diff options
| -rw-r--r-- | drivers/md/dm-vdo/block-map.h | 15 | ||||
| -rw-r--r-- | drivers/md/dm-vdo/dedupe.c | 5 | ||||
| -rw-r--r-- | drivers/md/dm-vdo/recovery-journal.h | 4 | ||||
| -rw-r--r-- | drivers/md/dm-vdo/slab-depot.h | 16 |
4 files changed, 35 insertions, 5 deletions
diff --git a/drivers/md/dm-vdo/block-map.h b/drivers/md/dm-vdo/block-map.h index c574bd524bc2..b662c318c2ea 100644 --- a/drivers/md/dm-vdo/block-map.h +++ b/drivers/md/dm-vdo/block-map.h @@ -19,6 +19,21 @@ #include "vio.h" #include "wait-queue.h" +/* + * The block map is responsible for tracking all the logical to physical mappings of a VDO. It + * consists of a collection of 60 radix trees gradually allocated as logical addresses are used. + * Each tree is assigned to a logical zone such that it is easy to compute which zone must handle + * each logical address. Each logical zone also has a dedicated portion of the leaf page cache. + * + * Each logical zone has a single dedicated queue and thread for performing all updates to the + * radix trees assigned to that zone. The concurrency guarantees of this single-threaded model + * allow the code to omit more fine-grained locking for the block map structures. + * + * Load operations must be performed on the admin thread. Normal operations, such as reading and + * updating mappings, must be performed on the appropriate logical zone thread. Save operations + * must be launched from the same admin thread as the original load operation. + */ + enum { BLOCK_MAP_VIO_POOL_SIZE = 64, }; diff --git a/drivers/md/dm-vdo/dedupe.c b/drivers/md/dm-vdo/dedupe.c index 4b00135511dd..d81065a0951c 100644 --- a/drivers/md/dm-vdo/dedupe.c +++ b/drivers/md/dm-vdo/dedupe.c @@ -14,6 +14,11 @@ * deduplicate against a single block instead of being serialized through a PBN read lock. Only one * index query is needed for each hash_lock, instead of one for every data_vio. * + * Hash_locks are assigned to hash_zones by computing a modulus on the hash itself. Each hash_zone + * has a single dedicated queue and thread for performing all operations on the hash_locks assigned + * to that zone. The concurrency guarantees of this single-threaded model allow the code to omit + * more fine-grained locking for the hash_lock structures. + * * A hash_lock acts like a state machine perhaps more than as a lock. Other than the starting and * ending states INITIALIZING and BYPASSING, every state represents and is held for the duration of * an asynchronous operation. All state transitions are performed on the thread of the hash_zone diff --git a/drivers/md/dm-vdo/recovery-journal.h b/drivers/md/dm-vdo/recovery-journal.h index 19fa7ed9648a..d78c6c7da4ea 100644 --- a/drivers/md/dm-vdo/recovery-journal.h +++ b/drivers/md/dm-vdo/recovery-journal.h @@ -26,6 +26,10 @@ * write amplification of writes by providing amortization of slab journal and block map page * updates. * + * The recovery journal has a single dedicated queue and thread for performing all journal updates. + * The concurrency guarantees of this single-threaded model allow the code to omit more + * fine-grained locking for recovery journal structures. + * * The journal consists of a set of on-disk blocks arranged as a circular log with monotonically * increasing sequence numbers. Three sequence numbers serve to define the active extent of the * journal. The 'head' is the oldest active block in the journal. The 'tail' is the end of the diff --git a/drivers/md/dm-vdo/slab-depot.h b/drivers/md/dm-vdo/slab-depot.h index efdef566709a..fba293f9713e 100644 --- a/drivers/md/dm-vdo/slab-depot.h +++ b/drivers/md/dm-vdo/slab-depot.h @@ -29,11 +29,17 @@ * a single array of slabs in order to eliminate the need for additional math in order to compute * which physical zone a PBN is in. It also has a block_allocator per zone. * - * Load operations are required to be performed on a single thread. Normal operations are assumed - * to be performed in the appropriate zone. Allocations and reference count updates must be done - * from the thread of their physical zone. Requests to commit slab journal tail blocks from the - * recovery journal must be done on the journal zone thread. Save operations are required to be - * launched from the same thread as the original load operation. + * Each physical zone has a single dedicated queue and thread for performing all updates to the + * slabs assigned to that zone. The concurrency guarantees of this single-threaded model allow the + * code to omit more fine-grained locking for the various slab structures. Each physical zone + * maintains a separate copy of the slab summary to remove the need for explicit locking on that + * structure as well. + * + * Load operations must be performed on the admin thread. Normal operations, such as allocations + * and reference count updates, must be performed on the appropriate physical zone thread. Requests + * from the recovery journal to commit slab journal tail blocks must be scheduled from the recovery + * journal thread to run on the appropriate physical zone thread. Save operations must be launched + * from the same admin thread as the original load operation. */ enum { |