Re: [dm-devel] DM brokeness with NOWAIT

[Jens Axboe <axboe@kernel.dk>]

On 9/15/23 1:13 PM, Mikulas Patocka wrote:
> 
> 
> On Fri, 15 Sep 2023, Mike Snitzer wrote:
> 
>> On Fri, Sep 15 2023 at 12:14P -0400,
>> Jens Axboe <axboe@kernel.dk> wrote:
>>
>>> On 9/15/23 10:04 AM, Jens Axboe wrote:
>>>> Hi,
>>>>
>>>> Threw some db traffic into my testing mix, and that ended in tears
>>>> very quickly:
>>>>
>>>> CPU: 7 PID: 49609 Comm: ringbuf-read.t Tainted: G        W          6.6.0-rc1-g39956d2dcd81 #129
>>>> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.2-debian-1.16.2-1 04/01/2014
>>>> Call Trace:
>>>>  <TASK>
>>>>  dump_stack_lvl+0x11d/0x1b0
>>>>  __might_resched+0x3c3/0x5e0
>>>>  ? preempt_count_sub+0x150/0x150
>>>>  mempool_alloc+0x1e2/0x390
>>>>  ? sanity_check_pinned_pages+0x23/0x1010
>>>>  ? mempool_resize+0x7d0/0x7d0
>>>>  bio_alloc_bioset+0x417/0x8c0
>>>>  ? bvec_alloc+0x200/0x200
>>>>  ? __gup_device_huge+0x900/0x900
>>>>  bio_alloc_clone+0x53/0x100
>>>>  dm_submit_bio+0x27f/0x1a20
>>>>  ? lock_release+0x4b7/0x670
>>>>  ? pin_user_pages_fast+0xb6/0xf0
>>>>  ? blk_try_enter_queue+0x1a0/0x4d0
>>>>  ? dm_dax_direct_access+0x260/0x260
>>>>  ? rcu_is_watching+0x12/0xb0
>>>>  ? blk_try_enter_queue+0x1cc/0x4d0
>>>>  __submit_bio+0x239/0x310
>>>>  ? __bio_queue_enter+0x700/0x700
>>>>  ? kvm_clock_get_cycles+0x40/0x60
>>>>  ? ktime_get+0x285/0x470
>>>>  submit_bio_noacct_nocheck+0x4d9/0xb80
>>>>  ? should_fail_request+0x80/0x80
>>>>  ? preempt_count_sub+0x150/0x150
>>>>  ? folio_flags+0x6c/0x1e0
>>>>  submit_bio_noacct+0x53e/0x1b30
>>>>  blkdev_direct_IO.part.0+0x833/0x1810
>>>>  ? rcu_is_watching+0x12/0xb0
>>>>  ? lock_release+0x4b7/0x670
>>>>  ? blkdev_read_iter+0x40d/0x530
>>>>  ? reacquire_held_locks+0x4e0/0x4e0
>>>>  ? __blkdev_direct_IO_simple+0x780/0x780
>>>>  ? rcu_is_watching+0x12/0xb0
>>>>  ? __mark_inode_dirty+0x297/0xd50
>>>>  ? preempt_count_add+0x72/0x140
>>>>  blkdev_read_iter+0x2a4/0x530
>>>>  ? blkdev_write_iter+0xc40/0xc40
>>>>  io_read+0x369/0x1490
>>>>  ? rcu_is_watching+0x12/0xb0
>>>>  ? io_writev_prep_async+0x260/0x260
>>>>  ? __fget_files+0x279/0x410
>>>>  ? rcu_is_watching+0x12/0xb0
>>>>  io_issue_sqe+0x18a/0xd90
>>>>  io_submit_sqes+0x970/0x1ed0
>>>>  __do_sys_io_uring_enter+0x14d4/0x2650
>>>>  ? io_submit_sqes+0x1ed0/0x1ed0
>>>>  ? rcu_is_watching+0x12/0xb0
>>>>  ? __do_sys_io_uring_register+0x3f6/0x2190
>>>>  ? io_req_caches_free+0x500/0x500
>>>>  ? ksys_mmap_pgoff+0x85/0x5b0
>>>>  ? rcu_is_watching+0x12/0xb0
>>>>  ? trace_irq_enable.constprop.0+0xd0/0x100
>>>>  do_syscall_64+0x39/0xb0
>>>>  entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>>>
>>>> which seems to demonstrate a misunderstanding on what REQ_NOWAIT is
>>>> about. In particulary, it seems to assume you can then submit with
>>>> atomic context? DM does an rcu_read_lock() and happily proceeds to
>>>> attempt to submit IO under RCU being disabled.
>>>
>>> Did a quick check to see where this came from, and it got added with:
>>>
>>> commit 563a225c9fd207326c2a2af9d59b4097cb31ce70
>>> Author: Mike Snitzer <snitzer@kernel.org>
>>> Date:   Sat Mar 26 21:08:36 2022 -0400
>>>
>>>     dm: introduce dm_{get,put}_live_table_bio called from dm_submit_bio
>>>
>>> which conspiciously doesn't include any numbers on why this is necessary
>>> or a good thing, and notably probably wasn't tested? This landed in 5.19
>>> fwiw.
>>
>> Don't recall what I was thinking, and I clearly didn't properly test
>> either... should've consulted Mikulas.  Sorry for the trouble.
>>
>> Would you like to send a formal patch with your Signed-off-by and I'll
>> mark it for stable@ and get it to Linus?
>>
>> Mike
> 
> We could revert that commit or we could change the all the remaining 
> GFP_NOIOs in drivers/md/dm.c to "bio_opf & REQ_NOWAIT ? GFP_NOWAIT : 
> GFP_NOIO". I'm not sure which one of these possibilities is better.
> 
> Converting GFP_NOIOs would complicate dm.c a bit, but it would make sure 
> that requests with REQ_NOWAIT don't really sleep. What do you think?

I've sent out a patch for this now. Getting rid of SRCU for NOWAIT may
indeed make sense, but I think that should get introduced separately and
with actual numbers demonstrating it is a win and by how much. IMHO it
doesn't necessarily need to be a big win, the main benefit here would be
that NOWAIT is supported a lot better.

-- 
Jens Axboe

--
dm-devel mailing list
dm-devel@redhat.com
https://listman.redhat.com/mailman/listinfo/dm-devel