[PATCH 5.10] io_uring: fix files grab/cancel race

[PATCH 5.10] io_uring: fix files grab/cancel race

[Pavel Begunkov]

When one task is in io_uring_cancel_files() and another is doing
io_prep_async_work() a race may happen. That's because after accounting
a request inflight in first call to io_grab_identity() it still may fail
and go to io_identity_cow(), which migh briefly keep dangling
work.identity and not only.

Grab files last, so io_prep_async_work() won't fail if it did get into
->inflight_list.

note: the bug shouldn't exist after making io_uring_cancel_files() not
poking into other tasks' requests.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 fs/io_uring.c | 31 +++++++++++++++----------------
 1 file changed, 15 insertions(+), 16 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index ff6deffe5aa9..1023f7b44cea 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -1313,22 +1313,6 @@ static bool io_grab_identity(struct io_kiocb *req)
 			return false;
 		req->work.flags |= IO_WQ_WORK_FSIZE;
 	}
-
-	if (!(req->work.flags & IO_WQ_WORK_FILES) &&
-	    (def->work_flags & IO_WQ_WORK_FILES) &&
-	    !(req->flags & REQ_F_NO_FILE_TABLE)) {
-		if (id->files != current->files ||
-		    id->nsproxy != current->nsproxy)
-			return false;
-		atomic_inc(&id->files->count);
-		get_nsproxy(id->nsproxy);
-		req->flags |= REQ_F_INFLIGHT;
-
-		spin_lock_irq(&ctx->inflight_lock);
-		list_add(&req->inflight_entry, &ctx->inflight_list);
-		spin_unlock_irq(&ctx->inflight_lock);
-		req->work.flags |= IO_WQ_WORK_FILES;
-	}
 #ifdef CONFIG_BLK_CGROUP
 	if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
 	    (def->work_flags & IO_WQ_WORK_BLKCG)) {
@@ -1370,6 +1354,21 @@ static bool io_grab_identity(struct io_kiocb *req)
 		}
 		spin_unlock(&current->fs->lock);
 	}
+	if (!(req->work.flags & IO_WQ_WORK_FILES) &&
+	    (def->work_flags & IO_WQ_WORK_FILES) &&
+	    !(req->flags & REQ_F_NO_FILE_TABLE)) {
+		if (id->files != current->files ||
+		    id->nsproxy != current->nsproxy)
+			return false;
+		atomic_inc(&id->files->count);
+		get_nsproxy(id->nsproxy);
+		req->flags |= REQ_F_INFLIGHT;
+
+		spin_lock_irq(&ctx->inflight_lock);
+		list_add(&req->inflight_entry, &ctx->inflight_list);
+		spin_unlock_irq(&ctx->inflight_lock);
+		req->work.flags |= IO_WQ_WORK_FILES;
+	}
 
 	return true;
 }
-- 
2.24.0

Re: [PATCH 5.10] io_uring: fix files grab/cancel race

[Jens Axboe]

On 11/25/20 11:41 AM, Pavel Begunkov wrote:
> When one task is in io_uring_cancel_files() and another is doing
> io_prep_async_work() a race may happen. That's because after accounting
> a request inflight in first call to io_grab_identity() it still may fail
> and go to io_identity_cow(), which migh briefly keep dangling
> work.identity and not only.
> 
> Grab files last, so io_prep_async_work() won't fail if it did get into
> ->inflight_list.
> 
> note: the bug shouldn't exist after making io_uring_cancel_files() not
> poking into other tasks' requests.

Applied, thanks.

-- 
Jens Axboe