Re: [PATCH 3/5] io_uring: let fast poll support multishot

[Hao Xu <haoxu.linux@gmail.com>]

在 2022/5/7 上午1:19, Pavel Begunkov 写道:
> On 5/6/22 08:01, Hao Xu wrote:
>> From: Hao Xu <howeyxu@tencent.com>
>>
>> For operations like accept, multishot is a useful feature, since we can
>> reduce a number of accept sqe. Let's integrate it to fast poll, it may
>> be good for other operations in the future.
>>
>> Signed-off-by: Hao Xu <howeyxu@tencent.com>
>> ---
>>   fs/io_uring.c | 41 ++++++++++++++++++++++++++---------------
>>   1 file changed, 26 insertions(+), 15 deletions(-)
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 8ebb1a794e36..d33777575faf 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -5952,7 +5952,7 @@ static void io_poll_remove_entries(struct 
>> io_kiocb *req)
>>    * either spurious wakeup or multishot CQE is served. 0 when it's 
>> done with
>>    * the request, then the mask is stored in req->cqe.res.
>>    */
>> -static int io_poll_check_events(struct io_kiocb *req, bool locked)
>> +static int io_poll_check_events(struct io_kiocb *req, bool *locked)
>>   {
>>       struct io_ring_ctx *ctx = req->ctx;
>>       int v;
>> @@ -5981,17 +5981,26 @@ static int io_poll_check_events(struct 
>> io_kiocb *req, bool locked)
>>           /* multishot, just fill an CQE and proceed */
>>           if (req->cqe.res && !(req->apoll_events & EPOLLONESHOT)) {
>> -            __poll_t mask = mangle_poll(req->cqe.res & 
>> req->apoll_events);
>> -            bool filled;
>> -
>> -            spin_lock(&ctx->completion_lock);
>> -            filled = io_fill_cqe_aux(ctx, req->cqe.user_data, mask,
>> -                         IORING_CQE_F_MORE);
>> -            io_commit_cqring(ctx);
>> -            spin_unlock(&ctx->completion_lock);
>> -            if (unlikely(!filled))
>> -                return -ECANCELED;
>> -            io_cqring_ev_posted(ctx);
>> +            if (req->flags & REQ_F_APOLL_MULTISHOT) {
>> +                io_tw_lock(req->ctx, locked);
>> +                if (likely(!(req->task->flags & PF_EXITING)))
>> +                    io_queue_sqe(req);
> 
> That looks dangerous, io_queue_sqe() usually takes the request ownership
> and doesn't expect that someone, i.e. io_poll_check_events(), may still be
> actively using it.
> 
> E.g. io_accept() fails on fd < 0, return an error,
> io_queue_sqe() -> io_queue_async() -> io_req_complete_failed()
> kills it. Then io_poll_check_events() and polling in general
> carry on using the freed request => UAF. Didn't look at it
> too carefully, but there might other similar cases.
> 
I checked this when I did the coding, it seems the only case is
while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs));
uses req again after req recycled in io_queue_sqe() path like you
pointed out above, but this case should be ok since we haven't
reuse the struct req{} at that point.
In my first version, I skiped the do while{} in io_poll_check_events()
for multishot apoll and do the reap in io_req_task_submit()

static void io_apoll_task_func(struct io_kiocb *req, bool *locked) 

   { 

           int ret; 

 

           ret = io_poll_check_events(req, locked); 

           if (ret > 0) 

                   return; 

 

           __io_poll_clean(req); 

 

           if (!ret) 

                   io_req_task_submit(req, locked);   <------here 

           else 

                   io_req_complete_failed(req, ret); 

   }

But the disadvantage is in high frequent workloads case, it may loop in
io_poll_check_events for long time, then finally generating cqes in the
above io_req_task_submit() which is not good in terms of latency. 

>