DM-RAID: Fix RAID10's check for sufficient redundancy
authorJonathan Brassow <jbrassow@redhat.com>
Wed, 23 Jan 2013 03:42:18 +0000 (21:42 -0600)
committerNeilBrown <neilb@suse.de>
Thu, 24 Jan 2013 01:02:36 +0000 (12:02 +1100)
Before attempting to activate a RAID array, it is checked for sufficient
redundancy.  That is, we make sure that there are not too many failed
devices - or devices specified for rebuild - to undermine our ability to
activate the array.  The current code performs this check twice - once to
ensure there were not too many devices specified for rebuild by the user
('validate_rebuild_devices') and again after possibly experiencing a failure
to read the superblock ('analyse_superblocks').  Neither of these checks are
sufficient.  The first check is done properly but with insufficient
information about the possible failure state of the devices to make a good
determination if the array can be activated.  The second check is simply
done wrong in the case of RAID10 because it doesn't account for the
independence of the stripes (i.e. mirror sets).  The solution is to use the
properly written check ('validate_rebuild_devices'), but perform the check
after the superblocks have been read and we know which devices have failed.
This gives us one check instead of two and performs it in a location where
it can be done right.

Only RAID10 was affected and it was affected in the following ways:
- the code did not properly catch the condition where a user specified
  a device for rebuild that already had a failed device in the same mirror
  set.  (This condition would, however, be caught at a deeper level in MD.)
- the code triggers a false positive and denies activation when devices in
  independent mirror sets have failed - counting the failures as though they
  were all in the same set.

The most likely place this error was introduced (or this patch should have
been included) is in commit 4ec1e369 - first introduced in v3.7-rc1.
Consequently this fix should also go in v3.7.y, however there is a
small conflict on the .version in raid_target, so I'll submit a
separate patch to -stable.

Cc: stable@vger.kernel.org
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Documentation/device-mapper/dm-raid.txt
drivers/md/dm-raid.c

index 728c38c242d631b88117958360c6f8441a72be4b..56fb62b09fc59ad757fc81de6ad6e478b3b0184b 100644 (file)
@@ -141,3 +141,4 @@ Version History
 1.2.0  Handle creation of arrays that contain failed devices.
 1.3.0  Added support for RAID 10
 1.3.1  Allow device replacement/rebuild for RAID 10
+1.3.2   Fix/improve redundancy checking for RAID10
index 3d8984edeff79b4dda6910670d95d34802411f6e..9e58dbd8d8cba839da9613e5f4e4060795cd0dc1 100644 (file)
@@ -340,24 +340,22 @@ static int validate_region_size(struct raid_set *rs, unsigned long region_size)
 }
 
 /*
- * validate_rebuild_devices
+ * validate_raid_redundancy
  * @rs
  *
- * Determine if the devices specified for rebuild can result in a valid
- * usable array that is capable of rebuilding the given devices.
+ * Determine if there are enough devices in the array that haven't
+ * failed (or are being rebuilt) to form a usable array.
  *
  * Returns: 0 on success, -EINVAL on failure.
  */
-static int validate_rebuild_devices(struct raid_set *rs)
+static int validate_raid_redundancy(struct raid_set *rs)
 {
        unsigned i, rebuild_cnt = 0;
        unsigned rebuilds_per_group, copies, d;
 
-       if (!(rs->print_flags & DMPF_REBUILD))
-               return 0;
-
        for (i = 0; i < rs->md.raid_disks; i++)
-               if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
+               if (!test_bit(In_sync, &rs->dev[i].rdev.flags) ||
+                   !rs->dev[i].rdev.sb_page)
                        rebuild_cnt++;
 
        switch (rs->raid_type->level) {
@@ -393,27 +391,24 @@ static int validate_rebuild_devices(struct raid_set *rs)
                 *          A    A    B    B    C
                 *          C    D    D    E    E
                 */
-               rebuilds_per_group = 0;
                for (i = 0; i < rs->md.raid_disks * copies; i++) {
+                       if (!(i % copies))
+                               rebuilds_per_group = 0;
                        d = i % rs->md.raid_disks;
-                       if (!test_bit(In_sync, &rs->dev[d].rdev.flags) &&
+                       if ((!rs->dev[d].rdev.sb_page ||
+                            !test_bit(In_sync, &rs->dev[d].rdev.flags)) &&
                            (++rebuilds_per_group >= copies))
                                goto too_many;
-                       if (!((i + 1) % copies))
-                               rebuilds_per_group = 0;
                }
                break;
        default:
-               DMERR("The rebuild parameter is not supported for %s",
-                     rs->raid_type->name);
-               rs->ti->error = "Rebuild not supported for this RAID type";
-               return -EINVAL;
+               if (rebuild_cnt)
+                       return -EINVAL;
        }
 
        return 0;
 
 too_many:
-       rs->ti->error = "Too many rebuild devices specified";
        return -EINVAL;
 }
 
@@ -664,9 +659,6 @@ static int parse_raid_params(struct raid_set *rs, char **argv,
        }
        rs->md.dev_sectors = sectors_per_dev;
 
-       if (validate_rebuild_devices(rs))
-               return -EINVAL;
-
        /* Assume there are no metadata devices until the drives are parsed */
        rs->md.persistent = 0;
        rs->md.external = 1;
@@ -995,28 +987,10 @@ static int super_validate(struct mddev *mddev, struct md_rdev *rdev)
 static int analyse_superblocks(struct dm_target *ti, struct raid_set *rs)
 {
        int ret;
-       unsigned redundancy = 0;
        struct raid_dev *dev;
        struct md_rdev *rdev, *tmp, *freshest;
        struct mddev *mddev = &rs->md;
 
-       switch (rs->raid_type->level) {
-       case 1:
-               redundancy = rs->md.raid_disks - 1;
-               break;
-       case 4:
-       case 5:
-       case 6:
-               redundancy = rs->raid_type->parity_devs;
-               break;
-       case 10:
-               redundancy = raid10_md_layout_to_copies(mddev->layout) - 1;
-               break;
-       default:
-               ti->error = "Unknown RAID type";
-               return -EINVAL;
-       }
-
        freshest = NULL;
        rdev_for_each_safe(rdev, tmp, mddev) {
                /*
@@ -1045,44 +1019,43 @@ static int analyse_superblocks(struct dm_target *ti, struct raid_set *rs)
                        break;
                default:
                        dev = container_of(rdev, struct raid_dev, rdev);
-                       if (redundancy--) {
-                               if (dev->meta_dev)
-                                       dm_put_device(ti, dev->meta_dev);
-
-                               dev->meta_dev = NULL;
-                               rdev->meta_bdev = NULL;
+                       if (dev->meta_dev)
+                               dm_put_device(ti, dev->meta_dev);
 
-                               if (rdev->sb_page)
-                                       put_page(rdev->sb_page);
+                       dev->meta_dev = NULL;
+                       rdev->meta_bdev = NULL;
 
-                               rdev->sb_page = NULL;
+                       if (rdev->sb_page)
+                               put_page(rdev->sb_page);
 
-                               rdev->sb_loaded = 0;
+                       rdev->sb_page = NULL;
 
-                               /*
-                                * We might be able to salvage the data device
-                                * even though the meta device has failed.  For
-                                * now, we behave as though '- -' had been
-                                * set for this device in the table.
-                                */
-                               if (dev->data_dev)
-                                       dm_put_device(ti, dev->data_dev);
+                       rdev->sb_loaded = 0;
 
-                               dev->data_dev = NULL;
-                               rdev->bdev = NULL;
+                       /*
+                        * We might be able to salvage the data device
+                        * even though the meta device has failed.  For
+                        * now, we behave as though '- -' had been
+                        * set for this device in the table.
+                        */
+                       if (dev->data_dev)
+                               dm_put_device(ti, dev->data_dev);
 
-                               list_del(&rdev->same_set);
+                       dev->data_dev = NULL;
+                       rdev->bdev = NULL;
 
-                               continue;
-                       }
-                       ti->error = "Failed to load superblock";
-                       return ret;
+                       list_del(&rdev->same_set);
                }
        }
 
        if (!freshest)
                return 0;
 
+       if (validate_raid_redundancy(rs)) {
+               rs->ti->error = "Insufficient redundancy to activate array";
+               return -EINVAL;
+       }
+
        /*
         * Validation of the freshest device provides the source of
         * validation for the remaining devices.
@@ -1432,7 +1405,7 @@ static void raid_resume(struct dm_target *ti)
 
 static struct target_type raid_target = {
        .name = "raid",
-       .version = {1, 4, 0},
+       .version = {1, 4, 1},
        .module = THIS_MODULE,
        .ctr = raid_ctr,
        .dtr = raid_dtr,