Re: [PATCH] bdi: Fix another oops in wb_workfn()

From: Jan Kara
Date: Thu Jun 14 2018 - 06:11:14 EST


On Thu 14-06-18 06:04:04, Tetsuo Handa wrote:
> On 2018/06/14 1:45, Jan Kara wrote:
> > On Wed 13-06-18 09:25:03, Linus Torvalds wrote:
> >> On Wed, Jun 13, 2018 at 9:21 AM Tetsuo Handa
> >> <penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote:
> >>>
> >>> Since multiple addresses share bit_wait_table[256], isn't it possible that
> >>> cgwb_start_shutdown() prematurely returns false due to wake_up_bit() by
> >>> hash-conflicting addresses (i.e. not limited to clear_and_wake_up_bit() from
> >>> wb_shutdown())? I think that we cannot be sure without confirming that
> >>> test_bit(WB_shutting_down, &wb->state) == false after returning from schedule().
> >>
> >> Right.
> >>
> >> That's _always_ true, btw. Something else entirely could have woken
> >> you up. TASK_UNINTERRUPTIBLE does not mean "nothing else wakes me", it
> >> just means "_signals_ don't wake me".
> >>
> >> So every single sleep always needs to be in a loop. Always.
> >
> > Agreed and in my patch it actually is in a loop - the one iterating the
> > list of active writeback structures. If we get a false wakeup, we find the
> > same structure in the list again and wait again...
>
> Indeed. I overlooked that wb = list_first_entry() will select same wb again
> if cgwb_remove_from_bdi_list() is not yet called. Well, we could update
> "(in which case we also wait for it to finish)" part or move the body of
> cgwb_start_shutdown() to cgwb_bdi_unregister() so that it becomes clear
> that false wake-up is not a problem in this case.

I prefer to keep the wb shutdown in a separate function but I've added some
comments to explain that.

Honza

--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR