kernel/io_uring: cancel io_uring before task works
authorPavel Begunkov <asml.silence@gmail.com>
Wed, 30 Dec 2020 21:34:16 +0000 (21:34 +0000)
committerJens Axboe <axboe@kernel.dk>
Thu, 31 Dec 2020 02:36:54 +0000 (19:36 -0700)
For cancelling io_uring requests it needs either to be able to run
currently enqueued task_works or having it shut down by that moment.
Otherwise io_uring_cancel_files() may be waiting for requests that won't
ever complete.

Go with the first way and do cancellations before setting PF_EXITING and
so before putting the task_work infrastructure into a transition state
where task_work_run() would better not be called.

Cc: stable@vger.kernel.org # 5.5+
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
fs/file.c
kernel/exit.c

index c0b60961c67222fe95e60641313192f494dcde48..dab120b71e44d9faeedf56371793067ab92984e6 100644 (file)
--- a/fs/file.c
+++ b/fs/file.c
@@ -21,7 +21,6 @@
 #include <linux/rcupdate.h>
 #include <linux/close_range.h>
 #include <net/sock.h>
-#include <linux/io_uring.h>
 
 unsigned int sysctl_nr_open __read_mostly = 1024*1024;
 unsigned int sysctl_nr_open_min = BITS_PER_LONG;
@@ -428,7 +427,6 @@ void exit_files(struct task_struct *tsk)
        struct files_struct * files = tsk->files;
 
        if (files) {
-               io_uring_files_cancel(files);
                task_lock(tsk);
                tsk->files = NULL;
                task_unlock(tsk);
index 3594291a854286545de8c33048bb57396871b083..04029e35e69af40942815b96e39ebb576f8394c3 100644 (file)
@@ -63,6 +63,7 @@
 #include <linux/random.h>
 #include <linux/rcuwait.h>
 #include <linux/compat.h>
+#include <linux/io_uring.h>
 
 #include <linux/uaccess.h>
 #include <asm/unistd.h>
@@ -776,6 +777,7 @@ void __noreturn do_exit(long code)
                schedule();
        }
 
+       io_uring_files_cancel(tsk->files);
        exit_signals(tsk);  /* sets PF_EXITING */
 
        /* sync mm's RSS info before statistics gathering */