]> git.itanic.dy.fi Git - linux-stable/commit
net/sched: Fix mirred deadlock on device recursion
authorEric Dumazet <edumazet@google.com>
Mon, 15 Apr 2024 21:07:28 +0000 (18:07 -0300)
committerJakub Kicinski <kuba@kernel.org>
Thu, 18 Apr 2024 01:22:52 +0000 (18:22 -0700)
commit0f022d32c3eca477fbf79a205243a6123ed0fe11
treefe0a49f92c3127dae86f34b3328733447c3fca9b
parent83781384a96b95e2b6403d3c8a002b2c89031770
net/sched: Fix mirred deadlock on device recursion

When the mirred action is used on a classful egress qdisc and a packet is
mirrored or redirected to self we hit a qdisc lock deadlock.
See trace below.

[..... other info removed for brevity....]
[   82.890906]
[   82.890906] ============================================
[   82.890906] WARNING: possible recursive locking detected
[   82.890906] 6.8.0-05205-g77fadd89fe2d-dirty #213 Tainted: G        W
[   82.890906] --------------------------------------------
[   82.890906] ping/418 is trying to acquire lock:
[   82.890906] ffff888006994110 (&sch->q.lock){+.-.}-{3:3}, at:
__dev_queue_xmit+0x1778/0x3550
[   82.890906]
[   82.890906] but task is already holding lock:
[   82.890906] ffff888006994110 (&sch->q.lock){+.-.}-{3:3}, at:
__dev_queue_xmit+0x1778/0x3550
[   82.890906]
[   82.890906] other info that might help us debug this:
[   82.890906]  Possible unsafe locking scenario:
[   82.890906]
[   82.890906]        CPU0
[   82.890906]        ----
[   82.890906]   lock(&sch->q.lock);
[   82.890906]   lock(&sch->q.lock);
[   82.890906]
[   82.890906]  *** DEADLOCK ***
[   82.890906]
[..... other info removed for brevity....]

Example setup (eth0->eth0) to recreate
tc qdisc add dev eth0 root handle 1: htb default 30
tc filter add dev eth0 handle 1: protocol ip prio 2 matchall \
     action mirred egress redirect dev eth0

Another example(eth0->eth1->eth0) to recreate
tc qdisc add dev eth0 root handle 1: htb default 30
tc filter add dev eth0 handle 1: protocol ip prio 2 matchall \
     action mirred egress redirect dev eth1

tc qdisc add dev eth1 root handle 1: htb default 30
tc filter add dev eth1 handle 1: protocol ip prio 2 matchall \
     action mirred egress redirect dev eth0

We fix this by adding an owner field (CPU id) to struct Qdisc set after
root qdisc is entered. When the softirq enters it a second time, if the
qdisc owner is the same CPU, the packet is dropped to break the loop.

Reported-by: Mingshuai Ren <renmingshuai@huawei.com>
Closes: https://lore.kernel.org/netdev/20240314111713.5979-1-renmingshuai@huawei.com/
Fixes: 3bcb846ca4cf ("net: get rid of spin_trylock() in net_tx_action()")
Fixes: e578d9c02587 ("net: sched: use counter to break reclassify loops")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Reviewed-by: Pedro Tammela <pctammela@mojatatu.com>
Tested-by: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://lore.kernel.org/r/20240415210728.36949-1-victor@mojatatu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
include/net/sch_generic.h
net/core/dev.c
net/sched/sch_generic.c