Re: [bug report] io_uring/poll: get rid of unlocked cancel hash

[Jens Axboe <axboe@kernel.dk>]

On 10/4/24 8:54 AM, Dan Carpenter wrote:
> On Fri, Oct 04, 2024 at 07:54:32AM -0600, Jens Axboe wrote:
>> On 10/4/24 7:50 AM, Jens Axboe wrote:
>>> On 10/4/24 3:00 AM, Dan Carpenter wrote:
>>>> Hello Jens Axboe,
>>>>
>>>> Commit 313314db5bcb ("io_uring/poll: get rid of unlocked cancel
>>>> hash") from Sep 30, 2024 (linux-next), leads to the following Smatch
>>>> static checker warning:
>>>>
>>>> 	io_uring/poll.c:932 io_poll_remove()
>>>> 	warn: duplicate check 'ret2' (previous on line 930)
>>>>
>>>> io_uring/poll.c
>>>>     919 int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
>>>>     920 {
>>>>     921         struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
>>>>     922         struct io_ring_ctx *ctx = req->ctx;
>>>>     923         struct io_cancel_data cd = { .ctx = ctx, .data = poll_update->old_user_data, };
>>>>     924         struct io_kiocb *preq;
>>>>     925         int ret2, ret = 0;
>>>>     926 
>>>>     927         io_ring_submit_lock(ctx, issue_flags);
>>>>     928         preq = io_poll_find(ctx, true, &cd);
>>>>     929         ret2 = io_poll_disarm(preq);
>>>>     930         if (!ret2)
>>>>     931                 goto found;
>>>> --> 932         if (ret2) {
>>>>     933                 ret = ret2;
>>>>     934                 goto out;
>>>>     935         }
>>>>
>>>> A lot of the function is dead code now.  ;)
>>>
>>> Thanks, will revisit and fold in a fix!
>>
>> Should just need this incremental. There's no dead code as far as I can
>> see, just a needless found label and jump.
>>
>> diff --git a/io_uring/poll.c b/io_uring/poll.c
>> index 69382da48c00..217d667e0622 100644
>> --- a/io_uring/poll.c
>> +++ b/io_uring/poll.c
>> @@ -940,13 +940,10 @@ int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
>>  	ret2 = io_poll_disarm(preq);
>>  	if (bucket)
>>  		spin_unlock(&bucket->lock);
>> -	if (!ret2)
>> -		goto found;
> 
> Oh.  I thought this was a goto out.  That explains how the code was passing
> tests.  That was an easy fix.

Yeah, that check was just dead code, but the rest was fine. Thanks for
letting me know!

-- 
Jens Axboe