On Tue, Nov 12, 2024 at 12:32:29PM +0100, Marek Szyprowski wrote: > On 12.11.2024 11:15, Ming Lei wrote: > > On Tue, Nov 12, 2024 at 09:36:40AM +0100, Marek Szyprowski wrote: > >> On 29.10.2024 16:58, Ming Lei wrote: > >>> On Tue, Oct 29, 2024 at 12:13:35PM +0100, Marek Szyprowski wrote: > >>>> On 25.10.2024 02:37, Ming Lei wrote: > >>>>> Recently we got several deadlock report[1][2][3] caused by > >>>>> blk_mq_freeze_queue and blk_enter_queue(). > >>>>> > >>>>> Turns out the two are just like acquiring read/write lock, so model them > >>>>> as read/write lock for supporting lockdep: > >>>>> > >>>>> 1) model q->q_usage_counter as two locks(io and queue lock) > >>>>> > >>>>> - queue lock covers sync with blk_enter_queue() > >>>>> > >>>>> - io lock covers sync with bio_enter_queue() > >>>>> > >>>>> 2) make the lockdep class/key as per-queue: > >>>>> > >>>>> - different subsystem has very different lock use pattern, shared lock > >>>>> class causes false positive easily > >>>>> > >>>>> - freeze_queue degrades to no lock in case that disk state becomes DEAD > >>>>> because bio_enter_queue() won't be blocked any more > >>>>> > >>>>> - freeze_queue degrades to no lock in case that request queue becomes dying > >>>>> because blk_enter_queue() won't be blocked any more > >>>>> > >>>>> 3) model blk_mq_freeze_queue() as acquire_exclusive & try_lock > >>>>> - it is exclusive lock, so dependency with blk_enter_queue() is covered > >>>>> > >>>>> - it is trylock because blk_mq_freeze_queue() are allowed to run > >>>>> concurrently > >>>>> > >>>>> 4) model blk_enter_queue() & bio_enter_queue() as acquire_read() > >>>>> - nested blk_enter_queue() are allowed > >>>>> > >>>>> - dependency with blk_mq_freeze_queue() is covered > >>>>> > >>>>> - blk_queue_exit() is often called from other contexts(such as irq), and > >>>>> it can't be annotated as lock_release(), so simply do it in > >>>>> blk_enter_queue(), this way still covered cases as many as possible > >>>>> > >>>>> With lockdep support, such kind of reports may be reported asap and > >>>>> needn't wait until the real deadlock is triggered. > >>>>> > >>>>> For example, lockdep report can be triggered in the report[3] with this > >>>>> patch applied. > >>>>> > >>>>> [1] occasional block layer hang when setting 'echo noop > /sys/block/sda/queue/scheduler' > >>>>> https://bugzilla.kernel.org/show_bug.cgi?id=219166 > >>>>> > >>>>> [2] del_gendisk() vs blk_queue_enter() race condition > >>>>> https://lore.kernel.org/linux-block/20241003085610.GK11458@xxxxxxxxxx/ > >>>>> > >>>>> [3] queue_freeze & queue_enter deadlock in scsi > >>>>> https://lore.kernel.org/linux-block/ZxG38G9BuFdBpBHZ@fedora/T/#u > >>>>> > >>>>> Reviewed-by: Christoph Hellwig <hch@xxxxxx> > >>>>> Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx> > >>>> This patch landed yesterday in linux-next as commit f1be1788a32e > >>>> ("block: model freeze & enter queue as lock for supporting lockdep"). > >>>> In my tests I found that it introduces the following 2 lockdep warnings: > >>>> > >>>>> ... > >>>> > >>>> 2. On QEMU's ARM64 virt machine, observed during system suspend/resume > >>>> cycle: > >>>> > >>>> # time rtcwake -s10 -mmem > >>>> rtcwake: wakeup from "mem" using /dev/rtc0 at Tue Oct 29 11:54:30 2024 > >>>> PM: suspend entry (s2idle) > >>>> Filesystems sync: 0.004 seconds > >>>> Freezing user space processes > >>>> Freezing user space processes completed (elapsed 0.007 seconds) > >>>> OOM killer disabled. > >>>> Freezing remaining freezable tasks > >>>> Freezing remaining freezable tasks completed (elapsed 0.004 seconds) > >>>> > >>>> ====================================================== > >>>> WARNING: possible circular locking dependency detected > >>>> 6.12.0-rc4+ #9291 Not tainted > >>>> ------------------------------------------------------ > >>>> rtcwake/1299 is trying to acquire lock: > >>>> ffff80008358a7f8 (rtnl_mutex){+.+.}-{3:3}, at: rtnl_lock+0x1c/0x28 > >>>> > >>>> but task is already holding lock: > >>>> ffff000006136d68 (&q->q_usage_counter(io)#5){++++}-{0:0}, at: > >>>> virtblk_freeze+0x24/0x60 > >>>> > >>>> which lock already depends on the new lock. > >>>> > >>>> > >>>> the existing dependency chain (in reverse order) is: > >>> This one looks a real thing, at least the added lockdep code works as > >>> expected, also the blk_mq_freeze_queue() use in virtio-blk's ->suspend() > >>> is questionable. I will take a further look. > >> Did you find a way to fix this one? I still observe such warnings in my > >> tests, even though your lockdep fixes are already merged to -next: > >> https://lore.kernel.org/all/20241031133723.303835-1-ming.lei@xxxxxxxxxx/ > > The lockdep fixes in ->next is just for making the added lockdep work > > correctly, and virtio-blk is another story. > > > > It might be fine to annotate it with blk_mq_freeze_queue_no_owner(), > > but it looks very fragile to call freeze queue in ->suspend(), and the lock > > is just kept as being grabbed in the whole suspend code path. > > > > Can you try the following patch? > > Yes, this hides this lockdep warning, but imho it looks like a > workaround, not a final fix. > > Reported-by: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> > Tested-by: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> Thanks for the test! It is actually not workaround, because what virtblk_freeze() needs is to drain all in-flight IOs. One thing missed is to mark the queue as quiesced, and I will post one formal patch with queue quiesce covered. Thanks, Ming
