[PATCH v3] workqueue: mark pool->nr_running as __data_racy

[Nguyen Ngoc Thang <ngocthang2710.1999@gmail.com>]

wq_worker_running() increments pool->nr_running with only preemption
disabled, while kick_pool_pick() reads it via need_more_worker() under
pool->lock, so the two can run concurrently on different CPUs. KCSAN
reports:

  BUG: KCSAN: data-race in kick_pool_pick / wq_worker_running

  write to 0xffff888059c2af24 of 4 bytes by task 3007 on cpu 0:
   wq_worker_running+0x114/0x130
   schedule_timeout+0x16e/0x180
   io_schedule_timeout+0x4c/0x70
   submit_bio_wait+0xd7/0x110
   blkdev_issue_flush+0x4c/0x70
   ext4_sync_file+0x60d/0x6c0
   aio_fsync_work+0x7f/0x270
   process_one_work+0x4ae/0xa00

  read to 0xffff888059c2af24 of 4 bytes by task 3001 on cpu 1:
   kick_pool_pick+0x7a/0x3a0
   __queue_work+0x72d/0xcf0
   queue_work_on+0x9d/0xb0
   schedule_on_each_cpu+0x11a/0x2d0
   vmstat_refresh+0x23/0x140
   proc_sys_call_handler+0x2b9/0x4f0

The race is benign. The increment in wq_worker_running() is the only
lockless write to nr_running; decrements and the reset are done under
pool->lock, so a reader holding the lock sees either the old or the new
count. A stale 0 costs at most a spurious wakeup of an idle worker. A
nonzero count is correct because the incrementing worker is running and
re-checks the worklist under pool->lock before it stops counting. The
increment can also go 1 -> 2, e.g. when a worker sleeping in a work item
wakes up after kick_pool() woke another worker.

The lockless increment is part of how nr_running is designed to work,
as the comment above the field documents. Mark the field __data_racy
so that every reader is covered, not just need_more_worker().

Reported-by: syzbot+8922e6b8343f70be92e8@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=8922e6b8343f70be92e8
Signed-off-by: Nguyen Ngoc Thang <ngocthang2710.1999@gmail.com>
---
v3: (Breno)
 - Annotate the field with __data_racy instead of using data_race() at
   the read in need_more_worker(). That covers every reader, and the
   lockless increment is already documented above the field.
 - Retested with KCSAN in QEMU: the baseline gives 13 reports in 1 run,
   v3 gives 0 in 2 runs.
v2: https://lore.kernel.org/all/20261013112905.9673-1-ngocthang2710.1999@gmail.com/
v1: https://lore.kernel.org/all/20261012143605.281358-1-ngocthang2710.1999@gmail.com/

 kernel/workqueue.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/workqueue.c b/kernel/workqueue.c
index 1ae3732a2c51..f285e68d6ee2 100644
--- a/kernel/workqueue.c
+++ b/kernel/workqueue.c
@@ -208,7 +208,7 @@ struct worker_pool {
 	 * but w/ pool->lock held. The readers grab pool->lock and are
 	 * guaranteed to see if the counter reached zero.
 	 */
-	int			nr_running;
+	int __data_racy nr_running;
 
 	struct list_head	worklist;	/* L: list of pending works */
 
-- 
2.43.0