Re: [NET]: gen_estimator deadlock fix

[Patrick McHardy <kaber@trash.net>]

Jarek Poplawski wrote:
> This patch looks fine, but while checking for this lock I've found
> another strange thing: for actions tcfc_stats_lock is used here, which
> is equivalent to tcfc_lock; so, in gen_kill_estimator we get this lock
> sometimes after dev->queue_lock; this order is also possible during
> tc_classify if actions are used; on the other hand act_mirred calls
> dev_queue_xmit under this lock, so dev->queue_lock is taken in another
> order. I hope it's with different devs, and there is no real deadlock
> possible, but this all is a bit queer...


It *should* be a different device, but AFAIK nothing enforces this.
There are quite a few possible deadlocks with TC actions, mid-term
things like the mirred action need to be redesigned to inject packet
from a different context.

> I don't know actions enough, but it seems, if it's possible that they
> are always run only from tc_classify, with dev->queue_lock, maybe it
> would be simpler to use this lock for actions' stats with
> gen_estimator too.


The same action can be shared between devices, so they need seperate
locks.