Eric Dumazet says:
====================
net_sched: sch_fq: enable in-kernel pacing for QUIC servers
Willem added GSO support to UDP stack, greatly improving performance
of QUIC servers.
We also want to enable in-kernel pacing, which is possible thanks to EDT
model, since each sendmsg() can provide a timestamp for the skbs.
We have to change sch_fq to enable feeding packets in arbitrary EDT order,
and make sure that packet classification do not trust unconnected sockets.
Note that this patch series also is a prereq for a future TCP change
enabling per-flow delays/reorders/losses to implement high performance
TCP emulators.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
FQ packet scheduler assumed that packets could be classified
based on their owning socket.
This means that if a UDP server uses one UDP socket to send
packets to different destinations, packets all land
in one FQ flow.
This is unfair, since each TCP flow has a unique bucket, meaning
that in case of pressure (fully utilised uplink), TCP flows
have more share of the bandwidth.
If we instead detect unconnected sockets, we can use a stochastic
hash based on the 4-tuple hash.
This also means a QUIC server using one UDP socket will properly
spread the outgoing packets to different buckets, and in-kernel
pacing based on EDT model will no longer risk having big rb-tree on
one flow.
Note that UDP application might provide the skb->hash in an
ancillary message at sendmsg() time to avoid the cost of a dissection
in fq packet scheduler.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP stack makes sure packets for a given flow are monotically
increasing, but we want to allow UDP packets to use EDT as
well, so that QUIC servers can use in-kernel pacing.
This patch adds a per-flow rb-tree on which packets might
be stored. We still try to use the linear list for the
typical cases where packets are queued with monotically
increasing skb->tstamp, since queue/dequeue packets on
a standard list is O(1).
Note that the ability to store packets in arbitrary EDT
order will allow us to implement later a per TCP socket
mechanism adding delays (with jitter eventually) and reorders,
to implement convenient network emulators.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Huazhong Tan says:
====================
cleanup & optimizations & bugfixes for HNS3 driver
This patchset contains some cleanup related to hns3_enet_ring
struct and tx bd filling process, optimizations related
to rx page reusing, barrier using and tso handling process,
bugfixes related to tunnel type handling and error handling for
desc filling.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch uses devm_kcalloc instead of kcalloc when allocating
ring->desc_cb, because devm_kcalloc not only ensure to free the
memory when the dev is deallocted, but also allocate the memory
from it's device memory node.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes some unused field in struct hns3_enet_ring,
use ring->dev for ring_to_dev macro, and use dev consistently
in hns3_fill_desc.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When page size is 64K, RX buffer is currently not reused when the
page_offset is moved to last buffer. This patch adds checking to
decide whether the buffer page can be reused when last_offset is
moved beyond last offset.
If the driver is the only user of page when page_offset is moved
to beyond last offset, then buffer can be reused and page_offset
is set to zero.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, a barrier is used when cleaning each TX BD, which may
cause performance degradation.
This patch optimizes it to use one barrier when cleaning TX BD
each round.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When desc filling fails in hns3_nic_net_xmit, it will call
hns3_clear_desc to unmap the dma mapping. But currently the
ring->next_to_use points to the desc where the desc filling
or dma mapping return error, which means the desc that
ring->next_to_use points to has not done the dma mapping,
the desc that need unmapping is before the ring->next_to_use.
This patch fixes it by calling ring_ptr_move_bw(next_to_use)
before doing unmapping operation, and set desc_cb->dma to
zero to avoid freeing it again when unloading.
Also, when filling skb head or frag fails, both need to unmap
all the way back to next_to_use_head, so remove one desc filling
error handling.
Fixes: 76ad4f0ee7 ("net: hns3: Add support of HNS3 Ethernet Driver for hip08 SoC")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When filling len and checksum info to description, there is some
similar checking or calculation.
So this patch adds hns3_set_l2l3l4 to fill the inner(/normal)
header's len and checksum info. If it is a encapsulation skb, it
calls hns3_set_outer_l2l3l4 to handle the outer header's len and
checksum info, in order to avoid some similar checking or
calculation.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch separates the inner and outer l2l3l4 len handling in
hns3_set_l2l3l4_len, this is a preparation to combine the l2l3l4
len and checksum handling for inner and outer header.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
According to hardware user manual, the tunnel packet type is
available in the rx.ol_info field of struct hns3_desc. Currently
the tunnel packet type is decided by the rx.l234_info, which may
cause RX checksum handling error.
This patch fixes it by using the correct field in struct hns3_desc
to decide the tunnel packet type.
Fixes: 76ad4f0ee7 ("net: hns3: Add support of HNS3 Ethernet Driver for hip08 SoC")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
HW requires every continuous 8 buffer data to be larger than MSS,
we simplify it by ensuring skb_headlen + the first continuous
7 frags to to be larger than GSO header len + mss, and the
remaining continuous 7 frags to be larger than MSS except the
last 7 frags.
This patch adds hns3_skb_need_linearized to handle it for TSO
case.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, using "ethtool --statistics" can show how many time RX
page have been reused, but there is no counter for RX page not
being reused.
This patch adds non_reuse_pg counter to better debug the performance
issue, because it is hard to determine when the RX page is reused
or not if there is no such counter.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
napi_schedule_irqoff is introduced to be used from hard interrupts
handlers or when irqs are already masked, see:
https://lists.openwall.net/netdev/2014/10/29/2
So this patch replaces napi_schedule with napi_schedule_irqoff.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, maybe_stop_tx ops for TSO and non-TSO case share some BD
calculation code, so this patch unifies the maybe_stop_tx by removing
the maybe_stop_tx ops. skb_is_gso() can be used to differentiate the
case between TSO and non-TSO case if there is need to handle special
case for TSO case.
This patch also add tx_copy field in "ethtool --statistics" to help
better debug the performance issue caused by calling skb_copy.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Hauke Mehrtens says:
====================
net: dsa: lantiq: Add bridge offloading
This adds bridge offloading for the Intel / Lantiq GSWIP 2.1 switch.
Changes since:
v2:
- Added Fixes tag to patch 1
- Fixed typo
- added GSWIP_TABLE_MAC_BRIDGE_STATIC and made use of it
- used GSWIP_TABLE_MAC_BRIDGE in more places
v1:
- fix typo signle -> single
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds functions to add and remove static entries to and from the
forwarding database and dump the full forwarding database.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fast aging per port is not supported directly by the hardware, it is
only possible to configure a global aging time.
Do the fast aging by iterating over the MAC forwarding table and remove
all dynamic entries for a given port.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The VLAN aware bridge offloading is similar to the VLAN unaware
offloading, this makes it possible to offload the VLAN bridge
functionalities.
The hardware supports up to 64 VLAN bridge entries, we already use one
entry for each LAN port to prevent forwarding of packets between the
ports when the ports are not in a bridge, so in the end we have 57
possible VLANs.
The VLAN filtering is currently only active when the ports are in a
bridge, VLAN filtering for ports not in a bridge is not implemented.
It is currently not possible to change between VLAN filtering and not
filtering while the port is already in a bridge, this would make the
driver more complicated.
The VLANs are only defined on bridge entries, so we will not add
anything into the hardware when the port joins a bridge if it is doing
VLAN filtering, but only when an allowed VLAN is added.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This allows to offload bridges with DSA to the switch hardware and do
the packet forwarding in hardware.
This implements generic functions to access the switch hardware tables,
which are used to control many features of the switch.
This patch activates the MAC learning by removing the MAC address table
lock, to prevent uncontrolled forwarding of packets between all the LAN
ports, they are added into individual bridge tables entries with
individual flow ids and the switch will do the MAC learning for each
port separately before they are added to a real bridge.
Each bridge consist of an entry in the active VLAN table and the VLAN
mapping table, table entries with the same index are matching. In the
VLAN unaware mode we configure everything with VLAN ID 0, but we use
different flow IDs, the switch should handle all VLANs as normal payload
and ignore them. When the hardware looks for the port of the destination
MAC address it only takes the entries which have the same flow ID of the
ingress packet.
The bridges are configured with 64 possible entries with these
information:
Table Index, 0...63
VLAN ID, 0...4095: VLAN ID 0 is untagged
flow ID, 0..63: Same flow IDs share entries in MAC learning table
port map, one bit for each port number
tagged port map, one bit for each port number
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow the special tag in ingress only on the CPU port and not on all
ports. A packet with a special tag could circumvent the hardware
forwarding and should only be allowed on the CPU port where Linux
controls the port.
Fixes: 14fceff477 ("net: dsa: Add Lantiq / Intel DSA driver for vrx200)"
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Daniel Borkmann says:
====================
pull-request: bpf-next 2019-05-06
The following pull-request contains BPF updates for your *net-next* tree.
The main changes are:
1) Two AF_XDP libbpf fixes for socket teardown; first one an invalid
munmap and the other one an invalid skmap cleanup, both from Björn.
2) More graceful CONFIG_DEBUG_INFO_BTF handling when pahole is not
present in the system to generate vmlinux btf info, from Andrii.
3) Fix libbpf and thus fix perf build error with uClibc on arc
architecture, from Vineet.
4) Fix missing libbpf_util.h header install in libbpf, from William.
5) Exclude bash-completion/bpftool from .gitignore pattern, from Masahiro.
6) Fix up rlimit in test_libbpf_open kselftest test case, from Yonghong.
7) Minor misc cleanups.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
The patches with fixes tags added a cast-to-void in the places when
the return value of a function was ignored.
This is not common practice in the kernel, and is therefore removed in
this patch.
Reported-by: Daniel Borkmann <daniel@iogearbox.net>
Fixes: 5750902a6e ("libbpf: proper XSKMAP cleanup")
Fixes: 0e6741f092 ("libbpf: fix invalid munmap call")
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
When BTF generation is enabled through CONFIG_DEBUG_INFO_BTF,
scripts/link-vmlinux.sh detects if pahole version is too old and
gracefully continues build process, skipping BTF generation build step.
But if pahole is not available, build will still fail. This patch adds
check for whether pahole exists at all and bails out gracefully, if not.
Cc: Alexei Starovoitov <ast@fb.com>
Reported-by: Yonghong Song <yhs@fb.com>
Fixes: e83b9f5544 ("kbuild: add ability to generate BTF type info for vmlinux")
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Heiner Kallweit says:
====================
r8169: replace some magic with more speaking functions
Based on info from Realtek replace some magic with speaking functions
even though the exact meaning of certain values isn't known.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Based on info from Realtek add a function for defining the thresholds
controlling ethernet flow control.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Based on info from Realtek replace FIFO size config magic with
a function.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Ido Schimmel says:
====================
mlxsw: spectrum: Implement loopback ethtool feature
This patchset from Jiri allows users to enable loopback feature for
individual ports using ethtool. The loopback feature is useful for
testing purposes and will also be used by upcoming patchsets to enable
the monitoring of buffer drops.
Patch #1 adds the relevant device register.
Patch #2 Implements support in the driver.
Patch #3 adds a selftest.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Add selftest for loopback feature
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow user to enable loopback feature for individual ports using ethtool.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The PPLR register allows configuration of the port's loopback mode.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit 3c75f6ee13 ("net_sched: sch_htb: add per class overlimits counter")
we added an overlimits counter for each HTB class which could
properly reflect how many times we use up all the bandwidth
on each class. However, the overlimits counter in HTB qdisc
does not, it is way bigger than the sum of each HTB class.
In fact, this qdisc overlimits counter increases when we have
no skb to dequeue, which happens more often than we run out of
bandwidth.
It makes more sense to make this qdisc overlimits counter just
be a sum of each HTB class, in case people still get confused.
I have verified this patch with one single HTB class, where HTB
qdisc counters now always match HTB class counters as expected.
Eric suggested we could fold this field into 'direct_pkts' as
we only use its 32bit on 64bit CPU, this saves one cache line.
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move adjusting the EEE LED frequency to rtl8168_config_eee_mac.
Exclude RTL8411 (version 38) like in the existing code.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make both functions macros to allow omitting the ARRAY_SIZE(x) argument.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Vladimir Oltean says:
====================
Traffic support for SJA1105 DSA driver
This patch set is a continuation of the "NXP SJA1105 DSA driver" v3
series, which was split in multiple pieces for easier review.
Supporting a fully-featured (traffic-capable) driver for this switch
requires some rework in DSA and also leaves behind a more generic
infrastructure for other dumb switches that rely on 802.1Q pseudo-switch
tagging for port separation. Among the DSA changes required are:
* Generic xmit and rcv functions for pushing/popping 802.1Q tags on
skb's. These are modeled as a tagging protocol of its own but which
must be customized by drivers to fit their own hardware possibilities.
* Permitting the .setup callback to invoke switchdev operations that
will loop back into the driver through the switchdev notifier chain.
The SJA1105 driver then proceeds to extend this 8021q switch tagging
protocol while adding its own (tag_sja1105). This is done because
the driver actually implements a "dual tagger":
* For normal traffic it uses 802.1Q tags
* For management (multicast DMAC) frames the switch has native support
for recognizing and annotating these with source port and switch id
information.
Because this is a "dual tagger", decoding of management frames should
still function when regular traffic can't (under a bridge with VLAN
filtering).
There was intervention in the DSA receive hotpath, where a new
filtering function called from eth_type_trans() is needed. This is
useful in the general sense for switches that might actually have some
limited means of source port decoding, such as only for management
traffic, but not for everything.
In order for the 802.1Q tagging protocol (which cannot be enabled under
all conditions, unlike the management traffic decoding) to not be an
all-or-nothing choice, the filtering function matches everything that
can be decoded, and everything else is left to pass to the master
netdevice.
Lastly, DSA core support was added for drivers to request skb deferral.
SJA1105 needs this for SPI intervention during transmission of link-local
traffic. This is not done from within the tagger.
Some patches were carried over unchanged from the previous patchset
(01/09). Others were slightly reworked while adapting to the recent
changes in "Make DSA tag drivers kernel modules" (02/09).
The introduction of some structures (DSA_SKB_CB, dp->priv) may seem a
little premature at this point and the new structures under-utilized.
The reason is that traffic support has been rewritten with PTP
timestamping in mind, and then I removed the timestamping code from the
current submission (1. it is a different topic, 2. it does not work very
well yet). On demand I can provide the timestamping patchset as a RFC
though.
"NXP SJA1105 DSA driver" v3 patchset can be found at:
https://lkml.org/lkml/2019/4/12/978
v1 patchset can be found at:
https://lkml.org/lkml/2019/5/3/877
Changes in v2:
* Made the deferred xmit workqueue also be drained on the netdev suspend
callback, not just on ndo_stop.
* Added clarification about how other netdevices may be bridged with the
switch ports.
v2 patchset can be found at:
https://www.spinics.net/lists/netdev/msg568818.html
Changes in v3:
* Exported the dsa_port_vid_add and dsa_port_vid_del symbols to fix an
error reported by the kbuild test robot
* Fixed the following checkpatch warnings in 05/10:
Macro argument reuse 'skb' - possible side-effects?
Macro argument reuse 'clone' - possible side-effects?
* Added a commit description to the documentation patch (10/10)
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds a table which illustrates what combinations of management /
regular traffic work depending on the state the switch ports are in.
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While not explicitly documented as supported in UM10944, compliance with
the STP states can be obtained by manipulating 3 settings at the
(per-port) MAC config level: dynamic learning, inhibiting reception of
regular traffic, and inhibiting transmission of regular traffic.
In all these modes, transmission and reception of special BPDU frames
from the stack is still enabled (not inhibited by the MAC-level
settings).
On ingress, BPDUs are classified by the MAC filter as link-local
(01-80-C2-00-00-00) and forwarded to the CPU port. This mechanism works
under all conditions (even without the custom 802.1Q tagging) because
the switch hardware inserts the source port and switch ID into bytes 4
and 5 of the MAC-filtered frames. Then the DSA .rcv handler needs to put
back zeroes into the MAC address after decoding the source port
information.
On egress, BPDUs are transmitted using management routes from the xmit
worker thread. Again this does not require switch tagging, as the switch
port is programmed through SPI to hold a temporary (single-fire) route
for a frame with the programmed destination MAC (01-80-C2-00-00-00).
STP is activated using the following commands and was tested by
connecting two front-panel ports together and noticing that switching
loops were prevented (one port remains in the blocking state):
$ ip link add name br0 type bridge stp_state 1 && ip link set br0 up
$ for eth in $(ls /sys/devices/platform/soc/2100000.spi/spi_master/spi0/spi0.1/net/);
do ip link set ${eth} master br0 && ip link set ${eth} up; done
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In order to support this, we are creating a make-shift switch tag out of
a VLAN trunk configured on the CPU port. Termination of normal traffic
on switch ports only works when not under a vlan_filtering bridge.
Termination of management (PTP, BPDU) traffic works under all
circumstances because it uses a different tagging mechanism
(incl_srcpt). We are making use of the generic CONFIG_NET_DSA_TAG_8021Q
code and leveraging it from our own CONFIG_NET_DSA_TAG_SJA1105.
There are two types of traffic: regular and link-local.
The link-local traffic received on the CPU port is trapped from the
switch's regular forwarding decisions because it matched one of the two
DMAC filters for management traffic.
On transmission, the switch requires special massaging for these
link-local frames. Due to a weird implementation of the switching IP, by
default it drops link-local frames that originate on the CPU port.
It needs to be told where to forward them to, through an SPI command
("management route") that is valid for only a single frame.
So when we're sending link-local traffic, we are using the
dsa_defer_xmit mechanism.
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is supposed to share information between the driver and the tagger,
or used by the tagger to keep some state. Its use is optional.
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Reviewed-by: Vivien Didelot <vivien.didelot@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some hardware needs to take work to get convinced to receive frames on
the CPU port (such as the sja1105 which takes temporary L2 forwarding
rules over SPI that last for a single frame). Such work needs a
sleepable context, and because the regular .ndo_start_xmit is atomic,
this cannot be done in the tagger. So introduce a generic DSA mechanism
that sets up a transmit skb queue and a workqueue for deferred
transmission.
The new driver callback (.port_deferred_xmit) is in dsa_switch and not
in the tagger because the operations that require sleeping typically
also involve interacting with the hardware, and not simply skb
manipulations. Therefore having it there simplifies the structure a bit
and makes it unnecessary to export functions from the driver to the
tagger.
The driver is responsible of calling dsa_enqueue_skb which transfers it
to the master netdevice. This is so that it has a chance of performing
some more work afterwards, such as cleanup or TX timestamping.
To tell DSA that skb xmit deferral is required, I have thought about
changing the return type of the tagger .xmit from struct sk_buff * into
a enum dsa_tx_t that could potentially encode a DSA_XMIT_DEFER value.
But the trailer tagger is reallocating every skb on xmit and therefore
making a valid use of the pointer return value. So instead of reworking
the API in complicated ways, right now a boolean property in the newly
introduced DSA_SKB_CB is set.
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Map a DSA structure over the 48-byte control block that will hold
skb info on transmit and receive. This is only for use within the DSA
processing layer (e.g. communicating between DSA core and tagger) and
not for passing info around with other layers such as the master net
device.
Also add a DSA_SKB_CB_PRIV() macro which retrieves a pointer to the
space up to 48 bytes that the DSA structure does not use. This space can
be used for drivers to add their own private info.
One use is for the PTP timestamping code path. When cloning a skb,
annotate the original with a pointer to the clone, which the driver can
then find easily and place the timestamp to. This avoids the need of a
separate queue to hold clones and a way to match an original to a cloned
skb.
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Frames get processed by DSA and redirected to switch port net devices
based on the ETH_P_XDSA multiplexed packet_type handler found by the
network stack when calling eth_type_trans().
The running assumption is that once the DSA .rcv function is called, DSA
is always able to decode the switch tag in order to change the skb->dev
from its master.
However there are tagging protocols (such as the new DSA_TAG_PROTO_SJA1105,
user of DSA_TAG_PROTO_8021Q) where this assumption is not completely
true, since switch tagging piggybacks on the absence of a vlan_filtering
bridge. Moreover, management traffic (BPDU, PTP) for this switch doesn't
rely on switch tagging, but on a different mechanism. So it would make
sense to at least be able to terminate that.
Having DSA receive traffic it can't decode would put it in an impossible
situation: the eth_type_trans() function would invoke the DSA .rcv(),
which could not change skb->dev, then eth_type_trans() would be invoked
again, which again would call the DSA .rcv, and the packet would never
be able to exit the DSA filter and would spiral in a loop until the
whole system dies.
This happens because eth_type_trans() doesn't actually look at the skb
(so as to identify a potential tag) when it deems it as being
ETH_P_XDSA. It just checks whether skb->dev has a DSA private pointer
installed (therefore it's a DSA master) and that there exists a .rcv
callback (everybody except DSA_TAG_PROTO_NONE has that). This is
understandable as there are many switch tags out there, and exhaustively
checking for all of them is far from ideal.
The solution lies in introducing a filtering function for each tagging
protocol. In the absence of a filtering function, all traffic is passed
to the .rcv DSA callback. The tagging protocol should see the filtering
function as a pre-validation that it can decode the incoming skb. The
traffic that doesn't match the filter will bypass the DSA .rcv callback
and be left on the master netdevice, which wasn't previously possible.
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch provides generic DSA code for using VLAN (802.1Q) tags for
the same purpose as a dedicated switch tag for injection/extraction.
It is based on the discussions and interest that has been so far
expressed in https://www.spinics.net/lists/netdev/msg556125.html.
Unlike all other DSA-supported tagging protocols, CONFIG_NET_DSA_TAG_8021Q
does not offer a complete solution for drivers (nor can it). Instead, it
provides generic code that driver can opt into calling:
- dsa_8021q_xmit: Inserts a VLAN header with the specified contents.
Can be called from another tagging protocol's xmit function.
Currently the LAN9303 driver is inserting headers that are simply
802.1Q with custom fields, so this is an opportunity for code reuse.
- dsa_8021q_rcv: Retrieves the TPID and TCI from a VLAN-tagged skb.
Removing the VLAN header is left as a decision for the caller to make.
- dsa_port_setup_8021q_tagging: For each user port, installs an Rx VID
and a Tx VID, for proper untagged traffic identification on ingress
and steering on egress. Also sets up the VLAN trunk on the upstream
(CPU or DSA) port. Drivers are intentionally left to call this
function explicitly, depending on the context and hardware support.
The expected switch behavior and VLAN semantics should not be violated
under any conditions. That is, after calling
dsa_port_setup_8021q_tagging, the hardware should still pass all
ingress traffic, be it tagged or untagged.
For uniformity with the other tagging protocols, a module for the
dsa_8021q_netdev_ops structure is registered, but the typical usage is
to set up another tagging protocol which selects CONFIG_NET_DSA_TAG_8021Q,
and calls the API from tag_8021q.h. Null function definitions are also
provided so that a "depends on" is not forced in the Kconfig.
This tagging protocol only works when switch ports are standalone, or
when they are added to a VLAN-unaware bridge. It will probably remain
this way for the reasons below.
When added to a bridge that has vlan_filtering 1, the bridge core will
install its own VLANs and reset the pvids through switchdev. For the
bridge core, switchdev is a write-only pipe. All VLAN-related state is
kept in the bridge core and nothing is read from DSA/switchdev or from
the driver. So the bridge core will break this port separation because
it will install the vlan_default_pvid into all switchdev ports.
Even if we could teach the bridge driver about switchdev preference of a
certain vlan_default_pvid (task difficult in itself since the current
setting is per-bridge but we would need it per-port), there would still
exist many other challenges.
Firstly, in the DSA rcv callback, a driver would have to perform an
iterative reverse lookup to find the correct switch port. That is
because the port is a bridge slave, so its Rx VID (port PVID) is subject
to user configuration. How would we ensure that the user doesn't reset
the pvid to a different value (which would make an O(1) translation
impossible), or to a non-unique value within this DSA switch tree (which
would make any translation impossible)?
Finally, not all switch ports are equal in DSA, and that makes it
difficult for the bridge to be completely aware of this anyway.
The CPU port needs to transmit tagged packets (VLAN trunk) in order for
the DSA rcv code to be able to decode source information.
But the bridge code has absolutely no idea which switch port is the CPU
port, if nothing else then just because there is no netdevice registered
by DSA for the CPU port.
Also DSA does not currently allow the user to specify that they want the
CPU port to do VLAN trunking anyway. VLANs are added to the CPU port
using the same flags as they were added on the user port.
So the VLANs installed by dsa_port_setup_8021q_tagging per driver
request should remain private from the bridge's and user's perspective,
and should not alter the VLAN semantics observed by the user.
In the current implementation a VLAN range ending at 4095 (VLAN_N_VID)
is reserved for this purpose. Each port receives a unique Rx VLAN and a
unique Tx VLAN. Separate VLANs are needed for Rx and Tx because they
serve different purposes: on Rx the switch must process traffic as
untagged and process it with a port-based VLAN, but with care not to
hinder bridging. On the other hand, the Tx VLAN is where the
reachability restrictions are imposed, since by tagging frames in the
xmit callback we are telling the switch onto which port to steer the
frame.
Some general guidance on how this support might be employed for
real-life hardware (some comments made by Florian Fainelli):
- If the hardware supports VLAN tag stacking, it should somehow back
up its private VLAN settings when the bridge tries to override them.
Then the driver could re-apply them as outer tags. Dedicating an outer
tag per bridge device would allow identical inner tag VID numbers to
co-exist, yet preserve broadcast domain isolation.
- If the switch cannot handle VLAN tag stacking, it should disable this
port separation when added as slave to a vlan_filtering bridge, in
that case having reduced functionality.
- Drivers for old switches that don't support the entire VLAN_N_VID
range will need to rework the current range selection mechanism.
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Reviewed-by: Vivien Didelot <vivien.didelot@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is needed so that the newly introduced tag_8021q may access these
core DSA functions when built as a module.
Reported-by: kbuild test robot <lkp@intel.com>
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This allows the driver to perform some manipulations of its own during
setup, using generic switchdev calls. Having the notifiers registered at
setup time is important because otherwise any switchdev transaction
emitted during this time would be ignored (dispatched to an empty call
chain).
One current usage scenario is for the driver to request DSA to set up
802.1Q based switch tagging for its ports.
There is no danger for the driver setup code to start racing now with
switchdev events emitted from the network stack (such as bridge core)
even if the notifier is registered earlier. This is because the network
stack needs a net_device as a vehicle to perform switchdev operations,
and the slave net_devices are registered later than the core driver
setup anyway (ds->ops->setup in dsa_switch_setup vs dsa_port_setup).
Luckily DSA doesn't need a net_device to carry out switchdev callbacks,
and therefore drivers shouldn't assume either that net_devices are
available at the time their switchdev callbacks get invoked.
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Reviewed-by: Vivien Didelot <vivien.didelot@gmail.com>-
Signed-off-by: David S. Miller <davem@davemloft.net>
The Marvell SOHO switches have several ways to access the internal
registers. One of them being the System Management Interface (SMI),
using the MDC and MDIO pins, with direct and indirect variants.
In preparation for adding support for other register accesses, move
the SMI code into its own files. At the same time, refine the code
to make it clear that the indirect variant is implemented using the
direct variant accessing only two registers for command and data.
Signed-off-by: Vivien Didelot <vivien.didelot@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jakub Kicinski says:
===================
net: act_police offload support
this set starts by converting cls_matchall to the new flow offload
infrastructure. It so happens that all drivers implementing cls_matchall
offload today also offload cls_flower, so its a little easier for
them to handle the actions in unified flow_rule format, even though
in cls_matchall there is no flow to speak of. If a driver ever appears
which would prefer the old, direct access to TC exts, we can add the
pointer in the offload structure back and support both.
Next the act_police is added to actions supported by flow offload API.
NFP support for act_police offload is added as the final step. The flower
firmware is configured to perform TX rate limiting in a way which matches
act_police's behaviour. It does not use DMA.IN back pressure, and
instead drops packets after they had been already DMAed into the NIC.
IOW it uses our standard traffic policing implementation, future patches
will extend it to other ports and traffic directions.
===================
Signed-off-by: David S. Miller <davem@davemloft.net>
Add stats request function that sends a stats request message to hw for
a specific police-filter. Process stats reply from hw and update the
stored qos structure.
Signed-off-by: Pieter Jansen van Vuuren <pieter.jansenvanvuuren@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add install and remove offload functionality for qos offloads. We
first check that a police filter can be implemented by the VF rate
limiting feature in hw, then we install the filter via the qos
infrastructure. Finally we implement the mechanism for removing
these types of filters.
Signed-off-by: Pieter Jansen van Vuuren <pieter.jansenvanvuuren@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
