Skip to contentSkip to conversation

io_uring/fdinfo: remove need for sqpoll lock for thread/pid retrieval

PATCH Superseded Jens Axboe 2 participants 3 messages Nov 14, 2023 – Nov 15, 2023 io_uring

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Jens Axboe

author maintainer
#0 lore raw

PATCH v2 1/1 io_uring/fdinfo: remove need for sqpoll lock for thread/pid retrieval

A previous commit added a trylock for getting the SQPOLL thread info via fdinfo, but this introduced a regression where we often fail to get it if the thread is busy. For that case, we end up not printing the current CPU and PID info.

Rather than rely on this lock, just print the pid we already stored in the io_sq_data struct, and ensure we update the current CPU every time we are going to sleep. The latter won't potentially be 100% accurate, but that wasn't the case before either as the task can get migrated at any time unless it has been pinned at creation time.

We retain keeping the io_sq_data dereference inside the ctx->uring_lock, as it has always been, as destruction of the thread and data happen below that. We could make this RCU safe, but there's little point in doing that.

With this, we always print the last valid information we had, rather than have spurious outputs with missing information.

Fixes 7644b1a1c9a7 ("io_uring/fdinfo: lock SQ thread while retrieving thread cpu/pid")

---

v2: actually remember to use the cached values... also update ->sq_cpu
when we initially set it up, if it's not pinned to a given CPU.

io_uring/fdinfo.c +2 -7
index f04a43044d91..976e9500f651 100644
@@ -145,13 +145,8 @@ __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
 	if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
 		struct io_sq_data *sq = ctx->sq_data;
 
-		if (mutex_trylock(&sq->lock)) {
-			if (sq->thread) {
-				sq_pid = task_pid_nr(sq->thread);
-				sq_cpu = task_cpu(sq->thread);
-			}
-			mutex_unlock(&sq->lock);
-		}
+		sq_pid = sq->task_pid;
+		sq_cpu = sq->sq_cpu;
 	}
 
 	seq_printf(m, "SqThread:\t%d\n", sq_pid);
io_uring/sqpoll.c +5 -2
index bd6c2c7959a5..ecb00322a4e5 100644
@@ -229,10 +229,12 @@ static int io_sq_thread(void *data)
 	snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
 	set_task_comm(current, buf);
 
-	if (sqd->sq_cpu != -1)
+	if (sqd->sq_cpu != -1) {
 		set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
-	else
+	} else {
 		set_cpus_allowed_ptr(current, cpu_online_mask);
+		sqd->sq_cpu = task_cpu(current);
+	}
 
 	mutex_lock(&sqd->lock);
 	while (1) {
@@ -291,6 +293,7 @@ static int io_sq_thread(void *data)
 			}
 
 			if (needs_sched) {
+				sqd->sq_cpu = task_cpu(current);
 				mutex_unlock(&sqd->lock);
 				schedule();
 				mutex_lock(&sqd->lock);
Signature · 1 line
Jens Axboe
#1 lore raw
Quoted 66 lines · Jens Axboe

A previous commit added a trylock for getting the SQPOLL thread info via fdinfo, but this introduced a regression where we often fail to get it if the thread is busy. For that case, we end up not printing the current CPU and PID info.

Rather than rely on this lock, just print the pid we already stored in the io_sq_data struct, and ensure we update the current CPU every time we are going to sleep. The latter won't potentially be 100% accurate, but that wasn't the case before either as the task can get migrated at any time unless it has been pinned at creation time.

We retain keeping the io_sq_data dereference inside the ctx->uring_lock, as it has always been, as destruction of the thread and data happen below that. We could make this RCU safe, but there's little point in doing that.

With this, we always print the last valid information we had, rather than have spurious outputs with missing information.

Fixes 7644b1a1c9a7 ("io_uring/fdinfo: lock SQ thread while retrieving thread cpu/pid")

---

v2: actually remember to use the cached values... also update ->sq_cpu
when we initially set it up, if it's not pinned to a given CPU.

io_uring/fdinfo.c +2 -7
index f04a43044d91..976e9500f651 100644
@@ -145,13 +145,8 @@ __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
 	if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
 		struct io_sq_data *sq = ctx->sq_data;
 
-		if (mutex_trylock(&sq->lock)) {
-			if (sq->thread) {
-				sq_pid = task_pid_nr(sq->thread);
-				sq_cpu = task_cpu(sq->thread);
-			}
-			mutex_unlock(&sq->lock);
-		}
+		sq_pid = sq->task_pid;
+		sq_cpu = sq->sq_cpu;
 	}
 
 	seq_printf(m, "SqThread:\t%d\n", sq_pid);
io_uring/sqpoll.c +4 -2
index bd6c2c7959a5..ecb00322a4e5 100644
@@ -229,10 +229,12 @@ static int io_sq_thread(void *data)
 	snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
 	set_task_comm(current, buf);
 
-	if (sqd->sq_cpu != -1)
+	if (sqd->sq_cpu != -1) {
 		set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
-	else
+	} else {
 		set_cpus_allowed_ptr(current, cpu_online_mask);
+		sqd->sq_cpu = task_cpu(current);
+	}
 
 	mutex_lock(&sqd->lock);
 	while (1) {

io_uring/sqpoll.c static int io_sq_thread(void *data) · line 296

Quoted 4 lines
io_uring/sqpoll.c +1 -0
@@ -291,6 +293,7 @@ static int io_sq_thread(void *data)
 			}
 
 			if (needs_sched) {
+				sqd->sq_cpu = task_cpu(current);

Don't you also need to update sqd->sq_cpu in io_sqd_handle_event before releasing the lock? sqpoll might get migrated after the following schedule and then parked, in which case sqd->sq_cpu will not be up to date for a while.

Other than that, I think the patch is fine.

Quoted 3 lines
io_uring/sqpoll.c +0 -0
 				mutex_unlock(&sqd->lock);
 				schedule();
 				mutex_lock(&sqd->lock);
Signature · 1 line
Gabriel Krisman Bertazi
author maintainer
#2 lore raw
Quoted 78 lines · Gabriel Krisman Bertazi
Quoted 66 lines · Jens Axboe

A previous commit added a trylock for getting the SQPOLL thread info via fdinfo, but this introduced a regression where we often fail to get it if the thread is busy. For that case, we end up not printing the current CPU and PID info.

Rather than rely on this lock, just print the pid we already stored in the io_sq_data struct, and ensure we update the current CPU every time we are going to sleep. The latter won't potentially be 100% accurate, but that wasn't the case before either as the task can get migrated at any time unless it has been pinned at creation time.

We retain keeping the io_sq_data dereference inside the ctx->uring_lock, as it has always been, as destruction of the thread and data happen below that. We could make this RCU safe, but there's little point in doing that.

With this, we always print the last valid information we had, rather than have spurious outputs with missing information.

Fixes 7644b1a1c9a7 ("io_uring/fdinfo: lock SQ thread while retrieving thread cpu/pid")

---

v2: actually remember to use the cached values... also update ->sq_cpu
when we initially set it up, if it's not pinned to a given CPU.

io_uring/fdinfo.c +2 -7
index f04a43044d91..976e9500f651 100644
@@ -145,13 +145,8 @@ __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
 	if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
 		struct io_sq_data *sq = ctx->sq_data;
 
-		if (mutex_trylock(&sq->lock)) {
-			if (sq->thread) {
-				sq_pid = task_pid_nr(sq->thread);
-				sq_cpu = task_cpu(sq->thread);
-			}
-			mutex_unlock(&sq->lock);
-		}
+		sq_pid = sq->task_pid;
+		sq_cpu = sq->sq_cpu;
 	}
 
 	seq_printf(m, "SqThread:\t%d\n", sq_pid);
io_uring/sqpoll.c +4 -2
index bd6c2c7959a5..ecb00322a4e5 100644
@@ -229,10 +229,12 @@ static int io_sq_thread(void *data)
 	snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
 	set_task_comm(current, buf);
 
-	if (sqd->sq_cpu != -1)
+	if (sqd->sq_cpu != -1) {
 		set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
-	else
+	} else {
 		set_cpus_allowed_ptr(current, cpu_online_mask);
+		sqd->sq_cpu = task_cpu(current);
+	}
 
 	mutex_lock(&sqd->lock);
 	while (1) {

io_uring/sqpoll.c static int io_sq_thread(void *data) · line 296

Quoted 4 lines
io_uring/sqpoll.c +1 -0
@@ -291,6 +293,7 @@ static int io_sq_thread(void *data)
 			}
 
 			if (needs_sched) {
+				sqd->sq_cpu = task_cpu(current);

Don't you also need to update sqd->sq_cpu in io_sqd_handle_event before releasing the lock? sqpoll might get migrated after the following schedule and then parked, in which case sqd->sq_cpu will not be up to date for a while.

Quoted 1 line

Other than that, I think the patch is fine.

We probably should, and I also forgot that we can just use raw_smp_processor_id() at this point. I'll send out a v3.

Signature · 1 line
Jens Axboe