===================================================================
@@ -601,6 +601,38 @@ void dm_bm_unlock(struct dm_block *b)
}
EXPORT_SYMBOL_GPL(dm_bm_unlock);
+void dm_bm_fast_lock(struct dm_block_manager *bm)
+{
+ dm_bufio_lock_read(bm->bufio);
+}
+
+void dm_bm_fast_unlock(struct dm_block_manager *bm)
+{
+ dm_bufio_unlock_read(bm->bufio);
+}
+
+int dm_bm_fast_get_block(struct dm_block_manager *bm,
+ dm_block_t b, struct dm_block_validator *v,
+ struct dm_block **result)
+{
+ int r;
+ struct buffer_aux *aux;
+ void *p;
+
+ p = dm_bufio_get_unlocked(bm->bufio, b, (struct dm_buffer **) result);
+ if (IS_ERR(p))
+ return PTR_ERR(p);
+ if (unlikely(!p))
+ return -EWOULDBLOCK;
+
+ aux = dm_bufio_get_aux_data(to_buffer(*result));
+ r = dm_bm_validate_buffer(bm, to_buffer(*result), aux, v);
+ if (unlikely(r))
+ return r;
+
+ return 0;
+}
+
int dm_bm_flush(struct dm_block_manager *bm)
{
if (dm_bm_is_read_only(bm))
===================================================================
@@ -96,6 +96,12 @@ int dm_bm_write_lock_zero(struct dm_bloc
void dm_bm_unlock(struct dm_block *b);
+void dm_bm_fast_lock(struct dm_block_manager *bm);
+void dm_bm_fast_unlock(struct dm_block_manager *bm);
+int dm_bm_fast_get_block(struct dm_block_manager *bm,
+ dm_block_t b, struct dm_block_validator *v,
+ struct dm_block **result);
+
/*
* It's a common idiom to have a superblock that should be committed last.
*
===================================================================
@@ -64,6 +64,8 @@ struct ro_spine {
struct dm_btree_info *info;
struct dm_block *node;
+ bool fast_locked;
+ bool fast_lock_failed;
};
void init_ro_spine(struct ro_spine *s, struct dm_btree_info *info);
===================================================================
@@ -118,27 +118,60 @@ void unlock_block(struct dm_btree_info *
dm_tm_unlock(info->tm, b);
}
+static void bn_fast_lock(struct dm_btree_info *info)
+{
+ dm_tm_fast_lock(info->tm);
+}
+
+static void bn_fast_unlock(struct dm_btree_info *info)
+{
+ dm_tm_fast_unlock(info->tm);
+}
+
+static int bn_fast_get_block(struct dm_btree_info *info, dm_block_t b,
+ struct dm_block **result)
+{
+ return dm_tm_fast_get_block(info->tm, b, &btree_node_validator, result);
+}
+
/*----------------------------------------------------------------*/
void init_ro_spine(struct ro_spine *s, struct dm_btree_info *info)
{
s->info = info;
s->node = NULL;
+ s->fast_locked = false;
+ s->fast_lock_failed = false;
}
void exit_ro_spine(struct ro_spine *s)
{
- if (s->node)
+ if (s->fast_locked)
+ bn_fast_unlock(s->info);
+ else if (s->node)
unlock_block(s->info, s->node);
}
int ro_step(struct ro_spine *s, dm_block_t new_child)
{
if (s->node) {
- unlock_block(s->info, s->node);
+ if (unlikely(!s->fast_locked))
+ unlock_block(s->info, s->node);
s->node = NULL;
}
-
+ if (likely(!s->fast_lock_failed)) {
+ int r;
+ if (!s->fast_locked) {
+ bn_fast_lock(s->info);
+ s->fast_locked = true;
+ }
+ r = bn_fast_get_block(s->info, new_child, &s->node);
+ if (likely(!r))
+ return 0;
+ s->fast_lock_failed = true;
+ s->fast_locked = false;
+ bn_fast_unlock(s->info);
+ }
return bn_read_lock(s->info, new_child, &s->node);
}
===================================================================
@@ -348,6 +348,23 @@ void dm_tm_unlock(struct dm_transaction_
}
EXPORT_SYMBOL_GPL(dm_tm_unlock);
+void dm_tm_fast_lock(struct dm_transaction_manager *tm)
+{
+ dm_bm_fast_lock(tm->is_clone ? tm->real->bm : tm->bm);
+}
+
+void dm_tm_fast_unlock(struct dm_transaction_manager *tm)
+{
+ dm_bm_fast_unlock(tm->is_clone ? tm->real->bm : tm->bm);
+}
+
+int dm_tm_fast_get_block(struct dm_transaction_manager *tm, dm_block_t b,
+ struct dm_block_validator *v,
+ struct dm_block **blk)
+{
+ return dm_bm_fast_get_block(tm->is_clone ? tm->real->bm : tm->bm, b, v, blk);
+}
+
void dm_tm_inc(struct dm_transaction_manager *tm, dm_block_t b)
{
/*
===================================================================
@@ -96,6 +96,12 @@ int dm_tm_read_lock(struct dm_transactio
void dm_tm_unlock(struct dm_transaction_manager *tm, struct dm_block *b);
+void dm_tm_fast_lock(struct dm_transaction_manager *tm);
+void dm_tm_fast_unlock(struct dm_transaction_manager *tm);
+int dm_tm_fast_get_block(struct dm_transaction_manager *tm, dm_block_t b,
+ struct dm_block_validator *v,
+ struct dm_block **blk);
+
/*
* Functions for altering the reference count of a block directly.
*/
This patch reduces lock contention in btree walks. We modify the functions init_ro_wpin, exit_ro_spine and ro_step so that they use dm_bufio_lock_read/dm_bufio_get_unlocked/dm_bufio_unlock_read. If dm_bm_fast_get_block fails, we fallback to normal locking. When doing tests on pmem and fully provisioned thin volume, it improves thoughput from 286MiB/s to 442MiB/s. fio --ioengine=psync --iodepth=1 --rw=randrw --bs=4k --direct=1 --numjobs=12 --time_based --runtime=10 --group_reporting --name=/dev/vg/thin before: READ: bw=286MiB/s WRITE: bw=286MiB/s after: READ: bw=442MiB/s WRITE: bw=442MiB/s Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> --- drivers/md/persistent-data/dm-block-manager.c | 32 ++++++++++++++++ drivers/md/persistent-data/dm-block-manager.h | 6 +++ drivers/md/persistent-data/dm-btree-internal.h | 2 + drivers/md/persistent-data/dm-btree-spine.c | 39 ++++++++++++++++++-- drivers/md/persistent-data/dm-transaction-manager.c | 17 ++++++++ drivers/md/persistent-data/dm-transaction-manager.h | 6 +++ 6 files changed, 99 insertions(+), 3 deletions(-) -- dm-devel mailing list dm-devel@redhat.com https://listman.redhat.com/mailman/listinfo/dm-devel