brassow
We used to take a best effort approach to recovery - failures during
a resync were completely ignored. This meant that regions where
marked 'in-sync' and 'clean' and removed from the hash list. Future
reads and writes that query the region would 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.
Note that if the DM_RAID1_HANDLE_ERRORS flag is not present, the region
will still be marked 'in-sync' (although it remains 'dirty'). 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, you can not simultaneously have
read-balancing and not have the DM_FEATURE_HANDLE_ERRORS flag. This
has been noted in a comment.
Index: linux-2.6.22-rc1-mm1/drivers/md/dm-raid1.c
===================================================================
--- linux-2.6.22-rc1-mm1.orig/drivers/md/dm-raid1.c
+++ linux-2.6.22-rc1-mm1/drivers/md/dm-raid1.c
@@ -85,6 +85,7 @@ 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 +205,7 @@ static int rh_init(struct region_hash *r
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 +370,7 @@ static void rh_update_states(struct regi
LIST_HEAD(clean);
LIST_HEAD(recovered);
+ LIST_HEAD(failed_recovered);
/*
* Quickly grab the lists.
@@ -389,6 +392,15 @@ static void rh_update_states(struct regi
list_for_each_entry (reg, &recovered, list)
list_del(®->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(®->hash_list);
+ }
+
spin_unlock(&rh->region_lock);
write_unlock_irq(&rh->hash_lock);
@@ -403,6 +415,14 @@ static void rh_update_states(struct regi
mempool_free(reg, rh->region_pool);
}
+ list_for_each_entry_safe (reg, next, &failed_recovered, list) {
+ if (rh->ms->features & DM_RAID1_HANDLE_ERRORS)
+ complete_resync_work(reg, 0);
+ else
+ complete_resync_work(reg, 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 +575,17 @@ static struct region *rh_recovery_start(
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(®->list, ®->rh->recovered_regions);
+ if (success)
+ list_add(®->list, ®->rh->recovered_regions);
+ else {
+ reg->state = RH_NOSYNC;
+ list_add(®->list, ®->rh->failed_recovered_regions);
+ }
spin_unlock_irq(&rh->region_lock);
wake(rh->ms);
@@ -633,7 +657,14 @@ static void recovery_complete(int read_e
{
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 +1176,15 @@ static int mirror_ctr(struct dm_target *
argv += args_used;
argc -= args_used;
+ /*
+ * Note: 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 not properly be reflected.
+ */
+
if (argc) {
ti->error = "Too many mirror arguments";
free_context(ms, ti, ms->nr_mirrors);
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
