Does need memory barrier to synchronize req->result with req->iopoll_completed

Does need memory barrier to synchronize req->result with req->iopoll_completed

[Xiaoguang Wang]

hi,

I have taken some further thoughts about previous IPOLL race fix patch,
if io_complete_rw_iopoll() is called in interrupt context, "req->result = res"
and "WRITE_ONCE(req->iopoll_completed, 1);" are independent store operations.
So in io_do_iopoll(), if iopoll_completed is ture, can we make sure that
req->result has already been perceived by the cpu executing io_do_iopoll()?

Regards,
Xiaoguang Wang

Re: Does need memory barrier to synchronize req->result with req->iopoll_completed

[Jens Axboe]

On 6/14/20 8:10 AM, Xiaoguang Wang wrote:
> hi,
> 
> I have taken some further thoughts about previous IPOLL race fix patch,
> if io_complete_rw_iopoll() is called in interrupt context, "req->result = res"
> and "WRITE_ONCE(req->iopoll_completed, 1);" are independent store operations.
> So in io_do_iopoll(), if iopoll_completed is ture, can we make sure that
> req->result has already been perceived by the cpu executing io_do_iopoll()?

Good point, I think if we do something like the below, we should be
totally safe against an IRQ completion. Since we batch the completions,
we can get by with just a single smp_rmb() on the completion side.

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 155f3d830ddb..74c2a4709b63 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -1736,6 +1736,9 @@ static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
 	struct req_batch rb;
 	struct io_kiocb *req;
 
+	/* order with ->result store in io_complete_rw_iopoll() */
+	smp_rmb();
+
 	rb.to_free = rb.need_iter = 0;
 	while (!list_empty(done)) {
 		int cflags = 0;
@@ -1976,6 +1979,8 @@ static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
 	if (res != req->result)
 		req_set_fail_links(req);
 	req->result = res;
+	/* order with io_poll_complete() checking ->result */
+	smp_wmb();
 	if (res != -EAGAIN)
 		WRITE_ONCE(req->iopoll_completed, 1);
 }

-- 
Jens Axboe

Re: Does need memory barrier to synchronize req->result with req->iopoll_completed

[Xiaoguang Wang]

hi,

> On 6/14/20 8:10 AM, Xiaoguang Wang wrote:
>> hi,
>>
>> I have taken some further thoughts about previous IPOLL race fix patch,
>> if io_complete_rw_iopoll() is called in interrupt context, "req->result = res"
>> and "WRITE_ONCE(req->iopoll_completed, 1);" are independent store operations.
>> So in io_do_iopoll(), if iopoll_completed is ture, can we make sure that
>> req->result has already been perceived by the cpu executing io_do_iopoll()?
> 
> Good point, I think if we do something like the below, we should be
> totally safe against an IRQ completion. Since we batch the completions,
> we can get by with just a single smp_rmb() on the completion side.
Yes, agree.
And thanks for confirming this issue, I'll make a formal patch with proper commit log.

Regards,
Xiaoguang Wang
> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index 155f3d830ddb..74c2a4709b63 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -1736,6 +1736,9 @@ static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
>   	struct req_batch rb;
>   	struct io_kiocb *req;
>   
> +	/* order with ->result store in io_complete_rw_iopoll() */
> +	smp_rmb();
> +
>   	rb.to_free = rb.need_iter = 0;
>   	while (!list_empty(done)) {
>   		int cflags = 0;
> @@ -1976,6 +1979,8 @@ static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
>   	if (res != req->result)
>   		req_set_fail_links(req);
>   	req->result = res;
> +	/* order with io_poll_complete() checking ->result */
> +	smp_wmb();
>   	if (res != -EAGAIN)
>   		WRITE_ONCE(req->iopoll_completed, 1);
>   }
> 

Re: Does need memory barrier to synchronize req->result with req->iopoll_completed

[Bijan Mottahedeh]

On 6/14/2020 8:36 AM, Jens Axboe wrote:
> On 6/14/20 8:10 AM, Xiaoguang Wang wrote:
>> hi,
>>
>> I have taken some further thoughts about previous IPOLL race fix patch,
>> if io_complete_rw_iopoll() is called in interrupt context, "req->result = res"
>> and "WRITE_ONCE(req->iopoll_completed, 1);" are independent store operations.
>> So in io_do_iopoll(), if iopoll_completed is ture, can we make sure that
>> req->result has already been perceived by the cpu executing io_do_iopoll()?
> Good point, I think if we do something like the below, we should be
> totally safe against an IRQ completion. Since we batch the completions,
> we can get by with just a single smp_rmb() on the completion side.
>
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index 155f3d830ddb..74c2a4709b63 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -1736,6 +1736,9 @@ static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
>   	struct req_batch rb;
>   	struct io_kiocb *req;
>   
> +	/* order with ->result store in io_complete_rw_iopoll() */
> +	smp_rmb();
> +
>   	rb.to_free = rb.need_iter = 0;
>   	while (!list_empty(done)) {
>   		int cflags = 0;
> @@ -1976,6 +1979,8 @@ static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
>   	if (res != req->result)
>   		req_set_fail_links(req);
>   	req->result = res;
> +	/* order with io_poll_complete() checking ->result */
> +	smp_wmb();
>   	if (res != -EAGAIN)
>   		WRITE_ONCE(req->iopoll_completed, 1);
>   }
>
I'm just trying to understand how the above smp_rmb() works. When 
io_complete_rw_iopoll() is called, all requests on the done list have 
already had ->iopoll_completed checked, and given the smp_wmb(),we know 
the two writes were ordered, so what does the smp_rmb() achieve here 
exactly? What ordering does it perform?

Thanks.

--bijan

Re: Does need memory barrier to synchronize req->result with req->iopoll_completed

[Jens Axboe]

On 6/16/20 11:31 AM, Bijan Mottahedeh wrote:
> On 6/14/2020 8:36 AM, Jens Axboe wrote:
>> On 6/14/20 8:10 AM, Xiaoguang Wang wrote:
>>> hi,
>>>
>>> I have taken some further thoughts about previous IPOLL race fix patch,
>>> if io_complete_rw_iopoll() is called in interrupt context, "req->result = res"
>>> and "WRITE_ONCE(req->iopoll_completed, 1);" are independent store operations.
>>> So in io_do_iopoll(), if iopoll_completed is ture, can we make sure that
>>> req->result has already been perceived by the cpu executing io_do_iopoll()?
>> Good point, I think if we do something like the below, we should be
>> totally safe against an IRQ completion. Since we batch the completions,
>> we can get by with just a single smp_rmb() on the completion side.
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 155f3d830ddb..74c2a4709b63 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -1736,6 +1736,9 @@ static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
>>   	struct req_batch rb;
>>   	struct io_kiocb *req;
>>   
>> +	/* order with ->result store in io_complete_rw_iopoll() */
>> +	smp_rmb();
>> +
>>   	rb.to_free = rb.need_iter = 0;
>>   	while (!list_empty(done)) {
>>   		int cflags = 0;
>> @@ -1976,6 +1979,8 @@ static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
>>   	if (res != req->result)
>>   		req_set_fail_links(req);
>>   	req->result = res;
>> +	/* order with io_poll_complete() checking ->result */
>> +	smp_wmb();
>>   	if (res != -EAGAIN)
>>   		WRITE_ONCE(req->iopoll_completed, 1);
>>   }
>>
> I'm just trying to understand how the above smp_rmb() works. When 
> io_complete_rw_iopoll() is called, all requests on the done list have 
> already had ->iopoll_completed checked, and given the smp_wmb(),we know 
> the two writes were ordered, so what does the smp_rmb() achieve here 
> exactly? What ordering does it perform?

Documentation/memory-barriers.txt actually has a good example of that,
skip to line 2219 or so.

-- 
Jens Axboe