Re: [PATCH] io_uring: leave clean req to be done in flush overflow

[Pavel Begunkov <asml.silence@gmail.com>]

On 21/01/2021 01:37, Joseph Qi wrote:
> 
> 
> On 1/20/21 8:35 PM, Pavel Begunkov wrote:
>> On 20/01/2021 08:11, Joseph Qi wrote:
>>> Abaci reported the following BUG:
>>>
>>> [   27.629441] BUG: sleeping function called from invalid context at fs/file.c:402
>>> [   27.631317] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 1012, name: io_wqe_worker-0
>>> [   27.633220] 1 lock held by io_wqe_worker-0/1012:
>>> [   27.634286]  #0: ffff888105e26c98 (&ctx->completion_lock){....}-{2:2}, at: __io_req_complete.part.102+0x30/0x70
>>> [   27.636487] irq event stamp: 66658
>>> [   27.637302] hardirqs last  enabled at (66657): [<ffffffff8144ba02>] kmem_cache_free+0x1f2/0x3b0
>>> [   27.639211] hardirqs last disabled at (66658): [<ffffffff82003a77>] _raw_spin_lock_irqsave+0x17/0x50
>>> [   27.641196] softirqs last  enabled at (64686): [<ffffffff824003c5>] __do_softirq+0x3c5/0x5aa
>>> [   27.643062] softirqs last disabled at (64681): [<ffffffff8220108f>] asm_call_irq_on_stack+0xf/0x20
>>> [   27.645029] CPU: 1 PID: 1012 Comm: io_wqe_worker-0 Not tainted 5.11.0-rc4+ #68
>>> [   27.646651] Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
>>> [   27.649249] Call Trace:
>>> [   27.649874]  dump_stack+0xac/0xe3
>>> [   27.650666]  ___might_sleep+0x284/0x2c0
>>> [   27.651566]  put_files_struct+0xb8/0x120
>>> [   27.652481]  __io_clean_op+0x10c/0x2a0
>>> [   27.653362]  __io_cqring_fill_event+0x2c1/0x350
>>> [   27.654399]  __io_req_complete.part.102+0x41/0x70
>>> [   27.655464]  io_openat2+0x151/0x300
>>> [   27.656297]  io_issue_sqe+0x6c/0x14e0
>>> [   27.657170]  ? lock_acquire+0x31a/0x440
>>> [   27.658068]  ? io_worker_handle_work+0x24e/0x8a0
>>> [   27.659119]  ? find_held_lock+0x28/0xb0
>>> [   27.660026]  ? io_wq_submit_work+0x7f/0x240
>>> [   27.660991]  io_wq_submit_work+0x7f/0x240
>>> [   27.661915]  ? trace_hardirqs_on+0x46/0x110
>>> [   27.662890]  io_worker_handle_work+0x501/0x8a0
>>> [   27.663917]  ? io_wqe_worker+0x135/0x520
>>> [   27.664836]  io_wqe_worker+0x158/0x520
>>> [   27.665719]  ? __kthread_parkme+0x96/0xc0
>>> [   27.666663]  ? io_worker_handle_work+0x8a0/0x8a0
>>> [   27.667726]  kthread+0x134/0x180
>>> [   27.668506]  ? kthread_create_worker_on_cpu+0x90/0x90
>>> [   27.669641]  ret_from_fork+0x1f/0x30
>>>
>>> It blames we call cond_resched() with completion_lock when clean
>>> request. In fact we will do it during flush overflow and it seems we
>>> have no reason to do it before. So just remove io_clean_op() in
>>> __io_cqring_fill_event() to fix this BUG.
>>
>> Nope, it would be broken. You may override, e.g. iov pointer
>> that is dynamically allocated, and the function makes sure all
>> those are deleted and freed. Most probably there will be problems
>> on flush side as well.
>>
>> Looks like the problem is that we do spin_lock_irqsave() in
>> __io_req_complete() and then just spin_lock() for put_files_struct().
>> Jens, is it a real problem?
>>
> From the code, it is because it might sleep in close_files():

Makes sense. The diff should handle it, but let's see if
it would ever be applicable after some other bug fixes.

> 
> ...
> if (file) {
> 	filp_close(file, files);
> 	cond_resched();
> }
> 
> 
> Thanks,
> Joseph
> 
>> At least for 5.12 there is a cleanup as below, moving drop_files()
>> into io_req_clean_work/io_free_req(), which is out of locks. Depends
>> on that don't-cancel-by-files patch, but I guess can be for 5.11

-- 
Pavel Begunkov