dm raid1: handle resync failures

Device-mapper mirroring currently takes a best effort approach to
recovery - failures during mirror synchronization are completely ignored.
This means that regions are marked 'in-sync' and 'clean' and removed
from the hash list.  Future reads and writes that query the region
will incorrectly interpret the region as in-sync.

This patch handles failures during the recovery process.  If a failure
occurs, the region is marked as 'not-in-sync' (aka RH_NOSYNC) and added
to a new list 'failed_recovered_regions'.

Regions on the 'failed_recovered_regions' list are not marked as 'clean'
upon removal from the list.  Furthermore, if the DM_RAID1_HANDLE_ERRORS
flag is set, the region is marked as 'not-in-sync'.  This action prevents
any future read-balancing from choosing an invalid device because of the
'not-in-sync' status.

If "handle_errors" is not specified when creating a mirror (leaving the
DM_RAID1_HANDLE_ERRORS flag unset), failures will be ignored exactly as they
would be without this patch.  This is to preserve backwards compatibility with
user-space tools, such as 'pvmove'.  However, since future read-balancing
policies will rely on the correct sync status of a region, a user must choose
"handle_errors" when using read-balancing.

Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 41 insertions, 3 deletions

diff --git a/drivers/md/dm-raid1.c b/drivers/md/dm-raid1.c
index 04dce7a66494..dee4221caa73 100644
--- a/drivers/md/dm-raid1.c
+++ b/drivers/md/dm-raid1.c
@@ -24,6 +24,7 @@
 #define DM_IO_PAGES 64
 
 #define DM_RAID1_HANDLE_ERRORS 0x01
+#define errors_handled(p)       ((p)->features & DM_RAID1_HANDLE_ERRORS)
 
 static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
 
@@ -85,6 +86,7 @@ struct region_hash {
         struct list_head clean_regions;
         struct list_head quiesced_regions;
         struct list_head recovered_regions;
+        struct list_head failed_recovered_regions;
 };
 
 enum {
@@ -204,6 +206,7 @@ static int rh_init(struct region_hash *rh, struct mirror_set *ms,
         INIT_LIST_HEAD(&rh->clean_regions);
         INIT_LIST_HEAD(&rh->quiesced_regions);
         INIT_LIST_HEAD(&rh->recovered_regions);
+        INIT_LIST_HEAD(&rh->failed_recovered_regions);
 
         rh->region_pool = mempool_create_kmalloc_pool(MIN_REGIONS,
                                                       sizeof(struct region));
@@ -368,6 +371,7 @@ static void rh_update_states(struct region_hash *rh)
 
         LIST_HEAD(clean);
         LIST_HEAD(recovered);
+        LIST_HEAD(failed_recovered);
 
         /*
          * Quickly grab the lists.
@@ -389,6 +393,15 @@ static void rh_update_states(struct region_hash *rh)
                 list_for_each_entry (reg, &recovered, list)
                         list_del(&reg->hash_list);
         }
+
+        if (!list_empty(&rh->failed_recovered_regions)) {
+                list_splice(&rh->failed_recovered_regions, &failed_recovered);
+                INIT_LIST_HEAD(&rh->failed_recovered_regions);
+
+                list_for_each_entry(reg, &failed_recovered, list)
+                        list_del(&reg->hash_list);
+        }
+
         spin_unlock(&rh->region_lock);
         write_unlock_irq(&rh->hash_lock);
 
@@ -403,6 +416,11 @@ static void rh_update_states(struct region_hash *rh)
                 mempool_free(reg, rh->region_pool);
         }
 
+        list_for_each_entry_safe(reg, next, &failed_recovered, list) {
+                complete_resync_work(reg, errors_handled(rh->ms) ? 0 : 1);
+                mempool_free(reg, rh->region_pool);
+        }
+
         list_for_each_entry_safe(reg, next, &clean, list) {
                 rh->log->type->clear_region(rh->log, reg->key);
                 mempool_free(reg, rh->region_pool);
@@ -555,13 +573,17 @@ static struct region *rh_recovery_start(struct region_hash *rh)
         return reg;
 }
 
-/* FIXME: success ignored for now */
 static void rh_recovery_end(struct region *reg, int success)
 {
         struct region_hash *rh = reg->rh;
 
         spin_lock_irq(&rh->region_lock);
-        list_add(&reg->list, &reg->rh->recovered_regions);
+        if (success)
+                list_add(&reg->list, &reg->rh->recovered_regions);
+        else {
+                reg->state = RH_NOSYNC;
+                list_add(&reg->list, &reg->rh->failed_recovered_regions);
+        }
         spin_unlock_irq(&rh->region_lock);
 
         wake(rh->ms);
@@ -633,7 +655,14 @@ static void recovery_complete(int read_err, unsigned int write_err,
 {
         struct region *reg = (struct region *) context;
 
-        /* FIXME: better error handling */
+        if (read_err)
+                /* Read error means the failure of default mirror. */
+                DMERR_LIMIT("Unable to read primary mirror during recovery");
+
+        if (write_err)
+                DMERR_LIMIT("Write error during recovery (error = 0x%x)",
+                            write_err);
+
         rh_recovery_end(reg, !(read_err || write_err));
 }
 
@@ -1145,6 +1174,15 @@ static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
         argv += args_used;
         argc -= args_used;
 
+        /*
+         * Any read-balancing addition depends on the
+         * DM_RAID1_HANDLE_ERRORS flag being present.
+         * This is because the decision to balance depends
+         * on the sync state of a region.  If the above
+         * flag is not present, we ignore errors; and
+         * the sync state may be inaccurate.
+         */
+
         if (argc) {
                 ti->error = "Too many mirror arguments";
                 free_context(ms, ti, ms->nr_mirrors);