Re: [PATCH] io_uring: simplify the SQPOLL thread check when cancelling requests

[Bui Quang Minh <minhquangbui99@gmail.com>]

On 1/12/25 22:45, lizetao wrote:
> Hi,
> 
>> -----Original Message-----
>> From: Bui Quang Minh <minhquangbui99@gmail.com>
>> Sent: Sunday, January 12, 2025 10:34 PM
>> To: linux-kernel@vger.kernel.org
>> Cc: Bui Quang Minh <minhquangbui99@gmail.com>; Jens Axboe
>> <axboe@kernel.dk>; Pavel Begunkov <asml.silence@gmail.com>; io-
>> uring@vger.kernel.org;
>> syzbot+3c750be01dab672c513d@syzkaller.appspotmail.com; lizetao
>> <lizetao1@huawei.com>
>> Subject: [PATCH] io_uring: simplify the SQPOLL thread check when cancelling
>> requests
>>
>> In io_uring_try_cancel_requests, we check whether sq_data->thread ==
>> current to determine if the function is called by the SQPOLL thread to do iopoll
>> when IORING_SETUP_SQPOLL is set. This check can race with the SQPOLL
>> thread termination.
>>
>> io_uring_cancel_generic is used in 2 places: io_uring_cancel_generic and
>> io_ring_exit_work. In io_uring_cancel_generic, we have the information
>> whether the current is SQPOLL thread already. In io_ring_exit_work, in case
>> the SQPOLL thread reaches this path, we don't need to iopoll and leave that for
>> io_uring_cancel_generic to handle.
>>
>> So to avoid the racy check, this commit adds a boolean flag to
>> io_uring_try_cancel_requests to determine if we need to do iopoll inside the
>> function and only sets this flag in io_uring_cancel_generic when the current is
>> SQPOLL thread.
>>
>> Reported-by: syzbot+3c750be01dab672c513d@syzkaller.appspotmail.com
>> Reported-by: Li Zetao <lizetao1@huawei.com>
>> Signed-off-by: Bui Quang Minh <minhquangbui99@gmail.com>
>> ---
>>   io_uring/io_uring.c | 21 +++++++++++++++------
>>   1 file changed, 15 insertions(+), 6 deletions(-)
>>
>> diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c index
>> ff691f37462c..f28ea1254143 100644
>> --- a/io_uring/io_uring.c
>> +++ b/io_uring/io_uring.c
>> @@ -143,7 +143,8 @@ struct io_defer_entry {
>>
>>   static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
>>   					 struct io_uring_task *tctx,
>> -					 bool cancel_all);
>> +					 bool cancel_all,
>> +					 bool force_iopoll);
>>
>>   static void io_queue_sqe(struct io_kiocb *req);
>>
>> @@ -2898,7 +2899,12 @@ static __cold void io_ring_exit_work(struct
>> work_struct *work)
>>   		if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
>>   			io_move_task_work_from_local(ctx);
>>
>> -		while (io_uring_try_cancel_requests(ctx, NULL, true))
>> +		/*
>> +		 * Even if SQPOLL thread reaches this path, don't force
>> +		 * iopoll here, let the io_uring_cancel_generic handle
>> +		 * it.
> 
> Just curious, will sq_thread enter this io_ring_exit_work path?

AFAIK, yes. The SQPOLL thread is created with create_io_thread, this 
function creates a new task with CLONE_FILES. So all the open files is 
shared. There will be case that the parent closes its io_uring file and 
SQPOLL thread become the only owner of that file. So it can reach this 
path when terminating.


>> +		 */
>> +		while (io_uring_try_cancel_requests(ctx, NULL, true, false))
>>   			cond_resched();
>>
>>   		if (ctx->sq_data) {
>> @@ -3066,7 +3072,8 @@ static __cold bool io_uring_try_cancel_iowq(struct
>> io_ring_ctx *ctx)
>>
>>   static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
>>   						struct io_uring_task *tctx,
>> -						bool cancel_all)
>> +						bool cancel_all,
>> +						bool force_iopoll)
>>   {
>>   	struct io_task_cancel cancel = { .tctx = tctx, .all = cancel_all, };
>>   	enum io_wq_cancel cret;
>> @@ -3096,7 +3103,7 @@ static __cold bool
>> io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
>>
>>   	/* SQPOLL thread does its own polling */
>>   	if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
>> -	    (ctx->sq_data && ctx->sq_data->thread == current)) {
>> +	    force_iopoll) {
>>   		while (!wq_list_empty(&ctx->iopoll_list)) {
>>   			io_iopoll_try_reap_events(ctx);
>>   			ret = true;
>> @@ -3169,13 +3176,15 @@ __cold void io_uring_cancel_generic(bool
>> cancel_all, struct io_sq_data *sqd)
>>   					continue;
>>   				loop |= io_uring_try_cancel_requests(node-
>>> ctx,
>>   							current->io_uring,
>> -							cancel_all);
>> +							cancel_all,
>> +							false);
>>   			}
>>   		} else {
>>   			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
>>   				loop |= io_uring_try_cancel_requests(ctx,
>>   								     current-
>>> io_uring,
>> -								     cancel_all);
>> +								     cancel_all,
>> +								     true);
>>   		}
>>
>>   		if (loop) {
>> --
>> 2.43.0
>>
> 
> Maybe you miss something, just like Begunkov mentioned in your last version patch:
> 
>    io_uring_cancel_generic
>      WARN_ON_ONCE(sqd && sqd->thread != current);
> 
> This WARN_ON_ONCE will never be triggered, so you could remove it.

He meant that we don't need to annotate sqd->thread access in this debug 
check. The io_uring_cancel_generic function has assumption that the sgd 
is not NULL only when it's called by a SQPOLL thread. So the check means 
to ensure this assumption. A data race happens only when this function 
is called by other tasks than the SQPOLL thread, so it can race with the 
SQPOLL termination. However, the sgd is not NULL only when this function 
is called by SQPOLL thread. In normal situation following the 
io_uring_cancel_generic's assumption, the data race cannot happen. And 
in case the assumption is broken, the warning almost always is triggered 
even if data race happens. So we can ignore the race here.

Thanks,
Quang Minh.