lkml.org 
[lkml]   [2010]   [Dec]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH net-next-2.6] bnx2: remove cancel_work_sync() from remove_one
From
From: Tejun Heo <tj@kernel.org>
Date: Wed, 15 Dec 2010 14:52:29 +0100

> After looking through the code, I don't think this is necessarily
> correct. ->ndo_close() doesn't guarantee that the watchdog timer has
> finished running (the timer is deleted with del_timer() not
> del_timer_sync()). ie. the watchdog timer could still be running
> after ->ndo_close() and may schedule reset_task. If remove_one
> doesn't flush the task, it may still be running when remove_one() is
> called.
>
> David, am I missing something? Wouldn't it cleaner to guarantee that
> ->ndo_close() is called with the guarantee that the watchdog timer is
> not running anymore?

It would but we can't just make the change over to del_timer_sync()
otherwise we'd deadlock on netif_tx_lock().

But I think things might be OK as-is.

The timer is deleted by dev_deactivate_many() which resets the qdisc
to the no-op qdisc. Then it deletes the timer.

Any running timer will complete or see the no-op qdisc attached and
return immediately.

synchronize_rcu() is then executed which guarentees completion.

Since both the watchdog timer itself and the del_timer() call run
with netif_tx_lock() held, this makes sure the timer, once deleted,
will only see the no-op qdisc and return immediately if it is
amidst running, else it has already returned when the timer delete
completes.

So we might be OK here.


\
 
 \ /
  Last update: 2010-12-20 22:13    [W:0.048 / U:1.040 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site