lkml.org 
[lkml]   [2019]   [Jul]   [31]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] bfq: Check if bfqq is NULL in bfq_insert_request
From
Date
On 7/31/19 3:11 AM, Paolo Valente wrote:
>
>
>> Il giorno 30 lug 2019, alle ore 15:35, Guenter Roeck <linux@roeck-us.net> ha scritto:
>>
>> On 7/30/19 1:55 AM, Paolo Valente wrote:
>>> Hi Guenter,
>>> sorry for the delay (Dolomiti's fault).
>>> I didn't consider that rq->elv-icq might have been NULL also
>>> because of OOM. Thanks for spotting this issue.
>>> As for the other places where the return value of bfq_init_rq is used,
>>> unfortunately I think they matter too. Those other places are related
>>> to request merging, which is the alternative destiny of requests
>>> (instead of being just inserted). But, regardless of whether a
>>> request is to be merged or inserted, that request may be destined to a
>>> bfq_queue (possibly merged with a request already in a bfq_queue), and
>>> a NULL return value by bfq_init_rq leads to a crash. I guess you can
>>> reproduce your failure also for the merge case, by generating
>>> sequential, direct I/O with queue depth > 1, and of course by enabling
>>> failslab.
>> My assumption was that requests would only be merged if they are associated
>> with the same io context. In that case, that IO context isn't reallocated
>> with ioc_create_icq() but reused, and icq would thus never be NULL.
>> I guess that assumption was wrong.
>
> I don't remember such a filtering. I had a look again, but didn't
> find anything relevant. However, more competent people see these

Me not either, when I had a closer look yesterday. My conclusion was
that your analysis is correct.

Thanks,
Guenter

> emails. Maybe someone can give us better advice. Otherwise, to stay
> on the safe side, I'd propose to handle any possible NULL return.
>
> And I'll manage it, as per your request.
>
> Thanks,
> Paolo
>
>>
>>> If my considerations above are correct, do you want to propose a
>>> complete fix yourself?
>>
>> Sure, I'll send an updated patch.
>>
>> Thanks,
>> Guenter
>>
>>> Thanks,
>>> Paolo
>>>> Il giorno 28 lug 2019, alle ore 17:19, Guenter Roeck <linux@roeck-us.net> ha scritto:
>>>>
>>>> ping ... just in case this patch got lost in Paolo's queue.
>>>>
>>>> Guenter
>>>>
>>>> On Mon, Jul 22, 2019 at 10:30:48AM -0700, Guenter Roeck wrote:
>>>>> In bfq_insert_request(), bfqq is initialized with:
>>>>> bfqq = bfq_init_rq(rq);
>>>>> In bfq_init_rq(), we find:
>>>>> if (unlikely(!rq->elv.icq))
>>>>> return NULL;
>>>>> Indeed, rq->elv.icq can be NULL if the memory allocation in
>>>>> create_task_io_context() failed.
>>>>>
>>>>> A comment in bfq_insert_request() suggests that bfqq is supposed to be
>>>>> non-NULL if 'at_head || blk_rq_is_passthrough(rq)' is false. Yet, as
>>>>> debugging and practical experience shows, this is not the case in the
>>>>> above situation.
>>>>>
>>>>> This results in the following crash.
>>>>>
>>>>> Unable to handle kernel NULL pointer dereference
>>>>> at virtual address 00000000000001b0
>>>>> ...
>>>>> Call trace:
>>>>> bfq_setup_cooperator+0x44/0x134
>>>>> bfq_insert_requests+0x10c/0x630
>>>>> blk_mq_sched_insert_requests+0x60/0xb4
>>>>> blk_mq_flush_plug_list+0x290/0x2d4
>>>>> blk_flush_plug_list+0xe0/0x230
>>>>> blk_finish_plug+0x30/0x40
>>>>> generic_writepages+0x60/0x94
>>>>> blkdev_writepages+0x24/0x30
>>>>> do_writepages+0x74/0xac
>>>>> __filemap_fdatawrite_range+0x94/0xc8
>>>>> file_write_and_wait_range+0x44/0xa0
>>>>> blkdev_fsync+0x38/0x68
>>>>> vfs_fsync_range+0x68/0x80
>>>>> do_fsync+0x44/0x80
>>>>>
>>>>> The problem is relatively easy to reproduce by running an image with
>>>>> failslab enabled, such as with:
>>>>>
>>>>> cd /sys/kernel/debug/failslab
>>>>> echo 10 > probability
>>>>> echo 300 > times
>>>>>
>>>>> Avoid the problem by checking if bfqq is NULL before using it. With the
>>>>> NULL check in place, requests with missing io context are queued
>>>>> immediately, and the crash is no longer seen.
>>>>>
>>>>> Fixes: 18e5a57d79878 ("block, bfq: postpone rq preparation to insert or merge")
>>>>> Reported-by: Hsin-Yi Wang <hsinyi@google.com>
>>>>> Cc: Hsin-Yi Wang <hsinyi@google.com>
>>>>> Cc: Nicolas Boichat <drinkcat@chromium.org>
>>>>> Cc: Doug Anderson <dianders@chromium.org>
>>>>> Signed-off-by: Guenter Roeck <linux@roeck-us.net>
>>>>> ---
>>>>> block/bfq-iosched.c | 2 +-
>>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
>>>>> index 72860325245a..56f3f4227010 100644
>>>>> --- a/block/bfq-iosched.c
>>>>> +++ b/block/bfq-iosched.c
>>>>> @@ -5417,7 +5417,7 @@ static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
>>>>>
>>>>> spin_lock_irq(&bfqd->lock);
>>>>> bfqq = bfq_init_rq(rq);
>>>>> - if (at_head || blk_rq_is_passthrough(rq)) {
>>>>> + if (!bfqq || at_head || blk_rq_is_passthrough(rq)) {
>>>>> if (at_head)
>>>>> list_add(&rq->queuelist, &bfqd->dispatch);
>>>>> else
>>>>> --
>>>>> 2.7.4
>>>>>
>>
>
>

\
 
 \ /
  Last update: 2019-07-31 15:20    [W:0.084 / U:0.176 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site