|
|
.update = csum_partial_ext,
|
2013-10-30 04:50:51 -06:00
|
|
|
.combine = csum_block_add_ext,
|
|
|
|
};
|
|
|
|
|
|
|
|
return __skb_checksum(skb, offset, len, csum, &ops);
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_checksum);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/* Both of above in one bottle. */
|
|
|
|
|
2006-11-14 22:37:33 -07:00
|
|
|
__wsum skb_copy_and_csum_bits(const struct sk_buff *skb, int offset,
|
|
|
|
u8 *to, int len, __wsum csum)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2007-04-27 16:21:23 -06:00
|
|
|
int start = skb_headlen(skb);
|
|
|
|
int i, copy = start - offset;
|
2009-06-09 01:18:59 -06:00
|
|
|
struct sk_buff *frag_iter;
|
2005-04-16 16:20:36 -06:00
|
|
|
int pos = 0;
|
|
|
|
|
|
|
|
/* Copy header. */
|
|
|
|
if (copy > 0) {
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
|
|
|
csum = csum_partial_copy_nocheck(skb->data + offset, to,
|
|
|
|
copy, csum);
|
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return csum;
|
|
|
|
offset += copy;
|
|
|
|
to += copy;
|
|
|
|
pos = copy;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
|
2007-04-27 16:21:23 -06:00
|
|
|
int end;
|
|
|
|
|
2008-07-25 22:43:18 -06:00
|
|
|
WARN_ON(start > offset + len);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2011-10-18 15:00:24 -06:00
|
|
|
end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
|
2005-04-16 16:20:36 -06:00
|
|
|
if ((copy = end - offset) > 0) {
|
2006-11-14 22:36:34 -07:00
|
|
|
__wsum csum2;
|
2005-04-16 16:20:36 -06:00
|
|
|
u8 *vaddr;
|
|
|
|
skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
|
|
|
|
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
2012-04-05 03:35:15 -06:00
|
|
|
vaddr = kmap_atomic(skb_frag_page(frag));
|
2005-04-16 16:20:36 -06:00
|
|
|
csum2 = csum_partial_copy_nocheck(vaddr +
|
2007-04-27 16:21:23 -06:00
|
|
|
frag->page_offset +
|
|
|
|
offset - start, to,
|
|
|
|
copy, 0);
|
2012-04-05 03:35:15 -06:00
|
|
|
kunmap_atomic(vaddr);
|
2005-04-16 16:20:36 -06:00
|
|
|
csum = csum_block_add(csum, csum2, pos);
|
|
|
|
if (!(len -= copy))
|
|
|
|
return csum;
|
|
|
|
offset += copy;
|
|
|
|
to += copy;
|
|
|
|
pos += copy;
|
|
|
|
}
|
2007-04-27 16:21:23 -06:00
|
|
|
start = end;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
skb_walk_frags(skb, frag_iter) {
|
|
|
|
__wsum csum2;
|
|
|
|
int end;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
WARN_ON(start > offset + len);
|
|
|
|
|
|
|
|
end = start + frag_iter->len;
|
|
|
|
if ((copy = end - offset) > 0) {
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
|
|
|
csum2 = skb_copy_and_csum_bits(frag_iter,
|
|
|
|
offset - start,
|
|
|
|
to, copy, 0);
|
|
|
|
csum = csum_block_add(csum, csum2, pos);
|
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return csum;
|
|
|
|
offset += copy;
|
|
|
|
to += copy;
|
|
|
|
pos += copy;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2009-06-09 01:18:59 -06:00
|
|
|
start = end;
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2006-01-08 23:24:28 -07:00
|
|
|
BUG_ON(len);
|
2005-04-16 16:20:36 -06:00
|
|
|
return csum;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_copy_and_csum_bits);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2013-12-13 07:22:17 -07:00
|
|
|
/**
|
|
|
|
* skb_zerocopy_headlen - Calculate headroom needed for skb_zerocopy()
|
|
|
|
* @from: source buffer
|
|
|
|
*
|
|
|
|
* Calculates the amount of linear headroom needed in the 'to' skb passed
|
|
|
|
* into skb_zerocopy().
|
|
|
|
*/
|
|
|
|
unsigned int
|
|
|
|
skb_zerocopy_headlen(const struct sk_buff *from)
|
|
|
|
{
|
|
|
|
unsigned int hlen = 0;
|
|
|
|
|
|
|
|
if (!from->head_frag ||
|
|
|
|
skb_headlen(from) < L1_CACHE_BYTES ||
|
|
|
|
skb_shinfo(from)->nr_frags >= MAX_SKB_FRAGS)
|
|
|
|
hlen = skb_headlen(from);
|
|
|
|
|
|
|
|
if (skb_has_frag_list(from))
|
|
|
|
hlen = from->len;
|
|
|
|
|
|
|
|
return hlen;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_zerocopy_headlen);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_zerocopy - Zero copy skb to skb
|
|
|
|
* @to: destination buffer
|
2014-01-28 09:05:28 -07:00
|
|
|
* @from: source buffer
|
2013-12-13 07:22:17 -07:00
|
|
|
* @len: number of bytes to copy from source buffer
|
|
|
|
* @hlen: size of linear headroom in destination buffer
|
|
|
|
*
|
|
|
|
* Copies up to `len` bytes from `from` to `to` by creating references
|
|
|
|
* to the frags in the source buffer.
|
|
|
|
*
|
|
|
|
* The `hlen` as calculated by skb_zerocopy_headlen() specifies the
|
|
|
|
* headroom in the `to` buffer.
|
2014-03-26 16:37:45 -06:00
|
|
|
*
|
|
|
|
* Return value:
|
|
|
|
* 0: everything is OK
|
|
|
|
* -ENOMEM: couldn't orphan frags of @from due to lack of memory
|
|
|
|
* -EFAULT: skb_copy_bits() found some problem with skb geometry
|
2013-12-13 07:22:17 -07:00
|
|
|
*/
|
2014-03-26 16:37:45 -06:00
|
|
|
int
|
|
|
|
skb_zerocopy(struct sk_buff *to, struct sk_buff *from, int len, int hlen)
|
2013-12-13 07:22:17 -07:00
|
|
|
{
|
|
|
|
int i, j = 0;
|
|
|
|
int plen = 0; /* length of skb->head fragment */
|
2014-03-26 16:37:45 -06:00
|
|
|
int ret;
|
2013-12-13 07:22:17 -07:00
|
|
|
struct page *page;
|
|
|
|
unsigned int offset;
|
|
|
|
|
|
|
|
BUG_ON(!from->head_frag && !hlen);
|
|
|
|
|
|
|
|
/* dont bother with small payloads */
|
2014-03-26 16:37:45 -06:00
|
|
|
if (len <= skb_tailroom(to))
|
|
|
|
return skb_copy_bits(from, 0, skb_put(to, len), len);
|
2013-12-13 07:22:17 -07:00
|
|
|
|
|
|
|
if (hlen) {
|
2014-03-26 16:37:45 -06:00
|
|
|
ret = skb_copy_bits(from, 0, skb_put(to, hlen), hlen);
|
|
|
|
if (unlikely(ret))
|
|
|
|
return ret;
|
2013-12-13 07:22:17 -07:00
|
|
|
len -= hlen;
|
|
|
|
} else {
|
|
|
|
plen = min_t(int, skb_headlen(from), len);
|
|
|
|
if (plen) {
|
|
|
|
page = virt_to_head_page(from->head);
|
|
|
|
offset = from->data - (unsigned char *)page_address(page);
|
|
|
|
__skb_fill_page_desc(to, 0, page, offset, plen);
|
|
|
|
get_page(page);
|
|
|
|
j = 1;
|
|
|
|
len -= plen;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
to->truesize += len + plen;
|
|
|
|
to->len += len + plen;
|
|
|
|
to->data_len += len + plen;
|
|
|
|
|
2014-03-26 16:37:45 -06:00
|
|
|
if (unlikely(skb_orphan_frags(from, GFP_ATOMIC))) {
|
|
|
|
skb_tx_error(from);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
|
2013-12-13 07:22:17 -07:00
|
|
|
for (i = 0; i < skb_shinfo(from)->nr_frags; i++) {
|
|
|
|
if (!len)
|
|
|
|
break;
|
|
|
|
skb_shinfo(to)->frags[j] = skb_shinfo(from)->frags[i];
|
|
|
|
skb_shinfo(to)->frags[j].size = min_t(int, skb_shinfo(to)->frags[j].size, len);
|
|
|
|
len -= skb_shinfo(to)->frags[j].size;
|
|
|
|
skb_frag_ref(to, j);
|
|
|
|
j++;
|
|
|
|
}
|
|
|
|
skb_shinfo(to)->nr_frags = j;
|
2014-03-26 16:37:45 -06:00
|
|
|
|
|
|
|
return 0;
|
2013-12-13 07:22:17 -07:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_zerocopy);
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
void skb_copy_and_csum_dev(const struct sk_buff *skb, u8 *to)
|
|
|
|
{
|
2006-11-14 22:24:49 -07:00
|
|
|
__wsum csum;
|
2005-04-16 16:20:36 -06:00
|
|
|
long csstart;
|
|
|
|
|
2006-08-29 17:44:56 -06:00
|
|
|
if (skb->ip_summed == CHECKSUM_PARTIAL)
|
2010-12-14 08:24:08 -07:00
|
|
|
csstart = skb_checksum_start_offset(skb);
|
2005-04-16 16:20:36 -06:00
|
|
|
else
|
|
|
|
csstart = skb_headlen(skb);
|
|
|
|
|
2006-01-08 23:24:28 -07:00
|
|
|
BUG_ON(csstart > skb_headlen(skb));
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2007-03-27 15:55:52 -06:00
|
|
|
skb_copy_from_linear_data(skb, to, csstart);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
csum = 0;
|
|
|
|
if (csstart != skb->len)
|
|
|
|
csum = skb_copy_and_csum_bits(skb, csstart, to + csstart,
|
|
|
|
skb->len - csstart, 0);
|
|
|
|
|
2006-08-29 17:44:56 -06:00
|
|
|
if (skb->ip_summed == CHECKSUM_PARTIAL) {
|
2006-11-20 19:07:29 -07:00
|
|
|
long csstuff = csstart + skb->csum_offset;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2006-11-14 22:24:49 -07:00
|
|
|
*((__sum16 *)(to + csstuff)) = csum_fold(csum);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_copy_and_csum_dev);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_dequeue - remove from the head of the queue
|
|
|
|
* @list: list to dequeue from
|
|
|
|
*
|
|
|
|
* Remove the head of the list. The list lock is taken so the function
|
|
|
|
* may be used safely with other locking list functions. The head item is
|
|
|
|
* returned or %NULL if the list is empty.
|
|
|
|
*/
|
|
|
|
|
|
|
|
struct sk_buff *skb_dequeue(struct sk_buff_head *list)
|
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
struct sk_buff *result;
|
|
|
|
|
|
|
|
spin_lock_irqsave(&list->lock, flags);
|
|
|
|
result = __skb_dequeue(list);
|
|
|
|
spin_unlock_irqrestore(&list->lock, flags);
|
|
|
|
return result;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_dequeue);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_dequeue_tail - remove from the tail of the queue
|
|
|
|
* @list: list to dequeue from
|
|
|
|
*
|
|
|
|
* Remove the tail of the list. The list lock is taken so the function
|
|
|
|
* may be used safely with other locking list functions. The tail item is
|
|
|
|
* returned or %NULL if the list is empty.
|
|
|
|
*/
|
|
|
|
struct sk_buff *skb_dequeue_tail(struct sk_buff_head *list)
|
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
struct sk_buff *result;
|
|
|
|
|
|
|
|
spin_lock_irqsave(&list->lock, flags);
|
|
|
|
result = __skb_dequeue_tail(list);
|
|
|
|
spin_unlock_irqrestore(&list->lock, flags);
|
|
|
|
return result;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_dequeue_tail);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_queue_purge - empty a list
|
|
|
|
* @list: list to empty
|
|
|
|
*
|
|
|
|
* Delete all buffers on an &sk_buff list. Each buffer is removed from
|
|
|
|
* the list and one reference dropped. This function takes the list
|
|
|
|
* lock and is atomic with respect to other list locking functions.
|
|
|
|
*/
|
|
|
|
void skb_queue_purge(struct sk_buff_head *list)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb;
|
|
|
|
while ((skb = skb_dequeue(list)) != NULL)
|
|
|
|
kfree_skb(skb);
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_queue_purge);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
tcp: use an RB tree for ooo receive queue
Over the years, TCP BDP has increased by several orders of magnitude,
and some people are considering to reach the 2 Gbytes limit.
Even with current window scale limit of 14, ~1 Gbytes maps to ~740,000
MSS.
In presence of packet losses (or reorders), TCP stores incoming packets
into an out of order queue, and number of skbs sitting there waiting for
the missing packets to be received can be in the 10^5 range.
Most packets are appended to the tail of this queue, and when
packets can finally be transferred to receive queue, we scan the queue
from its head.
However, in presence of heavy losses, we might have to find an arbitrary
point in this queue, involving a linear scan for every incoming packet,
throwing away cpu caches.
This patch converts it to a RB tree, to get bounded latencies.
Yaogong wrote a preliminary patch about 2 years ago.
Eric did the rebase, added ofo_last_skb cache, polishing and tests.
Tested with network dropping between 1 and 10 % packets, with good
success (about 30 % increase of throughput in stress tests)
Next step would be to also use an RB tree for the write queue at sender
side ;)
Signed-off-by: Yaogong Wang <wygivan@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Acked-By: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-09-07 15:49:28 -06:00
|
|
|
/**
|
|
|
|
* skb_rbtree_purge - empty a skb rbtree
|
|
|
|
* @root: root of the rbtree to empty
|
|
|
|
*
|
|
|
|
* Delete all buffers on an &sk_buff rbtree. Each buffer is removed from
|
|
|
|
* the list and one reference dropped. This function does not take
|
|
|
|
* any lock. Synchronization should be handled by the caller (e.g., TCP
|
|
|
|
* out-of-order queue is protected by the socket lock).
|
|
|
|
*/
|
|
|
|
void skb_rbtree_purge(struct rb_root *root)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb, *next;
|
|
|
|
|
|
|
|
rbtree_postorder_for_each_entry_safe(skb, next, root, rbnode)
|
|
|
|
kfree_skb(skb);
|
|
|
|
|
|
|
|
*root = RB_ROOT;
|
|
|
|
}
|
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
/**
|
|
|
|
* skb_queue_head - queue a buffer at the list head
|
|
|
|
* @list: list to use
|
|
|
|
* @newsk: buffer to queue
|
|
|
|
*
|
|
|
|
* Queue a buffer at the start of the list. This function takes the
|
|
|
|
* list lock and can be used safely with other locking &sk_buff functions
|
|
|
|
* safely.
|
|
|
|
*
|
|
|
|
* A buffer cannot be placed on two lists at the same time.
|
|
|
|
*/
|
|
|
|
void skb_queue_head(struct sk_buff_head *list, struct sk_buff *newsk)
|
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
spin_lock_irqsave(&list->lock, flags);
|
|
|
|
__skb_queue_head(list, newsk);
|
|
|
|
spin_unlock_irqrestore(&list->lock, flags);
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_queue_head);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_queue_tail - queue a buffer at the list tail
|
|
|
|
* @list: list to use
|
|
|
|
* @newsk: buffer to queue
|
|
|
|
*
|
|
|
|
* Queue a buffer at the tail of the list. This function takes the
|
|
|
|
* list lock and can be used safely with other locking &sk_buff functions
|
|
|
|
* safely.
|
|
|
|
*
|
|
|
|
* A buffer cannot be placed on two lists at the same time.
|
|
|
|
*/
|
|
|
|
void skb_queue_tail(struct sk_buff_head *list, struct sk_buff *newsk)
|
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
spin_lock_irqsave(&list->lock, flags);
|
|
|
|
__skb_queue_tail(list, newsk);
|
|
|
|
spin_unlock_irqrestore(&list->lock, flags);
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_queue_tail);
|
2005-08-09 20:25:21 -06:00
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
/**
|
|
|
|
* skb_unlink - remove a buffer from a list
|
|
|
|
* @skb: buffer to remove
|
2005-08-09 20:25:21 -06:00
|
|
|
* @list: list to use
|
2005-04-16 16:20:36 -06:00
|
|
|
*
|
2005-08-09 20:25:21 -06:00
|
|
|
* Remove a packet from a list. The list locks are taken and this
|
|
|
|
* function is atomic with respect to other list locked calls
|
2005-04-16 16:20:36 -06:00
|
|
|
*
|
2005-08-09 20:25:21 -06:00
|
|
|
* You must know what list the SKB is on.
|
2005-04-16 16:20:36 -06:00
|
|
|
*/
|
2005-08-09 20:25:21 -06:00
|
|
|
void skb_unlink(struct sk_buff *skb, struct sk_buff_head *list)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
2005-08-09 20:25:21 -06:00
|
|
|
unsigned long flags;
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2005-08-09 20:25:21 -06:00
|
|
|
spin_lock_irqsave(&list->lock, flags);
|
|
|
|
__skb_unlink(skb, list);
|
|
|
|
spin_unlock_irqrestore(&list->lock, flags);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_unlink);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_append - append a buffer
|
|
|
|
* @old: buffer to insert after
|
|
|
|
* @newsk: buffer to insert
|
2005-08-09 20:25:21 -06:00
|
|
|
* @list: list to use
|
2005-04-16 16:20:36 -06:00
|
|
|
*
|
|
|
|
* Place a packet after a given packet in a list. The list locks are taken
|
|
|
|
* and this function is atomic with respect to other list locked calls.
|
|
|
|
* A buffer cannot be placed on two lists at the same time.
|
|
|
|
*/
|
2005-08-09 20:25:21 -06:00
|
|
|
void skb_append(struct sk_buff *old, struct sk_buff *newsk, struct sk_buff_head *list)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
|
2005-08-09 20:25:21 -06:00
|
|
|
spin_lock_irqsave(&list->lock, flags);
|
2008-04-14 01:05:09 -06:00
|
|
|
__skb_queue_after(list, old, newsk);
|
2005-08-09 20:25:21 -06:00
|
|
|
spin_unlock_irqrestore(&list->lock, flags);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_append);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_insert - insert a buffer
|
|
|
|
* @old: buffer to insert before
|
|
|
|
* @newsk: buffer to insert
|
2005-08-09 20:25:21 -06:00
|
|
|
* @list: list to use
|
|
|
|
*
|
|
|
|
* Place a packet before a given packet in a list. The list locks are
|
|
|
|
* taken and this function is atomic with respect to other list locked
|
|
|
|
* calls.
|
2005-04-16 16:20:36 -06:00
|
|
|
*
|
|
|
|
* A buffer cannot be placed on two lists at the same time.
|
|
|
|
*/
|
2005-08-09 20:25:21 -06:00
|
|
|
void skb_insert(struct sk_buff *old, struct sk_buff *newsk, struct sk_buff_head *list)
|
2005-04-16 16:20:36 -06:00
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
|
2005-08-09 20:25:21 -06:00
|
|
|
spin_lock_irqsave(&list->lock, flags);
|
|
|
|
__skb_insert(newsk, old->prev, old, list);
|
|
|
|
spin_unlock_irqrestore(&list->lock, flags);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_insert);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
static inline void skb_split_inside_header(struct sk_buff *skb,
|
|
|
|
struct sk_buff* skb1,
|
|
|
|
const u32 len, const int pos)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
2007-03-27 15:55:52 -06:00
|
|
|
skb_copy_from_linear_data_offset(skb, len, skb_put(skb1, pos - len),
|
|
|
|
pos - len);
|
2005-04-16 16:20:36 -06:00
|
|
|
/* And move data appendix as is. */
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
|
|
|
|
skb_shinfo(skb1)->frags[i] = skb_shinfo(skb)->frags[i];
|
|
|
|
|
|
|
|
skb_shinfo(skb1)->nr_frags = skb_shinfo(skb)->nr_frags;
|
|
|
|
skb_shinfo(skb)->nr_frags = 0;
|
|
|
|
skb1->data_len = skb->data_len;
|
|
|
|
skb1->len += skb1->data_len;
|
|
|
|
skb->data_len = 0;
|
|
|
|
skb->len = len;
|
2007-04-19 21:29:13 -06:00
|
|
|
skb_set_tail_pointer(skb, len);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline void skb_split_no_header(struct sk_buff *skb,
|
|
|
|
struct sk_buff* skb1,
|
|
|
|
const u32 len, int pos)
|
|
|
|
{
|
|
|
|
int i, k = 0;
|
|
|
|
const int nfrags = skb_shinfo(skb)->nr_frags;
|
|
|
|
|
|
|
|
skb_shinfo(skb)->nr_frags = 0;
|
|
|
|
skb1->len = skb1->data_len = skb->len - len;
|
|
|
|
skb->len = len;
|
|
|
|
skb->data_len = len - pos;
|
|
|
|
|
|
|
|
for (i = 0; i < nfrags; i++) {
|
2011-10-18 15:00:24 -06:00
|
|
|
int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
|
|
|
if (pos + size > len) {
|
|
|
|
skb_shinfo(skb1)->frags[k] = skb_shinfo(skb)->frags[i];
|
|
|
|
|
|
|
|
if (pos < len) {
|
|
|
|
/* Split frag.
|
|
|
|
* We have two variants in this case:
|
|
|
|
* 1. Move all the frag to the second
|
|
|
|
* part, if it is possible. F.e.
|
|
|
|
* this approach is mandatory for TUX,
|
|
|
|
* where splitting is expensive.
|
|
|
|
* 2. Split is accurately. We make this.
|
|
|
|
*/
|
2011-08-22 17:44:58 -06:00
|
|
|
skb_frag_ref(skb, i);
|
2005-04-16 16:20:36 -06:00
|
|
|
skb_shinfo(skb1)->frags[0].page_offset += len - pos;
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_sub(&skb_shinfo(skb1)->frags[0], len - pos);
|
|
|
|
skb_frag_size_set(&skb_shinfo(skb)->frags[i], len - pos);
|
2005-04-16 16:20:36 -06:00
|
|
|
skb_shinfo(skb)->nr_frags++;
|
|
|
|
}
|
|
|
|
k++;
|
|
|
|
} else
|
|
|
|
skb_shinfo(skb)->nr_frags++;
|
|
|
|
pos += size;
|
|
|
|
}
|
|
|
|
skb_shinfo(skb1)->nr_frags = k;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_split - Split fragmented skb to two parts at length len.
|
|
|
|
* @skb: the buffer to split
|
|
|
|
* @skb1: the buffer to receive the second part
|
|
|
|
* @len: new length for skb
|
|
|
|
*/
|
|
|
|
void skb_split(struct sk_buff *skb, struct sk_buff *skb1, const u32 len)
|
|
|
|
{
|
|
|
|
int pos = skb_headlen(skb);
|
|
|
|
|
2013-02-19 15:51:30 -07:00
|
|
|
skb_shinfo(skb1)->tx_flags = skb_shinfo(skb)->tx_flags & SKBTX_SHARED_FRAG;
|
2005-04-16 16:20:36 -06:00
|
|
|
if (len < pos) /* Split line is inside header. */
|
|
|
|
skb_split_inside_header(skb, skb1, len, pos);
|
|
|
|
else /* Second chunk has no header, nothing to copy. */
|
|
|
|
skb_split_no_header(skb, skb1, len, pos);
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_split);
|
2005-04-16 16:20:36 -06:00
|
|
|
|
2008-11-25 14:57:01 -07:00
|
|
|
/* Shifting from/to a cloned skb is a no-go.
|
|
|
|
*
|
|
|
|
* Caller cannot keep skb_shinfo related pointers past calling here!
|
|
|
|
*/
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
static int skb_prepare_for_shift(struct sk_buff *skb)
|
|
|
|
{
|
2008-11-24 22:30:21 -07:00
|
|
|
return skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_shift - Shifts paged data partially from skb to another
|
|
|
|
* @tgt: buffer into which tail data gets added
|
|
|
|
* @skb: buffer from which the paged data comes from
|
|
|
|
* @shiftlen: shift up to this many bytes
|
|
|
|
*
|
|
|
|
* Attempts to shift up to shiftlen worth of bytes, which may be less than
|
2011-11-20 18:47:11 -07:00
|
|
|
* the length of the skb, from skb to tgt. Returns number bytes shifted.
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
* It's up to caller to free skb if everything was shifted.
|
|
|
|
*
|
|
|
|
* If @tgt runs out of frags, the whole operation is aborted.
|
|
|
|
*
|
|
|
|
* Skb cannot include anything else but paged data while tgt is allowed
|
|
|
|
* to have non-paged data as well.
|
|
|
|
*
|
|
|
|
* TODO: full sized shift could be optimized but that would need
|
|
|
|
* specialized skb free'er to handle frags without up-to-date nr_frags.
|
|
|
|
*/
|
|
|
|
int skb_shift(struct sk_buff *tgt, struct sk_buff *skb, int shiftlen)
|
|
|
|
{
|
|
|
|
int from, to, merge, todo;
|
|
|
|
struct skb_frag_struct *fragfrom, *fragto;
|
|
|
|
|
|
|
|
BUG_ON(shiftlen > skb->len);
|
tcp: enhance tcp_collapse_retrans() with skb_shift()
In commit 2331ccc5b323 ("tcp: enhance tcp collapsing"),
we made a first step allowing copying right skb to left skb head.
Since all skbs in socket write queue are headless (but possibly the very
first one), this strategy often does not work.
This patch extends tcp_collapse_retrans() to perform frag shifting,
thanks to skb_shift() helper.
This helper needs to not BUG on non headless skbs, as callers are ok
with that.
Tested:
Following packetdrill test now passes :
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0 bind(3, ..., ...) = 0
+0 listen(3, 1) = 0
+0 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 8>
+0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8>
+.100 < . 1:1(0) ack 1 win 257
+0 accept(3, ..., ...) = 4
+0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0
+0 write(4, ..., 200) = 200
+0 > P. 1:201(200) ack 1
+.001 write(4, ..., 200) = 200
+0 > P. 201:401(200) ack 1
+.001 write(4, ..., 200) = 200
+0 > P. 401:601(200) ack 1
+.001 write(4, ..., 200) = 200
+0 > P. 601:801(200) ack 1
+.001 write(4, ..., 200) = 200
+0 > P. 801:1001(200) ack 1
+.001 write(4, ..., 100) = 100
+0 > P. 1001:1101(100) ack 1
+.001 write(4, ..., 100) = 100
+0 > P. 1101:1201(100) ack 1
+.001 write(4, ..., 100) = 100
+0 > P. 1201:1301(100) ack 1
+.001 write(4, ..., 100) = 100
+0 > P. 1301:1401(100) ack 1
+.099 < . 1:1(0) ack 201 win 257
+.001 < . 1:1(0) ack 201 win 257 <nop,nop,sack 1001:1401>
+0 > P. 201:1001(800) ack 1
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 13:51:50 -07:00
|
|
|
|
|
|
|
if (skb_headlen(skb))
|
|
|
|
return 0;
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
|
|
|
|
todo = shiftlen;
|
|
|
|
from = 0;
|
|
|
|
to = skb_shinfo(tgt)->nr_frags;
|
|
|
|
fragfrom = &skb_shinfo(skb)->frags[from];
|
|
|
|
|
|
|
|
/* Actual merge is delayed until the point when we know we can
|
|
|
|
* commit all, so that we don't have to undo partial changes
|
|
|
|
*/
|
|
|
|
if (!to ||
|
2011-08-22 17:44:58 -06:00
|
|
|
!skb_can_coalesce(tgt, to, skb_frag_page(fragfrom),
|
|
|
|
fragfrom->page_offset)) {
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
merge = -1;
|
|
|
|
} else {
|
|
|
|
merge = to - 1;
|
|
|
|
|
2011-10-18 15:00:24 -06:00
|
|
|
todo -= skb_frag_size(fragfrom);
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
if (todo < 0) {
|
|
|
|
if (skb_prepare_for_shift(skb) ||
|
|
|
|
skb_prepare_for_shift(tgt))
|
|
|
|
return 0;
|
|
|
|
|
2008-11-25 14:57:01 -07:00
|
|
|
/* All previous frag pointers might be stale! */
|
|
|
|
fragfrom = &skb_shinfo(skb)->frags[from];
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
fragto = &skb_shinfo(tgt)->frags[merge];
|
|
|
|
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_add(fragto, shiftlen);
|
|
|
|
skb_frag_size_sub(fragfrom, shiftlen);
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
fragfrom->page_offset += shiftlen;
|
|
|
|
|
|
|
|
goto onlymerged;
|
|
|
|
}
|
|
|
|
|
|
|
|
from++;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Skip full, not-fitting skb to avoid expensive operations */
|
|
|
|
if ((shiftlen == skb->len) &&
|
|
|
|
(skb_shinfo(skb)->nr_frags - from) > (MAX_SKB_FRAGS - to))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (skb_prepare_for_shift(skb) || skb_prepare_for_shift(tgt))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
while ((todo > 0) && (from < skb_shinfo(skb)->nr_frags)) {
|
|
|
|
if (to == MAX_SKB_FRAGS)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
fragfrom = &skb_shinfo(skb)->frags[from];
|
|
|
|
fragto = &skb_shinfo(tgt)->frags[to];
|
|
|
|
|
2011-10-18 15:00:24 -06:00
|
|
|
if (todo >= skb_frag_size(fragfrom)) {
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
*fragto = *fragfrom;
|
2011-10-18 15:00:24 -06:00
|
|
|
todo -= skb_frag_size(fragfrom);
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
from++;
|
|
|
|
to++;
|
|
|
|
|
|
|
|
} else {
|
2011-08-22 17:44:58 -06:00
|
|
|
__skb_frag_ref(fragfrom);
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
fragto->page = fragfrom->page;
|
|
|
|
fragto->page_offset = fragfrom->page_offset;
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_set(fragto, todo);
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
|
|
|
|
fragfrom->page_offset += todo;
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_sub(fragfrom, todo);
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
todo = 0;
|
|
|
|
|
|
|
|
to++;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Ready to "commit" this state change to tgt */
|
|
|
|
skb_shinfo(tgt)->nr_frags = to;
|
|
|
|
|
|
|
|
if (merge >= 0) {
|
|
|
|
fragfrom = &skb_shinfo(skb)->frags[0];
|
|
|
|
fragto = &skb_shinfo(tgt)->frags[merge];
|
|
|
|
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_add(fragto, skb_frag_size(fragfrom));
|
2011-08-22 17:44:58 -06:00
|
|
|
__skb_frag_unref(fragfrom);
|
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-24 22:20:15 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Reposition in the original skb */
|
|
|
|
to = 0;
|
|
|
|
while (from < skb_shinfo(skb)->nr_frags)
|
|
|
|
skb_shinfo(skb)->frags[to++] = skb_shinfo(skb)->frags[from++];
|
|
|
|
skb_shinfo(skb)->nr_frags = to;
|
|
|
|
|
|
|
|
BUG_ON(todo > 0 && !skb_shinfo(skb)->nr_frags);
|
|
|
|
|
|
|
|
onlymerged:
|
|
|
|
/* Most likely the tgt won't ever need its checksum anymore, skb on
|
|
|
|
* the other hand might need it if it needs to be resent
|
|
|
|
*/
|
|
|
|
tgt->ip_summed = CHECKSUM_PARTIAL;
|
|
|
|
skb->ip_summed = CHECKSUM_PARTIAL;
|
|
|
|
|
|
|
|
/* Yak, is it really working this way? Some helper please? */
|
|
|
|
skb->len -= shiftlen;
|
|
|
|
skb->data_len -= shiftlen;
|
|
|
|
skb->truesize -= shiftlen;
|
|
|
|
tgt->len += shiftlen;
|
|
|
|
tgt->data_len += shiftlen;
|
|
|
|
tgt->truesize += shiftlen;
|
|
|
|
|
|
|
|
return shiftlen;
|
|
|
|
}
|
|
|
|
|
2005-06-23 21:59:51 -06:00
|
|
|
/**
|
|
|
|
* skb_prepare_seq_read - Prepare a sequential read of skb data
|
|
|
|
* @skb: the buffer to read
|
|
|
|
* @from: lower offset of data to be read
|
|
|
|
* @to: upper offset of data to be read
|
|
|
|
* @st: state variable
|
|
|
|
*
|
|
|
|
* Initializes the specified state variable. Must be called before
|
|
|
|
* invoking skb_seq_read() for the first time.
|
|
|
|
*/
|
|
|
|
void skb_prepare_seq_read(struct sk_buff *skb, unsigned int from,
|
|
|
|
unsigned int to, struct skb_seq_state *st)
|
|
|
|
{
|
|
|
|
st->lower_offset = from;
|
|
|
|
st->upper_offset = to;
|
|
|
|
st->root_skb = st->cur_skb = skb;
|
|
|
|
st->frag_idx = st->stepped_offset = 0;
|
|
|
|
st->frag_data = NULL;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_prepare_seq_read);
|
2005-06-23 21:59:51 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_seq_read - Sequentially read skb data
|
|
|
|
* @consumed: number of bytes consumed by the caller so far
|
|
|
|
* @data: destination pointer for data to be returned
|
|
|
|
* @st: state variable
|
|
|
|
*
|
2013-11-07 06:18:26 -07:00
|
|
|
* Reads a block of skb data at @consumed relative to the
|
2005-06-23 21:59:51 -06:00
|
|
|
* lower offset specified to skb_prepare_seq_read(). Assigns
|
2013-11-07 06:18:26 -07:00
|
|
|
* the head of the data block to @data and returns the length
|
2005-06-23 21:59:51 -06:00
|
|
|
* of the block or 0 if the end of the skb data or the upper
|
|
|
|
* offset has been reached.
|
|
|
|
*
|
|
|
|
* The caller is not required to consume all of the data
|
2013-11-07 06:18:26 -07:00
|
|
|
* returned, i.e. @consumed is typically set to the number
|
2005-06-23 21:59:51 -06:00
|
|
|
* of bytes already consumed and the next call to
|
|
|
|
* skb_seq_read() will return the remaining part of the block.
|
|
|
|
*
|
2011-03-30 19:57:33 -06:00
|
|
|
* Note 1: The size of each block of data returned can be arbitrary,
|
2014-09-04 08:44:36 -06:00
|
|
|
* this limitation is the cost for zerocopy sequential
|
2005-06-23 21:59:51 -06:00
|
|
|
* reads of potentially non linear data.
|
|
|
|
*
|
2008-02-13 16:03:25 -07:00
|
|
|
* Note 2: Fragment lists within fragments are not implemented
|
2005-06-23 21:59:51 -06:00
|
|
|
* at the moment, state->root_skb could be replaced with
|
|
|
|
* a stack for this purpose.
|
|
|
|
*/
|
|
|
|
unsigned int skb_seq_read(unsigned int consumed, const u8 **data,
|
|
|
|
struct skb_seq_state *st)
|
|
|
|
{
|
|
|
|
unsigned int block_limit, abs_offset = consumed + st->lower_offset;
|
|
|
|
skb_frag_t *frag;
|
|
|
|
|
2013-06-24 00:33:48 -06:00
|
|
|
if (unlikely(abs_offset >= st->upper_offset)) {
|
|
|
|
if (st->frag_data) {
|
|
|
|
kunmap_atomic(st->frag_data);
|
|
|
|
st->frag_data = NULL;
|
|
|
|
}
|
2005-06-23 21:59:51 -06:00
|
|
|
return 0;
|
2013-06-24 00:33:48 -06:00
|
|
|
}
|
2005-06-23 21:59:51 -06:00
|
|
|
|
|
|
|
next_skb:
|
2009-01-29 17:07:52 -07:00
|
|
|
block_limit = skb_headlen(st->cur_skb) + st->stepped_offset;
|
2005-06-23 21:59:51 -06:00
|
|
|
|
2009-05-18 22:43:27 -06:00
|
|
|
if (abs_offset < block_limit && !st->frag_data) {
|
2009-01-29 17:07:52 -07:00
|
|
|
*data = st->cur_skb->data + (abs_offset - st->stepped_offset);
|
2005-06-23 21:59:51 -06:00
|
|
|
return block_limit - abs_offset;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (st->frag_idx == 0 && !st->frag_data)
|
|
|
|
st->stepped_offset += skb_headlen(st->cur_skb);
|
|
|
|
|
|
|
|
while (st->frag_idx < skb_shinfo(st->cur_skb)->nr_frags) {
|
|
|
|
frag = &skb_shinfo(st->cur_skb)->frags[st->frag_idx];
|
2011-10-18 15:00:24 -06:00
|
|
|
block_limit = skb_frag_size(frag) + st->stepped_offset;
|
2005-06-23 21:59:51 -06:00
|
|
|
|
|
|
|
if (abs_offset < block_limit) {
|
|
|
|
if (!st->frag_data)
|
2012-04-05 03:35:15 -06:00
|
|
|
st->frag_data = kmap_atomic(skb_frag_page(frag));
|
2005-06-23 21:59:51 -06:00
|
|
|
|
|
|
|
*data = (u8 *) st->frag_data + frag->page_offset +
|
|
|
|
(abs_offset - st->stepped_offset);
|
|
|
|
|
|
|
|
return block_limit - abs_offset;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (st->frag_data) {
|
2012-04-05 03:35:15 -06:00
|
|
|
kunmap_atomic(st->frag_data);
|
2005-06-23 21:59:51 -06:00
|
|
|
st->frag_data = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
st->frag_idx++;
|
2011-10-18 15:00:24 -06:00
|
|
|
st->stepped_offset += skb_frag_size(frag);
|
2005-06-23 21:59:51 -06:00
|
|
|
}
|
|
|
|
|
2007-06-24 00:11:52 -06:00
|
|
|
if (st->frag_data) {
|
2012-04-05 03:35:15 -06:00
|
|
|
kunmap_atomic(st->frag_data);
|
2007-06-24 00:11:52 -06:00
|
|
|
st->frag_data = NULL;
|
|
|
|
}
|
|
|
|
|
2010-08-23 01:13:46 -06:00
|
|
|
if (st->root_skb == st->cur_skb && skb_has_frag_list(st->root_skb)) {
|
2009-01-29 17:12:42 -07:00
|
|
|
st->cur_skb = skb_shinfo(st->root_skb)->frag_list;
|
2005-06-23 21:59:51 -06:00
|
|
|
st->frag_idx = 0;
|
|
|
|
goto next_skb;
|
2009-01-29 17:12:42 -07:00
|
|
|
} else if (st->cur_skb->next) {
|
|
|
|
st->cur_skb = st->cur_skb->next;
|
2009-01-29 17:07:52 -07:00
|
|
|
st->frag_idx = 0;
|
2005-06-23 21:59:51 -06:00
|
|
|
goto next_skb;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_seq_read);
|
2005-06-23 21:59:51 -06:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_abort_seq_read - Abort a sequential read of skb data
|
|
|
|
* @st: state variable
|
|
|
|
*
|
|
|
|
* Must be called if skb_seq_read() was not called until it
|
|
|
|
* returned 0.
|
|
|
|
*/
|
|
|
|
void skb_abort_seq_read(struct skb_seq_state *st)
|
|
|
|
{
|
|
|
|
if (st->frag_data)
|
2012-04-05 03:35:15 -06:00
|
|
|
kunmap_atomic(st->frag_data);
|
2005-06-23 21:59:51 -06:00
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_abort_seq_read);
|
2005-06-23 21:59:51 -06:00
|
|
|
|
2005-06-23 22:00:17 -06:00
|
|
|
#define TS_SKB_CB(state) ((struct skb_seq_state *) &((state)->cb))
|
|
|
|
|
|
|
|
static unsigned int skb_ts_get_next_block(unsigned int offset, const u8 **text,
|
|
|
|
struct ts_config *conf,
|
|
|
|
struct ts_state *state)
|
|
|
|
{
|
|
|
|
return skb_seq_read(offset, text, TS_SKB_CB(state));
|
|
|
|
}
|
|
|
|
|
|
|
|
static void skb_ts_finish(struct ts_config *conf, struct ts_state *state)
|
|
|
|
{
|
|
|
|
skb_abort_seq_read(TS_SKB_CB(state));
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_find_text - Find a text pattern in skb data
|
|
|
|
* @skb: the buffer to look in
|
|
|
|
* @from: search offset
|
|
|
|
* @to: search limit
|
|
|
|
* @config: textsearch configuration
|
|
|
|
*
|
|
|
|
* Finds a pattern in the skb data according to the specified
|
|
|
|
* textsearch configuration. Use textsearch_next() to retrieve
|
|
|
|
* subsequent occurrences of the pattern. Returns the offset
|
|
|
|
* to the first occurrence or UINT_MAX if no match was found.
|
|
|
|
*/
|
|
|
|
unsigned int skb_find_text(struct sk_buff *skb, unsigned int from,
|
2015-02-22 03:46:35 -07:00
|
|
|
unsigned int to, struct ts_config *config)
|
2005-06-23 22:00:17 -06:00
|
|
|
{
|
2015-02-22 03:46:35 -07:00
|
|
|
struct ts_state state;
|
2006-06-26 01:00:57 -06:00
|
|
|
unsigned int ret;
|
|
|
|
|
2005-06-23 22:00:17 -06:00
|
|
|
config->get_next_block = skb_ts_get_next_block;
|
|
|
|
config->finish = skb_ts_finish;
|
|
|
|
|
2015-02-22 03:46:35 -07:00
|
|
|
skb_prepare_seq_read(skb, from, to, TS_SKB_CB(&state));
|
2005-06-23 22:00:17 -06:00
|
|
|
|
2015-02-22 03:46:35 -07:00
|
|
|
ret = textsearch_find(config, &state);
|
2006-06-26 01:00:57 -06:00
|
|
|
return (ret <= to - from ? ret : UINT_MAX);
|
2005-06-23 22:00:17 -06:00
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_find_text);
|
2005-06-23 22:00:17 -06:00
|
|
|
|
2005-10-18 16:46:41 -06:00
|
|
|
/**
|
2012-07-10 04:55:09 -06:00
|
|
|
* skb_append_datato_frags - append the user data to a skb
|
2005-10-18 16:46:41 -06:00
|
|
|
* @sk: sock structure
|
2014-09-04 08:44:36 -06:00
|
|
|
* @skb: skb structure to be appended with user data.
|
2005-10-18 16:46:41 -06:00
|
|
|
* @getfrag: call back function to be used for getting the user data
|
|
|
|
* @from: pointer to user message iov
|
|
|
|
* @length: length of the iov message
|
|
|
|
*
|
|
|
|
* Description: This procedure append the user data in the fragment part
|
|
|
|
* of the skb if any page alloc fails user this procedure returns -ENOMEM
|
|
|
|
*/
|
|
|
|
int skb_append_datato_frags(struct sock *sk, struct sk_buff *skb,
|
2005-12-05 14:40:12 -07:00
|
|
|
int (*getfrag)(void *from, char *to, int offset,
|
2005-10-18 16:46:41 -06:00
|
|
|
int len, int odd, struct sk_buff *skb),
|
|
|
|
void *from, int length)
|
|
|
|
{
|
2012-12-27 23:06:37 -07:00
|
|
|
int frg_cnt = skb_shinfo(skb)->nr_frags;
|
|
|
|
int copy;
|
2005-10-18 16:46:41 -06:00
|
|
|
int offset = 0;
|
|
|
|
int ret;
|
2012-12-27 23:06:37 -07:00
|
|
|
struct page_frag *pfrag = ¤t->task_frag;
|
2005-10-18 16:46:41 -06:00
|
|
|
|
|
|
|
do {
|
|
|
|
/* Return error if we don't have space for new frag */
|
|
|
|
if (frg_cnt >= MAX_SKB_FRAGS)
|
2012-12-27 23:06:37 -07:00
|
|
|
return -EMSGSIZE;
|
2005-10-18 16:46:41 -06:00
|
|
|
|
2012-12-27 23:06:37 -07:00
|
|
|
if (!sk_page_frag_refill(sk, pfrag))
|
2005-10-18 16:46:41 -06:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
/* copy the user data to page */
|
2012-12-27 23:06:37 -07:00
|
|
|
copy = min_t(int, length, pfrag->size - pfrag->offset);
|
2005-10-18 16:46:41 -06:00
|
|
|
|
2012-12-27 23:06:37 -07:00
|
|
|
ret = getfrag(from, page_address(pfrag->page) + pfrag->offset,
|
|
|
|
offset, copy, 0, skb);
|
2005-10-18 16:46:41 -06:00
|
|
|
if (ret < 0)
|
|
|
|
return -EFAULT;
|
|
|
|
|
|
|
|
/* copy was successful so update the size parameters */
|
2012-12-27 23:06:37 -07:00
|
|
|
skb_fill_page_desc(skb, frg_cnt, pfrag->page, pfrag->offset,
|
|
|
|
copy);
|
|
|
|
frg_cnt++;
|
|
|
|
pfrag->offset += copy;
|
|
|
|
get_page(pfrag->page);
|
|
|
|
|
|
|
|
skb->truesize += copy;
|
|
|
|
atomic_add(copy, &sk->sk_wmem_alloc);
|
2005-10-18 16:46:41 -06:00
|
|
|
skb->len += copy;
|
|
|
|
skb->data_len += copy;
|
|
|
|
offset += copy;
|
|
|
|
length -= copy;
|
|
|
|
|
|
|
|
} while (length > 0);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL(skb_append_datato_frags);
|
2005-10-18 16:46:41 -06:00
|
|
|
|
2015-05-21 08:59:58 -06:00
|
|
|
int skb_append_pagefrags(struct sk_buff *skb, struct page *page,
|
|
|
|
int offset, size_t size)
|
|
|
|
{
|
|
|
|
int i = skb_shinfo(skb)->nr_frags;
|
|
|
|
|
|
|
|
if (skb_can_coalesce(skb, i, page, offset)) {
|
|
|
|
skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
|
|
|
|
} else if (i < MAX_SKB_FRAGS) {
|
|
|
|
get_page(page);
|
|
|
|
skb_fill_page_desc(skb, i, page, offset, size);
|
|
|
|
} else {
|
|
|
|
return -EMSGSIZE;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_append_pagefrags);
|
|
|
|
|
2006-03-20 23:43:56 -07:00
|
|
|
/**
|
|
|
|
* skb_pull_rcsum - pull skb and update receive checksum
|
|
|
|
* @skb: buffer to update
|
|
|
|
* @len: length of data pulled
|
|
|
|
*
|
|
|
|
* This function performs an skb_pull on the packet and updates
|
2008-02-12 23:03:25 -07:00
|
|
|
* the CHECKSUM_COMPLETE checksum. It should be used on
|
2006-08-29 17:44:56 -06:00
|
|
|
* receive path processing instead of skb_pull unless you know
|
|
|
|
* that the checksum difference is zero (e.g., a valid IP header)
|
|
|
|
* or you are setting ip_summed to CHECKSUM_NONE.
|
2006-03-20 23:43:56 -07:00
|
|
|
*/
|
|
|
|
unsigned char *skb_pull_rcsum(struct sk_buff *skb, unsigned int len)
|
|
|
|
{
|
2015-09-28 18:24:25 -06:00
|
|
|
unsigned char *data = skb->data;
|
|
|
|
|
2006-03-20 23:43:56 -07:00
|
|
|
BUG_ON(len > skb->len);
|
2015-09-28 18:24:25 -06:00
|
|
|
__skb_pull(skb, len);
|
|
|
|
skb_postpull_rcsum(skb, data, len);
|
|
|
|
return skb->data;
|
2006-03-20 23:43:56 -07:00
|
|
|
}
|
2006-03-20 23:47:55 -07:00
|
|
|
EXPORT_SYMBOL_GPL(skb_pull_rcsum);
|
|
|
|
|
2006-06-22 04:02:40 -06:00
|
|
|
/**
|
|
|
|
* skb_segment - Perform protocol segmentation on skb.
|
2014-03-10 10:29:19 -06:00
|
|
|
* @head_skb: buffer to segment
|
2006-06-27 14:22:38 -06:00
|
|
|
* @features: features for the output path (see dev->features)
|
2006-06-22 04:02:40 -06:00
|
|
|
*
|
|
|
|
* This function performs segmentation on the given skb. It returns
|
2008-04-13 22:52:48 -06:00
|
|
|
* a pointer to the first in a list of new skbs for the segments.
|
|
|
|
* In case of error it returns ERR_PTR(err).
|
2006-06-22 04:02:40 -06:00
|
|
|
*/
|
2014-03-10 10:29:19 -06:00
|
|
|
struct sk_buff *skb_segment(struct sk_buff *head_skb,
|
|
|
|
netdev_features_t features)
|
2006-06-22 04:02:40 -06:00
|
|
|
{
|
|
|
|
struct sk_buff *segs = NULL;
|
|
|
|
struct sk_buff *tail = NULL;
|
2014-03-10 11:27:59 -06:00
|
|
|
struct sk_buff *list_skb = skb_shinfo(head_skb)->frag_list;
|
2014-03-10 10:29:19 -06:00
|
|
|
skb_frag_t *frag = skb_shinfo(head_skb)->frags;
|
|
|
|
unsigned int mss = skb_shinfo(head_skb)->gso_size;
|
|
|
|
unsigned int doffset = head_skb->data - skb_mac_header(head_skb);
|
2014-03-10 11:28:08 -06:00
|
|
|
struct sk_buff *frag_skb = head_skb;
|
2006-06-22 04:02:40 -06:00
|
|
|
unsigned int offset = doffset;
|
2014-03-10 10:29:19 -06:00
|
|
|
unsigned int tnl_hlen = skb_tnl_header_len(head_skb);
|
2016-04-10 19:45:03 -06:00
|
|
|
unsigned int partial_segs = 0;
|
2006-06-22 04:02:40 -06:00
|
|
|
unsigned int headroom;
|
2016-04-10 19:45:03 -06:00
|
|
|
unsigned int len = head_skb->len;
|
2013-03-07 02:28:01 -07:00
|
|
|
__be16 proto;
|
2016-05-02 10:38:18 -06:00
|
|
|
bool csum, sg;
|
2014-03-10 10:29:19 -06:00
|
|
|
int nfrags = skb_shinfo(head_skb)->nr_frags;
|
2006-06-22 04:02:40 -06:00
|
|
|
int err = -ENOMEM;
|
|
|
|
int i = 0;
|
|
|
|
int pos;
|
2014-03-27 15:26:18 -06:00
|
|
|
int dummy;
|
2006-06-22 04:02:40 -06:00
|
|
|
|
2014-06-09 00:48:54 -06:00
|
|
|
__skb_push(head_skb, doffset);
|
2014-03-27 15:26:18 -06:00
|
|
|
proto = skb_network_protocol(head_skb, &dummy);
|
2013-03-07 02:28:01 -07:00
|
|
|
if (unlikely(!proto))
|
|
|
|
return ERR_PTR(-EINVAL);
|
|
|
|
|
2016-05-02 10:38:18 -06:00
|
|
|
sg = !!(features & NETIF_F_SG);
|
2016-02-05 16:28:26 -07:00
|
|
|
csum = !!can_checksum_protocol(features, proto);
|
2014-06-04 18:20:02 -06:00
|
|
|
|
2016-09-19 04:58:47 -06:00
|
|
|
if (sg && csum && (mss != GSO_BY_FRAGS)) {
|
|
|
|
if (!(features & NETIF_F_GSO_PARTIAL)) {
|
|
|
|
struct sk_buff *iter;
|
|
|
|
|
|
|
|
if (!list_skb ||
|
|
|
|
!net_gso_ok(features, skb_shinfo(head_skb)->gso_type))
|
|
|
|
goto normal;
|
|
|
|
|
|
|
|
/* Split the buffer at the frag_list pointer.
|
|
|
|
* This is based on the assumption that all
|
|
|
|
* buffers in the chain excluding the last
|
|
|
|
* containing the same amount of data.
|
|
|
|
*/
|
|
|
|
skb_walk_frags(head_skb, iter) {
|
|
|
|
if (skb_headlen(iter))
|
|
|
|
goto normal;
|
|
|
|
|
|
|
|
len -= iter->len;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* GSO partial only requires that we trim off any excess that
|
|
|
|
* doesn't fit into an MSS sized block, so take care of that
|
|
|
|
* now.
|
|
|
|
*/
|
2016-04-10 19:45:03 -06:00
|
|
|
partial_segs = len / mss;
|
2016-05-02 10:38:12 -06:00
|
|
|
if (partial_segs > 1)
|
|
|
|
mss *= partial_segs;
|
|
|
|
else
|
|
|
|
partial_segs = 0;
|
2016-04-10 19:45:03 -06:00
|
|
|
}
|
|
|
|
|
2016-09-19 04:58:47 -06:00
|
|
|
normal:
|
2014-03-10 10:29:19 -06:00
|
|
|
headroom = skb_headroom(head_skb);
|
|
|
|
pos = skb_headlen(head_skb);
|
2006-06-22 04:02:40 -06:00
|
|
|
|
|
|
|
do {
|
|
|
|
struct sk_buff *nskb;
|
2014-03-10 10:29:04 -06:00
|
|
|
skb_frag_t *nskb_frag;
|
2006-10-29 16:59:41 -07:00
|
|
|
int hsize;
|
2006-06-22 04:02:40 -06:00
|
|
|
int size;
|
|
|
|
|
2016-06-02 12:05:40 -06:00
|
|
|
if (unlikely(mss == GSO_BY_FRAGS)) {
|
|
|
|
len = list_skb->len;
|
|
|
|
} else {
|
|
|
|
len = head_skb->len - offset;
|
|
|
|
if (len > mss)
|
|
|
|
len = mss;
|
|
|
|
}
|
2006-06-22 04:02:40 -06:00
|
|
|
|
2014-03-10 10:29:19 -06:00
|
|
|
hsize = skb_headlen(head_skb) - offset;
|
2006-06-22 04:02:40 -06:00
|
|
|
if (hsize < 0)
|
|
|
|
hsize = 0;
|
2006-10-29 16:59:41 -07:00
|
|
|
if (hsize > len || !sg)
|
|
|
|
hsize = len;
|
2006-06-22 04:02:40 -06:00
|
|
|
|
2014-03-10 11:27:59 -06:00
|
|
|
if (!hsize && i >= nfrags && skb_headlen(list_skb) &&
|
|
|
|
(skb_headlen(list_skb) == len || sg)) {
|
|
|
|
BUG_ON(skb_headlen(list_skb) > len);
|
2013-11-21 12:10:04 -07:00
|
|
|
|
|
|
|
i = 0;
|
2014-03-10 11:27:59 -06:00
|
|
|
nfrags = skb_shinfo(list_skb)->nr_frags;
|
|
|
|
frag = skb_shinfo(list_skb)->frags;
|
2014-03-10 11:28:08 -06:00
|
|
|
frag_skb = list_skb;
|
2014-03-10 11:27:59 -06:00
|
|
|
pos += skb_headlen(list_skb);
|
2013-11-21 12:10:04 -07:00
|
|
|
|
|
|
|
while (pos < offset + len) {
|
|
|
|
BUG_ON(i >= nfrags);
|
|
|
|
|
2014-03-10 10:29:14 -06:00
|
|
|
size = skb_frag_size(frag);
|
2013-11-21 12:10:04 -07:00
|
|
|
if (pos + size > offset + len)
|
|
|
|
break;
|
|
|
|
|
|
|
|
i++;
|
|
|
|
pos += size;
|
2014-03-10 10:29:14 -06:00
|
|
|
frag++;
|
2013-11-21 12:10:04 -07:00
|
|
|
}
|
2008-12-16 00:26:06 -07:00
|
|
|
|
2014-03-10 11:27:59 -06:00
|
|
|
nskb = skb_clone(list_skb, GFP_ATOMIC);
|
|
|
|
list_skb = list_skb->next;
|
2008-12-16 00:26:06 -07:00
|
|
|
|
|
|
|
if (unlikely(!nskb))
|
|
|
|
goto err;
|
|
|
|
|
2013-11-21 12:10:04 -07:00
|
|
|
if (unlikely(pskb_trim(nskb, len))) {
|
|
|
|
kfree_skb(nskb);
|
|
|
|
goto err;
|
|
|
|
}
|
|
|
|
|
2012-05-04 08:26:56 -06:00
|
|
|
hsize = skb_end_offset(nskb);
|
2008-12-16 00:26:06 -07:00
|
|
|
if (skb_cow_head(nskb, doffset + headroom)) {
|
|
|
|
kfree_skb(nskb);
|
|
|
|
goto err;
|
|
|
|
}
|
|
|
|
|
2012-05-04 08:26:56 -06:00
|
|
|
nskb->truesize += skb_end_offset(nskb) - hsize;
|
2008-12-16 00:26:06 -07:00
|
|
|
skb_release_head_state(nskb);
|
|
|
|
__skb_push(nskb, doffset);
|
|
|
|
} else {
|
2012-07-31 17:44:19 -06:00
|
|
|
nskb = __alloc_skb(hsize + doffset + headroom,
|
2014-03-10 10:29:19 -06:00
|
|
|
GFP_ATOMIC, skb_alloc_rx_flag(head_skb),
|
2012-07-31 17:44:19 -06:00
|
|
|
NUMA_NO_NODE);
|
2008-12-16 00:26:06 -07:00
|
|
|
|
|
|
|
if (unlikely(!nskb))
|
|
|
|
goto err;
|
|
|
|
|
|
|
|
skb_reserve(nskb, headroom);
|
|
|
|
__skb_put(nskb, doffset);
|
|
|
|
}
|
2006-06-22 04:02:40 -06:00
|
|
|
|
|
|
|
if (segs)
|
|
|
|
tail->next = nskb;
|
|
|
|
else
|
|
|
|
segs = nskb;
|
|
|
|
tail = nskb;
|
|
|
|
|
2014-03-10 10:29:19 -06:00
|
|
|
__copy_skb_header(nskb, head_skb);
|
2006-06-22 04:02:40 -06:00
|
|
|
|
2013-10-19 12:42:54 -06:00
|
|
|
skb_headers_offset_update(nskb, skb_headroom(nskb) - headroom);
|
2014-07-31 08:33:06 -06:00
|
|
|
skb_reset_mac_len(nskb);
|
2013-02-14 07:02:41 -07:00
|
|
|
|
2014-03-10 10:29:19 -06:00
|
|
|
skb_copy_from_linear_data_offset(head_skb, -tnl_hlen,
|
2013-02-14 07:02:41 -07:00
|
|
|
nskb->data - tnl_hlen,
|
|
|
|
doffset + tnl_hlen);
|
2008-12-16 00:26:06 -07:00
|
|
|
|
2013-11-21 12:10:04 -07:00
|
|
|
if (nskb->len == len + doffset)
|
2013-05-19 09:46:49 -06:00
|
|
|
goto perform_csum_check;
|
2008-12-16 00:26:06 -07:00
|
|
|
|
2016-02-05 16:27:43 -07:00
|
|
|
if (!sg) {
|
|
|
|
if (!nskb->remcsum_offload)
|
|
|
|
nskb->ip_summed = CHECKSUM_NONE;
|
2016-02-05 16:27:37 -07:00
|
|
|
SKB_GSO_CB(nskb)->csum =
|
|
|
|
skb_copy_and_csum_bits(head_skb, offset,
|
|
|
|
skb_put(nskb, len),
|
|
|
|
len, 0);
|
2014-06-04 18:20:02 -06:00
|
|
|
SKB_GSO_CB(nskb)->csum_start =
|
2016-02-05 16:27:37 -07:00
|
|
|
skb_headroom(nskb) + doffset;
|
2006-06-22 04:02:40 -06:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2014-03-10 10:29:04 -06:00
|
|
|
nskb_frag = skb_shinfo(nskb)->frags;
|
2006-06-22 04:02:40 -06:00
|
|
|
|
2014-03-10 10:29:19 -06:00
|
|
|
skb_copy_from_linear_data_offset(head_skb, offset,
|
2007-03-27 15:55:52 -06:00
|
|
|
skb_put(nskb, hsize), hsize);
|
2006-06-22 04:02:40 -06:00
|
|
|
|
2014-03-10 10:29:19 -06:00
|
|
|
skb_shinfo(nskb)->tx_flags = skb_shinfo(head_skb)->tx_flags &
|
|
|
|
SKBTX_SHARED_FRAG;
|
2013-01-25 13:34:37 -07:00
|
|
|
|
2013-11-21 12:10:04 -07:00
|
|
|
while (pos < offset + len) {
|
|
|
|
if (i >= nfrags) {
|
2014-03-10 11:27:59 -06:00
|
|
|
BUG_ON(skb_headlen(list_skb));
|
2013-11-21 12:10:04 -07:00
|
|
|
|
|
|
|
i = 0;
|
2014-03-10 11:27:59 -06:00
|
|
|
nfrags = skb_shinfo(list_skb)->nr_frags;
|
|
|
|
frag = skb_shinfo(list_skb)->frags;
|
2014-03-10 11:28:08 -06:00
|
|
|
frag_skb = list_skb;
|
2013-11-21 12:10:04 -07:00
|
|
|
|
|
|
|
BUG_ON(!nfrags);
|
|
|
|
|
2014-03-10 11:27:59 -06:00
|
|
|
list_skb = list_skb->next;
|
2013-11-21 12:10:04 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
if (unlikely(skb_shinfo(nskb)->nr_frags >=
|
|
|
|
MAX_SKB_FRAGS)) {
|
|
|
|
net_warn_ratelimited(
|
|
|
|
"skb_segment: too many frags: %u %u\n",
|
|
|
|
pos, mss);
|
|
|
|
goto err;
|
|
|
|
}
|
|
|
|
|
2014-03-10 11:28:08 -06:00
|
|
|
if (unlikely(skb_orphan_frags(frag_skb, GFP_ATOMIC)))
|
|
|
|
goto err;
|
|
|
|
|
2014-03-10 10:29:14 -06:00
|
|
|
*nskb_frag = *frag;
|
2014-03-10 10:29:04 -06:00
|
|
|
__skb_frag_ref(nskb_frag);
|
|
|
|
size = skb_frag_size(nskb_frag);
|
2006-06-22 04:02:40 -06:00
|
|
|
|
|
|
|
if (pos < offset) {
|
2014-03-10 10:29:04 -06:00
|
|
|
nskb_frag->page_offset += offset - pos;
|
|
|
|
skb_frag_size_sub(nskb_frag, offset - pos);
|
2006-06-22 04:02:40 -06:00
|
|
|
}
|
|
|
|
|
2008-12-16 00:26:06 -07:00
|
|
|
skb_shinfo(nskb)->nr_frags++;
|
2006-06-22 04:02:40 -06:00
|
|
|
|
|
|
|
if (pos + size <= offset + len) {
|
|
|
|
i++;
|
2014-03-10 10:29:14 -06:00
|
|
|
frag++;
|
2006-06-22 04:02:40 -06:00
|
|
|
pos += size;
|
|
|
|
} else {
|
2014-03-10 10:29:04 -06:00
|
|
|
skb_frag_size_sub(nskb_frag, pos + size - (offset + len));
|
2008-12-16 00:26:06 -07:00
|
|
|
goto skip_fraglist;
|
2006-06-22 04:02:40 -06:00
|
|
|
}
|
|
|
|
|
2014-03-10 10:29:04 -06:00
|
|
|
nskb_frag++;
|
2006-06-22 04:02:40 -06:00
|
|
|
}
|
|
|
|
|
2008-12-16 00:26:06 -07:00
|
|
|
skip_fraglist:
|
2006-06-22 04:02:40 -06:00
|
|
|
nskb->data_len = len - hsize;
|
|
|
|
nskb->len += nskb->data_len;
|
|
|
|
nskb->truesize += nskb->data_len;
|
2013-03-07 02:28:01 -07:00
|
|
|
|
2013-05-19 09:46:49 -06:00
|
|
|
perform_csum_check:
|
2016-02-05 16:27:43 -07:00
|
|
|
if (!csum) {
|
2016-02-05 16:27:55 -07:00
|
|
|
if (skb_has_shared_frag(nskb)) {
|
|
|
|
err = __skb_linearize(nskb);
|
|
|
|
if (err)
|
|
|
|
goto err;
|
|
|
|
}
|
2016-02-05 16:27:43 -07:00
|
|
|
if (!nskb->remcsum_offload)
|
|
|
|
nskb->ip_summed = CHECKSUM_NONE;
|
2016-02-05 16:27:37 -07:00
|
|
|
SKB_GSO_CB(nskb)->csum =
|
|
|
|
skb_checksum(nskb, doffset,
|
|
|
|
nskb->len - doffset, 0);
|
2014-06-04 18:20:02 -06:00
|
|
|
SKB_GSO_CB(nskb)->csum_start =
|
2016-02-05 16:27:37 -07:00
|
|
|
skb_headroom(nskb) + doffset;
|
2013-03-07 02:28:01 -07:00
|
|
|
}
|
2014-03-10 10:29:19 -06:00
|
|
|
} while ((offset += len) < head_skb->len);
|
2006-06-22 04:02:40 -06:00
|
|
|
|
2014-10-03 21:59:19 -06:00
|
|
|
/* Some callers want to get the end of the list.
|
|
|
|
* Put it in segs->prev to avoid walking the list.
|
|
|
|
* (see validate_xmit_skb_list() for example)
|
|
|
|
*/
|
|
|
|
segs->prev = tail;
|
2014-10-27 11:30:51 -06:00
|
|
|
|
2016-04-10 19:45:03 -06:00
|
|
|
if (partial_segs) {
|
2016-09-19 04:58:47 -06:00
|
|
|
struct sk_buff *iter;
|
2016-04-10 19:45:03 -06:00
|
|
|
int type = skb_shinfo(head_skb)->gso_type;
|
2016-09-19 04:58:47 -06:00
|
|
|
unsigned short gso_size = skb_shinfo(head_skb)->gso_size;
|
2016-04-10 19:45:03 -06:00
|
|
|
|
|
|
|
/* Update type to add partial and then remove dodgy if set */
|
2016-09-19 04:58:47 -06:00
|
|
|
type |= (features & NETIF_F_GSO_PARTIAL) / NETIF_F_GSO_PARTIAL * SKB_GSO_PARTIAL;
|
2016-04-10 19:45:03 -06:00
|
|
|
type &= ~SKB_GSO_DODGY;
|
|
|
|
|
|
|
|
/* Update GSO info and prepare to start updating headers on
|
|
|
|
* our way back down the stack of protocols.
|
|
|
|
*/
|
2016-09-19 04:58:47 -06:00
|
|
|
for (iter = segs; iter; iter = iter->next) {
|
|
|
|
skb_shinfo(iter)->gso_size = gso_size;
|
|
|
|
skb_shinfo(iter)->gso_segs = partial_segs;
|
|
|
|
skb_shinfo(iter)->gso_type = type;
|
|
|
|
SKB_GSO_CB(iter)->data_offset = skb_headroom(iter) + doffset;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (tail->len - doffset <= gso_size)
|
|
|
|
skb_shinfo(tail)->gso_size = 0;
|
|
|
|
else if (tail != segs)
|
|
|
|
skb_shinfo(tail)->gso_segs = DIV_ROUND_UP(tail->len - doffset, gso_size);
|
2016-04-10 19:45:03 -06:00
|
|
|
}
|
|
|
|
|
2014-10-27 11:30:51 -06:00
|
|
|
/* Following permits correct backpressure, for protocols
|
|
|
|
* using skb_set_owner_w().
|
|
|
|
* Idea is to tranfert ownership from head_skb to last segment.
|
|
|
|
*/
|
|
|
|
if (head_skb->destructor == sock_wfree) {
|
|
|
|
swap(tail->truesize, head_skb->truesize);
|
|
|
|
swap(tail->destructor, head_skb->destructor);
|
|
|
|
swap(tail->sk, head_skb->sk);
|
|
|
|
}
|
2006-06-22 04:02:40 -06:00
|
|
|
return segs;
|
|
|
|
|
|
|
|
err:
|
2013-12-20 15:29:08 -07:00
|
|
|
kfree_skb_list(segs);
|
2006-06-22 04:02:40 -06:00
|
|
|
return ERR_PTR(err);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_segment);
|
|
|
|
|
2008-12-16 00:42:33 -07:00
|
|
|
int skb_gro_receive(struct sk_buff **head, struct sk_buff *skb)
|
|
|
|
{
|
2013-10-08 10:02:23 -06:00
|
|
|
struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb);
|
2009-05-26 12:50:22 -06:00
|
|
|
unsigned int offset = skb_gro_offset(skb);
|
|
|
|
unsigned int headlen = skb_headlen(skb);
|
2013-10-08 10:02:23 -06:00
|
|
|
unsigned int len = skb_gro_len(skb);
|
2015-03-05 14:47:48 -07:00
|
|
|
struct sk_buff *lp, *p = *head;
|
2012-05-02 17:33:21 -06:00
|
|
|
unsigned int delta_truesize;
|
2008-12-16 00:42:33 -07:00
|
|
|
|
2013-10-08 10:02:23 -06:00
|
|
|
if (unlikely(p->len + len >= 65536))
|
2008-12-16 00:42:33 -07:00
|
|
|
return -E2BIG;
|
|
|
|
|
2014-05-16 12:34:37 -06:00
|
|
|
lp = NAPI_GRO_CB(p)->last;
|
2013-10-08 10:02:23 -06:00
|
|
|
pinfo = skb_shinfo(lp);
|
|
|
|
|
|
|
|
if (headlen <= offset) {
|
2009-05-26 12:50:19 -06:00
|
|
|
skb_frag_t *frag;
|
2009-05-26 12:50:32 -06:00
|
|
|
skb_frag_t *frag2;
|
2009-05-26 12:50:33 -06:00
|
|
|
int i = skbinfo->nr_frags;
|
|
|
|
int nr_frags = pinfo->nr_frags + i;
|
2009-05-26 12:50:32 -06:00
|
|
|
|
|
|
|
if (nr_frags > MAX_SKB_FRAGS)
|
2013-10-08 10:02:23 -06:00
|
|
|
goto merge;
|
2009-01-29 07:19:51 -07:00
|
|
|
|
2013-10-08 10:02:23 -06:00
|
|
|
offset -= headlen;
|
2009-05-26 12:50:33 -06:00
|
|
|
pinfo->nr_frags = nr_frags;
|
|
|
|
skbinfo->nr_frags = 0;
|
2009-01-29 07:19:50 -07:00
|
|
|
|
2009-05-26 12:50:33 -06:00
|
|
|
frag = pinfo->frags + nr_frags;
|
|
|
|
frag2 = skbinfo->frags + i;
|
2009-05-26 12:50:32 -06:00
|
|
|
do {
|
|
|
|
*--frag = *--frag2;
|
|
|
|
} while (--i);
|
2009-01-04 17:13:40 -07:00
|
|
|
|
2009-05-26 12:50:32 -06:00
|
|
|
frag->page_offset += offset;
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_sub(frag, offset);
|
2009-01-14 21:40:03 -07:00
|
|
|
|
2012-05-02 17:33:21 -06:00
|
|
|
/* all fragments truesize : remove (head size + sk_buff) */
|
2012-05-04 08:26:56 -06:00
|
|
|
delta_truesize = skb->truesize -
|
|
|
|
SKB_TRUESIZE(skb_end_offset(skb));
|
2012-05-02 17:33:21 -06:00
|
|
|
|
2009-01-14 21:40:03 -07:00
|
|
|
skb->truesize -= skb->data_len;
|
|
|
|
skb->len -= skb->data_len;
|
|
|
|
skb->data_len = 0;
|
|
|
|
|
2012-05-02 17:33:21 -06:00
|
|
|
NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
|
2009-01-04 17:13:40 -07:00
|
|
|
goto done;
|
2012-04-30 02:10:34 -06:00
|
|
|
} else if (skb->head_frag) {
|
|
|
|
int nr_frags = pinfo->nr_frags;
|
|
|
|
skb_frag_t *frag = pinfo->frags + nr_frags;
|
|
|
|
struct page *page = virt_to_head_page(skb->head);
|
|
|
|
unsigned int first_size = headlen - offset;
|
|
|
|
unsigned int first_offset;
|
|
|
|
|
|
|
|
if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
|
2013-10-08 10:02:23 -06:00
|
|
|
goto merge;
|
2012-04-30 02:10:34 -06:00
|
|
|
|
|
|
|
first_offset = skb->data -
|
|
|
|
(unsigned char *)page_address(page) +
|
|
|
|
offset;
|
|
|
|
|
|
|
|
pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
|
|
|
|
|
|
|
|
frag->page.p = page;
|
|
|
|
frag->page_offset = first_offset;
|
|
|
|
skb_frag_size_set(frag, first_size);
|
|
|
|
|
|
|
|
memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
|
|
|
|
/* We dont need to clear skbinfo->nr_frags here */
|
|
|
|
|
2012-05-02 17:33:21 -06:00
|
|
|
delta_truesize = skb->truesize - SKB_DATA_ALIGN(sizeof(struct sk_buff));
|
2012-04-30 02:10:34 -06:00
|
|
|
NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
|
|
|
|
goto done;
|
2013-10-08 10:02:23 -06:00
|
|
|
}
|
2008-12-16 00:42:33 -07:00
|
|
|
|
|
|
|
merge:
|
2012-05-02 17:33:21 -06:00
|
|
|
delta_truesize = skb->truesize;
|
2009-05-26 12:50:22 -06:00
|
|
|
if (offset > headlen) {
|
2011-01-24 05:08:48 -07:00
|
|
|
unsigned int eat = offset - headlen;
|
|
|
|
|
|
|
|
skbinfo->frags[0].page_offset += eat;
|
2011-10-18 15:00:24 -06:00
|
|
|
skb_frag_size_sub(&skbinfo->frags[0], eat);
|
2011-01-24 05:08:48 -07:00
|
|
|
skb->data_len -= eat;
|
|
|
|
skb->len -= eat;
|
2009-05-26 12:50:22 -06:00
|
|
|
offset = headlen;
|
2009-02-05 22:26:52 -07:00
|
|
|
}
|
|
|
|
|
2009-05-26 12:50:22 -06:00
|
|
|
__skb_pull(skb, offset);
|
2009-02-05 22:26:52 -07:00
|
|
|
|
2014-05-16 12:34:37 -06:00
|
|
|
if (NAPI_GRO_CB(p)->last == p)
|
2013-10-08 10:02:23 -06:00
|
|
|
skb_shinfo(p)->frag_list = skb;
|
|
|
|
else
|
|
|
|
NAPI_GRO_CB(p)->last->next = skb;
|
2012-12-06 06:54:59 -07:00
|
|
|
NAPI_GRO_CB(p)->last = skb;
|
2014-09-22 17:29:32 -06:00
|
|
|
__skb_header_release(skb);
|
2013-10-08 10:02:23 -06:00
|
|
|
lp = p;
|
2008-12-16 00:42:33 -07:00
|
|
|
|
2009-01-04 17:13:40 -07:00
|
|
|
done:
|
|
|
|
NAPI_GRO_CB(p)->count++;
|
2009-01-17 12:48:13 -07:00
|
|
|
p->data_len += len;
|
2012-05-02 17:33:21 -06:00
|
|
|
p->truesize += delta_truesize;
|
2009-01-17 12:48:13 -07:00
|
|
|
p->len += len;
|
2013-10-08 10:02:23 -06:00
|
|
|
if (lp != p) {
|
|
|
|
lp->data_len += len;
|
|
|
|
lp->truesize += delta_truesize;
|
|
|
|
lp->len += len;
|
|
|
|
}
|
2008-12-16 00:42:33 -07:00
|
|
|
NAPI_GRO_CB(skb)->same_flow = 1;
|
|
|
|
return 0;
|
|
|
|
}
|
2016-06-02 12:05:39 -06:00
|
|
|
EXPORT_SYMBOL_GPL(skb_gro_receive);
|
2008-12-16 00:42:33 -07:00
|
|
|
|
2005-04-16 16:20:36 -06:00
|
|
|
void __init skb_init(void)
|
|
|
|
{
|
|
|
|
skbuff_head_cache = kmem_cache_create("skbuff_head_cache",
|
|
|
|
sizeof(struct sk_buff),
|
|
|
|
0,
|
2006-08-26 20:25:52 -06:00
|
|
|
SLAB_HWCACHE_ALIGN|SLAB_PANIC,
|
2007-07-19 19:11:58 -06:00
|
|
|
NULL);
|
2005-08-17 15:57:30 -06:00
|
|
|
skbuff_fclone_cache = kmem_cache_create("skbuff_fclone_cache",
|
2014-09-29 14:29:15 -06:00
|
|
|
sizeof(struct sk_buff_fclones),
|
2005-08-17 15:57:30 -06:00
|
|
|
0,
|
2006-08-26 20:25:52 -06:00
|
|
|
SLAB_HWCACHE_ALIGN|SLAB_PANIC,
|
2007-07-19 19:11:58 -06:00
|
|
|
NULL);
|
2005-04-16 16:20:36 -06:00
|
|
|
}
|
|
|
|
|
2007-04-02 21:19:53 -06:00
|
|
|
/**
|
|
|
|
* skb_to_sgvec - Fill a scatter-gather list from a socket buffer
|
|
|
|
* @skb: Socket buffer containing the buffers to be mapped
|
|
|
|
* @sg: The scatter-gather list to map into
|
|
|
|
* @offset: The offset into the buffer's contents to start mapping
|
|
|
|
* @len: Length of buffer space to be mapped
|
|
|
|
*
|
|
|
|
* Fill the specified scatter-gather list with mappings/pointers into a
|
|
|
|
* region of the buffer space attached to a socket buffer.
|
|
|
|
*/
|
2007-10-30 22:29:29 -06:00
|
|
|
static int
|
|
|
|
__skb_to_sgvec(struct sk_buff *skb, struct scatterlist *sg, int offset, int len)
|
2007-04-02 21:19:53 -06:00
|
|
|
{
|
2007-04-27 16:21:23 -06:00
|
|
|
int start = skb_headlen(skb);
|
|
|
|
int i, copy = start - offset;
|
2009-06-09 01:18:59 -06:00
|
|
|
struct sk_buff *frag_iter;
|
2007-04-02 21:19:53 -06:00
|
|
|
int elt = 0;
|
|
|
|
|
|
|
|
if (copy > 0) {
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
2007-10-24 03:20:47 -06:00
|
|
|
sg_set_buf(sg, skb->data + offset, copy);
|
2007-04-02 21:19:53 -06:00
|
|
|
elt++;
|
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return elt;
|
|
|
|
offset += copy;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
|
2007-04-27 16:21:23 -06:00
|
|
|
int end;
|
2007-04-02 21:19:53 -06:00
|
|
|
|
2008-07-25 22:43:18 -06:00
|
|
|
WARN_ON(start > offset + len);
|
2007-04-27 16:21:23 -06:00
|
|
|
|
2011-10-18 15:00:24 -06:00
|
|
|
end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
|
2007-04-02 21:19:53 -06:00
|
|
|
if ((copy = end - offset) > 0) {
|
|
|
|
skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
|
|
|
|
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
2011-08-22 17:44:58 -06:00
|
|
|
sg_set_page(&sg[elt], skb_frag_page(frag), copy,
|
2007-10-24 03:20:47 -06:00
|
|
|
frag->page_offset+offset-start);
|
2007-04-02 21:19:53 -06:00
|
|
|
elt++;
|
|
|
|
if (!(len -= copy))
|
|
|
|
return elt;
|
|
|
|
offset += copy;
|
|
|
|
}
|
2007-04-27 16:21:23 -06:00
|
|
|
start = end;
|
2007-04-02 21:19:53 -06:00
|
|
|
}
|
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
skb_walk_frags(skb, frag_iter) {
|
|
|
|
int end;
|
2007-04-27 16:21:23 -06:00
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
WARN_ON(start > offset + len);
|
2007-04-02 21:19:53 -06:00
|
|
|
|
2009-06-09 01:18:59 -06:00
|
|
|
end = start + frag_iter->len;
|
|
|
|
if ((copy = end - offset) > 0) {
|
|
|
|
if (copy > len)
|
|
|
|
copy = len;
|
|
|
|
elt += __skb_to_sgvec(frag_iter, sg+elt, offset - start,
|
|
|
|
copy);
|
|
|
|
if ((len -= copy) == 0)
|
|
|
|
return elt;
|
|
|
|
offset += copy;
|
2007-04-02 21:19:53 -06:00
|
|
|
}
|
2009-06-09 01:18:59 -06:00
|
|
|
start = end;
|
2007-04-02 21:19:53 -06:00
|
|
|
}
|
|
|
|
BUG_ON(len);
|
|
|
|
return elt;
|
|
|
|
}
|
|
|
|
|
2014-01-17 18:54:23 -07:00
|
|
|
/* As compared with skb_to_sgvec, skb_to_sgvec_nomark only map skb to given
|
|
|
|
* sglist without mark the sg which contain last skb data as the end.
|
|
|
|
* So the caller can mannipulate sg list as will when padding new data after
|
|
|
|
* the first call without calling sg_unmark_end to expend sg list.
|
|
|
|
*
|
|
|
|
* Scenario to use skb_to_sgvec_nomark:
|
|
|
|
* 1. sg_init_table
|
|
|
|
* 2. skb_to_sgvec_nomark(payload1)
|
|
|
|
* 3. skb_to_sgvec_nomark(payload2)
|
|
|
|
*
|
|
|
|
* This is equivalent to:
|
|
|
|
* 1. sg_init_table
|
|
|
|
* 2. skb_to_sgvec(payload1)
|
|
|
|
* 3. sg_unmark_end
|
|
|
|
* 4. skb_to_sgvec(payload2)
|
|
|
|
*
|
|
|
|
* When mapping mutilple payload conditionally, skb_to_sgvec_nomark
|
|
|
|
* is more preferable.
|
|
|
|
*/
|
|
|
|
int skb_to_sgvec_nomark(struct sk_buff *skb, struct scatterlist *sg,
|
|
|
|
int offset, int len)
|
|
|
|
{
|
|
|
|
return __skb_to_sgvec(skb, sg, offset, len);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_to_sgvec_nomark);
|
|
|
|
|
2007-10-30 22:29:29 -06:00
|
|
|
int skb_to_sgvec(struct sk_buff *skb, struct scatterlist *sg, int offset, int len)
|
|
|
|
{
|
|
|
|
int nsg = __skb_to_sgvec(skb, sg, offset, len);
|
|
|
|
|
2007-10-31 05:06:37 -06:00
|
|
|
sg_mark_end(&sg[nsg - 1]);
|
2007-10-30 22:29:29 -06:00
|
|
|
|
|
|
|
return nsg;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL_GPL(skb_to_sgvec);
|
2007-10-30 22:29:29 -06:00
|
|
|
|
2007-04-02 21:19:53 -06:00
|
|
|
/**
|
|
|
|
* skb_cow_data - Check that a socket buffer's data buffers are writable
|
|
|
|
* @skb: The socket buffer to check.
|
|
|
|
* @tailbits: Amount of trailing space to be added
|
|
|
|
* @trailer: Returned pointer to the skb where the @tailbits space begins
|
|
|
|
*
|
|
|
|
* Make sure that the data buffers attached to a socket buffer are
|
|
|
|
* writable. If they are not, private copies are made of the data buffers
|
|
|
|
* and the socket buffer is set to use these instead.
|
|
|
|
*
|
|
|
|
* If @tailbits is given, make sure that there is space to write @tailbits
|
|
|
|
* bytes of data beyond current end of socket buffer. @trailer will be
|
|
|
|
* set to point to the skb in which this space begins.
|
|
|
|
*
|
|
|
|
* The number of scatterlist elements required to completely map the
|
|
|
|
* COW'd and extended socket buffer will be returned.
|
|
|
|
*/
|
|
|
|
int skb_cow_data(struct sk_buff *skb, int tailbits, struct sk_buff **trailer)
|
|
|
|
{
|
|
|
|
int copyflag;
|
|
|
|
int elt;
|
|
|
|
struct sk_buff *skb1, **skb_p;
|
|
|
|
|
|
|
|
/* If skb is cloned or its head is paged, reallocate
|
|
|
|
* head pulling out all the pages (pages are considered not writable
|
|
|
|
* at the moment even if they are anonymous).
|
|
|
|
*/
|
|
|
|
if ((skb_cloned(skb) || skb_shinfo(skb)->nr_frags) &&
|
|
|
|
__pskb_pull_tail(skb, skb_pagelen(skb)-skb_headlen(skb)) == NULL)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
/* Easy case. Most of packets will go this way. */
|
2010-08-23 01:13:46 -06:00
|
|
|
if (!skb_has_frag_list(skb)) {
|
2007-04-02 21:19:53 -06:00
|
|
|
/* A little of trouble, not enough of space for trailer.
|
|
|
|
* This should not happen, when stack is tuned to generate
|
|
|
|
* good frames. OK, on miss we reallocate and reserve even more
|
|
|
|
* space, 128 bytes is fair. */
|
|
|
|
|
|
|
|
if (skb_tailroom(skb) < tailbits &&
|
|
|
|
pskb_expand_head(skb, 0, tailbits-skb_tailroom(skb)+128, GFP_ATOMIC))
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
/* Voila! */
|
|
|
|
*trailer = skb;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Misery. We are in troubles, going to mincer fragments... */
|
|
|
|
|
|
|
|
elt = 1;
|
|
|
|
skb_p = &skb_shinfo(skb)->frag_list;
|
|
|
|
copyflag = 0;
|
|
|
|
|
|
|
|
while ((skb1 = *skb_p) != NULL) {
|
|
|
|
int ntail = 0;
|
|
|
|
|
|
|
|
/* The fragment is partially pulled by someone,
|
|
|
|
* this can happen on input. Copy it and everything
|
|
|
|
* after it. */
|
|
|
|
|
|
|
|
if (skb_shared(skb1))
|
|
|
|
copyflag = 1;
|
|
|
|
|
|
|
|
/* If the skb is the last, worry about trailer. */
|
|
|
|
|
|
|
|
if (skb1->next == NULL && tailbits) {
|
|
|
|
if (skb_shinfo(skb1)->nr_frags ||
|
2010-08-23 01:13:46 -06:00
|
|
|
skb_has_frag_list(skb1) ||
|
2007-04-02 21:19:53 -06:00
|
|
|
skb_tailroom(skb1) < tailbits)
|
|
|
|
ntail = tailbits + 128;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (copyflag ||
|
|
|
|
skb_cloned(skb1) ||
|
|
|
|
ntail ||
|
|
|
|
skb_shinfo(skb1)->nr_frags ||
|
2010-08-23 01:13:46 -06:00
|
|
|
skb_has_frag_list(skb1)) {
|
2007-04-02 21:19:53 -06:00
|
|
|
struct sk_buff *skb2;
|
|
|
|
|
|
|
|
/* Fuck, we are miserable poor guys... */
|
|
|
|
if (ntail == 0)
|
|
|
|
skb2 = skb_copy(skb1, GFP_ATOMIC);
|
|
|
|
else
|
|
|
|
skb2 = skb_copy_expand(skb1,
|
|
|
|
skb_headroom(skb1),
|
|
|
|
ntail,
|
|
|
|
GFP_ATOMIC);
|
|
|
|
if (unlikely(skb2 == NULL))
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
if (skb1->sk)
|
|
|
|
skb_set_owner_w(skb2, skb1->sk);
|
|
|
|
|
|
|
|
/* Looking around. Are we still alive?
|
|
|
|
* OK, link new skb, drop old one */
|
|
|
|
|
|
|
|
skb2->next = skb1->next;
|
|
|
|
*skb_p = skb2;
|
|
|
|
kfree_skb(skb1);
|
|
|
|
skb1 = skb2;
|
|
|
|
}
|
|
|
|
elt++;
|
|
|
|
*trailer = skb1;
|
|
|
|
skb_p = &skb1->next;
|
|
|
|
}
|
|
|
|
|
|
|
|
return elt;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL_GPL(skb_cow_data);
|
2007-04-02 21:19:53 -06:00
|
|
|
|
2010-06-01 00:44:05 -06:00
|
|
|
static void sock_rmem_free(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
struct sock *sk = skb->sk;
|
|
|
|
|
|
|
|
atomic_sub(skb->truesize, &sk->sk_rmem_alloc);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Note: We dont mem charge error packets (no sk_forward_alloc changes)
|
|
|
|
*/
|
|
|
|
int sock_queue_err_skb(struct sock *sk, struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
if (atomic_read(&sk->sk_rmem_alloc) + skb->truesize >=
|
2012-04-14 23:58:06 -06:00
|
|
|
(unsigned int)sk->sk_rcvbuf)
|
2010-06-01 00:44:05 -06:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
skb_orphan(skb);
|
|
|
|
skb->sk = sk;
|
|
|
|
skb->destructor = sock_rmem_free;
|
|
|
|
atomic_add(skb->truesize, &sk->sk_rmem_alloc);
|
|
|
|
|
2011-05-18 00:21:31 -06:00
|
|
|
/* before exiting rcu section, make sure dst is refcounted */
|
|
|
|
skb_dst_force(skb);
|
|
|
|
|
2010-06-01 00:44:05 -06:00
|
|
|
skb_queue_tail(&sk->sk_error_queue, skb);
|
|
|
|
if (!sock_flag(sk, SOCK_DEAD))
|
2014-04-11 14:15:36 -06:00
|
|
|
sk->sk_data_ready(sk);
|
2010-06-01 00:44:05 -06:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(sock_queue_err_skb);
|
|
|
|
|
2016-11-30 12:01:08 -07:00
|
|
|
static bool is_icmp_err_skb(const struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
return skb && (SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP ||
|
|
|
|
SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP6);
|
|
|
|
}
|
|
|
|
|
2014-08-31 19:30:27 -06:00
|
|
|
struct sk_buff *sock_dequeue_err_skb(struct sock *sk)
|
|
|
|
{
|
|
|
|
struct sk_buff_head *q = &sk->sk_error_queue;
|
2016-11-30 12:01:08 -07:00
|
|
|
struct sk_buff *skb, *skb_next = NULL;
|
|
|
|
bool icmp_next = false;
|
2015-02-18 06:47:55 -07:00
|
|
|
unsigned long flags;
|
2014-08-31 19:30:27 -06:00
|
|
|
|
2015-02-18 06:47:55 -07:00
|
|
|
spin_lock_irqsave(&q->lock, flags);
|
2014-08-31 19:30:27 -06:00
|
|
|
skb = __skb_dequeue(q);
|
|
|
|
if (skb && (skb_next = skb_peek(q)))
|
2016-11-30 12:01:08 -07:00
|
|
|
icmp_next = is_icmp_err_skb(skb_next);
|
2015-02-18 06:47:55 -07:00
|
|
|
spin_unlock_irqrestore(&q->lock, flags);
|
2014-08-31 19:30:27 -06:00
|
|
|
|
2016-11-30 12:01:08 -07:00
|
|
|
if (is_icmp_err_skb(skb) && !icmp_next)
|
|
|
|
sk->sk_err = 0;
|
|
|
|
|
|
|
|
if (skb_next)
|
2014-08-31 19:30:27 -06:00
|
|
|
sk->sk_error_report(sk);
|
|
|
|
|
|
|
|
return skb;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(sock_dequeue_err_skb);
|
|
|
|
|
2014-09-10 16:05:26 -06:00
|
|
|
/**
|
|
|
|
* skb_clone_sk - create clone of skb, and take reference to socket
|
|
|
|
* @skb: the skb to clone
|
|
|
|
*
|
|
|
|
* This function creates a clone of a buffer that holds a reference on
|
|
|
|
* sk_refcnt. Buffers created via this function are meant to be
|
|
|
|
* returned using sock_queue_err_skb, or free via kfree_skb.
|
|
|
|
*
|
|
|
|
* When passing buffers allocated with this function to sock_queue_err_skb
|
|
|
|
* it is necessary to wrap the call with sock_hold/sock_put in order to
|
|
|
|
* prevent the socket from being released prior to being enqueued on
|
|
|
|
* the sk_error_queue.
|
|
|
|
*/
|
2014-09-04 11:31:35 -06:00
|
|
|
struct sk_buff *skb_clone_sk(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
struct sock *sk = skb->sk;
|
|
|
|
struct sk_buff *clone;
|
|
|
|
|
|
|
|
if (!sk || !atomic_inc_not_zero(&sk->sk_refcnt))
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
clone = skb_clone(skb, GFP_ATOMIC);
|
|
|
|
if (!clone) {
|
|
|
|
sock_put(sk);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
clone->sk = sk;
|
|
|
|
clone->destructor = sock_efree;
|
|
|
|
|
|
|
|
return clone;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_clone_sk);
|
|
|
|
|
2014-09-04 11:31:10 -06:00
|
|
|
static void __skb_complete_tx_timestamp(struct sk_buff *skb,
|
|
|
|
struct sock *sk,
|
|
|
|
int tstype)
|
2009-02-11 22:03:37 -07:00
|
|
|
{
|
|
|
|
struct sock_exterr_skb *serr;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
serr = SKB_EXT_ERR(skb);
|
|
|
|
memset(serr, 0, sizeof(*serr));
|
|
|
|
serr->ee.ee_errno = ENOMSG;
|
|
|
|
serr->ee.ee_origin = SO_EE_ORIGIN_TIMESTAMPING;
|
2014-08-04 20:11:48 -06:00
|
|
|
serr->ee.ee_info = tstype;
|
net-timestamp: TCP timestamping
TCP timestamping extends SO_TIMESTAMPING to bytestreams.
Bytestreams do not have a 1:1 relationship between send() buffers and
network packets. The feature interprets a send call on a bytestream as
a request for a timestamp for the last byte in that send() buffer.
The choice corresponds to a request for a timestamp when all bytes in
the buffer have been sent. That assumption depends on in-order kernel
transmission. This is the common case. That said, it is possible to
construct a traffic shaping tree that would result in reordering.
The guarantee is strong, then, but not ironclad.
This implementation supports send and sendpages (splice). GSO replaces
one large packet with multiple smaller packets. This patch also copies
the option into the correct smaller packet.
This patch does not yet support timestamping on data in an initial TCP
Fast Open SYN, because that takes a very different data path.
If ID generation in ee_data is enabled, bytestream timestamps return a
byte offset, instead of the packet counter for datagrams.
The implementation supports a single timestamp per packet. It silenty
replaces requests for previous timestamps. To avoid missing tstamps,
flush the tcp queue by disabling Nagle, cork and autocork. Missing
tstamps can be detected by offset when the ee_data ID is enabled.
Implementation details:
- On GSO, the timestamping code can be included in the main loop. I
moved it into its own loop to reduce the impact on the common case
to a single branch.
- To avoid leaking the absolute seqno to userspace, the offset
returned in ee_data must always be relative. It is an offset between
an skb and sk field. The first is always set (also for GSO & ACK).
The second must also never be uninitialized. Only allow the ID
option on sockets in the ESTABLISHED state, for which the seqno
is available. Never reset it to zero (instead, move it to the
current seqno when reenabling the option).
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 20:11:49 -06:00
|
|
|
if (sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID) {
|
2014-08-04 20:11:47 -06:00
|
|
|
serr->ee.ee_data = skb_shinfo(skb)->tskey;
|
2015-12-17 00:39:04 -07:00
|
|
|
if (sk->sk_protocol == IPPROTO_TCP &&
|
|
|
|
sk->sk_type == SOCK_STREAM)
|
net-timestamp: TCP timestamping
TCP timestamping extends SO_TIMESTAMPING to bytestreams.
Bytestreams do not have a 1:1 relationship between send() buffers and
network packets. The feature interprets a send call on a bytestream as
a request for a timestamp for the last byte in that send() buffer.
The choice corresponds to a request for a timestamp when all bytes in
the buffer have been sent. That assumption depends on in-order kernel
transmission. This is the common case. That said, it is possible to
construct a traffic shaping tree that would result in reordering.
The guarantee is strong, then, but not ironclad.
This implementation supports send and sendpages (splice). GSO replaces
one large packet with multiple smaller packets. This patch also copies
the option into the correct smaller packet.
This patch does not yet support timestamping on data in an initial TCP
Fast Open SYN, because that takes a very different data path.
If ID generation in ee_data is enabled, bytestream timestamps return a
byte offset, instead of the packet counter for datagrams.
The implementation supports a single timestamp per packet. It silenty
replaces requests for previous timestamps. To avoid missing tstamps,
flush the tcp queue by disabling Nagle, cork and autocork. Missing
tstamps can be detected by offset when the ee_data ID is enabled.
Implementation details:
- On GSO, the timestamping code can be included in the main loop. I
moved it into its own loop to reduce the impact on the common case
to a single branch.
- To avoid leaking the absolute seqno to userspace, the offset
returned in ee_data must always be relative. It is an offset between
an skb and sk field. The first is always set (also for GSO & ACK).
The second must also never be uninitialized. Only allow the ID
option on sockets in the ESTABLISHED state, for which the seqno
is available. Never reset it to zero (instead, move it to the
current seqno when reenabling the option).
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 20:11:49 -06:00
|
|
|
serr->ee.ee_data -= sk->sk_tskey;
|
|
|
|
}
|
2010-05-29 01:20:48 -06:00
|
|
|
|
2009-02-11 22:03:37 -07:00
|
|
|
err = sock_queue_err_skb(sk, skb);
|
2010-05-29 01:20:48 -06:00
|
|
|
|
2009-02-11 22:03:37 -07:00
|
|
|
if (err)
|
|
|
|
kfree_skb(skb);
|
|
|
|
}
|
2014-09-04 11:31:10 -06:00
|
|
|
|
2015-01-30 11:29:32 -07:00
|
|
|
static bool skb_may_tx_timestamp(struct sock *sk, bool tsonly)
|
|
|
|
{
|
|
|
|
bool ret;
|
|
|
|
|
|
|
|
if (likely(sysctl_tstamp_allow_data || tsonly))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
read_lock_bh(&sk->sk_callback_lock);
|
|
|
|
ret = sk->sk_socket && sk->sk_socket->file &&
|
|
|
|
file_ns_capable(sk->sk_socket->file, &init_user_ns, CAP_NET_RAW);
|
|
|
|
read_unlock_bh(&sk->sk_callback_lock);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2014-09-04 11:31:10 -06:00
|
|
|
void skb_complete_tx_timestamp(struct sk_buff *skb,
|
|
|
|
struct skb_shared_hwtstamps *hwtstamps)
|
|
|
|
{
|
|
|
|
struct sock *sk = skb->sk;
|
|
|
|
|
2015-01-30 11:29:32 -07:00
|
|
|
if (!skb_may_tx_timestamp(sk, false))
|
|
|
|
return;
|
|
|
|
|
2014-09-04 11:31:35 -06:00
|
|
|
/* take a reference to prevent skb_orphan() from freeing the socket */
|
|
|
|
sock_hold(sk);
|
2014-09-04 11:31:10 -06:00
|
|
|
|
2014-09-04 11:31:35 -06:00
|
|
|
*skb_hwtstamps(skb) = *hwtstamps;
|
|
|
|
__skb_complete_tx_timestamp(skb, sk, SCM_TSTAMP_SND);
|
2014-09-04 11:31:10 -06:00
|
|
|
|
|
|
|
sock_put(sk);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_complete_tx_timestamp);
|
|
|
|
|
|
|
|
void __skb_tstamp_tx(struct sk_buff *orig_skb,
|
|
|
|
struct skb_shared_hwtstamps *hwtstamps,
|
|
|
|
struct sock *sk, int tstype)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb;
|
2015-03-11 13:43:55 -06:00
|
|
|
bool tsonly;
|
2014-09-04 11:31:10 -06:00
|
|
|
|
2015-03-11 13:43:55 -06:00
|
|
|
if (!sk)
|
|
|
|
return;
|
|
|
|
|
|
|
|
tsonly = sk->sk_tsflags & SOF_TIMESTAMPING_OPT_TSONLY;
|
|
|
|
if (!skb_may_tx_timestamp(sk, tsonly))
|
2014-09-04 11:31:10 -06:00
|
|
|
return;
|
|
|
|
|
2016-11-28 00:07:18 -07:00
|
|
|
if (tsonly) {
|
|
|
|
#ifdef CONFIG_INET
|
|
|
|
if ((sk->sk_tsflags & SOF_TIMESTAMPING_OPT_STATS) &&
|
|
|
|
sk->sk_protocol == IPPROTO_TCP &&
|
|
|
|
sk->sk_type == SOCK_STREAM)
|
|
|
|
skb = tcp_get_timestamping_opt_stats(sk);
|
|
|
|
else
|
|
|
|
#endif
|
|
|
|
skb = alloc_skb(0, GFP_ATOMIC);
|
|
|
|
} else {
|
2015-01-30 11:29:31 -07:00
|
|
|
skb = skb_clone(orig_skb, GFP_ATOMIC);
|
2016-11-28 00:07:18 -07:00
|
|
|
}
|
2014-09-04 11:31:10 -06:00
|
|
|
if (!skb)
|
|
|
|
return;
|
|
|
|
|
2015-01-30 11:29:31 -07:00
|
|
|
if (tsonly) {
|
|
|
|
skb_shinfo(skb)->tx_flags = skb_shinfo(orig_skb)->tx_flags;
|
|
|
|
skb_shinfo(skb)->tskey = skb_shinfo(orig_skb)->tskey;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (hwtstamps)
|
|
|
|
*skb_hwtstamps(skb) = *hwtstamps;
|
|
|
|
else
|
|
|
|
skb->tstamp = ktime_get_real();
|
|
|
|
|
2014-09-04 11:31:10 -06:00
|
|
|
__skb_complete_tx_timestamp(skb, sk, tstype);
|
|
|
|
}
|
2014-08-04 20:11:48 -06:00
|
|
|
EXPORT_SYMBOL_GPL(__skb_tstamp_tx);
|
|
|
|
|
|
|
|
void skb_tstamp_tx(struct sk_buff *orig_skb,
|
|
|
|
struct skb_shared_hwtstamps *hwtstamps)
|
|
|
|
{
|
|
|
|
return __skb_tstamp_tx(orig_skb, hwtstamps, orig_skb->sk,
|
|
|
|
SCM_TSTAMP_SND);
|
|
|
|
}
|
2009-02-11 22:03:37 -07:00
|
|
|
EXPORT_SYMBOL_GPL(skb_tstamp_tx);
|
|
|
|
|
2011-11-09 02:15:42 -07:00
|
|
|
void skb_complete_wifi_ack(struct sk_buff *skb, bool acked)
|
|
|
|
{
|
|
|
|
struct sock *sk = skb->sk;
|
|
|
|
struct sock_exterr_skb *serr;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
skb->wifi_acked_valid = 1;
|
|
|
|
skb->wifi_acked = acked;
|
|
|
|
|
|
|
|
serr = SKB_EXT_ERR(skb);
|
|
|
|
memset(serr, 0, sizeof(*serr));
|
|
|
|
serr->ee.ee_errno = ENOMSG;
|
|
|
|
serr->ee.ee_origin = SO_EE_ORIGIN_TXSTATUS;
|
|
|
|
|
2014-09-10 16:05:42 -06:00
|
|
|
/* take a reference to prevent skb_orphan() from freeing the socket */
|
|
|
|
sock_hold(sk);
|
|
|
|
|
2011-11-09 02:15:42 -07:00
|
|
|
err = sock_queue_err_skb(sk, skb);
|
|
|
|
if (err)
|
|
|
|
kfree_skb(skb);
|
2014-09-10 16:05:42 -06:00
|
|
|
|
|
|
|
sock_put(sk);
|
2011-11-09 02:15:42 -07:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_complete_wifi_ack);
|
|
|
|
|
2008-02-04 21:49:54 -07:00
|
|
|
/**
|
|
|
|
* skb_partial_csum_set - set up and verify partial csum values for packet
|
|
|
|
* @skb: the skb to set
|
|
|
|
* @start: the number of bytes after skb->data to start checksumming.
|
|
|
|
* @off: the offset from start to place the checksum.
|
|
|
|
*
|
|
|
|
* For untrusted partially-checksummed packets, we need to make sure the values
|
|
|
|
* for skb->csum_start and skb->csum_offset are valid so we don't oops.
|
|
|
|
*
|
|
|
|
* This function checks and sets those values and skb->ip_summed: if this
|
|
|
|
* returns false you should drop the packet.
|
|
|
|
*/
|
|
|
|
bool skb_partial_csum_set(struct sk_buff *skb, u16 start, u16 off)
|
|
|
|
{
|
2009-06-03 19:22:01 -06:00
|
|
|
if (unlikely(start > skb_headlen(skb)) ||
|
|
|
|
unlikely((int)start + off > skb_headlen(skb) - 2)) {
|
2012-05-13 15:56:26 -06:00
|
|
|
net_warn_ratelimited("bad partial csum: csum=%u/%u len=%u\n",
|
|
|
|
start, off, skb_headlen(skb));
|
2008-02-04 21:49:54 -07:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
skb->ip_summed = CHECKSUM_PARTIAL;
|
|
|
|
skb->csum_start = skb_headroom(skb) + start;
|
|
|
|
skb->csum_offset = off;
|
2013-03-26 17:11:20 -06:00
|
|
|
skb_set_transport_header(skb, start);
|
2008-02-04 21:49:54 -07:00
|
|
|
return true;
|
|
|
|
}
|
2009-02-10 03:09:24 -07:00
|
|
|
EXPORT_SYMBOL_GPL(skb_partial_csum_set);
|
2008-02-04 21:49:54 -07:00
|
|
|
|
2014-01-09 03:02:46 -07:00
|
|
|
static int skb_maybe_pull_tail(struct sk_buff *skb, unsigned int len,
|
|
|
|
unsigned int max)
|
|
|
|
{
|
|
|
|
if (skb_headlen(skb) >= len)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
/* If we need to pullup then pullup to the max, so we
|
|
|
|
* won't need to do it again.
|
|
|
|
*/
|
|
|
|
if (max > skb->len)
|
|
|
|
max = skb->len;
|
|
|
|
|
|
|
|
if (__pskb_pull_tail(skb, max - skb_headlen(skb)) == NULL)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
if (skb_headlen(skb) < len)
|
|
|
|
return -EPROTO;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-03-11 07:56:05 -06:00
|
|
|
#define MAX_TCP_HDR_LEN (15 * 4)
|
|
|
|
|
|
|
|
static __sum16 *skb_checksum_setup_ip(struct sk_buff *skb,
|
|
|
|
typeof(IPPROTO_IP) proto,
|
|
|
|
unsigned int off)
|
|
|
|
{
|
|
|
|
switch (proto) {
|
|
|
|
int err;
|
|
|
|
|
|
|
|
case IPPROTO_TCP:
|
|
|
|
err = skb_maybe_pull_tail(skb, off + sizeof(struct tcphdr),
|
|
|
|
off + MAX_TCP_HDR_LEN);
|
|
|
|
if (!err && !skb_partial_csum_set(skb, off,
|
|
|
|
offsetof(struct tcphdr,
|
|
|
|
check)))
|
|
|
|
err = -EPROTO;
|
|
|
|
return err ? ERR_PTR(err) : &tcp_hdr(skb)->check;
|
|
|
|
|
|
|
|
case IPPROTO_UDP:
|
|
|
|
err = skb_maybe_pull_tail(skb, off + sizeof(struct udphdr),
|
|
|
|
off + sizeof(struct udphdr));
|
|
|
|
if (!err && !skb_partial_csum_set(skb, off,
|
|
|
|
offsetof(struct udphdr,
|
|
|
|
check)))
|
|
|
|
err = -EPROTO;
|
|
|
|
return err ? ERR_PTR(err) : &udp_hdr(skb)->check;
|
|
|
|
}
|
|
|
|
|
|
|
|
return ERR_PTR(-EPROTO);
|
|
|
|
}
|
|
|
|
|
2014-01-09 03:02:46 -07:00
|
|
|
/* This value should be large enough to cover a tagged ethernet header plus
|
|
|
|
* maximally sized IP and TCP or UDP headers.
|
|
|
|
*/
|
|
|
|
#define MAX_IP_HDR_LEN 128
|
|
|
|
|
2014-03-11 07:56:05 -06:00
|
|
|
static int skb_checksum_setup_ipv4(struct sk_buff *skb, bool recalculate)
|
2014-01-09 03:02:46 -07:00
|
|
|
{
|
|
|
|
unsigned int off;
|
|
|
|
bool fragment;
|
2014-03-11 07:56:05 -06:00
|
|
|
__sum16 *csum;
|
2014-01-09 03:02:46 -07:00
|
|
|
int err;
|
|
|
|
|
|
|
|
fragment = false;
|
|
|
|
|
|
|
|
err = skb_maybe_pull_tail(skb,
|
|
|
|
sizeof(struct iphdr),
|
|
|
|
MAX_IP_HDR_LEN);
|
|
|
|
if (err < 0)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
if (ip_hdr(skb)->frag_off & htons(IP_OFFSET | IP_MF))
|
|
|
|
fragment = true;
|
|
|
|
|
|
|
|
off = ip_hdrlen(skb);
|
|
|
|
|
|
|
|
err = -EPROTO;
|
|
|
|
|
|
|
|
if (fragment)
|
|
|
|
goto out;
|
|
|
|
|
2014-03-11 07:56:05 -06:00
|
|
|
csum = skb_checksum_setup_ip(skb, ip_hdr(skb)->protocol, off);
|
|
|
|
if (IS_ERR(csum))
|
|
|
|
return PTR_ERR(csum);
|
2014-01-09 03:02:46 -07:00
|
|
|
|
2014-03-11 07:56:05 -06:00
|
|
|
if (recalculate)
|
|
|
|
*csum = ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
|
|
|
|
ip_hdr(skb)->daddr,
|
|
|
|
skb->len - off,
|
|
|
|
ip_hdr(skb)->protocol, 0);
|
2014-01-09 03:02:46 -07:00
|
|
|
err = 0;
|
|
|
|
|
|
|
|
out:
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* This value should be large enough to cover a tagged ethernet header plus
|
|
|
|
* an IPv6 header, all options, and a maximal TCP or UDP header.
|
|
|
|
*/
|
|
|
|
#define MAX_IPV6_HDR_LEN 256
|
|
|
|
|
|
|
|
#define OPT_HDR(type, skb, off) \
|
|
|
|
(type *)(skb_network_header(skb) + (off))
|
|
|
|
|
|
|
|
static int skb_checksum_setup_ipv6(struct sk_buff *skb, bool recalculate)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
u8 nexthdr;
|
|
|
|
unsigned int off;
|
|
|
|
unsigned int len;
|
|
|
|
bool fragment;
|
|
|
|
bool done;
|
2014-03-11 07:56:05 -06:00
|
|
|
__sum16 *csum;
|
2014-01-09 03:02:46 -07:00
|
|
|
|
|
|
|
fragment = false;
|
|
|
|
done = false;
|
|
|
|
|
|
|
|
off = sizeof(struct ipv6hdr);
|
|
|
|
|
|
|
|
err = skb_maybe_pull_tail(skb, off, MAX_IPV6_HDR_LEN);
|
|
|
|
if (err < 0)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
nexthdr = ipv6_hdr(skb)->nexthdr;
|
|
|
|
|
|
|
|
len = sizeof(struct ipv6hdr) + ntohs(ipv6_hdr(skb)->payload_len);
|
|
|
|
while (off <= len && !done) {
|
|
|
|
switch (nexthdr) {
|
|
|
|
case IPPROTO_DSTOPTS:
|
|
|
|
case IPPROTO_HOPOPTS:
|
|
|
|
case IPPROTO_ROUTING: {
|
|
|
|
struct ipv6_opt_hdr *hp;
|
|
|
|
|
|
|
|
err = skb_maybe_pull_tail(skb,
|
|
|
|
off +
|
|
|
|
sizeof(struct ipv6_opt_hdr),
|
|
|
|
MAX_IPV6_HDR_LEN);
|
|
|
|
if (err < 0)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
hp = OPT_HDR(struct ipv6_opt_hdr, skb, off);
|
|
|
|
nexthdr = hp->nexthdr;
|
|
|
|
off += ipv6_optlen(hp);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case IPPROTO_AH: {
|
|
|
|
struct ip_auth_hdr *hp;
|
|
|
|
|
|
|
|
err = skb_maybe_pull_tail(skb,
|
|
|
|
off +
|
|
|
|
sizeof(struct ip_auth_hdr),
|
|
|
|
MAX_IPV6_HDR_LEN);
|
|
|
|
if (err < 0)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
hp = OPT_HDR(struct ip_auth_hdr, skb, off);
|
|
|
|
nexthdr = hp->nexthdr;
|
|
|
|
off += ipv6_authlen(hp);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case IPPROTO_FRAGMENT: {
|
|
|
|
struct frag_hdr *hp;
|
|
|
|
|
|
|
|
err = skb_maybe_pull_tail(skb,
|
|
|
|
off +
|
|
|
|
sizeof(struct frag_hdr),
|
|
|
|
MAX_IPV6_HDR_LEN);
|
|
|
|
if (err < 0)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
hp = OPT_HDR(struct frag_hdr, skb, off);
|
|
|
|
|
|
|
|
if (hp->frag_off & htons(IP6_OFFSET | IP6_MF))
|
|
|
|
fragment = true;
|
|
|
|
|
|
|
|
nexthdr = hp->nexthdr;
|
|
|
|
off += sizeof(struct frag_hdr);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
done = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
err = -EPROTO;
|
|
|
|
|
|
|
|
if (!done || fragment)
|
|
|
|
goto out;
|
|
|
|
|
2014-03-11 07:56:05 -06:00
|
|
|
csum = skb_checksum_setup_ip(skb, nexthdr, off);
|
|
|
|
if (IS_ERR(csum))
|
|
|
|
return PTR_ERR(csum);
|
2014-01-09 03:02:46 -07:00
|
|
|
|
2014-03-11 07:56:05 -06:00
|
|
|
if (recalculate)
|
|
|
|
*csum = ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
|
|
|
|
&ipv6_hdr(skb)->daddr,
|
|
|
|
skb->len - off, nexthdr, 0);
|
2014-01-09 03:02:46 -07:00
|
|
|
err = 0;
|
|
|
|
|
|
|
|
out:
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_checksum_setup - set up partial checksum offset
|
|
|
|
* @skb: the skb to set up
|
|
|
|
* @recalculate: if true the pseudo-header checksum will be recalculated
|
|
|
|
*/
|
|
|
|
int skb_checksum_setup(struct sk_buff *skb, bool recalculate)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
|
|
|
|
switch (skb->protocol) {
|
|
|
|
case htons(ETH_P_IP):
|
2014-03-11 07:56:05 -06:00
|
|
|
err = skb_checksum_setup_ipv4(skb, recalculate);
|
2014-01-09 03:02:46 -07:00
|
|
|
break;
|
|
|
|
|
|
|
|
case htons(ETH_P_IPV6):
|
|
|
|
err = skb_checksum_setup_ipv6(skb, recalculate);
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
err = -EPROTO;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_checksum_setup);
|
|
|
|
|
2015-05-02 06:01:07 -06:00
|
|
|
/**
|
|
|
|
* skb_checksum_maybe_trim - maybe trims the given skb
|
|
|
|
* @skb: the skb to check
|
|
|
|
* @transport_len: the data length beyond the network header
|
|
|
|
*
|
|
|
|
* Checks whether the given skb has data beyond the given transport length.
|
|
|
|
* If so, returns a cloned skb trimmed to this transport length.
|
|
|
|
* Otherwise returns the provided skb. Returns NULL in error cases
|
|
|
|
* (e.g. transport_len exceeds skb length or out-of-memory).
|
|
|
|
*
|
2015-08-12 21:54:07 -06:00
|
|
|
* Caller needs to set the skb transport header and free any returned skb if it
|
|
|
|
* differs from the provided skb.
|
2015-05-02 06:01:07 -06:00
|
|
|
*/
|
|
|
|
static struct sk_buff *skb_checksum_maybe_trim(struct sk_buff *skb,
|
|
|
|
unsigned int transport_len)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb_chk;
|
|
|
|
unsigned int len = skb_transport_offset(skb) + transport_len;
|
|
|
|
int ret;
|
|
|
|
|
2015-08-12 21:54:07 -06:00
|
|
|
if (skb->len < len)
|
2015-05-02 06:01:07 -06:00
|
|
|
return NULL;
|
2015-08-12 21:54:07 -06:00
|
|
|
else if (skb->len == len)
|
2015-05-02 06:01:07 -06:00
|
|
|
return skb;
|
|
|
|
|
|
|
|
skb_chk = skb_clone(skb, GFP_ATOMIC);
|
|
|
|
if (!skb_chk)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
ret = pskb_trim_rcsum(skb_chk, len);
|
|
|
|
if (ret) {
|
|
|
|
kfree_skb(skb_chk);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
return skb_chk;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_checksum_trimmed - validate checksum of an skb
|
|
|
|
* @skb: the skb to check
|
|
|
|
* @transport_len: the data length beyond the network header
|
|
|
|
* @skb_chkf: checksum function to use
|
|
|
|
*
|
|
|
|
* Applies the given checksum function skb_chkf to the provided skb.
|
|
|
|
* Returns a checked and maybe trimmed skb. Returns NULL on error.
|
|
|
|
*
|
|
|
|
* If the skb has data beyond the given transport length, then a
|
|
|
|
* trimmed & cloned skb is checked and returned.
|
|
|
|
*
|
2015-08-12 21:54:07 -06:00
|
|
|
* Caller needs to set the skb transport header and free any returned skb if it
|
|
|
|
* differs from the provided skb.
|
2015-05-02 06:01:07 -06:00
|
|
|
*/
|
|
|
|
struct sk_buff *skb_checksum_trimmed(struct sk_buff *skb,
|
|
|
|
unsigned int transport_len,
|
|
|
|
__sum16(*skb_chkf)(struct sk_buff *skb))
|
|
|
|
{
|
|
|
|
struct sk_buff *skb_chk;
|
|
|
|
unsigned int offset = skb_transport_offset(skb);
|
2015-05-04 16:19:35 -06:00
|
|
|
__sum16 ret;
|
2015-05-02 06:01:07 -06:00
|
|
|
|
|
|
|
skb_chk = skb_checksum_maybe_trim(skb, transport_len);
|
|
|
|
if (!skb_chk)
|
2015-08-12 21:54:07 -06:00
|
|
|
goto err;
|
2015-05-02 06:01:07 -06:00
|
|
|
|
2015-08-12 21:54:07 -06:00
|
|
|
if (!pskb_may_pull(skb_chk, offset))
|
|
|
|
goto err;
|
2015-05-02 06:01:07 -06:00
|
|
|
|
2016-02-23 20:21:42 -07:00
|
|
|
skb_pull_rcsum(skb_chk, offset);
|
2015-05-02 06:01:07 -06:00
|
|
|
ret = skb_chkf(skb_chk);
|
2016-02-23 20:21:42 -07:00
|
|
|
skb_push_rcsum(skb_chk, offset);
|
2015-05-02 06:01:07 -06:00
|
|
|
|
2015-08-12 21:54:07 -06:00
|
|
|
if (ret)
|
|
|
|
goto err;
|
2015-05-02 06:01:07 -06:00
|
|
|
|
|
|
|
return skb_chk;
|
2015-08-12 21:54:07 -06:00
|
|
|
|
|
|
|
err:
|
|
|
|
if (skb_chk && skb_chk != skb)
|
|
|
|
kfree_skb(skb_chk);
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
|
2015-05-02 06:01:07 -06:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_checksum_trimmed);
|
|
|
|
|
2008-06-19 17:22:28 -06:00
|
|
|
void __skb_warn_lro_forwarding(const struct sk_buff *skb)
|
|
|
|
{
|
2012-05-13 15:56:26 -06:00
|
|
|
net_warn_ratelimited("%s: received packets cannot be forwarded while LRO is enabled\n",
|
|
|
|
skb->dev->name);
|
2008-06-19 17:22:28 -06:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(__skb_warn_lro_forwarding);
|
2012-05-18 21:02:02 -06:00
|
|
|
|
|
|
|
void kfree_skb_partial(struct sk_buff *skb, bool head_stolen)
|
|
|
|
{
|
2012-10-22 03:03:40 -06:00
|
|
|
if (head_stolen) {
|
|
|
|
skb_release_head_state(skb);
|
2012-05-18 21:02:02 -06:00
|
|
|
kmem_cache_free(skbuff_head_cache, skb);
|
2012-10-22 03:03:40 -06:00
|
|
|
} else {
|
2012-05-18 21:02:02 -06:00
|
|
|
__kfree_skb(skb);
|
2012-10-22 03:03:40 -06:00
|
|
|
}
|
2012-05-18 21:02:02 -06:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(kfree_skb_partial);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_try_coalesce - try to merge skb to prior one
|
|
|
|
* @to: prior buffer
|
|
|
|
* @from: buffer to add
|
|
|
|
* @fragstolen: pointer to boolean
|
2012-06-08 08:01:44 -06:00
|
|
|
* @delta_truesize: how much more was allocated than was requested
|
2012-05-18 21:02:02 -06:00
|
|
|
*/
|
|
|
|
bool skb_try_coalesce(struct sk_buff *to, struct sk_buff *from,
|
|
|
|
bool *fragstolen, int *delta_truesize)
|
|
|
|
{
|
|
|
|
int i, delta, len = from->len;
|
|
|
|
|
|
|
|
*fragstolen = false;
|
|
|
|
|
|
|
|
if (skb_cloned(to))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (len <= skb_tailroom(to)) {
|
2014-09-15 05:19:52 -06:00
|
|
|
if (len)
|
|
|
|
BUG_ON(skb_copy_bits(from, 0, skb_put(to, len), len));
|
2012-05-18 21:02:02 -06:00
|
|
|
*delta_truesize = 0;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (skb_has_frag_list(to) || skb_has_frag_list(from))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (skb_headlen(from) != 0) {
|
|
|
|
struct page *page;
|
|
|
|
unsigned int offset;
|
|
|
|
|
|
|
|
if (skb_shinfo(to)->nr_frags +
|
|
|
|
skb_shinfo(from)->nr_frags >= MAX_SKB_FRAGS)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (skb_head_is_locked(from))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
delta = from->truesize - SKB_DATA_ALIGN(sizeof(struct sk_buff));
|
|
|
|
|
|
|
|
page = virt_to_head_page(from->head);
|
|
|
|
offset = from->data - (unsigned char *)page_address(page);
|
|
|
|
|
|
|
|
skb_fill_page_desc(to, skb_shinfo(to)->nr_frags,
|
|
|
|
page, offset, skb_headlen(from));
|
|
|
|
*fragstolen = true;
|
|
|
|
} else {
|
|
|
|
if (skb_shinfo(to)->nr_frags +
|
|
|
|
skb_shinfo(from)->nr_frags > MAX_SKB_FRAGS)
|
|
|
|
return false;
|
|
|
|
|
2012-09-28 14:15:30 -06:00
|
|
|
delta = from->truesize - SKB_TRUESIZE(skb_end_offset(from));
|
2012-05-18 21:02:02 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
WARN_ON_ONCE(delta < len);
|
|
|
|
|
|
|
|
memcpy(skb_shinfo(to)->frags + skb_shinfo(to)->nr_frags,
|
|
|
|
skb_shinfo(from)->frags,
|
|
|
|
skb_shinfo(from)->nr_frags * sizeof(skb_frag_t));
|
|
|
|
skb_shinfo(to)->nr_frags += skb_shinfo(from)->nr_frags;
|
|
|
|
|
|
|
|
if (!skb_cloned(from))
|
|
|
|
skb_shinfo(from)->nr_frags = 0;
|
|
|
|
|
2012-09-18 10:53:21 -06:00
|
|
|
/* if the skb is not cloned this does nothing
|
|
|
|
* since we set nr_frags to 0.
|
|
|
|
*/
|
2012-05-18 21:02:02 -06:00
|
|
|
for (i = 0; i < skb_shinfo(from)->nr_frags; i++)
|
|
|
|
skb_frag_ref(from, i);
|
|
|
|
|
|
|
|
to->truesize += delta;
|
|
|
|
to->len += len;
|
|
|
|
to->data_len += len;
|
|
|
|
|
|
|
|
*delta_truesize = delta;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_try_coalesce);
|
2013-06-26 08:11:27 -06:00
|
|
|
|
|
|
|
/**
|
2013-09-02 07:34:56 -06:00
|
|
|
* skb_scrub_packet - scrub an skb
|
2013-06-26 08:11:27 -06:00
|
|
|
*
|
|
|
|
* @skb: buffer to clean
|
2013-09-02 07:34:56 -06:00
|
|
|
* @xnet: packet is crossing netns
|
|
|
|
*
|
|
|
|
* skb_scrub_packet can be used after encapsulating or decapsulting a packet
|
|
|
|
* into/from a tunnel. Some information have to be cleared during these
|
|
|
|
* operations.
|
|
|
|
* skb_scrub_packet can also be used to clean a skb before injecting it in
|
|
|
|
* another namespace (@xnet == true). We have to clear all information in the
|
|
|
|
* skb that could impact namespace isolation.
|
2013-06-26 08:11:27 -06:00
|
|
|
*/
|
2013-09-02 07:34:56 -06:00
|
|
|
void skb_scrub_packet(struct sk_buff *skb, bool xnet)
|
2013-06-26 08:11:27 -06:00
|
|
|
{
|
2016-12-25 03:38:40 -07:00
|
|
|
skb->tstamp = 0;
|
2013-06-26 08:11:27 -06:00
|
|
|
skb->pkt_type = PACKET_HOST;
|
|
|
|
skb->skb_iif = 0;
|
2014-05-04 17:39:18 -06:00
|
|
|
skb->ignore_df = 0;
|
2013-06-26 08:11:27 -06:00
|
|
|
skb_dst_drop(skb);
|
|
|
|
secpath_reset(skb);
|
|
|
|
nf_reset(skb);
|
|
|
|
nf_reset_trace(skb);
|
skbuff: Do not scrub skb mark within the same name space
On Wed, Apr 15, 2015 at 05:41:26PM +0200, Nicolas Dichtel wrote:
> Le 15/04/2015 15:57, Herbert Xu a écrit :
> >On Wed, Apr 15, 2015 at 06:22:29PM +0800, Herbert Xu wrote:
> [snip]
> >Subject: skbuff: Do not scrub skb mark within the same name space
> >
> >The commit ea23192e8e577dfc51e0f4fc5ca113af334edff9 ("tunnels:
> Maybe add a Fixes tag?
> Fixes: ea23192e8e57 ("tunnels: harmonize cleanup done on skb on rx path")
>
> >harmonize cleanup done on skb on rx path") broke anyone trying to
> >use netfilter marking across IPv4 tunnels. While most of the
> >fields that are cleared by skb_scrub_packet don't matter, the
> >netfilter mark must be preserved.
> >
> >This patch rearranges skb_scurb_packet to preserve the mark field.
> nit: s/scurb/scrub
>
> Else it's fine for me.
Sure.
PS I used the wrong email for James the first time around. So
let me repeat the question here. Should secmark be preserved
or cleared across tunnels within the same name space? In fact,
do our security models even support name spaces?
---8<---
The commit ea23192e8e577dfc51e0f4fc5ca113af334edff9 ("tunnels:
harmonize cleanup done on skb on rx path") broke anyone trying to
use netfilter marking across IPv4 tunnels. While most of the
fields that are cleared by skb_scrub_packet don't matter, the
netfilter mark must be preserved.
This patch rearranges skb_scrub_packet to preserve the mark field.
Fixes: ea23192e8e57 ("tunnels: harmonize cleanup done on skb on rx path")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-04-15 19:03:27 -06:00
|
|
|
|
|
|
|
if (!xnet)
|
|
|
|
return;
|
|
|
|
|
|
|
|
skb_orphan(skb);
|
|
|
|
skb->mark = 0;
|
2013-06-26 08:11:27 -06:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_scrub_packet);
|
2014-01-26 02:58:16 -07:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_gso_transport_seglen - Return length of individual segments of a gso packet
|
|
|
|
*
|
|
|
|
* @skb: GSO skb
|
|
|
|
*
|
|
|
|
* skb_gso_transport_seglen is used to determine the real size of the
|
|
|
|
* individual segments, including Layer4 headers (TCP/UDP).
|
|
|
|
*
|
|
|
|
* The MAC/L2 or network (IP, IPv6) headers are not accounted for.
|
|
|
|
*/
|
|
|
|
unsigned int skb_gso_transport_seglen(const struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
const struct skb_shared_info *shinfo = skb_shinfo(skb);
|
2014-10-20 05:49:18 -06:00
|
|
|
unsigned int thlen = 0;
|
2014-01-26 02:58:16 -07:00
|
|
|
|
2014-10-20 05:49:18 -06:00
|
|
|
if (skb->encapsulation) {
|
|
|
|
thlen = skb_inner_transport_header(skb) -
|
|
|
|
skb_transport_header(skb);
|
2014-04-09 02:28:50 -06:00
|
|
|
|
2014-10-20 05:49:18 -06:00
|
|
|
if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
|
|
|
|
thlen += inner_tcp_hdrlen(skb);
|
|
|
|
} else if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
|
|
|
|
thlen = tcp_hdrlen(skb);
|
2016-06-02 12:05:43 -06:00
|
|
|
} else if (unlikely(shinfo->gso_type & SKB_GSO_SCTP)) {
|
|
|
|
thlen = sizeof(struct sctphdr);
|
2014-10-20 05:49:18 -06:00
|
|
|
}
|
2014-04-09 02:28:50 -06:00
|
|
|
/* UFO sets gso_size to the size of the fragmentation
|
|
|
|
* payload, i.e. the size of the L4 (UDP) header is already
|
|
|
|
* accounted for.
|
|
|
|
*/
|
2014-10-20 05:49:18 -06:00
|
|
|
return thlen + shinfo->gso_size;
|
2014-01-26 02:58:16 -07:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_gso_transport_seglen);
|
net: Always untag vlan-tagged traffic on input.
Currently the functionality to untag traffic on input resides
as part of the vlan module and is build only when VLAN support
is enabled in the kernel. When VLAN is disabled, the function
vlan_untag() turns into a stub and doesn't really untag the
packets. This seems to create an interesting interaction
between VMs supporting checksum offloading and some network drivers.
There are some drivers that do not allow the user to change
tx-vlan-offload feature of the driver. These drivers also seem
to assume that any VLAN-tagged traffic they transmit will
have the vlan information in the vlan_tci and not in the vlan
header already in the skb. When transmitting skbs that already
have tagged data with partial checksum set, the checksum doesn't
appear to be updated correctly by the card thus resulting in a
failure to establish TCP connections.
The following is a packet trace taken on the receiver where a
sender is a VM with a VLAN configued. The host VM is running on
doest not have VLAN support and the outging interface on the
host is tg3:
10:12:43.503055 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27243,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x48d9), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294837885 ecr 0,nop,wscale 7], length 0
10:12:44.505556 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27244,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x44ee), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294838888 ecr 0,nop,wscale 7], length 0
This connection finally times out.
I've only access to the TG3 hardware in this configuration thus have
only tested this with TG3 driver. There are a lot of other drivers
that do not permit user changes to vlan acceleration features, and
I don't know if they all suffere from a similar issue.
The patch attempt to fix this another way. It moves the vlan header
stipping code out of the vlan module and always builds it into the
kernel network core. This way, even if vlan is not supported on
a virtualizatoin host, the virtual machines running on top of such
host will still work with VLANs enabled.
CC: Patrick McHardy <kaber@trash.net>
CC: Nithin Nayak Sujir <nsujir@broadcom.com>
CC: Michael Chan <mchan@broadcom.com>
CC: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Acked-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-08 12:42:13 -06:00
|
|
|
|
2016-06-02 12:05:41 -06:00
|
|
|
/**
|
|
|
|
* skb_gso_validate_mtu - Return in case such skb fits a given MTU
|
|
|
|
*
|
|
|
|
* @skb: GSO skb
|
2016-06-03 23:56:28 -06:00
|
|
|
* @mtu: MTU to validate against
|
2016-06-02 12:05:41 -06:00
|
|
|
*
|
|
|
|
* skb_gso_validate_mtu validates if a given skb will fit a wanted MTU
|
|
|
|
* once split.
|
|
|
|
*/
|
|
|
|
bool skb_gso_validate_mtu(const struct sk_buff *skb, unsigned int mtu)
|
|
|
|
{
|
|
|
|
const struct skb_shared_info *shinfo = skb_shinfo(skb);
|
|
|
|
const struct sk_buff *iter;
|
|
|
|
unsigned int hlen;
|
|
|
|
|
|
|
|
hlen = skb_gso_network_seglen(skb);
|
|
|
|
|
|
|
|
if (shinfo->gso_size != GSO_BY_FRAGS)
|
|
|
|
return hlen <= mtu;
|
|
|
|
|
|
|
|
/* Undo this so we can re-use header sizes */
|
|
|
|
hlen -= GSO_BY_FRAGS;
|
|
|
|
|
|
|
|
skb_walk_frags(skb, iter) {
|
|
|
|
if (hlen + skb_headlen(iter) > mtu)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(skb_gso_validate_mtu);
|
|
|
|
|
net: Always untag vlan-tagged traffic on input.
Currently the functionality to untag traffic on input resides
as part of the vlan module and is build only when VLAN support
is enabled in the kernel. When VLAN is disabled, the function
vlan_untag() turns into a stub and doesn't really untag the
packets. This seems to create an interesting interaction
between VMs supporting checksum offloading and some network drivers.
There are some drivers that do not allow the user to change
tx-vlan-offload feature of the driver. These drivers also seem
to assume that any VLAN-tagged traffic they transmit will
have the vlan information in the vlan_tci and not in the vlan
header already in the skb. When transmitting skbs that already
have tagged data with partial checksum set, the checksum doesn't
appear to be updated correctly by the card thus resulting in a
failure to establish TCP connections.
The following is a packet trace taken on the receiver where a
sender is a VM with a VLAN configued. The host VM is running on
doest not have VLAN support and the outging interface on the
host is tg3:
10:12:43.503055 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27243,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x48d9), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294837885 ecr 0,nop,wscale 7], length 0
10:12:44.505556 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27244,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x44ee), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294838888 ecr 0,nop,wscale 7], length 0
This connection finally times out.
I've only access to the TG3 hardware in this configuration thus have
only tested this with TG3 driver. There are a lot of other drivers
that do not permit user changes to vlan acceleration features, and
I don't know if they all suffere from a similar issue.
The patch attempt to fix this another way. It moves the vlan header
stipping code out of the vlan module and always builds it into the
kernel network core. This way, even if vlan is not supported on
a virtualizatoin host, the virtual machines running on top of such
host will still work with VLANs enabled.
CC: Patrick McHardy <kaber@trash.net>
CC: Nithin Nayak Sujir <nsujir@broadcom.com>
CC: Michael Chan <mchan@broadcom.com>
CC: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Acked-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-08 12:42:13 -06:00
|
|
|
static struct sk_buff *skb_reorder_vlan_header(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
if (skb_cow(skb, skb_headroom(skb)) < 0) {
|
|
|
|
kfree_skb(skb);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2015-12-14 15:44:10 -07:00
|
|
|
memmove(skb->data - ETH_HLEN, skb->data - skb->mac_len - VLAN_HLEN,
|
2015-11-16 13:43:44 -07:00
|
|
|
2 * ETH_ALEN);
|
net: Always untag vlan-tagged traffic on input.
Currently the functionality to untag traffic on input resides
as part of the vlan module and is build only when VLAN support
is enabled in the kernel. When VLAN is disabled, the function
vlan_untag() turns into a stub and doesn't really untag the
packets. This seems to create an interesting interaction
between VMs supporting checksum offloading and some network drivers.
There are some drivers that do not allow the user to change
tx-vlan-offload feature of the driver. These drivers also seem
to assume that any VLAN-tagged traffic they transmit will
have the vlan information in the vlan_tci and not in the vlan
header already in the skb. When transmitting skbs that already
have tagged data with partial checksum set, the checksum doesn't
appear to be updated correctly by the card thus resulting in a
failure to establish TCP connections.
The following is a packet trace taken on the receiver where a
sender is a VM with a VLAN configued. The host VM is running on
doest not have VLAN support and the outging interface on the
host is tg3:
10:12:43.503055 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27243,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x48d9), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294837885 ecr 0,nop,wscale 7], length 0
10:12:44.505556 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27244,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x44ee), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294838888 ecr 0,nop,wscale 7], length 0
This connection finally times out.
I've only access to the TG3 hardware in this configuration thus have
only tested this with TG3 driver. There are a lot of other drivers
that do not permit user changes to vlan acceleration features, and
I don't know if they all suffere from a similar issue.
The patch attempt to fix this another way. It moves the vlan header
stipping code out of the vlan module and always builds it into the
kernel network core. This way, even if vlan is not supported on
a virtualizatoin host, the virtual machines running on top of such
host will still work with VLANs enabled.
CC: Patrick McHardy <kaber@trash.net>
CC: Nithin Nayak Sujir <nsujir@broadcom.com>
CC: Michael Chan <mchan@broadcom.com>
CC: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Acked-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-08 12:42:13 -06:00
|
|
|
skb->mac_header += VLAN_HLEN;
|
|
|
|
return skb;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct sk_buff *skb_vlan_untag(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
struct vlan_hdr *vhdr;
|
|
|
|
u16 vlan_tci;
|
|
|
|
|
2015-01-13 09:13:44 -07:00
|
|
|
if (unlikely(skb_vlan_tag_present(skb))) {
|
net: Always untag vlan-tagged traffic on input.
Currently the functionality to untag traffic on input resides
as part of the vlan module and is build only when VLAN support
is enabled in the kernel. When VLAN is disabled, the function
vlan_untag() turns into a stub and doesn't really untag the
packets. This seems to create an interesting interaction
between VMs supporting checksum offloading and some network drivers.
There are some drivers that do not allow the user to change
tx-vlan-offload feature of the driver. These drivers also seem
to assume that any VLAN-tagged traffic they transmit will
have the vlan information in the vlan_tci and not in the vlan
header already in the skb. When transmitting skbs that already
have tagged data with partial checksum set, the checksum doesn't
appear to be updated correctly by the card thus resulting in a
failure to establish TCP connections.
The following is a packet trace taken on the receiver where a
sender is a VM with a VLAN configued. The host VM is running on
doest not have VLAN support and the outging interface on the
host is tg3:
10:12:43.503055 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27243,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x48d9), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294837885 ecr 0,nop,wscale 7], length 0
10:12:44.505556 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q
(0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27244,
offset 0, flags [DF], proto TCP (6), length 60)
10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect
-> 0x44ee), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val
4294838888 ecr 0,nop,wscale 7], length 0
This connection finally times out.
I've only access to the TG3 hardware in this configuration thus have
only tested this with TG3 driver. There are a lot of other drivers
that do not permit user changes to vlan acceleration features, and
I don't know if they all suffere from a similar issue.
The patch attempt to fix this another way. It moves the vlan header
stipping code out of the vlan module and always builds it into the
kernel network core. This way, even if vlan is not supported on
a virtualizatoin host, the virtual machines running on top of such
host will still work with VLANs enabled.
CC: Patrick McHardy <kaber@trash.net>
CC: Nithin Nayak Sujir <nsujir@broadcom.com>
CC: Michael Chan <mchan@broadcom.com>
CC: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Acked-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-08 12:42:13 -06:00
|
|
|
/* vlan_tci is already set-up so leave this for another time */
|
|
|
|
return skb;
|
|
|
|
}
|
|
|
|
|
|
|
|
skb = skb_share_check(skb, GFP_ATOMIC);
|
|
|
|
if (unlikely(!skb))
|
|
|
|
goto err_free;
|
|
|
|
|
|
|
|
if (unlikely(!pskb_may_pull(skb, VLAN_HLEN)))
|
|
|
|
goto err_free;
|
|
|
|
|
|
|
|
vhdr = (struct vlan_hdr *)skb->data;
|
|
|
|
vlan_tci = ntohs(vhdr->h_vlan_TCI);
|
|
|
|
__vlan_hwaccel_put_tag(skb, skb->protocol, vlan_tci);
|
|
|
|
|
|
|
|
skb_pull_rcsum(skb, VLAN_HLEN);
|
|
|
|
vlan_set_encap_proto(skb, vhdr);
|
|
|
|
|
|
|
|
skb = skb_reorder_vlan_header(skb);
|
|
|
|
if (unlikely(!skb))
|
|
|
|
goto err_free;
|
|
|
|
|
|
|
|
skb_reset_network_header(skb);
|
|
|
|
skb_reset_transport_header(skb);
|
|
|
|
skb_reset_mac_len(skb);
|
|
|
|
|
|
|
|
return skb;
|
|
|
|
|
|
|
|
err_free:
|
|
|
|
kfree_skb(skb);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_vlan_untag);
|
2014-09-17 05:49:49 -06:00
|
|
|
|
2014-11-19 06:05:01 -07:00
|
|
|
int skb_ensure_writable(struct sk_buff *skb, int write_len)
|
|
|
|
{
|
|
|
|
if (!pskb_may_pull(skb, write_len))
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
if (!skb_cloned(skb) || skb_clone_writable(skb, write_len))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
return pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_ensure_writable);
|
|
|
|
|
2016-09-19 10:11:09 -06:00
|
|
|
/* remove VLAN header from packet and update csum accordingly.
|
|
|
|
* expects a non skb_vlan_tag_present skb with a vlan tag payload
|
|
|
|
*/
|
|
|
|
int __skb_vlan_pop(struct sk_buff *skb, u16 *vlan_tci)
|
2014-11-19 06:05:02 -07:00
|
|
|
{
|
|
|
|
struct vlan_hdr *vhdr;
|
2016-09-29 03:10:41 -06:00
|
|
|
int offset = skb->data - skb_mac_header(skb);
|
2014-11-19 06:05:02 -07:00
|
|
|
int err;
|
|
|
|
|
2016-09-29 03:10:41 -06:00
|
|
|
if (WARN_ONCE(offset,
|
|
|
|
"__skb_vlan_pop got skb with skb->data not at mac header (offset %d)\n",
|
|
|
|
offset)) {
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2014-11-19 06:05:02 -07:00
|
|
|
err = skb_ensure_writable(skb, VLAN_ETH_HLEN);
|
|
|
|
if (unlikely(err))
|
2016-09-29 03:10:41 -06:00
|
|
|
return err;
|
2014-11-19 06:05:02 -07:00
|
|
|
|
|
|
|
skb_postpull_rcsum(skb, skb->data + (2 * ETH_ALEN), VLAN_HLEN);
|
|
|
|
|
|
|
|
vhdr = (struct vlan_hdr *)(skb->data + ETH_HLEN);
|
|
|
|
*vlan_tci = ntohs(vhdr->h_vlan_TCI);
|
|
|
|
|
|
|
|
memmove(skb->data + VLAN_HLEN, skb->data, 2 * ETH_ALEN);
|
|
|
|
__skb_pull(skb, VLAN_HLEN);
|
|
|
|
|
|
|
|
vlan_set_encap_proto(skb, vhdr);
|
|
|
|
skb->mac_header += VLAN_HLEN;
|
|
|
|
|
|
|
|
if (skb_network_offset(skb) < ETH_HLEN)
|
|
|
|
skb_set_network_header(skb, ETH_HLEN);
|
|
|
|
|
|
|
|
skb_reset_mac_len(skb);
|
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
2016-09-19 10:11:09 -06:00
|
|
|
EXPORT_SYMBOL(__skb_vlan_pop);
|
2014-11-19 06:05:02 -07:00
|
|
|
|
2016-09-29 03:10:41 -06:00
|
|
|
/* Pop a vlan tag either from hwaccel or from payload.
|
|
|
|
* Expects skb->data at mac header.
|
|
|
|
*/
|
2014-11-19 06:05:02 -07:00
|
|
|
int skb_vlan_pop(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
u16 vlan_tci;
|
|
|
|
__be16 vlan_proto;
|
|
|
|
int err;
|
|
|
|
|
2015-01-13 09:13:44 -07:00
|
|
|
if (likely(skb_vlan_tag_present(skb))) {
|
2014-11-19 06:05:02 -07:00
|
|
|
skb->vlan_tci = 0;
|
|
|
|
} else {
|
2016-09-20 03:48:37 -06:00
|
|
|
if (unlikely(!eth_type_vlan(skb->protocol)))
|
2014-11-19 06:05:02 -07:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
err = __skb_vlan_pop(skb, &vlan_tci);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
/* move next vlan tag to hw accel tag */
|
2016-09-20 03:48:37 -06:00
|
|
|
if (likely(!eth_type_vlan(skb->protocol)))
|
2014-11-19 06:05:02 -07:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
vlan_proto = skb->protocol;
|
|
|
|
err = __skb_vlan_pop(skb, &vlan_tci);
|
|
|
|
if (unlikely(err))
|
|
|
|
return err;
|
|
|
|
|
|
|
|
__vlan_hwaccel_put_tag(skb, vlan_proto, vlan_tci);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_vlan_pop);
|
|
|
|
|
2016-09-29 03:10:41 -06:00
|
|
|
/* Push a vlan tag either into hwaccel or into payload (if hwaccel tag present).
|
|
|
|
* Expects skb->data at mac header.
|
|
|
|
*/
|
2014-11-19 06:05:02 -07:00
|
|
|
int skb_vlan_push(struct sk_buff *skb, __be16 vlan_proto, u16 vlan_tci)
|
|
|
|
{
|
2015-01-13 09:13:44 -07:00
|
|
|
if (skb_vlan_tag_present(skb)) {
|
2016-09-29 03:10:41 -06:00
|
|
|
int offset = skb->data - skb_mac_header(skb);
|
2014-11-19 06:05:02 -07:00
|
|
|
int err;
|
|
|
|
|
2016-09-29 03:10:41 -06:00
|
|
|
if (WARN_ONCE(offset,
|
|
|
|
"skb_vlan_push got skb with skb->data not at mac header (offset %d)\n",
|
|
|
|
offset)) {
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2014-11-19 06:05:02 -07:00
|
|
|
err = __vlan_insert_tag(skb, skb->vlan_proto,
|
2015-01-13 09:13:44 -07:00
|
|
|
skb_vlan_tag_get(skb));
|
2016-09-29 03:10:41 -06:00
|
|
|
if (err)
|
2014-11-19 06:05:02 -07:00
|
|
|
return err;
|
2016-04-15 18:27:58 -06:00
|
|
|
|
2014-11-19 06:05:02 -07:00
|
|
|
skb->protocol = skb->vlan_proto;
|
|
|
|
skb->mac_len += VLAN_HLEN;
|
|
|
|
|
2016-02-19 16:29:30 -07:00
|
|
|
skb_postpush_rcsum(skb, skb->data + (2 * ETH_ALEN), VLAN_HLEN);
|
2014-11-19 06:05:02 -07:00
|
|
|
}
|
|
|
|
__vlan_hwaccel_put_tag(skb, vlan_proto, vlan_tci);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(skb_vlan_push);
|
|
|
|
|
2014-09-17 05:49:49 -06:00
|
|
|
/**
|
|
|
|
* alloc_skb_with_frags - allocate skb with page frags
|
|
|
|
*
|
2014-10-08 21:58:08 -06:00
|
|
|
* @header_len: size of linear part
|
|
|
|
* @data_len: needed length in frags
|
|
|
|
* @max_page_order: max page order desired.
|
|
|
|
* @errcode: pointer to error code if any
|
|
|
|
* @gfp_mask: allocation mask
|
2014-09-17 05:49:49 -06:00
|
|
|
*
|
|
|
|
* This can be used to allocate a paged skb, given a maximal order for frags.
|
|
|
|
*/
|
|
|
|
struct sk_buff *alloc_skb_with_frags(unsigned long header_len,
|
|
|
|
unsigned long data_len,
|
|
|
|
int max_page_order,
|
|
|
|
int *errcode,
|
|
|
|
gfp_t gfp_mask)
|
|
|
|
{
|
|
|
|
int npages = (data_len + (PAGE_SIZE - 1)) >> PAGE_SHIFT;
|
|
|
|
unsigned long chunk;
|
|
|
|
struct sk_buff *skb;
|
|
|
|
struct page *page;
|
|
|
|
gfp_t gfp_head;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
*errcode = -EMSGSIZE;
|
|
|
|
/* Note this test could be relaxed, if we succeed to allocate
|
|
|
|
* high order pages...
|
|
|
|
*/
|
|
|
|
if (npages > MAX_SKB_FRAGS)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
gfp_head = gfp_mask;
|
2015-11-06 17:28:21 -07:00
|
|
|
if (gfp_head & __GFP_DIRECT_RECLAIM)
|
2014-09-17 05:49:49 -06:00
|
|
|
gfp_head |= __GFP_REPEAT;
|
|
|
|
|
|
|
|
*errcode = -ENOBUFS;
|
|
|
|
skb = alloc_skb(header_len, gfp_head);
|
|
|
|
if (!skb)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
skb->truesize += npages << PAGE_SHIFT;
|
|
|
|
|
|
|
|
for (i = 0; npages > 0; i++) {
|
|
|
|
int order = max_page_order;
|
|
|
|
|
|
|
|
while (order) {
|
|
|
|
if (npages >= 1 << order) {
|
2015-11-06 17:28:21 -07:00
|
|
|
page = alloc_pages((gfp_mask & ~__GFP_DIRECT_RECLAIM) |
|
2014-09-17 05:49:49 -06:00
|
|
|
__GFP_COMP |
|
|
|
|
__GFP_NOWARN |
|
|
|
|
__GFP_NORETRY,
|
|
|
|
order);
|
|
|
|
if (page)
|
|
|
|
goto fill_page;
|
|
|
|
/* Do not retry other high order allocations */
|
|
|
|
order = 1;
|
|
|
|
max_page_order = 0;
|
|
|
|
}
|
|
|
|
order--;
|
|
|
|
}
|
|
|
|
page = alloc_page(gfp_mask);
|
|
|
|
if (!page)
|
|
|
|
goto failure;
|
|
|
|
fill_page:
|
|
|
|
chunk = min_t(unsigned long, data_len,
|
|
|
|
PAGE_SIZE << order);
|
|
|
|
skb_fill_page_desc(skb, i, page, 0, chunk);
|
|
|
|
data_len -= chunk;
|
|
|
|
npages -= 1 << order;
|
|
|
|
}
|
|
|
|
return skb;
|
|
|
|
|
|
|
|
failure:
|
|
|
|
kfree_skb(skb);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(alloc_skb_with_frags);
|
skbuff: Add pskb_extract() helper function
A pattern of skb usage seen in modules such as RDS-TCP is to
extract `to_copy' bytes from the received TCP segment, starting
at some offset `off' into a new skb `clone'. This is done in
the ->data_ready callback, where the clone skb is queued up for rx on
the PF_RDS socket, while the parent TCP segment is returned unchanged
back to the TCP engine.
The existing code uses the sequence
clone = skb_clone(..);
pskb_pull(clone, off, ..);
pskb_trim(clone, to_copy, ..);
with the intention of discarding the first `off' bytes. However,
skb_clone() + pskb_pull() implies pksb_expand_head(), which ends
up doing a redundant memcpy of bytes that will then get discarded
in __pskb_pull_tail().
To avoid this inefficiency, this commit adds pskb_extract() that
creates the clone, and memcpy's only the relevant header/frag/frag_list
to the start of `clone'. pskb_trim() is then invoked to trim clone
down to the requested to_copy bytes.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-22 19:36:35 -06:00
|
|
|
|
|
|
|
/* carve out the first off bytes from skb when off < headlen */
|
|
|
|
static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,
|
|
|
|
const int headlen, gfp_t gfp_mask)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
int size = skb_end_offset(skb);
|
|
|
|
int new_hlen = headlen - off;
|
|
|
|
u8 *data;
|
|
|
|
|
|
|
|
size = SKB_DATA_ALIGN(size);
|
|
|
|
|
|
|
|
if (skb_pfmemalloc(skb))
|
|
|
|
gfp_mask |= __GFP_MEMALLOC;
|
|
|
|
data = kmalloc_reserve(size +
|
|
|
|
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
|
|
|
|
gfp_mask, NUMA_NO_NODE, NULL);
|
|
|
|
if (!data)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
size = SKB_WITH_OVERHEAD(ksize(data));
|
|
|
|
|
|
|
|
/* Copy real data, and all frags */
|
|
|
|
skb_copy_from_linear_data_offset(skb, off, data, new_hlen);
|
|
|
|
skb->len -= off;
|
|
|
|
|
|
|
|
memcpy((struct skb_shared_info *)(data + size),
|
|
|
|
skb_shinfo(skb),
|
|
|
|
offsetof(struct skb_shared_info,
|
|
|
|
frags[skb_shinfo(skb)->nr_frags]));
|
|
|
|
if (skb_cloned(skb)) {
|
|
|
|
/* drop the old head gracefully */
|
|
|
|
if (skb_orphan_frags(skb, gfp_mask)) {
|
|
|
|
kfree(data);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
|
|
|
|
skb_frag_ref(skb, i);
|
|
|
|
if (skb_has_frag_list(skb))
|
|
|
|
skb_clone_fraglist(skb);
|
|
|
|
skb_release_data(skb);
|
|
|
|
} else {
|
|
|
|
/* we can reuse existing recount- all we did was
|
|
|
|
* relocate values
|
|
|
|
*/
|
|
|
|
skb_free_head(skb);
|
|
|
|
}
|
|
|
|
|
|
|
|
skb->head = data;
|
|
|
|
skb->data = data;
|
|
|
|
skb->head_frag = 0;
|
|
|
|
#ifdef NET_SKBUFF_DATA_USES_OFFSET
|
|
|
|
skb->end = size;
|
|
|
|
#else
|
|
|
|
skb->end = skb->head + size;
|
|
|
|
#endif
|
|
|
|
skb_set_tail_pointer(skb, skb_headlen(skb));
|
|
|
|
skb_headers_offset_update(skb, 0);
|
|
|
|
skb->cloned = 0;
|
|
|
|
skb->hdr_len = 0;
|
|
|
|
skb->nohdr = 0;
|
|
|
|
atomic_set(&skb_shinfo(skb)->dataref, 1);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int pskb_carve(struct sk_buff *skb, const u32 off, gfp_t gfp);
|
|
|
|
|
|
|
|
/* carve out the first eat bytes from skb's frag_list. May recurse into
|
|
|
|
* pskb_carve()
|
|
|
|
*/
|
|
|
|
static int pskb_carve_frag_list(struct sk_buff *skb,
|
|
|
|
struct skb_shared_info *shinfo, int eat,
|
|
|
|
gfp_t gfp_mask)
|
|
|
|
{
|
|
|
|
struct sk_buff *list = shinfo->frag_list;
|
|
|
|
struct sk_buff *clone = NULL;
|
|
|
|
struct sk_buff *insp = NULL;
|
|
|
|
|
|
|
|
do {
|
|
|
|
if (!list) {
|
|
|
|
pr_err("Not enough bytes to eat. Want %d\n", eat);
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
if (list->len <= eat) {
|
|
|
|
/* Eaten as whole. */
|
|
|
|
eat -= list->len;
|
|
|
|
list = list->next;
|
|
|
|
insp = list;
|
|
|
|
} else {
|
|
|
|
/* Eaten partially. */
|
|
|
|
if (skb_shared(list)) {
|
|
|
|
clone = skb_clone(list, gfp_mask);
|
|
|
|
if (!clone)
|
|
|
|
return -ENOMEM;
|
|
|
|
insp = list->next;
|
|
|
|
list = clone;
|
|
|
|
} else {
|
|
|
|
/* This may be pulled without problems. */
|
|
|
|
insp = list;
|
|
|
|
}
|
|
|
|
if (pskb_carve(list, eat, gfp_mask) < 0) {
|
|
|
|
kfree_skb(clone);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
} while (eat);
|
|
|
|
|
|
|
|
/* Free pulled out fragments. */
|
|
|
|
while ((list = shinfo->frag_list) != insp) {
|
|
|
|
shinfo->frag_list = list->next;
|
|
|
|
kfree_skb(list);
|
|
|
|
}
|
|
|
|
/* And insert new clone at head. */
|
|
|
|
if (clone) {
|
|
|
|
clone->next = list;
|
|
|
|
shinfo->frag_list = clone;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* carve off first len bytes from skb. Split line (off) is in the
|
|
|
|
* non-linear part of skb
|
|
|
|
*/
|
|
|
|
static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,
|
|
|
|
int pos, gfp_t gfp_mask)
|
|
|
|
{
|
|
|
|
int i, k = 0;
|
|
|
|
int size = skb_end_offset(skb);
|
|
|
|
u8 *data;
|
|
|
|
const int nfrags = skb_shinfo(skb)->nr_frags;
|
|
|
|
struct skb_shared_info *shinfo;
|
|
|
|
|
|
|
|
size = SKB_DATA_ALIGN(size);
|
|
|
|
|
|
|
|
if (skb_pfmemalloc(skb))
|
|
|
|
gfp_mask |= __GFP_MEMALLOC;
|
|
|
|
data = kmalloc_reserve(size +
|
|
|
|
SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
|
|
|
|
gfp_mask, NUMA_NO_NODE, NULL);
|
|
|
|
if (!data)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
size = SKB_WITH_OVERHEAD(ksize(data));
|
|
|
|
|
|
|
|
memcpy((struct skb_shared_info *)(data + size),
|
|
|
|
skb_shinfo(skb), offsetof(struct skb_shared_info,
|
|
|
|
frags[skb_shinfo(skb)->nr_frags]));
|
|
|
|
if (skb_orphan_frags(skb, gfp_mask)) {
|
|
|
|
kfree(data);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
shinfo = (struct skb_shared_info *)(data + size);
|
|
|
|
for (i = 0; i < nfrags; i++) {
|
|
|
|
int fsize = skb_frag_size(&skb_shinfo(skb)->frags[i]);
|
|
|
|
|
|
|
|
if (pos + fsize > off) {
|
|
|
|
shinfo->frags[k] = skb_shinfo(skb)->frags[i];
|
|
|
|
|
|
|
|
if (pos < off) {
|
|
|
|
/* Split frag.
|
|
|
|
* We have two variants in this case:
|
|
|
|
* 1. Move all the frag to the second
|
|
|
|
* part, if it is possible. F.e.
|
|
|
|
* this approach is mandatory for TUX,
|
|
|
|
* where splitting is expensive.
|
|
|
|
* 2. Split is accurately. We make this.
|
|
|
|
*/
|
|
|
|
shinfo->frags[0].page_offset += off - pos;
|
|
|
|
skb_frag_size_sub(&shinfo->frags[0], off - pos);
|
|
|
|
}
|
|
|
|
skb_frag_ref(skb, i);
|
|
|
|
k++;
|
|
|
|
}
|
|
|
|
pos += fsize;
|
|
|
|
}
|
|
|
|
shinfo->nr_frags = k;
|
|
|
|
if (skb_has_frag_list(skb))
|
|
|
|
skb_clone_fraglist(skb);
|
|
|
|
|
|
|
|
if (k == 0) {
|
|
|
|
/* split line is in frag list */
|
|
|
|
pskb_carve_frag_list(skb, shinfo, off - pos, gfp_mask);
|
|
|
|
}
|
|
|
|
skb_release_data(skb);
|
|
|
|
|
|
|
|
skb->head = data;
|
|
|
|
skb->head_frag = 0;
|
|
|
|
skb->data = data;
|
|
|
|
#ifdef NET_SKBUFF_DATA_USES_OFFSET
|
|
|
|
skb->end = size;
|
|
|
|
#else
|
|
|
|
skb->end = skb->head + size;
|
|
|
|
#endif
|
|
|
|
skb_reset_tail_pointer(skb);
|
|
|
|
skb_headers_offset_update(skb, 0);
|
|
|
|
skb->cloned = 0;
|
|
|
|
skb->hdr_len = 0;
|
|
|
|
skb->nohdr = 0;
|
|
|
|
skb->len -= off;
|
|
|
|
skb->data_len = skb->len;
|
|
|
|
atomic_set(&skb_shinfo(skb)->dataref, 1);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* remove len bytes from the beginning of the skb */
|
|
|
|
static int pskb_carve(struct sk_buff *skb, const u32 len, gfp_t gfp)
|
|
|
|
{
|
|
|
|
int headlen = skb_headlen(skb);
|
|
|
|
|
|
|
|
if (len < headlen)
|
|
|
|
return pskb_carve_inside_header(skb, len, headlen, gfp);
|
|
|
|
else
|
|
|
|
return pskb_carve_inside_nonlinear(skb, len, headlen, gfp);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Extract to_copy bytes starting at off from skb, and return this in
|
|
|
|
* a new skb
|
|
|
|
*/
|
|
|
|
struct sk_buff *pskb_extract(struct sk_buff *skb, int off,
|
|
|
|
int to_copy, gfp_t gfp)
|
|
|
|
{
|
|
|
|
struct sk_buff *clone = skb_clone(skb, gfp);
|
|
|
|
|
|
|
|
if (!clone)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
if (pskb_carve(clone, off, gfp) < 0 ||
|
|
|
|
pskb_trim(clone, to_copy)) {
|
|
|
|
kfree_skb(clone);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
return clone;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(pskb_extract);
|
udp: under rx pressure, try to condense skbs
Under UDP flood, many softirq producers try to add packets to
UDP receive queue, and one user thread is burning one cpu trying
to dequeue packets as fast as possible.
Two parts of the per packet cost are :
- copying payload from kernel space to user space,
- freeing memory pieces associated with skb.
If socket is under pressure, softirq handler(s) can try to pull in
skb->head the payload of the packet if it fits.
Meaning the softirq handler(s) can free/reuse the page fragment
immediately, instead of letting udp_recvmsg() do this hundreds of usec
later, possibly from another node.
Additional gains :
- We reduce skb->truesize and thus can store more packets per SO_RCVBUF
- We avoid cache line misses at copyout() time and consume_skb() time,
and avoid one put_page() with potential alien freeing on NUMA hosts.
This comes at the cost of a copy, bounded to available tail room, which
is usually small. (We might have to fix GRO_MAX_HEAD which looks bigger
than necessary)
This patch gave me about 5 % increase in throughput in my tests.
skb_condense() helper could probably used in other contexts.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-12-07 10:19:33 -07:00
|
|
|
|
|
|
|
/**
|
|
|
|
* skb_condense - try to get rid of fragments/frag_list if possible
|
|
|
|
* @skb: buffer
|
|
|
|
*
|
|
|
|
* Can be used to save memory before skb is added to a busy queue.
|
|
|
|
* If packet has bytes in frags and enough tail room in skb->head,
|
|
|
|
* pull all of them, so that we can free the frags right now and adjust
|
|
|
|
* truesize.
|
|
|
|
* Notes:
|
|
|
|
* We do not reallocate skb->head thus can not fail.
|
|
|
|
* Caller must re-evaluate skb->truesize if needed.
|
|
|
|
*/
|
|
|
|
void skb_condense(struct sk_buff *skb)
|
|
|
|
{
|
2016-12-09 09:02:05 -07:00
|
|
|
if (skb->data_len) {
|
|
|
|
if (skb->data_len > skb->end - skb->tail ||
|
|
|
|
skb_cloned(skb))
|
|
|
|
return;
|
udp: under rx pressure, try to condense skbs
Under UDP flood, many softirq producers try to add packets to
UDP receive queue, and one user thread is burning one cpu trying
to dequeue packets as fast as possible.
Two parts of the per packet cost are :
- copying payload from kernel space to user space,
- freeing memory pieces associated with skb.
If socket is under pressure, softirq handler(s) can try to pull in
skb->head the payload of the packet if it fits.
Meaning the softirq handler(s) can free/reuse the page fragment
immediately, instead of letting udp_recvmsg() do this hundreds of usec
later, possibly from another node.
Additional gains :
- We reduce skb->truesize and thus can store more packets per SO_RCVBUF
- We avoid cache line misses at copyout() time and consume_skb() time,
and avoid one put_page() with potential alien freeing on NUMA hosts.
This comes at the cost of a copy, bounded to available tail room, which
is usually small. (We might have to fix GRO_MAX_HEAD which looks bigger
than necessary)
This patch gave me about 5 % increase in throughput in my tests.
skb_condense() helper could probably used in other contexts.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-12-07 10:19:33 -07:00
|
|
|
|
2016-12-09 09:02:05 -07:00
|
|
|
/* Nice, we can free page frag(s) right now */
|
|
|
|
__pskb_pull_tail(skb, skb->data_len);
|
|
|
|
}
|
|
|
|
/* At this point, skb->truesize might be over estimated,
|
|
|
|
* because skb had a fragment, and fragments do not tell
|
|
|
|
* their truesize.
|
|
|
|
* When we pulled its content into skb->head, fragment
|
|
|
|
* was freed, but __pskb_pull_tail() could not possibly
|
|
|
|
* adjust skb->truesize, not knowing the frag truesize.
|
udp: under rx pressure, try to condense skbs
Under UDP flood, many softirq producers try to add packets to
UDP receive queue, and one user thread is burning one cpu trying
to dequeue packets as fast as possible.
Two parts of the per packet cost are :
- copying payload from kernel space to user space,
- freeing memory pieces associated with skb.
If socket is under pressure, softirq handler(s) can try to pull in
skb->head the payload of the packet if it fits.
Meaning the softirq handler(s) can free/reuse the page fragment
immediately, instead of letting udp_recvmsg() do this hundreds of usec
later, possibly from another node.
Additional gains :
- We reduce skb->truesize and thus can store more packets per SO_RCVBUF
- We avoid cache line misses at copyout() time and consume_skb() time,
and avoid one put_page() with potential alien freeing on NUMA hosts.
This comes at the cost of a copy, bounded to available tail room, which
is usually small. (We might have to fix GRO_MAX_HEAD which looks bigger
than necessary)
This patch gave me about 5 % increase in throughput in my tests.
skb_condense() helper could probably used in other contexts.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-12-07 10:19:33 -07:00
|
|
|
*/
|
|
|
|
skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
|
|
|
|
}
|