Local clients are not properly synchronized on 32-bit CPUs when
updating stats (3.10+). Now it is possible estimation_timer (timer),
a stats reader, to interrupt the local client in the middle of
write_seqcount_{begin,end} sequence leading to loop (DEADLOCK).
The same interrupt can happen from received packet (SoftIRQ)
which updates the same per-CPU stats.
Fix it by disabling BH while updating stats.
Found with debug:
WARNING: inconsistent lock state
4.17.0-rc2-00105-g35cb6d7-dirty #2 Not tainted
--------------------------------
inconsistent {IN-SOFTIRQ-R} -> {SOFTIRQ-ON-W} usage.
ftp/2545 [HC0[0]:SC0[0]:HE1:SE1] takes:
86845479 (&syncp->seq#6){+.+-}, at: ip_vs_schedule+0x1c5/0x59e [ip_vs]
{IN-SOFTIRQ-R} state was registered at:
lock_acquire+0x44/0x5b
estimation_timer+0x1b3/0x341 [ip_vs]
call_timer_fn+0x54/0xcd
run_timer_softirq+0x10c/0x12b
__do_softirq+0xc1/0x1a9
do_softirq_own_stack+0x1d/0x23
irq_exit+0x4a/0x64
smp_apic_timer_interrupt+0x63/0x71
apic_timer_interrupt+0x3a/0x40
default_idle+0xa/0xc
arch_cpu_idle+0x9/0xb
default_idle_call+0x21/0x23
do_idle+0xa0/0x167
cpu_startup_entry+0x19/0x1b
start_secondary+0x133/0x182
startup_32_smp+0x164/0x168
irq event stamp: 42213
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0
----
lock(&syncp->seq#6);
<Interrupt>
lock(&syncp->seq#6);
*** DEADLOCK ***
Fixes: ac69269a45 ("ipvs: do not disable bh for long time")
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Acked-by: Simon Horman <horms@verge.net.au>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
d5e032fc56