Re: [PATCH v3 3/3] io_uring: don't get completion_lock in io_poll_rewait()

[Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>]

hi,

> On 10/25/21 06:38, Xiaoguang Wang wrote:
>> In current implementation, if there are not available events,
>> io_poll_rewait() just gets completion_lock, and unlocks it in
>> io_poll_task_func() or io_async_task_func(), which isn't necessary.
>>
>> Change this logic to let io_poll_task_func() or io_async_task_func()
>> get the completion_lock lock.
>>
>> Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
>> ---
>>   fs/io_uring.c | 58 
>> ++++++++++++++++++++++++++--------------------------------
>>   1 file changed, 26 insertions(+), 32 deletions(-)
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index e4c779dac953..41ff8fdafe55 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -5248,10 +5248,7 @@ static inline int __io_async_wake(struct 
>> io_kiocb *req, struct io_poll_iocb *pol
>>   }
>>     static bool io_poll_rewait(struct io_kiocb *req, struct 
>> io_poll_iocb *poll)
>> -    __acquires(&req->ctx->completion_lock)
>>   {
>> -    struct io_ring_ctx *ctx = req->ctx;
>> -
>>       /* req->task == current here, checking PF_EXITING is safe */
>>       if (unlikely(req->task->flags & PF_EXITING))
>>           WRITE_ONCE(poll->canceled, true);
>> @@ -5262,7 +5259,6 @@ static bool io_poll_rewait(struct io_kiocb 
>> *req, struct io_poll_iocb *poll)
>>           req->result = vfs_poll(req->file, &pt) & poll->events;
>>       }
>>   -    spin_lock(&ctx->completion_lock);
>
> Don't remember poll sync too well but this was synchronising with the
> final section of __io_arm_poll_handler(), and I'm afraid it may go
> completely loose with races.
Yeah, I understand your concerns, and the final section of 
__io_arm_poll_handler() sees
very complicated, it maybe need better cleanup.

After checking my patch and __io_arm_poll_handler() again, I think the 
race which maybe
introduced in my patch is that:
   1)  __io_arm_poll_handler() calls list_del_init(&poll->wait.entry) 
under completion_lock.
   2) io_poll_rewait calls add_wait_queue() without completion_lock.

But __io_arm_poll_handler() only calls list_del_init(&poll->wait.entry)  
when mask isn't zero.
and io_poll_rewait() only calls add_wait_queue when no real event 
happens(mask is zero).
So 1) and 2) should not happen at the same time, seems that there's no race.

Regards,
Xiaoguang Wang

>
>
>>       if (!req->result && !READ_ONCE(poll->canceled)) {
>>           if (req->opcode == IORING_OP_POLL_ADD)
>>               WRITE_ONCE(poll->active, true);