[PATCH v2] io_uring: simplify the SQPOLL thread check when cancelling requests

[PATCH v2] io_uring: simplify the SQPOLL thread check when cancelling requests

[Bui Quang Minh]

In io_uring_try_cancel_requests, we check whether sq_data->thread ==
current to determine if the function is called by the SQPOLL thread to do
iopoll when IORING_SETUP_SQPOLL is set. This check can race with the SQPOLL
thread termination.

io_uring_cancel_generic is used in 2 places: io_uring_cancel_generic and
io_ring_exit_work. In io_uring_cancel_generic, we have the information
whether the current is SQPOLL thread already. And the SQPOLL thread never
reaches io_ring_exit_work.

So to avoid the racy check, this commit adds a boolean flag to
io_uring_try_cancel_requests to determine if the caller is SQPOLL thread.

Reported-by: syzbot+3c750be01dab672c513d@syzkaller.appspotmail.com
Reported-by: Li Zetao <lizetao1@huawei.com>
Reviewed-by: Li Zetao <lizetao1@huawei.com>
Signed-off-by: Bui Quang Minh <minhquangbui99@gmail.com>
---

Changes in v2
- Update the comment, commit message, change the name of new flag

 io_uring/io_uring.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index ff691f37462c..b529d6c8d781 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -143,7 +143,8 @@ struct io_defer_entry {
 
 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
 					 struct io_uring_task *tctx,
-					 bool cancel_all);
+					 bool cancel_all,
+					 bool is_sqpoll_thread);
 
 static void io_queue_sqe(struct io_kiocb *req);
 
@@ -2898,7 +2899,8 @@ static __cold void io_ring_exit_work(struct work_struct *work)
 		if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
 			io_move_task_work_from_local(ctx);
 
-		while (io_uring_try_cancel_requests(ctx, NULL, true))
+		/* The SQPOLL thread never reaches this path */
+		while (io_uring_try_cancel_requests(ctx, NULL, true, false))
 			cond_resched();
 
 		if (ctx->sq_data) {
@@ -3066,7 +3068,8 @@ static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
 
 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
 						struct io_uring_task *tctx,
-						bool cancel_all)
+						bool cancel_all,
+						bool is_sqpoll_thread)
 {
 	struct io_task_cancel cancel = { .tctx = tctx, .all = cancel_all, };
 	enum io_wq_cancel cret;
@@ -3096,7 +3099,7 @@ static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
 
 	/* SQPOLL thread does its own polling */
 	if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
-	    (ctx->sq_data && ctx->sq_data->thread == current)) {
+	    is_sqpoll_thread) {
 		while (!wq_list_empty(&ctx->iopoll_list)) {
 			io_iopoll_try_reap_events(ctx);
 			ret = true;
@@ -3169,13 +3172,15 @@ __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
 					continue;
 				loop |= io_uring_try_cancel_requests(node->ctx,
 							current->io_uring,
-							cancel_all);
+							cancel_all,
+							false);
 			}
 		} else {
 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 				loop |= io_uring_try_cancel_requests(ctx,
 								     current->io_uring,
-								     cancel_all);
+								     cancel_all,
+								     true);
 		}
 
 		if (loop) {
-- 
2.43.0

Re: [PATCH v2] io_uring: simplify the SQPOLL thread check when cancelling requests

[Pavel Begunkov]

On 1/13/25 16:03, Bui Quang Minh wrote:
> In io_uring_try_cancel_requests, we check whether sq_data->thread ==
> current to determine if the function is called by the SQPOLL thread to do
> iopoll when IORING_SETUP_SQPOLL is set. This check can race with the SQPOLL
> thread termination.
> 
> io_uring_cancel_generic is used in 2 places: io_uring_cancel_generic and
> io_ring_exit_work. In io_uring_cancel_generic, we have the information
> whether the current is SQPOLL thread already. And the SQPOLL thread never
> reaches io_ring_exit_work.
> 
> So to avoid the racy check, this commit adds a boolean flag to
> io_uring_try_cancel_requests to determine if the caller is SQPOLL thread.

I think the comment is excessive, but let's leave it at that if
you don't want to respin it.

Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>

-- 
Pavel Begunkov

Re: [PATCH v2] io_uring: simplify the SQPOLL thread check when cancelling requests

[Jens Axboe]

On Mon, 13 Jan 2025 23:03:31 +0700, Bui Quang Minh wrote:
> In io_uring_try_cancel_requests, we check whether sq_data->thread ==
> current to determine if the function is called by the SQPOLL thread to do
> iopoll when IORING_SETUP_SQPOLL is set. This check can race with the SQPOLL
> thread termination.
> 
> io_uring_cancel_generic is used in 2 places: io_uring_cancel_generic and
> io_ring_exit_work. In io_uring_cancel_generic, we have the information
> whether the current is SQPOLL thread already. And the SQPOLL thread never
> reaches io_ring_exit_work.
> 
> [...]

Applied, thanks!

[1/1] io_uring: simplify the SQPOLL thread check when cancelling requests
      commit: a13030fd194c88961be4679f87a1380f1bda0ebe

Best regards,
-- 
Jens Axboe