David Miller a écrit :

From: Eric Dumazet <redacted>
Date: Sun, 30 Nov 2008 17:36:56 +0100

quoted

[PATCH] net: percpu_counter_inc() should not be called in BH-disabled section

I checked all per_cpu_counter_xxx() usages in network tree, and I think
all call sites are BH enabled except one in inet_csk_listen_stop().

commit dd24c00191d5e4a1ae896aafe33c6b8095ab4bd1
(net: Use a percpu_counter for orphan_count)
replaced atomic_t orphan_count to a percpu_counter.

atomic_inc()/atomic_dec() can be called from any context, while percpu_counter_xxx()
should be called from a consistent state.

For orphan_count, this context can be the BH-enabled one.

Signed-off-by: Eric Dumazet <redacted>

I applied this preemptively even though Alexey hasn't given
test feedback yet, and I also added a mention of his report
in the commit message.

Hum, this patch was necessary but wont help Alexy case that was
not related, if I undertsnad well its oops.

Its oops was about nr_files and a network percpu_counter, one
always called in BH-enabled context, one always in BD-disabled context.

We need a core change here, so that lockdep dont assume all percpu_counter
have the same class.

I know nothing about lockdep so a fix could take me some time, one
can beat me easily ;)