Re: [PATCH] bfq: silence lockdep for bfqd/ioc lock inversion

[Paolo Valente <paolo.valente@linaro.org>]

Hi Jan, Khazhy,
sorry for my super delay.  Thanks Khazy for spotting this, and Jan for
proposing alternative solutions.

> Il giorno 15 apr 2021, alle ore 12:47, Jan Kara <jack@suse.cz> ha scritto:
> 
> On Wed 14-04-21 11:33:14, Khazhy Kumykov wrote:
>> On Wed, Apr 14, 2021 at 2:54 AM Jan Kara <jack@suse.cz> wrote:
>>> 
>>> On Thu 18-03-21 23:00:15, Khazhismel Kumykov wrote:
>>>> lockdep warns of circular locking due to inversion between
>>>> bfq_insert_requests and bfq_exit_icq. If we end freeing a request when
>>>> merging, we *may* grab an ioc->lock if that request is the last refcount
>>>> to that ioc. bfq_bio_merge also potentially could have this ordering.
>>>> bfq_exit_icq, conversely, grabs bfqd but is always called with ioc->lock
>>>> held.
>>>> 
>>>> bfq_exit_icq may either be called from put_io_context_active with ioc
>>>> refcount raised, ioc_release_fn after the last refcount was already
>>>> dropped, or ioc_clear_queue, which is only called while queue is
>>>> quiesced or exiting, so the inverted orderings should never conflict.
>>>> 
>>>> Fixes: aee69d78dec0 ("block, bfq: introduce the BFQ-v0 I/O scheduler as
>>>> an extra scheduler")
>>>> 
>>>> Signed-off-by: Khazhismel Kumykov <khazhy@google.com>
>>> 
>>> I've just hit the same lockdep complaint. When looking at this another
>>> option to solve this complaint seemed to be to modify bfq_bio_merge() like:
>>> 
>>>        ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
>>> 
>>> +       spin_unlock_irq(&bfqd->lock);
>>>        if (free)
>>>                blk_mq_free_request(free);
>>> -       spin_unlock_irq(&bfqd->lock);
>>> 
>>>        return ret;
>>> 
>>> to release request outside of bfqd->lock. Because AFAICT there's no good
>>> reason why we are actually freeing the request under bfqd->lock. And it
>>> would seem a bit safer than annotating-away the lockdep complaint (as much
>>> as I don't see a problem with your analysis). Paolo?
>> 
>> If we can re-order the locking so we don't need the annotation, that
>> seems better ("inversion is OK so long as either we're frozen or we
>> have ioc refcount, and we only grab ioc->lock normally if we drop the
>> last refcount" is a tad "clever"). Though we still need to deal with
>> blk_mq_sched_try_insert_merge which can potentially free a request.
> 
> I see, right.
> 

Trying to put pieces together:
1) Moving ahead the invocation of spin_unlock_irq(&bfqd->lock) in
bfq_bio_merge(), as suggested by Jan, seems ok to me as well.  I also
proposed this change several years ago, but received no feedback.  So
I followed the conservative approach of not touching what apparently
works :)
2) If I'm not missing anything, then also what Jan suggests below is
ok.  That is, in blk_mq_sched_try_insert_merge, ioc->lock gets grabbed
only in case blk_mq_free_request is invoked.  So it is ok to simply move
the invocation of blk_mq_free_request outside
blk_mq_sched_try_insert_merge, using the same approach as with
blk_mq_sched_try_merge in bfq_bio_merge.

Thanks,
Paolo

>> (See the first stacktrace). Something simple that I wasn't sure of is:
>> could we delay bfq_exit_icq work, then avoid the inversion? Simpler to
>> analyze then.
> 
> That's problematic because ICQ (referencing BFQQs etc.) is going to be
> freed after RCU grace period expires. So we cannot really postpone the
> teardown of bfq_io_cq. What we could do is to modify
> blk_mq_sched_try_insert_merge() so that it returns request to free
> similarly to blk_mq_sched_try_merge().  Then we can free the request after
> dropping bfqd->lock.
> 
> 								Honza
> 
>>>> ---
>>>> block/bfq-iosched.c | 9 ++++++++-
>>>> 1 file changed, 8 insertions(+), 1 deletion(-)
>>>> 
>>>> Noticed this lockdep running xfstests (generic/464) on top of a bfq
>>>> block device. I was also able to tease it out w/ binary trying to issue
>>>> requests that would end up merging while rapidly swapping the active
>>>> scheduler. As far as I could see, the deadlock would not actually occur,
>>>> so this patch opts to change lock class for the inverted case.
>>>> 
>>>> bfqd -> ioc :
>>>> [ 2995.524557] __lock_acquire+0x18f5/0x2660
>>>> [ 2995.524562] lock_acquire+0xb4/0x3a0
>>>> [ 2995.524565] _raw_spin_lock_irqsave+0x3f/0x60
>>>> [ 2995.524569] put_io_context+0x33/0x90.  -> ioc->lock grabbed
>>>> [ 2995.524573] blk_mq_free_request+0x51/0x140
>>>> [ 2995.524577] blk_put_request+0xe/0x10
>>>> [ 2995.524580] blk_attempt_req_merge+0x1d/0x30
>>>> [ 2995.524585] elv_attempt_insert_merge+0x56/0xa0
>>>> [ 2995.524590] blk_mq_sched_try_insert_merge+0x4b/0x60
>>>> [ 2995.524595] bfq_insert_requests+0x9e/0x18c0.    -> bfqd->lock grabbed
>>>> [ 2995.524598] blk_mq_sched_insert_requests+0xd6/0x2b0
>>>> [ 2995.524602] blk_mq_flush_plug_list+0x154/0x280
>>>> [ 2995.524606] blk_finish_plug+0x40/0x60
>>>> [ 2995.524609] ext4_writepages+0x696/0x1320
>>>> [ 2995.524614] do_writepages+0x1c/0x80
>>>> [ 2995.524621] __filemap_fdatawrite_range+0xd7/0x120
>>>> [ 2995.524625] sync_file_range+0xac/0xf0
>>>> [ 2995.524642] __x64_sys_sync_file_range+0x44/0x70
>>>> [ 2995.524646] do_syscall_64+0x31/0x40
>>>> [ 2995.524649] entry_SYSCALL_64_after_hwframe+0x44/0xae
>>>> 
>>>> ioc -> bfqd
>>>> [ 2995.524490] _raw_spin_lock_irqsave+0x3f/0x60
>>>> [ 2995.524498] bfq_exit_icq+0xa3/0xe0 -> bfqd->lock grabbed
>>>> [ 2995.524512] put_io_context_active+0x78/0xb0 -> ioc->lock grabbed
>>>> [ 2995.524516] exit_io_context+0x48/0x50
>>>> [ 2995.524519] do_exit+0x7e9/0xdd0
>>>> [ 2995.524526] do_group_exit+0x54/0xc0
>>>> [ 2995.524530] __x64_sys_exit_group+0x18/0x20
>>>> [ 2995.524534] do_syscall_64+0x31/0x40
>>>> [ 2995.524537] entry_SYSCALL_64_after_hwframe+0x44/0xae
>>>> 
>>>> Another trace where we grab ioc -> bfqd through bfq_exit_icq is when
>>>> changing elevator
>>>>               -> #1 (&(&bfqd->lock)->rlock){-.-.}:
>>>> [  646.890820]        lock_acquire+0x9b/0x140
>>>> [  646.894868]        _raw_spin_lock_irqsave+0x3b/0x50
>>>> [  646.899707]        bfq_exit_icq_bfqq+0x47/0x1f0
>>>> [  646.904196]        bfq_exit_icq+0x21/0x30
>>>> [  646.908160]        ioc_destroy_icq+0xf3/0x130
>>>> [  646.912466]        ioc_clear_queue+0xb8/0x140
>>>> [  646.916771]        elevator_switch_mq+0xa4/0x3c0
>>>> [  646.921333]        elevator_switch+0x5f/0x340
>>>> 
>>>> diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
>>>> index 95586137194e..cb50ac0ffe80 100644
>>>> --- a/block/bfq-iosched.c
>>>> +++ b/block/bfq-iosched.c
>>>> @@ -5027,7 +5027,14 @@ static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
>>>>      if (bfqq && bfqd) {
>>>>              unsigned long flags;
>>>> 
>>>> -             spin_lock_irqsave(&bfqd->lock, flags);
>>>> +             /* bfq_exit_icq is usually called with ioc->lock held, which is
>>>> +              * inverse order from elsewhere, which may grab ioc->lock
>>>> +              * under bfqd->lock if we merge requests and drop the last ioc
>>>> +              * refcount. Since exit_icq is either called with a refcount,
>>>> +              * or with queue quiesced, use a differnet lock class to
>>>> +              * silence lockdep
>>>> +              */
>>>> +             spin_lock_irqsave_nested(&bfqd->lock, flags, 1);
>>>>              bfqq->bic = NULL;
>>>>              bfq_exit_bfqq(bfqd, bfqq);
>>>>              bic_set_bfqq(bic, NULL, is_sync);
>>>> --
>>>> 2.31.0.rc2.261.g7f71774620-goog
>>>> 
>>> --
>>> Jan Kara <jack@suse.com>
>>> SUSE Labs, CR
> 
> 
> -- 
> Jan Kara <jack@suse.com>
> SUSE Labs, CR