[PATCH] io_uring: check sqring and iopoll_list before shedule

[PATCH] io_uring: check sqring and iopoll_list before shedule

[Hao Xu]

do this to avoid race below:

         userspace                         kernel

                               |  check sqring and iopoll_list
submit sqe                     |
check IORING_SQ_NEED_WAKEUP    |
(which is not set)    |        |
                               |  set IORING_SQ_NEED_WAKEUP
wait cqe                       |  schedule(never wakeup again)

Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
---

Hi all,
I'm doing some work to reduce cpu usage in low IO pression, and I
removed timeout logic in io_sq_thread() to do some test with fio-3.26,
I found that fio hangs in getevents, inifinitely trying to get a cqe,
While sq-thread is sleeping. It seems there is race situation, and it
is still there even after I fix the issue described above in the commit
message. I doubt it is something to do with memory barrier logic
between userspace and kernel, I'm trying to address it, not many clues
for now.
I'll send the fio config and kernel modification I did for test in
following mail soon.

Thanks,
Hao

 fs/io_uring.c | 36 +++++++++++++++++++-----------------
 1 file changed, 19 insertions(+), 17 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index dff34975d86b..042f1149db51 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -6802,27 +6802,29 @@ static int io_sq_thread(void *data)
 			continue;
 		}
 
-		needs_sched = true;
 		prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
-		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
-			if ((ctx->flags & IORING_SETUP_IOPOLL) &&
-			    !list_empty_careful(&ctx->iopoll_list)) {
-				needs_sched = false;
-				break;
-			}
-			if (io_sqring_entries(ctx)) {
-				needs_sched = false;
-				break;
-			}
-		}
-
-		if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
+		if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 				io_ring_set_wakeup_flag(ctx);
 
-			mutex_unlock(&sqd->lock);
-			schedule();
-			mutex_lock(&sqd->lock);
+			needs_sched = true;
+			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
+				if ((ctx->flags & IORING_SETUP_IOPOLL) &&
+				    !list_empty_careful(&ctx->iopoll_list)) {
+					needs_sched = false;
+					break;
+				}
+				if (io_sqring_entries(ctx)) {
+					needs_sched = false;
+					break;
+				}
+			}
+
+			if (needs_sched) {
+				mutex_unlock(&sqd->lock);
+				schedule();
+				mutex_lock(&sqd->lock);
+			}
 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 				io_ring_clear_wakeup_flag(ctx);
 		}
-- 
1.8.3.1

Re: [PATCH] io_uring: check sqring and iopoll_list before shedule

[Hao Xu]

在 2021/4/21 下午11:19, Hao Xu 写道:
> do this to avoid race below:
> 
>           userspace                         kernel
> 
>                                 |  check sqring and iopoll_list
> submit sqe                     |
> check IORING_SQ_NEED_WAKEUP    |
> (which is not set)    |        |
>                                 |  set IORING_SQ_NEED_WAKEUP
> wait cqe                       |  schedule(never wakeup again)
> 
> Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
> ---
> 
> Hi all,
> I'm doing some work to reduce cpu usage in low IO pression, and I
> removed timeout logic in io_sq_thread() to do some test with fio-3.26,
> I found that fio hangs in getevents, inifinitely trying to get a cqe,
> While sq-thread is sleeping. It seems there is race situation, and it
> is still there even after I fix the issue described above in the commit
> message. I doubt it is something to do with memory barrier logic
> between userspace and kernel, I'm trying to address it, not many clues
> for now.
> I'll send the fio config and kernel modification I did for test in
> following mail soon.
> 
fio test config:
[global]
ioengine=io_uring
sqthread_poll=1
hipri=1
thread=1
bs=4k
direct=1
rw=randread
time_based=1
runtime=30
group_reporting=1
filename=/dev/nvme1n1
sqthread_poll_cpu=30

[job0]
iodepth=1

the issue mainly occur when iodepth=1 during my test.
I removed timeout logic in io_sq_thread() like this:

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 042f1149db51..dd9c95016f7f 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -6739,7 +6739,6 @@ static int io_sq_thread(void *data)
  {
         struct io_sq_data *sqd = data;
         struct io_ring_ctx *ctx;
-       unsigned long timeout = 0;
         char buf[TASK_COMM_LEN];
         DEFINE_WAIT(wait);

@@ -6777,7 +6776,6 @@ static int io_sq_thread(void *data)
                         io_run_task_work_head(&sqd->park_task_work);
                         if (did_sig)
                                 break;
-                       timeout = jiffies + sqd->sq_thread_idle;
                         continue;
                 }
                 sqt_spin = false;
@@ -6794,11 +6792,9 @@ static int io_sq_thread(void *data)
                                 sqt_spin = true;
                 }

-               if (sqt_spin || !time_after(jiffies, timeout)) {
+               if (sqt_spin) {
                         io_run_task_work();
                         cond_resched();
-                       if (sqt_spin)
-                               timeout = jiffies + sqd->sq_thread_idle;
                         continue;
                 }

@@ -6831,7 +6827,6 @@ static int io_sq_thread(void *data)

                 finish_wait(&sqd->wait, &wait);
                 io_run_task_work_head(&sqd->park_task_work);
-               timeout = jiffies + sqd->sq_thread_idle;
         }

         list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
~
~
> Thanks,
> Hao
> 
>   fs/io_uring.c | 36 +++++++++++++++++++-----------------
>   1 file changed, 19 insertions(+), 17 deletions(-)
> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index dff34975d86b..042f1149db51 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -6802,27 +6802,29 @@ static int io_sq_thread(void *data)
>   			continue;
>   		}
>   
> -		needs_sched = true;
>   		prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
> -		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
> -			if ((ctx->flags & IORING_SETUP_IOPOLL) &&
> -			    !list_empty_careful(&ctx->iopoll_list)) {
> -				needs_sched = false;
> -				break;
> -			}
> -			if (io_sqring_entries(ctx)) {
> -				needs_sched = false;
> -				break;
> -			}
> -		}
> -
> -		if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
> +		if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
>   			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
>   				io_ring_set_wakeup_flag(ctx);
>   
> -			mutex_unlock(&sqd->lock);
> -			schedule();
> -			mutex_lock(&sqd->lock);
> +			needs_sched = true;
> +			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
> +				if ((ctx->flags & IORING_SETUP_IOPOLL) &&
> +				    !list_empty_careful(&ctx->iopoll_list)) {
> +					needs_sched = false;
> +					break;
> +				}
> +				if (io_sqring_entries(ctx)) {
> +					needs_sched = false;
> +					break;
> +				}
> +			}
> +
> +			if (needs_sched) {
> +				mutex_unlock(&sqd->lock);
> +				schedule();
> +				mutex_lock(&sqd->lock);
> +			}
>   			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
>   				io_ring_clear_wakeup_flag(ctx);
>   		}
> 

Re: [PATCH] io_uring: check sqring and iopoll_list before shedule

[Pavel Begunkov]

On 4/21/21 4:19 PM, Hao Xu wrote:
> do this to avoid race below:
> 
>          userspace                         kernel
> 
>                                |  check sqring and iopoll_list
> submit sqe                     |
> check IORING_SQ_NEED_WAKEUP    |
> (which is not set)    |        |
>                                |  set IORING_SQ_NEED_WAKEUP
> wait cqe                       |  schedule(never wakeup again)

Agree, the flag should be set first.
Haven't tried it, but the patch looks reasonable

> 
> Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
> ---
> 
> Hi all,
> I'm doing some work to reduce cpu usage in low IO pression, and I
> removed timeout logic in io_sq_thread() to do some test with fio-3.26,
> I found that fio hangs in getevents, inifinitely trying to get a cqe,
> While sq-thread is sleeping. It seems there is race situation, and it
> is still there even after I fix the issue described above in the commit
> message. I doubt it is something to do with memory barrier logic
> between userspace and kernel, I'm trying to address it, not many clues
> for now.
> I'll send the fio config and kernel modification I did for test in
> following mail soon.
> 
> Thanks,
> Hao
> 
>  fs/io_uring.c | 36 +++++++++++++++++++-----------------
>  1 file changed, 19 insertions(+), 17 deletions(-)
> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index dff34975d86b..042f1149db51 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -6802,27 +6802,29 @@ static int io_sq_thread(void *data)
>  			continue;
>  		}
>  
> -		needs_sched = true;
>  		prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
> -		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
> -			if ((ctx->flags & IORING_SETUP_IOPOLL) &&
> -			    !list_empty_careful(&ctx->iopoll_list)) {
> -				needs_sched = false;
> -				break;
> -			}
> -			if (io_sqring_entries(ctx)) {
> -				needs_sched = false;
> -				break;
> -			}
> -		}
> -
> -		if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
> +		if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
>  			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
>  				io_ring_set_wakeup_flag(ctx);
>  
> -			mutex_unlock(&sqd->lock);
> -			schedule();
> -			mutex_lock(&sqd->lock);
> +			needs_sched = true;
> +			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
> +				if ((ctx->flags & IORING_SETUP_IOPOLL) &&
> +				    !list_empty_careful(&ctx->iopoll_list)) {
> +					needs_sched = false;
> +					break;
> +				}
> +				if (io_sqring_entries(ctx)) {
> +					needs_sched = false;
> +					break;
> +				}
> +			}
> +
> +			if (needs_sched) {
> +				mutex_unlock(&sqd->lock);
> +				schedule();
> +				mutex_lock(&sqd->lock);
> +			}
>  			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
>  				io_ring_clear_wakeup_flag(ctx);
>  		}
> 

-- 
Pavel Begunkov

Re: [PATCH] io_uring: check sqring and iopoll_list before shedule

[Jens Axboe]

On 4/21/21 9:19 AM, Hao Xu wrote:
> do this to avoid race below:
> 
>          userspace                         kernel
> 
>                                |  check sqring and iopoll_list
> submit sqe                     |
> check IORING_SQ_NEED_WAKEUP    |
> (which is not set)    |        |
>                                |  set IORING_SQ_NEED_WAKEUP
> wait cqe                       |  schedule(never wakeup again)

Applied, thanks.

-- 
Jens Axboe