1 files changed, 57 insertions, 10 deletions

diff --git a/drivers/md/dm-thin-metadata.c b/drivers/md/dm-thin-metadata.c
index a5ed59eafdc5..92e646d597ea 100644
--- a/drivers/md/dm-thin-metadata.c
+++ b/drivers/md/dm-thin-metadata.c
@@ -701,6 +701,15 @@ static int __open_metadata(struct dm_pool_metadata *pmd)
 		goto bad_cleanup_data_sm;
 	}
 
+	/*
+	 * For pool metadata opening process, root setting is redundant
+	 * because it will be set again in __begin_transaction(). But dm
+	 * pool aborting process really needs to get last transaction's
+	 * root to avoid accessing broken btree.
+	 */
+	pmd->root = le64_to_cpu(disk_super->data_mapping_root);
+	pmd->details_root = le64_to_cpu(disk_super->device_details_root);
+
 	__setup_btree_details(pmd);
 	dm_bm_unlock(sblock);
 
@@ -753,13 +762,15 @@ static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool f
 	return r;
 }
 
-static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
+static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd,
+					      bool destroy_bm)
 {
 	dm_sm_destroy(pmd->data_sm);
 	dm_sm_destroy(pmd->metadata_sm);
 	dm_tm_destroy(pmd->nb_tm);
 	dm_tm_destroy(pmd->tm);
-	dm_block_manager_destroy(pmd->bm);
+	if (destroy_bm)
+		dm_block_manager_destroy(pmd->bm);
 }
 
 static int __begin_transaction(struct dm_pool_metadata *pmd)
@@ -966,7 +977,7 @@ int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 	}
 	pmd_write_unlock(pmd);
 	if (!pmd->fail_io)
-		__destroy_persistent_data_objects(pmd);
+		__destroy_persistent_data_objects(pmd, true);
 
 	kfree(pmd);
 	return 0;
@@ -1875,19 +1886,52 @@ static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
 {
 	int r = -EINVAL;
+	struct dm_block_manager *old_bm = NULL, *new_bm = NULL;
+
+	/* fail_io is double-checked with pmd->root_lock held below */
+	if (unlikely(pmd->fail_io))
+		return r;
+
+	/*
+	 * Replacement block manager (new_bm) is created and old_bm destroyed outside of
+	 * pmd root_lock to avoid ABBA deadlock that would result (due to life-cycle of
+	 * shrinker associated with the block manager's bufio client vs pmd root_lock).
+	 * - must take shrinker_rwsem without holding pmd->root_lock
+	 */
+	new_bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
+					 THIN_MAX_CONCURRENT_LOCKS);
 
 	pmd_write_lock(pmd);
-	if (pmd->fail_io)
+	if (pmd->fail_io) {
+		pmd_write_unlock(pmd);
 		goto out;
+	}
 
 	__set_abort_with_changes_flags(pmd);
-	__destroy_persistent_data_objects(pmd);
-	r = __create_persistent_data_objects(pmd, false);
+	__destroy_persistent_data_objects(pmd, false);
+	old_bm = pmd->bm;
+	if (IS_ERR(new_bm)) {
+		DMERR("could not create block manager during abort");
+		pmd->bm = NULL;
+		r = PTR_ERR(new_bm);
+		goto out_unlock;
+	}
+
+	pmd->bm = new_bm;
+	r = __open_or_format_metadata(pmd, false);
+	if (r) {
+		pmd->bm = NULL;
+		goto out_unlock;
+	}
+	new_bm = NULL;
+out_unlock:
 	if (r)
 		pmd->fail_io = true;
-
-out:
 	pmd_write_unlock(pmd);
+	dm_block_manager_destroy(old_bm);
+out:
+	if (new_bm && !IS_ERR(new_bm))
+		dm_block_manager_destroy(new_bm);
 
 	return r;
 }
@@ -2060,10 +2104,13 @@ int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
 					dm_sm_threshold_fn fn,
 					void *context)
 {
-	int r;
+	int r = -EINVAL;
 
 	pmd_write_lock_in_core(pmd);
-	r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
+	if (!pmd->fail_io) {
+		r = dm_sm_register_threshold_callback(pmd->metadata_sm,
+						      threshold, fn, context);
+	}
 	pmd_write_unlock(pmd);
 
 	return r;