md: resolve external metadata handling deadlock in md_allow_write

md_allow_write() marks the metadata dirty while holding mddev->lock and then
waits for the write to complete.  For externally managed metadata this causes a
deadlock as userspace needs to take the lock to communicate that the metadata
update has completed.

Change md_allow_write() in the 'external' case to start the 'mark active'
operation and then return -EAGAIN.  The expected side effects while waiting for
userspace to write 'active' to 'array_state' are holding off reshape (code
currently handles -ENOMEM), cause some 'stripe_cache_size' change requests to
fail, cause some GET_BITMAP_FILE ioctl requests to fall back to GFP_NOIO, and
cause updates to 'raid_disks' to fail.  Except for 'stripe_cache_size' changes
these failures can be mitigated by coordinating with mdmon.

md_write_start() still prevents writes from occurring until the metadata
handler has had a chance to take action as it unconditionally waits for
MD_CHANGE_CLEAN to be cleared.

[neilb@suse.de: return -EAGAIN, try GFP_NOIO]
Signed-off-by: Dan Williams <dan.j.williams@intel.com>

1 files changed, 16 insertions, 11 deletions

diff --git a/drivers/md/md.c b/drivers/md/md.c
index df1230af02cd..43d033d9a05a 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c
@@ -4172,9 +4172,11 @@ static int get_bitmap_file(mddev_t * mddev, void __user * arg)
         char *ptr, *buf = NULL;
         int err = -ENOMEM;
 
-        md_allow_write(mddev);
+        if (md_allow_write(mddev))
+                file = kmalloc(sizeof(*file), GFP_NOIO);
+        else
+                file = kmalloc(sizeof(*file), GFP_KERNEL);
 
-        file = kmalloc(sizeof(*file), GFP_KERNEL);
         if (!file)
                 goto out;
 
@@ -5667,15 +5669,18 @@ void md_write_end(mddev_t *mddev)
  * may proceed without blocking.  It is important to call this before
  * attempting a GFP_KERNEL allocation while holding the mddev lock.
  * Must be called with mddev_lock held.
+ *
+ * In the ->external case MD_CHANGE_CLEAN can not be cleared until mddev->lock
+ * is dropped, so return -EAGAIN after notifying userspace.
  */
-void md_allow_write(mddev_t *mddev)
+int md_allow_write(mddev_t *mddev)
 {
         if (!mddev->pers)
-                return;
+                return 0;
         if (mddev->ro)
-                return;
+                return 0;
         if (!mddev->pers->sync_request)
-                return;
+                return 0;
 
         spin_lock_irq(&mddev->write_lock);
         if (mddev->in_sync) {
@@ -5686,14 +5691,14 @@ void md_allow_write(mddev_t *mddev)
                         mddev->safemode = 1;
                 spin_unlock_irq(&mddev->write_lock);
                 md_update_sb(mddev, 0);
-
                 sysfs_notify(&mddev->kobj, NULL, "array_state");
-                /* wait for the dirty state to be recorded in the metadata */
-                wait_event(mddev->sb_wait,
-                           !test_bit(MD_CHANGE_CLEAN, &mddev->flags) &&
-                           !test_bit(MD_CHANGE_PENDING, &mddev->flags));
         } else
                 spin_unlock_irq(&mddev->write_lock);
+
+        if (test_bit(MD_CHANGE_CLEAN, &mddev->flags))
+                return -EAGAIN;
+        else
+                return 0;
 }
 EXPORT_SYMBOL_GPL(md_allow_write);