Re: [PATCH] io_uring/sqpoll: annotate data race for access in debug check

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

On 1/8/25 23:24, lizetao wrote:
> Hi,
> 
>> -----Original Message-----
>> From: Bui Quang Minh <minhquangbui99@gmail.com>
>> Sent: Wednesday, January 8, 2025 11:11 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+5988142e8a69a67b1418@syzkaller.appspotmail.com
>> Subject: [PATCH] io_uring/sqpoll: annotate data race for access in debug check
>>
>> sqd->thread must only be access while holding sqd->lock. In
>> io_sq_thread_stop, the sqd->thread access to wake up the sq thread is placed
>> while holding sqd->lock, but the access in debug check is not. As this access if
>> for debug check only, we can safely ignore the data race here. So we annotate
>> this access with data_race to silence KCSAN.
>>
>> Reported-by: syzbot+5988142e8a69a67b1418@syzkaller.appspotmail.com
>> Signed-off-by: Bui Quang Minh <minhquangbui99@gmail.com>
>> ---
>>   io_uring/sqpoll.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/io_uring/sqpoll.c b/io_uring/sqpoll.c index
>> 9e5bd79fd2b5..2088c56dbaa0 100644
>> --- a/io_uring/sqpoll.c
>> +++ b/io_uring/sqpoll.c
>> @@ -57,7 +57,7 @@ void io_sq_thread_park(struct io_sq_data *sqd)
>>
>>   void io_sq_thread_stop(struct io_sq_data *sqd)  {
>> -	WARN_ON_ONCE(sqd->thread == current);
>> +	WARN_ON_ONCE(data_race(sqd->thread) == current);
>>   	WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd-
>>> state));
>>
>>   	set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
>> --
>> 2.43.0
>>
> 
> The modification of this patch itself is fine, but there are two other things I need to confirm.
> 1、Does the io_uring_cancel_generic() require the same modification?

I think yes, there is another syzbot's bug report on data race on the 
io_uring_cancel_generic I'm currently looking at. Here is the link: 
https://syzkaller.appspot.com/bug?extid=3c750be01dab672c513d

> 2、It is not holding sqd->lock in io_req_normal_work_add(), is it safe?

This is a valid point, I think we should add lock here too. I will try 
to write a proof-of-concept to validate this.

Thanks,
Quang Minh.