Re: [mmotm][BUG] lockdep warning block I/O (Was Re: mmotm2009-08-27-16-51 uploaded

[Andrew Morton]

On Tue, 1 Sep 2009 18:07:17 +0900
KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:

> 
> Here is mmont-Aug27's lockdep wanring. This was printed out when oom-kill happens.
> I'm sorry if already fixed.

My life's project is to hunt down the guy who invented mail client
wordwrapping, set him on fire then dance on his ashes.

> =
> Sep  1 18:01:16 localhost kernel: [ 3012.503035] ======================================================
> Sep  1 18:01:16 localhost kernel: [ 3012.503039] [ INFO: SOFTIRQ-safe -> SOFTIRQ-unsafe lock order detected ]
> Sep  1 18:01:16 localhost kernel: [ 3012.503042] 2.6.31-rc7-mm1 #3
> Sep  1 18:01:16 localhost kernel: [ 3012.503049] ------------------------------------------------------
> Sep  1 18:01:16 localhost kernel: [ 3012.503052] kblockd/7/350 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire:
> Sep  1 18:01:16 localhost kernel: [ 3012.503058]  (bdev_lock){+.+...}, at: [<ffffffff811458c7>] nr_blockdev_pages+0x1
> 7/0x80
> Sep  1 18:01:16 localhost kernel: [ 3012.503069]
> Sep  1 18:01:16 localhost kernel: [ 3012.503070] and this task is already holding:
> Sep  1 18:01:16 localhost kernel: [ 3012.503075]  (&q->__queue_lock){..-.-.}, at: [<ffffffff811e9ff8>] cfq_kick_queue
> +0x28/0x50
> Sep  1 18:01:16 localhost kernel: [ 3012.503083] which would create a new lock dependency:
> Sep  1 18:01:16 localhost kernel: [ 3012.503087]  (&q->__queue_lock){..-.-.} -> (bdev_lock){+.+...}
> Sep  1 18:01:16 localhost kernel: [ 3012.503100]

I'd say the core problem here is that __alloc_pages_slowpath() is
calling show_mem().  Because show_mem() is a "high level" function which
takes "high level" locks.  ie: bdev_lock.

It's inappropriate that alloc_pages() is assuming that it is safe to
call show_mem() from all contexts in which alloc_pages() might be
called.

That show_mem() call has been there since 2005, so I don't know what
caused this to be revealed now.

It's not at all a serious bug and the chances of us deadlocking the
kernel here are close to zero.  An appropriate fix would be to replace
that show_mem() call with something which can be safely called from all
contexts in which the page allocator can be called.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/