Re: 2.4.1-pre10 deadlock (Re: ps hang in 241-pre10)

[Lorenzo Allegrucci]

At 15.40 27/01/01 -0800, you wrote:
>
>
>On Sat, 27 Jan 2001, Lorenzo Allegrucci wrote:
>> 
>> A trivial "while(1) fork()" is enough to trigger it.
>> "mem=32M" by lilo, ulimit -u is 1024.
>
>Hmm.. This does not look like a VM deadlock - it looks like some IO
>request is waiting forever on "__get_request_wait()". In fact, it looks
>like a _lot_ of people are waiting for requests.
>
>So what happens is that somebody takes a page fault (and gets the mm
>lock), tries to read something in, and never gets anything back, thus
>leaving the MM locked.
>
>Jens: this looks suspiciously like somebody isn't waking things up when
>they add requests back to the request lists. Alternatively, maybe the
>unplugging isn't properly done, so that we have a lot of pending IO that
>doesn't get started..
>
>Ho humm. Jens: imagine that you have more people waiting for requests than
>"batchcount". Further, imagine that you have multiple requests finishing
>at the same time. Not unlikely. Now, imagine that one request finishes,
>and causes "batchcount" users to wake up, and immediately another request
>finishes but THAT one doesn't wake anybody up because it notices that the
>freelist isn't empty - so it thinks that it doesn't need to wake anybody.
>
>Lorenzo, does the problem go away for you if you remove the
>
>	if (!list_empty(&q->request_freelist[rw])) {
>		...
>	}
>
>code from blkdev_release_request() in drivers/block/ll_rw_block.c?

Yes, it does.

--
Lorenzo