]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
dm cache: fix race when issuing a POLICY_REPLACE operation
authorJoe Thornber <ejt@redhat.com>
Wed, 20 May 2015 09:30:32 +0000 (10:30 +0100)
committerMike Snitzer <snitzer@redhat.com>
Fri, 29 May 2015 18:19:03 +0000 (14:19 -0400)
There is a race between a policy deciding to replace a cache entry,
the core target writing back any dirty data from this block, and other
IO threads doing IO to the same block.

This sort of problem is avoided most of the time by the core target
grabbing a bio prison cell before making the request to the policy.
But for a demotion the core target doesn't know which block will be
demoted, so can't do this in advance.

Fix this demotion race by introducing a callback to the policy interface
that allows the policy to grab the cell on behalf of the core target.

Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org
drivers/md/dm-cache-policy-cleaner.c
drivers/md/dm-cache-policy-internal.h
drivers/md/dm-cache-policy-mq.c
drivers/md/dm-cache-policy.h
drivers/md/dm-cache-target.c

index b04d1f904d0763c43870e38c857fcd25baad4b7e..004e463c9423cd7fe81d8f2256b63877b02aebf0 100644 (file)
@@ -171,7 +171,8 @@ static void remove_cache_hash_entry(struct wb_cache_entry *e)
 /* Public interface (see dm-cache-policy.h */
 static int wb_map(struct dm_cache_policy *pe, dm_oblock_t oblock,
                  bool can_block, bool can_migrate, bool discarded_oblock,
-                 struct bio *bio, struct policy_result *result)
+                 struct bio *bio, struct policy_locker *locker,
+                 struct policy_result *result)
 {
        struct policy *p = to_policy(pe);
        struct wb_cache_entry *e;
index 2256a1f24f73e1e3fc0271f0fbd39b1f4c1b67b0..c198e6defb9cd53009b2c61ba0dd1ae3c30d6483 100644 (file)
  */
 static inline int policy_map(struct dm_cache_policy *p, dm_oblock_t oblock,
                             bool can_block, bool can_migrate, bool discarded_oblock,
-                            struct bio *bio, struct policy_result *result)
+                            struct bio *bio, struct policy_locker *locker,
+                            struct policy_result *result)
 {
-       return p->map(p, oblock, can_block, can_migrate, discarded_oblock, bio, result);
+       return p->map(p, oblock, can_block, can_migrate, discarded_oblock, bio, locker, result);
 }
 
 static inline int policy_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock)
index 3ddd1162334df3bcefe551ca86a5287b9f3ea4ca..515d44bf24d3de874a3a58f521de7446710c4527 100644 (file)
@@ -693,9 +693,10 @@ static void requeue(struct mq_policy *mq, struct entry *e)
  * - set the hit count to a hard coded value other than 1, eg, is it better
  *   if it goes in at level 2?
  */
-static int demote_cblock(struct mq_policy *mq, dm_oblock_t *oblock)
+static int demote_cblock(struct mq_policy *mq,
+                        struct policy_locker *locker, dm_oblock_t *oblock)
 {
-       struct entry *demoted = pop(mq, &mq->cache_clean);
+       struct entry *demoted = peek(&mq->cache_clean);
 
        if (!demoted)
                /*
@@ -707,6 +708,13 @@ static int demote_cblock(struct mq_policy *mq, dm_oblock_t *oblock)
                 */
                return -ENOSPC;
 
+       if (locker->fn(locker, demoted->oblock))
+               /*
+                * We couldn't lock the demoted block.
+                */
+               return -EBUSY;
+
+       del(mq, demoted);
        *oblock = demoted->oblock;
        free_entry(&mq->cache_pool, demoted);
 
@@ -795,6 +803,7 @@ static int cache_entry_found(struct mq_policy *mq,
  * finding which cache block to use.
  */
 static int pre_cache_to_cache(struct mq_policy *mq, struct entry *e,
+                             struct policy_locker *locker,
                              struct policy_result *result)
 {
        int r;
@@ -803,11 +812,12 @@ static int pre_cache_to_cache(struct mq_policy *mq, struct entry *e,
        /* Ensure there's a free cblock in the cache */
        if (epool_empty(&mq->cache_pool)) {
                result->op = POLICY_REPLACE;
-               r = demote_cblock(mq, &result->old_oblock);
+               r = demote_cblock(mq, locker, &result->old_oblock);
                if (r) {
                        result->op = POLICY_MISS;
                        return 0;
                }
+
        } else
                result->op = POLICY_NEW;
 
@@ -829,7 +839,8 @@ static int pre_cache_to_cache(struct mq_policy *mq, struct entry *e,
 
 static int pre_cache_entry_found(struct mq_policy *mq, struct entry *e,
                                 bool can_migrate, bool discarded_oblock,
-                                int data_dir, struct policy_result *result)
+                                int data_dir, struct policy_locker *locker,
+                                struct policy_result *result)
 {
        int r = 0;
 
@@ -842,7 +853,7 @@ static int pre_cache_entry_found(struct mq_policy *mq, struct entry *e,
 
        else {
                requeue(mq, e);
-               r = pre_cache_to_cache(mq, e, result);
+               r = pre_cache_to_cache(mq, e, locker, result);
        }
 
        return r;
@@ -872,6 +883,7 @@ static void insert_in_pre_cache(struct mq_policy *mq,
 }
 
 static void insert_in_cache(struct mq_policy *mq, dm_oblock_t oblock,
+                           struct policy_locker *locker,
                            struct policy_result *result)
 {
        int r;
@@ -879,7 +891,7 @@ static void insert_in_cache(struct mq_policy *mq, dm_oblock_t oblock,
 
        if (epool_empty(&mq->cache_pool)) {
                result->op = POLICY_REPLACE;
-               r = demote_cblock(mq, &result->old_oblock);
+               r = demote_cblock(mq, locker, &result->old_oblock);
                if (unlikely(r)) {
                        result->op = POLICY_MISS;
                        insert_in_pre_cache(mq, oblock);
@@ -907,11 +919,12 @@ static void insert_in_cache(struct mq_policy *mq, dm_oblock_t oblock,
 
 static int no_entry_found(struct mq_policy *mq, dm_oblock_t oblock,
                          bool can_migrate, bool discarded_oblock,
-                         int data_dir, struct policy_result *result)
+                         int data_dir, struct policy_locker *locker,
+                         struct policy_result *result)
 {
        if (adjusted_promote_threshold(mq, discarded_oblock, data_dir) <= 1) {
                if (can_migrate)
-                       insert_in_cache(mq, oblock, result);
+                       insert_in_cache(mq, oblock, locker, result);
                else
                        return -EWOULDBLOCK;
        } else {
@@ -928,7 +941,8 @@ static int no_entry_found(struct mq_policy *mq, dm_oblock_t oblock,
  */
 static int map(struct mq_policy *mq, dm_oblock_t oblock,
               bool can_migrate, bool discarded_oblock,
-              int data_dir, struct policy_result *result)
+              int data_dir, struct policy_locker *locker,
+              struct policy_result *result)
 {
        int r = 0;
        struct entry *e = hash_lookup(mq, oblock);
@@ -942,11 +956,11 @@ static int map(struct mq_policy *mq, dm_oblock_t oblock,
 
        else if (e)
                r = pre_cache_entry_found(mq, e, can_migrate, discarded_oblock,
-                                         data_dir, result);
+                                         data_dir, locker, result);
 
        else
                r = no_entry_found(mq, oblock, can_migrate, discarded_oblock,
-                                  data_dir, result);
+                                  data_dir, locker, result);
 
        if (r == -EWOULDBLOCK)
                result->op = POLICY_MISS;
@@ -1012,7 +1026,8 @@ static void copy_tick(struct mq_policy *mq)
 
 static int mq_map(struct dm_cache_policy *p, dm_oblock_t oblock,
                  bool can_block, bool can_migrate, bool discarded_oblock,
-                 struct bio *bio, struct policy_result *result)
+                 struct bio *bio, struct policy_locker *locker,
+                 struct policy_result *result)
 {
        int r;
        struct mq_policy *mq = to_mq_policy(p);
@@ -1028,7 +1043,7 @@ static int mq_map(struct dm_cache_policy *p, dm_oblock_t oblock,
 
        iot_examine_bio(&mq->tracker, bio);
        r = map(mq, oblock, can_migrate, discarded_oblock,
-               bio_data_dir(bio), result);
+               bio_data_dir(bio), locker, result);
 
        mutex_unlock(&mq->lock);
 
index f50fe360c5462928522101c2497ee0ef57542cf0..5524e21e483654d5dc44b8b8944f9404b84b0292 100644 (file)
@@ -69,6 +69,18 @@ enum policy_operation {
        POLICY_REPLACE
 };
 
+/*
+ * When issuing a POLICY_REPLACE the policy needs to make a callback to
+ * lock the block being demoted.  This doesn't need to occur during a
+ * writeback operation since the block remains in the cache.
+ */
+struct policy_locker;
+typedef int (*policy_lock_fn)(struct policy_locker *l, dm_oblock_t oblock);
+
+struct policy_locker {
+       policy_lock_fn fn;
+};
+
 /*
  * This is the instruction passed back to the core target.
  */
@@ -122,7 +134,8 @@ struct dm_cache_policy {
         */
        int (*map)(struct dm_cache_policy *p, dm_oblock_t oblock,
                   bool can_block, bool can_migrate, bool discarded_oblock,
-                  struct bio *bio, struct policy_result *result);
+                  struct bio *bio, struct policy_locker *locker,
+                  struct policy_result *result);
 
        /*
         * Sometimes we want to see if a block is in the cache, without
index 41b2594a80c63010c5282ac5d61079fc61fc986e..d5982480630b0dc4df0d67cb5dece46e266caeaa 100644 (file)
@@ -1439,16 +1439,43 @@ static void inc_miss_counter(struct cache *cache, struct bio *bio)
                   &cache->stats.read_miss : &cache->stats.write_miss);
 }
 
+/*----------------------------------------------------------------*/
+
+struct old_oblock_lock {
+       struct policy_locker locker;
+       struct cache *cache;
+       struct prealloc *structs;
+       struct dm_bio_prison_cell *cell;
+};
+
+static int null_locker(struct policy_locker *locker, dm_oblock_t b)
+{
+       /* This should never be called */
+       BUG();
+       return 0;
+}
+
+static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
+{
+       struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
+       struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
+
+       return bio_detain(l->cache, b, NULL, cell_prealloc,
+                         (cell_free_fn) prealloc_put_cell,
+                         l->structs, &l->cell);
+}
+
 static void process_bio(struct cache *cache, struct prealloc *structs,
                        struct bio *bio)
 {
        int r;
        bool release_cell = true;
        dm_oblock_t block = get_bio_block(cache, bio);
-       struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
+       struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
        struct policy_result lookup_result;
        bool passthrough = passthrough_mode(&cache->features);
        bool discarded_block, can_migrate;
+       struct old_oblock_lock ool;
 
        /*
         * Check to see if that block is currently migrating.
@@ -1463,8 +1490,12 @@ static void process_bio(struct cache *cache, struct prealloc *structs,
        discarded_block = is_discarded_oblock(cache, block);
        can_migrate = !passthrough && (discarded_block || spare_migration_bandwidth(cache));
 
+       ool.locker.fn = cell_locker;
+       ool.cache = cache;
+       ool.structs = structs;
+       ool.cell = NULL;
        r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
-                      bio, &lookup_result);
+                      bio, &ool.locker, &lookup_result);
 
        if (r == -EWOULDBLOCK)
                /* migration has been denied */
@@ -1521,27 +1552,11 @@ static void process_bio(struct cache *cache, struct prealloc *structs,
                break;
 
        case POLICY_REPLACE:
-               cell_prealloc = prealloc_get_cell(structs);
-               r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
-                              (cell_free_fn) prealloc_put_cell,
-                              structs, &old_ocell);
-               if (r > 0) {
-                       /*
-                        * We have to be careful to avoid lock inversion of
-                        * the cells.  So we back off, and wait for the
-                        * old_ocell to become free.
-                        */
-                       policy_force_mapping(cache->policy, block,
-                                            lookup_result.old_oblock);
-                       atomic_inc(&cache->stats.cache_cell_clash);
-                       break;
-               }
                atomic_inc(&cache->stats.demotion);
                atomic_inc(&cache->stats.promotion);
-
                demote_then_promote(cache, structs, lookup_result.old_oblock,
                                    block, lookup_result.cblock,
-                                   old_ocell, new_ocell);
+                                   ool.cell, new_ocell);
                release_cell = false;
                break;
 
@@ -2589,6 +2604,9 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
        bool discarded_block;
        struct policy_result lookup_result;
        struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
+       struct old_oblock_lock ool;
+
+       ool.locker.fn = null_locker;
 
        if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
                /*
@@ -2627,7 +2645,7 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
        discarded_block = is_discarded_oblock(cache, block);
 
        r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
-                      bio, &lookup_result);
+                      bio, &ool.locker, &lookup_result);
        if (r == -EWOULDBLOCK) {
                cell_defer(cache, *cell, true);
                return DM_MAPIO_SUBMITTED;