[PATCH] md: don't flush md_misc_wq from md_alloc()
COOLING14d
From: Li Youhong <hidden>
Date: 2026-09-15 08:34:38
Also in:
lkml
Subsystem:
software raid (multiple disks) support, the rest · Maintainers:
Song Liu, Yu Kuai, Linus Torvalds
From: Li Youhong <redacted> md_alloc() is called from md_probe(), while blk_probe_dev() still holds major_names_lock. The flush of md_misc_wq in md_alloc() is only meant to wait for the previous mddev_delayed_delete() to finish. md_misc_wq also runs sync_work (md_start_sync), and md_start_sync takes reconfig_mutex. On the other path, md_ioctl() already holds reconfig_mutex when md_import_device() opens a bdev and takes major_names_lock. Flushing md_misc_wq under major_names_lock therefore creates a lockdep cycle: major_names_lock -> md_misc_wq -> reconfig_mutex -> major_names_lock Move del_work to a dedicated workqueue and flush only that queue from md_alloc(). Leave sync_work on md_misc_wq. Reported-by: syzbot+68e1f51046d68329940f@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=68e1f51046d68329940f Fixes: e804ac780e2f ("md: fix and update workqueue usage") Signed-off-by: Li Youhong <redacted> --- drivers/md/md.c | 28 +++++++++++++++++++++------- 1 file changed, 21 insertions(+), 7 deletions(-)
diff --git a/drivers/md/md.c b/drivers/md/md.c
index 680b34a63cb3..87a851466892 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c@@ -86,13 +86,16 @@ static const struct kobj_type md_ktype; static DECLARE_WAIT_QUEUE_HEAD(resync_wait); /* - * This workqueue is used for sync_work to register new sync_thread, and for - * del_work to remove rdev, and for event_work that is only set by dm-raid. + * md_misc_wq runs sync_work (new sync_thread) and event_work (dm-raid). + * sync_work grabs reconfig_mutex, so this queue must not be flushed while + * holding reconfig_mutex, nor while holding major_names_lock (md_probe / + * md_alloc is called from blk_probe_dev with that lock held). * - * Noted that sync_work will grab reconfig_mutex, hence never flush this - * workqueue whith reconfig_mutex grabbed. + * md_del_wq runs only del_work (mddev_delayed_delete). md_alloc() flushes + * this queue to wait for a previous instance of the same device to go away. */ static struct workqueue_struct *md_misc_wq; +static struct workqueue_struct *md_del_wq; static int remove_and_add_spares(struct mddev *mddev, struct md_rdev *this);
@@ -651,7 +654,7 @@ static void __mddev_put(struct mddev *mddev) * Call queue_work inside the spinlock so that flush_workqueue() after * mddev_find will succeed in waiting for the work to be done. */ - queue_work(md_misc_wq, &mddev->del_work); + queue_work(md_del_wq, &mddev->del_work); } static void mddev_put_locked(struct mddev *mddev)
@@ -6329,9 +6332,12 @@ struct mddev *md_alloc(dev_t dev, char *name) /* * Wait for any previous instance of this device to be completely - * removed (mddev_delayed_delete). + * removed (mddev_delayed_delete). Do not flush md_misc_wq: it also + * runs sync_work, which takes reconfig_mutex and can take + * major_names_lock via md_import_device(). md_alloc() is called from + * md_probe() while blk_probe_dev() already holds major_names_lock. */ - flush_workqueue(md_misc_wq); + flush_workqueue(md_del_wq); mutex_lock(&disks_mutex); mddev = mddev_alloc(dev);
@@ -10765,6 +10771,10 @@ static int __init md_init(void) if (!md_misc_wq) goto err_misc_wq; + md_del_wq = alloc_workqueue("md_del", WQ_PERCPU, 0); + if (!md_del_wq) + goto err_del_wq; + ret = __register_blkdev(MD_MAJOR, "md", md_probe); if (ret < 0) goto err_md;
@@ -10783,6 +10793,8 @@ static int __init md_init(void) err_mdp: unregister_blkdev(MD_MAJOR, "md"); err_md: + destroy_workqueue(md_del_wq); +err_del_wq: destroy_workqueue(md_misc_wq); err_misc_wq: md_llbitmap_exit();
@@ -11093,6 +11105,7 @@ static __exit void md_exit(void) spin_unlock(&all_mddevs_lock); destroy_workqueue(md_misc_wq); + destroy_workqueue(md_del_wq); md_bitmap_exit(); }
--
2.43.0