Re: io_uring: io_fail_links() should only consider first linked timeout

[Bob Liu <bob.liu@oracle.com>]

On 11/20/19 7:07 PM, Pavel Begunkov wrote:
> On 11/20/2019 1:22 PM, Bob Liu wrote:
>> On 11/20/19 4:44 PM, Pavel Begunkov wrote:
>>> On 11/20/2019 1:33 AM, Jens Axboe wrote:
>>>> We currently clear the linked timeout field if we cancel such a timeout,
>>>> but we should only attempt to cancel if it's the first one we see.
>>>> Others should simply be freed like other requests, as they haven't
>>>> been started yet.
>>>>
>>>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>>>
>>>> ---
>>>>
>>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>>> index a79ef43367b1..d1085e4e8ae9 100644
>>>> --- a/fs/io_uring.c
>>>> +++ b/fs/io_uring.c
>>>> @@ -937,12 +937,12 @@ static void io_fail_links(struct io_kiocb *req)
>>>>  		if ((req->flags & REQ_F_LINK_TIMEOUT) &&
>>>>  		    link->submit.sqe->opcode == IORING_OP_LINK_TIMEOUT) {
>>>>  			io_link_cancel_timeout(link);
>>>> -			req->flags &= ~REQ_F_LINK_TIMEOUT;
>>>>  		} else {
>>>>  			io_cqring_fill_event(link, -ECANCELED);
>>>>  			__io_double_put_req(link);
>>>>  		}
>>>>  		kfree(sqe_to_free);
>>>> +		req->flags &= ~REQ_F_LINK_TIMEOUT;
>>>
>>> That's not necessary, but maybe would safer to keep. If
>>> REQ_F_LINK_TIMEOUT is set, than there was a link timeout request,
>>> and for it and only for it io_link_cancel_timeout() will be called.
>>>
>>> However, this is only true if linked timeout isn't fired. Otherwise,
>>> there is another bug, which isn't fixed by either of the patches. We
>>> need to clear REQ_F_LINK_TIMEOUT in io_link_timeout_fn() as well.
>>>
>>> Let: REQ -> L_TIMEOUT1 -> L_TIMEOUT2
>>> 1. L_TIMEOUT1 fired before REQ is completed
>>>
>>> 2. io_link_timeout_fn() removes L_TIMEOUT1 from the list:
>>> REQ|REQ_F_LINK_TIMEOUT -> L_TIMEOUT2
>>>
>>> 3. free_req(REQ) then call io_link_cancel_timeout(L_TIMEOUT2)
>>> leaking it (as described in my patch).
>>>
>>> P.S. haven't tried to test nor reproduce it yet.
>>>
>>
>> Off topic... I'm reading the code regarding IORING_OP_LINK_TIMEOUT.
>> But confused by what's going to happen if userspace submit a request with IORING_OP_LINK_TIMEOUT but not IOSQE_IO_LINK.
>>
> It fails in __io_submit_sqe() with -EINVAL. (see default branch in the
> switch). As for me, it's better to do it late, as it will generically
> handle dependant links (e.g. fail them properly).
> 

I see, thanks.
As for me, it may better return -EINVAL in advance so as to skip a lot unnecessary code for those reqs.

@@ -3176,6 +3176,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,

                if (!io_get_sqring(ctx, &req->submit)) {
		}
...
+               if (unlikely(req_is_invalid(req)))
+                       return -EINVAL;