Re: [PATCH v6 0/6] task work optimization

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

On 12/5/21 15:02, Hao Xu wrote:
> 在 2021/12/3 下午10:21, Pavel Begunkov 写道:
>> On 12/3/21 07:30, Hao Xu wrote:
>>> 在 2021/12/3 上午10:01, Pavel Begunkov 写道:
>>>> On 12/3/21 01:39, Pavel Begunkov wrote:
>>>>> On 11/26/21 10:07, Hao Xu wrote:
>>>>>> v4->v5
>>>>>> - change the implementation of merge_wq_list
>>>>>>
>> [...]
>>>> But testing with liburing tests I'm getting the stuff below,
>>>> e.g. cq-overflow hits it every time. Double checked that
>>>> I took [RESEND] version of 6/6.
>>>>
>>>> [   30.360370] BUG: scheduling while atomic: cq-overflow/2082/0x00000000
>>>> [   30.360520] Call Trace:
>>>> [   30.360523]  <TASK>
>>>> [   30.360527]  dump_stack_lvl+0x4c/0x63
>>>> [   30.360536]  dump_stack+0x10/0x12
>>>> [   30.360540]  __schedule_bug.cold+0x50/0x5e
>>>> [   30.360545]  __schedule+0x754/0x900
>>>> [   30.360551]  ? __io_cqring_overflow_flush+0xb6/0x200
>>>> [   30.360558]  schedule+0x55/0xd0
>>>> [   30.360563]  schedule_timeout+0xf8/0x140
>>>> [   30.360567]  ? prepare_to_wait_exclusive+0x58/0xa0
>>>> [   30.360573]  __x64_sys_io_uring_enter+0x69c/0x8e0
>>>> [   30.360578]  ? io_rsrc_buf_put+0x30/0x30
>>>> [   30.360582]  do_syscall_64+0x3b/0x80
>>>> [   30.360588]  entry_SYSCALL_64_after_hwframe+0x44/0xae
>>>> [   30.360592] RIP: 0033:0x7f9f9680118d
>>>> [   30.360618]  </TASK>
>>>> [   30.362295] BUG: scheduling while atomic: cq-overflow/2082/0x7ffffffe
>>>> [   30.362396] Call Trace:
>>>> [   30.362397]  <TASK>
>>>> [   30.362399]  dump_stack_lvl+0x4c/0x63
>>>> [   30.362406]  dump_stack+0x10/0x12
>>>> [   30.362409]  __schedule_bug.cold+0x50/0x5e
>>>> [   30.362413]  __schedule+0x754/0x900
>>>> [   30.362419]  schedule+0x55/0xd0
>>>> [   30.362423]  schedule_timeout+0xf8/0x140
>>>> [   30.362427]  ? prepare_to_wait_exclusive+0x58/0xa0
>>>> [   30.362431]  __x64_sys_io_uring_enter+0x69c/0x8e0
>>>> [   30.362437]  ? io_rsrc_buf_put+0x30/0x30
>>>> [   30.362440]  do_syscall_64+0x3b/0x80
>>>> [   30.362445]  entry_SYSCALL_64_after_hwframe+0x44/0xae
>>>> [   30.362449] RIP: 0033:0x7f9f9680118d
>>>> [   30.362470]  </TASK>
>>>> <repeated>
>>>>
>>> cannot repro this, all the liburing tests work well on my side..
>>
>> One problem is when on the first iteration tctx_task_work doen't
>> have anything in prior_task_list, it goes to handle_tw_list(),
>> which sets up @ctx but leaves @locked=false (say there is
>> contention). And then on the second iteration it goes to
>> handle_prior_tw_list() with non-NULL @ctx and @locked=false,
>> and tries to unlock not locked spin.
>>
>> Not sure that's the exactly the problem from traces, but at
>> least a quick hack resetting the ctx at the beginning of
>> handle_prior_tw_list() heals it.
> Good catch, thanks.
>>
>> note: apart from the quick fix the diff below includes
>> a couple of lines to force it to go through the new path.
>>
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 66d119ac4424..3868123eef87 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -2272,6 +2272,9 @@ static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
>>   static void handle_prior_tw_list(struct io_wq_work_node *node, struct io_ring_ctx **ctx,
>>                                   bool *locked)
>>   {
>> +       ctx_flush_and_put(*ctx, locked);
>> +       *ctx = NULL;
>> +
>>          do {
>>                  struct io_wq_work_node *next = node->next;
>>                  struct io_kiocb *req = container_of(node, struct io_kiocb,
>> @@ -2283,7 +2286,8 @@ static void handle_prior_tw_list(struct io_wq_work_node *node, struct io_ring_ct
>>                          ctx_flush_and_put(*ctx, locked);
>>                          *ctx = req->ctx;
>>                          /* if not contended, grab and improve batching */
>> -                       *locked = mutex_trylock(&(*ctx)->uring_lock);
>> +                       *locked = false;
>> +                       // *locked = mutex_trylock(&(*ctx)->uring_lock);
> I believe this one is your debug code which I shouldn't take, should I?

Right, just for debug, helped to catch the issue. FWIW, it doesn't seem
ctx_flush_and_put() is a good solution but was good enough to verify
my assumptions.

>>                          percpu_ref_get(&(*ctx)->refs);
>>                          if (unlikely(!*locked))
>>                                  spin_lock(&(*ctx)->completion_lock);
>> @@ -2840,7 +2844,7 @@ static void io_complete_rw(struct kiocb *kiocb, long res)
>>                  return;
>>          req->result = res;
>>          req->io_task_work.func = io_req_task_complete;
>> -       io_req_task_work_add(req, !!(req->ctx->flags & IORING_SETUP_SQPOLL));
>> +       io_req_task_work_add(req, true);
>>   }
>>
>>
>>
> 

-- 
Pavel Begunkov