1
0
Fork 0
remarkable-linux/net/tipc/name_distr.c

398 lines
11 KiB
C
Raw Normal View History

/*
* net/tipc/name_distr.c: TIPC name distribution code
*
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
* Copyright (c) 2000-2006, 2014, Ericsson AB
* Copyright (c) 2005, 2010-2011, Wind River Systems
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. Neither the names of the copyright holders nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* Alternatively, this software may be distributed under the terms of the
* GNU General Public License ("GPL") version 2 as published by the Free
* Software Foundation.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#include "core.h"
#include "link.h"
#include "name_distr.h"
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
int sysctl_tipc_named_timeout __read_mostly = 2000;
struct distr_queue_item {
struct distr_item i;
u32 dtype;
u32 node;
unsigned long expires;
struct list_head next;
};
/**
* publ_to_item - add publication info to a publication message
*/
static void publ_to_item(struct distr_item *i, struct publication *p)
{
i->type = htonl(p->type);
i->lower = htonl(p->lower);
i->upper = htonl(p->upper);
i->ref = htonl(p->ref);
i->key = htonl(p->key);
}
/**
* named_prepare_buf - allocate & initialize a publication message
*
* The buffer returned is of size INT_H_SIZE + payload size
*/
static struct sk_buff *named_prepare_buf(struct net *net, u32 type, u32 size,
u32 dest)
{
2015-02-05 06:36:36 -07:00
struct tipc_net *tn = net_generic(net, tipc_net_id);
struct sk_buff *buf = tipc_buf_acquire(INT_H_SIZE + size, GFP_ATOMIC);
struct tipc_msg *msg;
if (buf != NULL) {
msg = buf_msg(buf);
2015-02-05 06:36:36 -07:00
tipc_msg_init(tn->own_addr, msg, NAME_DISTRIBUTOR, type,
INT_H_SIZE, dest);
msg_set_size(msg, INT_H_SIZE + size);
}
return buf;
}
/**
* tipc_named_publish - tell other nodes about a new publication by this node
*/
struct sk_buff *tipc_named_publish(struct net *net, struct publication *publ)
{
struct tipc_net *tn = net_generic(net, tipc_net_id);
struct sk_buff *buf;
struct distr_item *item;
list_add_tail_rcu(&publ->local_list,
&tn->nametbl->publ_list[publ->scope]);
if (publ->scope == TIPC_NODE_SCOPE)
return NULL;
buf = named_prepare_buf(net, PUBLICATION, ITEM_SIZE, 0);
if (!buf) {
pr_warn("Publication distribution failure\n");
return NULL;
}
item = (struct distr_item *)msg_data(buf_msg(buf));
publ_to_item(item, publ);
return buf;
}
/**
* tipc_named_withdraw - tell other nodes about a withdrawn publication by this node
*/
struct sk_buff *tipc_named_withdraw(struct net *net, struct publication *publ)
{
struct sk_buff *buf;
struct distr_item *item;
list_del(&publ->local_list);
if (publ->scope == TIPC_NODE_SCOPE)
return NULL;
buf = named_prepare_buf(net, WITHDRAWAL, ITEM_SIZE, 0);
if (!buf) {
pr_warn("Withdrawal distribution failure\n");
return NULL;
}
item = (struct distr_item *)msg_data(buf_msg(buf));
publ_to_item(item, publ);
return buf;
}
/**
* named_distribute - prepare name info for bulk distribution to another node
* @list: list of messages (buffers) to be returned from this function
* @dnode: node to be updated
* @pls: linked list of publication items to be packed into buffer chain
*/
static void named_distribute(struct net *net, struct sk_buff_head *list,
u32 dnode, struct list_head *pls)
{
struct publication *publ;
struct sk_buff *skb = NULL;
struct distr_item *item = NULL;
u32 msg_dsz = ((tipc_node_get_mtu(net, dnode, 0) - INT_H_SIZE) /
ITEM_SIZE) * ITEM_SIZE;
u32 msg_rem = msg_dsz;
list_for_each_entry(publ, pls, local_list) {
/* Prepare next buffer: */
if (!skb) {
skb = named_prepare_buf(net, PUBLICATION, msg_rem,
dnode);
if (!skb) {
pr_warn("Bulk publication failure\n");
return;
}
tipc: fix broadcast link synchronization problem In commit 2d18ac4ba745 ("tipc: extend broadcast link initialization criteria") we tried to fix a problem with the initial synchronization of broadcast link acknowledge values. Unfortunately that solution is not sufficient to solve the issue. We have seen it happen that LINK_PROTOCOL/STATE packets with a valid non-zero unicast acknowledge number may bypass BCAST_PROTOCOL initialization, NAME_DISTRIBUTOR and other STATE packets with invalid broadcast acknowledge numbers, leading to premature opening of the broadcast link. When the bypassed packets finally arrive, they are inadvertently accepted, and the already correctly initialized acknowledge number in the broadcast receive link is overwritten by the invalid (zero) value of the said packets. After this the broadcast link goes stale. We now fix this by marking the packets where we know the acknowledge value is or may be invalid, and then ignoring the acks from those. To this purpose, we claim an unused bit in the header to indicate that the value is invalid. We set the bit to 1 in the initial BCAST_PROTOCOL synchronization packet and all initial ("bulk") NAME_DISTRIBUTOR packets, plus those LINK_PROTOCOL packets sent out before the broadcast links are fully synchronized. This minor protocol update is fully backwards compatible. Reported-by: John Thompson <thompa.atl@gmail.com> Tested-by: John Thompson <thompa.atl@gmail.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-10-27 16:51:55 -06:00
msg_set_bc_ack_invalid(buf_msg(skb), true);
item = (struct distr_item *)msg_data(buf_msg(skb));
}
/* Pack publication into message: */
publ_to_item(item, publ);
item++;
msg_rem -= ITEM_SIZE;
/* Append full buffer to list: */
if (!msg_rem) {
__skb_queue_tail(list, skb);
skb = NULL;
msg_rem = msg_dsz;
}
}
if (skb) {
msg_set_size(buf_msg(skb), INT_H_SIZE + (msg_dsz - msg_rem));
skb_trim(skb, INT_H_SIZE + (msg_dsz - msg_rem));
__skb_queue_tail(list, skb);
}
}
/**
* tipc_named_node_up - tell specified node about all publications by this node
*/
void tipc_named_node_up(struct net *net, u32 dnode)
{
struct tipc_net *tn = net_generic(net, tipc_net_id);
struct sk_buff_head head;
__skb_queue_head_init(&head);
rcu_read_lock();
named_distribute(net, &head, dnode,
&tn->nametbl->publ_list[TIPC_CLUSTER_SCOPE]);
named_distribute(net, &head, dnode,
&tn->nametbl->publ_list[TIPC_ZONE_SCOPE]);
rcu_read_unlock();
tipc_node_xmit(net, &head, dnode, 0);
}
/**
* tipc_publ_purge - remove publication associated with a failed node
*
* Invoked for each publication issued by a newly failed node.
* Removes publication structure from name table & deletes it.
*/
static void tipc_publ_purge(struct net *net, struct publication *publ, u32 addr)
{
struct tipc_net *tn = net_generic(net, tipc_net_id);
struct publication *p;
spin_lock_bh(&tn->nametbl_lock);
p = tipc_nametbl_remove_publ(net, publ->type, publ->lower,
publ->node, publ->ref, publ->key);
if (p)
tipc_node_unsubscribe(net, &p->nodesub_list, addr);
spin_unlock_bh(&tn->nametbl_lock);
if (p != publ) {
pr_err("Unable to remove publication from failed node\n"
" (type=%u, lower=%u, node=0x%x, ref=%u, key=%u)\n",
publ->type, publ->lower, publ->node, publ->ref,
publ->key);
}
kfree_rcu(p, rcu);
}
/**
* tipc_dist_queue_purge - remove deferred updates from a node that went down
*/
static void tipc_dist_queue_purge(struct net *net, u32 addr)
{
struct tipc_net *tn = net_generic(net, tipc_net_id);
struct distr_queue_item *e, *tmp;
spin_lock_bh(&tn->nametbl_lock);
list_for_each_entry_safe(e, tmp, &tn->dist_queue, next) {
if (e->node != addr)
continue;
list_del(&e->next);
kfree(e);
}
spin_unlock_bh(&tn->nametbl_lock);
}
void tipc_publ_notify(struct net *net, struct list_head *nsub_list, u32 addr)
{
struct publication *publ, *tmp;
list_for_each_entry_safe(publ, tmp, nsub_list, nodesub_list)
tipc_publ_purge(net, publ, addr);
tipc_dist_queue_purge(net, addr);
}
/**
* tipc_update_nametbl - try to process a nametable update and notify
* subscribers
*
* tipc_nametbl_lock must be held.
* Returns the publication item if successful, otherwise NULL.
*/
static bool tipc_update_nametbl(struct net *net, struct distr_item *i,
u32 node, u32 dtype)
{
struct publication *publ = NULL;
if (dtype == PUBLICATION) {
publ = tipc_nametbl_insert_publ(net, ntohl(i->type),
ntohl(i->lower),
ntohl(i->upper),
TIPC_CLUSTER_SCOPE, node,
ntohl(i->ref), ntohl(i->key));
if (publ) {
tipc_node_subscribe(net, &publ->nodesub_list, node);
return true;
}
} else if (dtype == WITHDRAWAL) {
publ = tipc_nametbl_remove_publ(net, ntohl(i->type),
ntohl(i->lower),
node, ntohl(i->ref),
ntohl(i->key));
if (publ) {
tipc_node_unsubscribe(net, &publ->nodesub_list, node);
kfree_rcu(publ, rcu);
return true;
}
} else {
pr_warn("Unrecognized name table message received\n");
}
return false;
}
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
/**
* tipc_named_add_backlog - add a failed name table update to the backlog
*
*/
static void tipc_named_add_backlog(struct net *net, struct distr_item *i,
u32 type, u32 node)
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
{
struct distr_queue_item *e;
struct tipc_net *tn = net_generic(net, tipc_net_id);
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
unsigned long now = get_jiffies_64();
e = kzalloc(sizeof(*e), GFP_ATOMIC);
if (!e)
return;
e->dtype = type;
e->node = node;
e->expires = now + msecs_to_jiffies(sysctl_tipc_named_timeout);
memcpy(e, i, sizeof(*i));
list_add_tail(&e->next, &tn->dist_queue);
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
}
/**
* tipc_named_process_backlog - try to process any pending name table updates
* from the network.
*/
void tipc_named_process_backlog(struct net *net)
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
{
struct distr_queue_item *e, *tmp;
struct tipc_net *tn = net_generic(net, tipc_net_id);
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
char addr[16];
unsigned long now = get_jiffies_64();
list_for_each_entry_safe(e, tmp, &tn->dist_queue, next) {
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
if (time_after(e->expires, now)) {
if (!tipc_update_nametbl(net, &e->i, e->node, e->dtype))
tipc: add name distributor resiliency queue TIPC name table updates are distributed asynchronously in a cluster, entailing a risk of certain race conditions. E.g., if two nodes simultaneously issue conflicting (overlapping) publications, this may not be detected until both publications have reached a third node, in which case one of the publications will be silently dropped on that node. Hence, we end up with an inconsistent name table. In most cases this conflict is just a temporary race, e.g., one node is issuing a publication under the assumption that a previous, conflicting, publication has already been withdrawn by the other node. However, because of the (rtt related) distributed update delay, this may not yet hold true on all nodes. The symptom of this failure is a syslog message: "tipc: Cannot publish {%u,%u,%u}, overlap error". In this commit we add a resiliency queue at the receiving end of the name table distributor. When insertion of an arriving publication fails, we retain it in this queue for a short amount of time, assuming that another update will arrive very soon and clear the conflict. If so happens, we insert the publication, otherwise we drop it. The (configurable) retention value defaults to 2000 ms. Knowing from experience that the situation described above is extremely rare, there is no risk that the queue will accumulate any large number of items. Signed-off-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-28 01:08:47 -06:00
continue;
} else {
tipc_addr_string_fill(addr, e->node);
pr_warn_ratelimited("Dropping name table update (%d) of {%u, %u, %u} from %s key=%u\n",
e->dtype, ntohl(e->i.type),
ntohl(e->i.lower),
ntohl(e->i.upper),
addr, ntohl(e->i.key));
}
list_del(&e->next);
kfree(e);
}
}
/**
tipc: resolve race problem at unicast message reception TIPC handles message cardinality and sequencing at the link layer, before passing messages upwards to the destination sockets. During the upcall from link to socket no locks are held. It is therefore possible, and we see it happen occasionally, that messages arriving in different threads and delivered in sequence still bypass each other before they reach the destination socket. This must not happen, since it violates the sequentiality guarantee. We solve this by adding a new input buffer queue to the link structure. Arriving messages are added safely to the tail of that queue by the link, while the head of the queue is consumed, also safely, by the receiving socket. Sequentiality is secured per socket by only allowing buffers to be dequeued inside the socket lock. Since there may be multiple simultaneous readers of the queue, we use a 'filter' parameter to reduce the risk that they peek the same buffer from the queue, hence also reducing the risk of contention on the receiving socket locks. This solves the sequentiality problem, and seems to cause no measurable performance degradation. A nice side effect of this change is that lock handling in the functions tipc_rcv() and tipc_bcast_rcv() now becomes uniform, something that will enable future simplifications of those functions. Reviewed-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-05 06:36:41 -07:00
* tipc_named_rcv - process name table update messages sent by another node
*/
tipc: resolve race problem at unicast message reception TIPC handles message cardinality and sequencing at the link layer, before passing messages upwards to the destination sockets. During the upcall from link to socket no locks are held. It is therefore possible, and we see it happen occasionally, that messages arriving in different threads and delivered in sequence still bypass each other before they reach the destination socket. This must not happen, since it violates the sequentiality guarantee. We solve this by adding a new input buffer queue to the link structure. Arriving messages are added safely to the tail of that queue by the link, while the head of the queue is consumed, also safely, by the receiving socket. Sequentiality is secured per socket by only allowing buffers to be dequeued inside the socket lock. Since there may be multiple simultaneous readers of the queue, we use a 'filter' parameter to reduce the risk that they peek the same buffer from the queue, hence also reducing the risk of contention on the receiving socket locks. This solves the sequentiality problem, and seems to cause no measurable performance degradation. A nice side effect of this change is that lock handling in the functions tipc_rcv() and tipc_bcast_rcv() now becomes uniform, something that will enable future simplifications of those functions. Reviewed-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-05 06:36:41 -07:00
void tipc_named_rcv(struct net *net, struct sk_buff_head *inputq)
{
struct tipc_net *tn = net_generic(net, tipc_net_id);
tipc: resolve race problem at unicast message reception TIPC handles message cardinality and sequencing at the link layer, before passing messages upwards to the destination sockets. During the upcall from link to socket no locks are held. It is therefore possible, and we see it happen occasionally, that messages arriving in different threads and delivered in sequence still bypass each other before they reach the destination socket. This must not happen, since it violates the sequentiality guarantee. We solve this by adding a new input buffer queue to the link structure. Arriving messages are added safely to the tail of that queue by the link, while the head of the queue is consumed, also safely, by the receiving socket. Sequentiality is secured per socket by only allowing buffers to be dequeued inside the socket lock. Since there may be multiple simultaneous readers of the queue, we use a 'filter' parameter to reduce the risk that they peek the same buffer from the queue, hence also reducing the risk of contention on the receiving socket locks. This solves the sequentiality problem, and seems to cause no measurable performance degradation. A nice side effect of this change is that lock handling in the functions tipc_rcv() and tipc_bcast_rcv() now becomes uniform, something that will enable future simplifications of those functions. Reviewed-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-05 06:36:41 -07:00
struct tipc_msg *msg;
struct distr_item *item;
uint count;
u32 node;
struct sk_buff *skb;
int mtype;
spin_lock_bh(&tn->nametbl_lock);
tipc: resolve race problem at unicast message reception TIPC handles message cardinality and sequencing at the link layer, before passing messages upwards to the destination sockets. During the upcall from link to socket no locks are held. It is therefore possible, and we see it happen occasionally, that messages arriving in different threads and delivered in sequence still bypass each other before they reach the destination socket. This must not happen, since it violates the sequentiality guarantee. We solve this by adding a new input buffer queue to the link structure. Arriving messages are added safely to the tail of that queue by the link, while the head of the queue is consumed, also safely, by the receiving socket. Sequentiality is secured per socket by only allowing buffers to be dequeued inside the socket lock. Since there may be multiple simultaneous readers of the queue, we use a 'filter' parameter to reduce the risk that they peek the same buffer from the queue, hence also reducing the risk of contention on the receiving socket locks. This solves the sequentiality problem, and seems to cause no measurable performance degradation. A nice side effect of this change is that lock handling in the functions tipc_rcv() and tipc_bcast_rcv() now becomes uniform, something that will enable future simplifications of those functions. Reviewed-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-05 06:36:41 -07:00
for (skb = skb_dequeue(inputq); skb; skb = skb_dequeue(inputq)) {
skb_linearize(skb);
tipc: resolve race problem at unicast message reception TIPC handles message cardinality and sequencing at the link layer, before passing messages upwards to the destination sockets. During the upcall from link to socket no locks are held. It is therefore possible, and we see it happen occasionally, that messages arriving in different threads and delivered in sequence still bypass each other before they reach the destination socket. This must not happen, since it violates the sequentiality guarantee. We solve this by adding a new input buffer queue to the link structure. Arriving messages are added safely to the tail of that queue by the link, while the head of the queue is consumed, also safely, by the receiving socket. Sequentiality is secured per socket by only allowing buffers to be dequeued inside the socket lock. Since there may be multiple simultaneous readers of the queue, we use a 'filter' parameter to reduce the risk that they peek the same buffer from the queue, hence also reducing the risk of contention on the receiving socket locks. This solves the sequentiality problem, and seems to cause no measurable performance degradation. A nice side effect of this change is that lock handling in the functions tipc_rcv() and tipc_bcast_rcv() now becomes uniform, something that will enable future simplifications of those functions. Reviewed-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-05 06:36:41 -07:00
msg = buf_msg(skb);
mtype = msg_type(msg);
item = (struct distr_item *)msg_data(msg);
count = msg_data_sz(msg) / ITEM_SIZE;
node = msg_orignode(msg);
while (count--) {
if (!tipc_update_nametbl(net, item, node, mtype))
tipc_named_add_backlog(net, item, mtype, node);
tipc: resolve race problem at unicast message reception TIPC handles message cardinality and sequencing at the link layer, before passing messages upwards to the destination sockets. During the upcall from link to socket no locks are held. It is therefore possible, and we see it happen occasionally, that messages arriving in different threads and delivered in sequence still bypass each other before they reach the destination socket. This must not happen, since it violates the sequentiality guarantee. We solve this by adding a new input buffer queue to the link structure. Arriving messages are added safely to the tail of that queue by the link, while the head of the queue is consumed, also safely, by the receiving socket. Sequentiality is secured per socket by only allowing buffers to be dequeued inside the socket lock. Since there may be multiple simultaneous readers of the queue, we use a 'filter' parameter to reduce the risk that they peek the same buffer from the queue, hence also reducing the risk of contention on the receiving socket locks. This solves the sequentiality problem, and seems to cause no measurable performance degradation. A nice side effect of this change is that lock handling in the functions tipc_rcv() and tipc_bcast_rcv() now becomes uniform, something that will enable future simplifications of those functions. Reviewed-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-05 06:36:41 -07:00
item++;
}
kfree_skb(skb);
tipc_named_process_backlog(net);
}
spin_unlock_bh(&tn->nametbl_lock);
}
/**
* tipc_named_reinit - re-initialize local publications
*
* This routine is called whenever TIPC networking is enabled.
* All name table entries published by this node are updated to reflect
* the node's new network address.
*/
void tipc_named_reinit(struct net *net)
{
struct tipc_net *tn = net_generic(net, tipc_net_id);
struct publication *publ;
int scope;
spin_lock_bh(&tn->nametbl_lock);
for (scope = TIPC_ZONE_SCOPE; scope <= TIPC_NODE_SCOPE; scope++)
list_for_each_entry_rcu(publ, &tn->nametbl->publ_list[scope],
local_list)
publ->node = tn->own_addr;
spin_unlock_bh(&tn->nametbl_lock);
}