dm space map metadata: fix refcount decrement below 0 which caused corruption
authorJoe Thornber <ejt@redhat.com>
Fri, 7 Mar 2014 14:57:19 +0000 (14:57 +0000)
committerMike Snitzer <snitzer@redhat.com>
Fri, 7 Mar 2014 17:02:47 +0000 (12:02 -0500)
This has been a relatively long-standing issue that wasn't nailed down
until Teng-Feng Yang's meticulous bug report to dm-devel on 3/7/2014,
see: http://www.redhat.com/archives/dm-devel/2014-March/msg00021.html

From that report:
  "When decreasing the reference count of a metadata block with its
  reference count equals 3, we will call dm_btree_remove() to remove
  this enrty from the B+tree which keeps the reference count info in
  metadata device.

  The B+tree will try to rebalance the entry of the child nodes in each
  node it traversed, and the rebalance process contains the following
  steps.

  (1) Finding the corresponding children in current node (shadow_current(s))
  (2) Shadow the children block (issue BOP_INC)
  (3) redistribute keys among children, and free children if necessary (issue BOP_DEC)

  Since the update of a metadata block's reference count could be
  recursive, we will stash these reference count update operations in
  smm->uncommitted and then process them in a FILO fashion.

  The problem is that step(3) could free the children which is created
  in step(2), so the BOP_DEC issued in step(3) will be carried out
  before the BOP_INC issued in step(2) since these BOPs will be
  processed in FILO fashion. Once the BOP_DEC from step(3) tries to
  decrease the reference count of newly shadow block, it will report
  failure for its reference equals 0 before decreasing. It looks like we
  can solve this issue by processing these BOPs in a FIFO fashion
  instead of FILO."

Commit 5b564d80 ("dm space map: disallow decrementing a reference count
below zero") changed the code to report an error for this temporary
refcount decrement below zero.  So what was previously a harmless
invalid refcount became a hard failure due to the new error path:

 device-mapper: space map common: unable to decrement a reference count below 0
 device-mapper: thin: 253:6: dm_thin_insert_block() failed: error = -22
 device-mapper: thin: 253:6: switching pool to read-only mode

This bug is in dm persistent-data code that is common to the DM thin and
cache targets.  So any users of those targets should apply this fix.

Fix this by applying recursive space map operations in FIFO order rather
than FILO.

Resolves: https://bugzilla.kernel.org/show_bug.cgi?id=68801

Reported-by: Apollon Oikonomopoulos <apoikos@debian.org>
Reported-by: edwillam1007@gmail.com
Reported-by: Teng-Feng Yang <shinrairis@gmail.com>
Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org # 3.13+
drivers/md/persistent-data/dm-space-map-metadata.c

index e9bdd46..786b689 100644 (file)
@@ -91,6 +91,69 @@ struct block_op {
        dm_block_t block;
 };
 
+struct bop_ring_buffer {
+       unsigned begin;
+       unsigned end;
+       struct block_op bops[MAX_RECURSIVE_ALLOCATIONS + 1];
+};
+
+static void brb_init(struct bop_ring_buffer *brb)
+{
+       brb->begin = 0;
+       brb->end = 0;
+}
+
+static bool brb_empty(struct bop_ring_buffer *brb)
+{
+       return brb->begin == brb->end;
+}
+
+static unsigned brb_next(struct bop_ring_buffer *brb, unsigned old)
+{
+       unsigned r = old + 1;
+       return (r >= (sizeof(brb->bops) / sizeof(*brb->bops))) ? 0 : r;
+}
+
+static int brb_push(struct bop_ring_buffer *brb,
+                   enum block_op_type type, dm_block_t b)
+{
+       struct block_op *bop;
+       unsigned next = brb_next(brb, brb->end);
+
+       /*
+        * We don't allow the last bop to be filled, this way we can
+        * differentiate between full and empty.
+        */
+       if (next == brb->begin)
+               return -ENOMEM;
+
+       bop = brb->bops + brb->end;
+       bop->type = type;
+       bop->block = b;
+
+       brb->end = next;
+
+       return 0;
+}
+
+static int brb_pop(struct bop_ring_buffer *brb, struct block_op *result)
+{
+       struct block_op *bop;
+
+       if (brb_empty(brb))
+               return -ENODATA;
+
+       bop = brb->bops + brb->begin;
+       result->type = bop->type;
+       result->block = bop->block;
+
+       brb->begin = brb_next(brb, brb->begin);
+
+       return 0;
+}
+
+/*----------------------------------------------------------------*/
+
 struct sm_metadata {
        struct dm_space_map sm;
 
@@ -101,25 +164,20 @@ struct sm_metadata {
 
        unsigned recursion_count;
        unsigned allocated_this_transaction;
-       unsigned nr_uncommitted;
-       struct block_op uncommitted[MAX_RECURSIVE_ALLOCATIONS];
+       struct bop_ring_buffer uncommitted;
 
        struct threshold threshold;
 };
 
 static int add_bop(struct sm_metadata *smm, enum block_op_type type, dm_block_t b)
 {
-       struct block_op *op;
+       int r = brb_push(&smm->uncommitted, type, b);
 
-       if (smm->nr_uncommitted == MAX_RECURSIVE_ALLOCATIONS) {
+       if (r) {
                DMERR("too many recursive allocations");
                return -ENOMEM;
        }
 
-       op = smm->uncommitted + smm->nr_uncommitted++;
-       op->type = type;
-       op->block = b;
-
        return 0;
 }
 
@@ -158,11 +216,17 @@ static int out(struct sm_metadata *smm)
                return -ENOMEM;
        }
 
-       if (smm->recursion_count == 1 && smm->nr_uncommitted) {
-               while (smm->nr_uncommitted && !r) {
-                       smm->nr_uncommitted--;
-                       r = commit_bop(smm, smm->uncommitted +
-                                      smm->nr_uncommitted);
+       if (smm->recursion_count == 1) {
+               while (!brb_empty(&smm->uncommitted)) {
+                       struct block_op bop;
+
+                       r = brb_pop(&smm->uncommitted, &bop);
+                       if (r) {
+                               DMERR("bug in bop ring buffer");
+                               break;
+                       }
+
+                       r = commit_bop(smm, &bop);
                        if (r)
                                break;
                }
@@ -217,7 +281,8 @@ static int sm_metadata_get_nr_free(struct dm_space_map *sm, dm_block_t *count)
 static int sm_metadata_get_count(struct dm_space_map *sm, dm_block_t b,
                                 uint32_t *result)
 {
-       int r, i;
+       int r;
+       unsigned i;
        struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
        unsigned adjustment = 0;
 
@@ -225,8 +290,10 @@ static int sm_metadata_get_count(struct dm_space_map *sm, dm_block_t b,
         * We may have some uncommitted adjustments to add.  This list
         * should always be really short.
         */
-       for (i = 0; i < smm->nr_uncommitted; i++) {
-               struct block_op *op = smm->uncommitted + i;
+       for (i = smm->uncommitted.begin;
+            i != smm->uncommitted.end;
+            i = brb_next(&smm->uncommitted, i)) {
+               struct block_op *op = smm->uncommitted.bops + i;
 
                if (op->block != b)
                        continue;
@@ -254,7 +321,8 @@ static int sm_metadata_get_count(struct dm_space_map *sm, dm_block_t b,
 static int sm_metadata_count_is_more_than_one(struct dm_space_map *sm,
                                              dm_block_t b, int *result)
 {
-       int r, i, adjustment = 0;
+       int r, adjustment = 0;
+       unsigned i;
        struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
        uint32_t rc;
 
@@ -262,8 +330,11 @@ static int sm_metadata_count_is_more_than_one(struct dm_space_map *sm,
         * We may have some uncommitted adjustments to add.  This list
         * should always be really short.
         */
-       for (i = 0; i < smm->nr_uncommitted; i++) {
-               struct block_op *op = smm->uncommitted + i;
+       for (i = smm->uncommitted.begin;
+            i != smm->uncommitted.end;
+            i = brb_next(&smm->uncommitted, i)) {
+
+               struct block_op *op = smm->uncommitted.bops + i;
 
                if (op->block != b)
                        continue;
@@ -671,7 +742,7 @@ int dm_sm_metadata_create(struct dm_space_map *sm,
        smm->begin = superblock + 1;
        smm->recursion_count = 0;
        smm->allocated_this_transaction = 0;
-       smm->nr_uncommitted = 0;
+       brb_init(&smm->uncommitted);
        threshold_init(&smm->threshold);
 
        memcpy(&smm->sm, &bootstrap_ops, sizeof(smm->sm));
@@ -715,7 +786,7 @@ int dm_sm_metadata_open(struct dm_space_map *sm,
        smm->begin = 0;
        smm->recursion_count = 0;
        smm->allocated_this_transaction = 0;
-       smm->nr_uncommitted = 0;
+       brb_init(&smm->uncommitted);
        threshold_init(&smm->threshold);
 
        memcpy(&smm->old_ll, &smm->ll, sizeof(smm->old_ll));