md: resolve external metadata handling deadlock in md_allow_write
Dan Williams [Sat, 28 Jun 2008 04:44:04 +0000 (21:44 -0700)]
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>

drivers/md/md.c
drivers/md/raid1.c
drivers/md/raid5.c
include/linux/raid/md.h

index df1230a..43d033d 100644 (file)
@@ -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);
 
index f05d598..491dc2d 100644 (file)
@@ -2136,7 +2136,7 @@ static int raid1_reshape(mddev_t *mddev)
        conf_t *conf = mddev_to_conf(mddev);
        int cnt, raid_disks;
        unsigned long flags;
-       int d, d2;
+       int d, d2, err;
 
        /* Cannot change chunk_size, layout, or level */
        if (mddev->chunk_size != mddev->new_chunk ||
@@ -2148,7 +2148,9 @@ static int raid1_reshape(mddev_t *mddev)
                return -EINVAL;
        }
 
-       md_allow_write(mddev);
+       err = md_allow_write(mddev);
+       if (err)
+               return err;
 
        raid_disks = mddev->raid_disks + mddev->delta_disks;
 
index 4426220..8f4c70a 100644 (file)
@@ -911,14 +911,16 @@ static int resize_stripes(raid5_conf_t *conf, int newsize)
        struct stripe_head *osh, *nsh;
        LIST_HEAD(newstripes);
        struct disk_info *ndisks;
-       int err = 0;
+       int err;
        struct kmem_cache *sc;
        int i;
 
        if (newsize <= conf->pool_size)
                return 0; /* never bother to shrink */
 
-       md_allow_write(conf->mddev);
+       err = md_allow_write(conf->mddev);
+       if (err)
+               return err;
 
        /* Step 1 */
        sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
@@ -3843,6 +3845,8 @@ raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
 {
        raid5_conf_t *conf = mddev_to_conf(mddev);
        unsigned long new;
+       int err;
+
        if (len >= PAGE_SIZE)
                return -EINVAL;
        if (!conf)
@@ -3858,7 +3862,9 @@ raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
                else
                        break;
        }
-       md_allow_write(mddev);
+       err = md_allow_write(mddev);
+       if (err)
+               return err;
        while (new > conf->max_nr_stripes) {
                if (grow_one_stripe(conf))
                        conf->max_nr_stripes++;
index b7386ae..dc0e3fc 100644 (file)
@@ -95,7 +95,7 @@ extern int sync_page_io(struct block_device *bdev, sector_t sector, int size,
                        struct page *page, int rw);
 extern void md_do_sync(mddev_t *mddev);
 extern void md_new_event(mddev_t *mddev);
-extern void md_allow_write(mddev_t *mddev);
+extern int md_allow_write(mddev_t *mddev);
 extern void md_wait_for_blocked_rdev(mdk_rdev_t *rdev, mddev_t *mddev);
 
 #endif /* CONFIG_MD */