Re: [PATCH v4 1/2] io_uring: avoid whole io_wq_work copy for requests completed inline

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

hi,

>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>> index 12296ce3e8b9..2a3a02838f7b 100644
>>> --- a/fs/io_uring.c
>>> +++ b/fs/io_uring.c
>>> @@ -907,9 +907,10 @@ static void io_file_put_work(struct work_struct *work);
>>>    static inline void io_req_init_async(struct io_kiocb *req,
>>>                           void (*func)(struct io_wq_work **))
>>>    {
>>> -       if (req->flags & REQ_F_WORK_INITIALIZED)
>>> -               req->work.func = func;
>>> -       else {
>>> +       if (req->flags & REQ_F_WORK_INITIALIZED) {
>>> +               if (!req->work.func)
>>> +                       req->work.func = func;
>>> +       } else {
>>>                   req->work = (struct io_wq_work){ .func = func };
>>>                   req->flags |= REQ_F_WORK_INITIALIZED;
>>>           }
>>> @@ -2920,6 +2921,8 @@ static int __io_splice_prep(struct io_kiocb *req,
>>>                   return ret;
>>>           req->flags |= REQ_F_NEED_CLEANUP;
>>>
>>> +       /* Splice will be punted aync, so initialize io_wq_work firstly_*/
>>> +       io_req_init_async(req, io_wq_submit_work);
>>>           if (!S_ISREG(file_inode(sp->file_in)->i_mode))
>>>                   req->work.flags |= IO_WQ_WORK_UNBOUND;
>>>
>>> @@ -3592,6 +3595,9 @@ static int io_statx(struct io_kiocb *req, bool force_nonblock)
>>>
>>>    static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>>    {
>>> +        /* Close may be punted aync, so initialize io_wq_work firstly */
>>> +       io_req_init_async(req, io_wq_submit_work);
>>> +
>> For splice and close requests, these two about io_req_init_async() calls should be
>> io_req_init_async(req, NULL), because they change req->work.flags firstly.
> 
> Please no. Such assumptions/dependencies are prone to break.
> It'll get us subtle bugs in no time.
> 
> BTW, why not io_wq_submit_work in place of NULL?
In the begin of __io_splice_prep or io_close_prep, current io_uring mainline codes will
modify req->work.flags firstly, so we need to call io_req_init_async to initialize
io_wq_work before the work.flags modification.
For below codes:
static inline void io_req_init_async(struct io_kiocb *req,
                         void (*func)(struct io_wq_work **))
{
         if (req->flags & REQ_F_WORK_INITIALIZED) {
                 if (!req->work.func)
                         req->work.func = func;
         } else {
                 req->work = (struct io_wq_work){ .func = func };
                 req->flags |= REQ_F_WORK_INITIALIZED;
         }
}

if we not pass NULL to parameter 'func', e.g. pass io_wq_submit_work, then
we can not use io_req_init_async() to pass io_close_finish again.

Now I'm confused how to write better codes based on current io_uring mainline codes :)
If you have some free time, please have a deeper look, thanks.

Regards,
Xiaoguang Wang


>