Skip to contentSkip to conversation

io_uring/io-wq: Fix a small time window for reading work->flags

PATCH In review lizetao 2 participants 4 messages Jan 14, 2025 – Jan 15, 2025 io_uring

lore
Share & tools
lore

lizetao

author
#0 lore raw

PATCH 1/1 io_uring/io-wq: Fix a small time window for reading work->flags

There is a small time window that is modified by other tasks after reading work->flags. It is changed to read before use, which is more in line with the semantics of atoms.

Fixes 3474d1b93f89 ("io_uring/io-wq: make io_wq_work flags atomic")
---
 io_uring/io-wq.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
io_uring/io-wq.c +2 -3
index a38f36b68060..75096e77b1fe 100644
@@ -932,7 +932,6 @@ static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 {
 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
-	unsigned int work_flags = atomic_read(&work->flags);
 	struct io_cb_cancel_data match = {
 		.fn		= io_wq_work_match_item,
 		.data		= work,
@@ -945,7 +944,7 @@ void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 	 * been marked as one that should not get executed, cancel it here.
 	 */
 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
-	    (work_flags & IO_WQ_WORK_CANCEL)) {
+	    (atomic_read(&work->flags) & IO_WQ_WORK_CANCEL)) {
 		io_run_cancel(work, wq);
 		return;
 	}
@@ -959,7 +958,7 @@ void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 	do_create = !io_wq_activate_free_worker(wq, acct);
 	rcu_read_unlock();
 
-	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
+	if (do_create && ((atomic_read(&work->flags) & IO_WQ_WORK_CONCURRENT) ||
 	    !atomic_read(&acct->nr_running))) {
 		bool did_create;
 
Signature · 1 line
2.33.0
#1 lore raw
Quoted 2 lines · lizetao

There is a small time window that is modified by other tasks after reading work->flags. It is changed to read before use, which is more

Can you elaborate on what races with what? I don't immediately see any race here.

Quoted 37 lines

in line with the semantics of atoms.

Fixes 3474d1b93f89 ("io_uring/io-wq: make io_wq_work flags atomic")
---
  io_uring/io-wq.c | 5 ++---
  1 file changed, 2 insertions(+), 3 deletions(-)
io_uring/io-wq.c +2 -3
index a38f36b68060..75096e77b1fe 100644
@@ -932,7 +932,6 @@ static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
  void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
  {
  	struct io_wq_acct *acct = io_work_get_acct(wq, work);
-	unsigned int work_flags = atomic_read(&work->flags);
  	struct io_cb_cancel_data match = {
  		.fn		= io_wq_work_match_item,
  		.data		= work,
@@ -945,7 +944,7 @@ void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
  	 * been marked as one that should not get executed, cancel it here.
  	 */
  	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
-	    (work_flags & IO_WQ_WORK_CANCEL)) {
+	    (atomic_read(&work->flags) & IO_WQ_WORK_CANCEL)) {
  		io_run_cancel(work, wq);
  		return;
  	}
@@ -959,7 +958,7 @@ void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
  	do_create = !io_wq_activate_free_worker(wq, acct);
  	rcu_read_unlock();
 
-	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
+	if (do_create && ((atomic_read(&work->flags) & IO_WQ_WORK_CONCURRENT) ||
  	    !atomic_read(&acct->nr_running))) {
  		bool did_create;
 
Signature · 1 line
Pavel Begunkov
author
#2 lore raw

Hi,

Quoted 6 lines

-----Original Message-----
From: Pavel Begunkov <[email protected]>
Sent: Wednesday, January 15, 2025 12:22 AM
To: lizetao <[email protected]>; Jens Axboe <[email protected]>
Cc: [email protected]
Subject: Re: [PATCH] io_uring/io-wq: Fix a small time window for reading work-

Quoted 8 lines
Quoted 1 line

flags

Quoted 2 lines · lizetao

There is a small time window that is modified by other tasks after reading work->flags. It is changed to read before use, which is more

Can you elaborate on what races with what? I don't immediately see any race here.

There is such a race context:

	worker												process
io_worker_handle_work:										IORING_OP_ASYNC_CANCEL
	io_wq_enqueue										__io_wq_worker_cancel
		work_flags = atomic_read(&work->flags);	// no IO_WQ_WORK_CANCEL		
													atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
		if (work_flags & IO_WQ_WORK_CANCEL)	// false

There seems to be a small time window here, resulting in the latest flags not being used.

Quoted 36 lines

io_uring/io-wq.c static bool io_wq_work_match_item(struct

Quoted 13 lines

in line with the semantics of atoms.

Fixes 3474d1b93f89 ("io_uring/io-wq: make io_wq_work flags atomic")
---
  io_uring/io-wq.c | 5 ++---
  1 file changed, 2 insertions(+), 3 deletions(-)
io_uring/io-wq.c index renamed from io_uring/io-wq.c +0 -0

a38f36b68060..75096e77b1fe 100644

io_uring/io-wq.c +0 -0
@@ -932,7 +932,6 @@ static bool io_wq_work_match_item(struct

io_wq_work *work, void *data)

io_uring/io-wq.c void io_wq_enqueue(struct io_wq *wq, struct

Quoted 8 lines
io_uring/io-wq.c +0 -1
  void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
  {
  	struct io_wq_acct *acct = io_work_get_acct(wq, work);
-	unsigned int work_flags = atomic_read(&work->flags);
  	struct io_cb_cancel_data match = {
  		.fn		= io_wq_work_match_item,
  		.data		= work,
@@ -945,7 +944,7 @@ void io_wq_enqueue(struct io_wq *wq, struct

io_wq_work *work)

io_uring/io-wq.c void io_wq_enqueue(struct io_wq *wq, struct

Quoted 9 lines
io_uring/io-wq.c +1 -1
  	 * been marked as one that should not get executed, cancel it here.
  	 */
  	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
-	    (work_flags & IO_WQ_WORK_CANCEL)) {
+	    (atomic_read(&work->flags) & IO_WQ_WORK_CANCEL)) {
  		io_run_cancel(work, wq);
  		return;
  	}
@@ -959,7 +958,7 @@ void io_wq_enqueue(struct io_wq *wq, struct

io_wq_work *work)

Quoted 3 lines
io_uring/io-wq.c +0 -0
  	do_create = !io_wq_activate_free_worker(wq, acct);
  	rcu_read_unlock();
 
Quoted 9 lines

io_uring/io-wq.c void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) · line 963

Quoted 6 lines
io_uring/io-wq.c +2 -1
-	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
+	if (do_create && ((atomic_read(&work->flags) &
+IO_WQ_WORK_CONCURRENT) ||
  	    !atomic_read(&acct->nr_running))) {
  		bool did_create;

--
Pavel Begunkov

---
Li Zetao

#3 lore raw
Quoted 17 lines · lizetao

Hi,

Quoted 6 lines

-----Original Message-----
From: Pavel Begunkov <[email protected]>
Sent: Wednesday, January 15, 2025 12:22 AM
To: lizetao <[email protected]>; Jens Axboe <[email protected]>
Cc: [email protected]
Subject: Re: [PATCH] io_uring/io-wq: Fix a small time window for reading work-

Quoted 8 lines
Quoted 1 line

flags

Quoted 2 lines · lizetao

There is a small time window that is modified by other tasks after reading work->flags. It is changed to read before use, which is more

Can you elaborate on what races with what? I don't immediately see any race here.

Quoted 7 lines

There is such a race context:

	worker												process
io_worker_handle_work:										IORING_OP_ASYNC_CANCEL
	io_wq_enqueue										__io_wq_worker_cancel
		work_flags = atomic_read(&work->flags);	// no IO_WQ_WORK_CANCEL		
													atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
													^^^

That can't happen, the request is not discoverable via iowq yet.

Quoted 1 line

if (work_flags & IO_WQ_WORK_CANCEL) // false

This check is for requests that came with the flag already set.

Signature · 1 line
Pavel Begunkov