[dm-devel] A patch to reduce dm-bufio locking

[Mikulas Patocka]

Hi

Regarding that dm-bufio overhead, I created this patch.

The patch exports dm_bufio_lock, dm_bufio_trylock and dm_bufio_unlock and 
introduces a new function dm_bufio_get_unlocked.

dm_bufio_get_unlocked is like dm_bufio_get, except that it doesn't take 
the lock and it assumes that the client structure is already locked with 
dm_bufio_lock.

When you walk the thinp metadata, it is recommended that you call 
dm_bufio_lock, then you call dm_bufio_get_unlocked repeatedly for each 
b-tree level and finally you call dm_bufio_unlock. So, you can walk the 
b-tree with just one lock/unlock pair.

When dm_bufio_get_unlocked returns NULL (the block is not in cache), you 
must call dm_bufio_unlock and then read the data with dm_bufio_read as 
usual - in this situation, the I/O will be bottleneck rather than locking.

Mikulas



Signed-off-by: Mikulas Patocka <mpatocka at redhat.com>

---
 drivers/md/dm-bufio.c    |   29 ++++++++++++++++++++++-------
 include/linux/dm-bufio.h |   15 +++++++++++++++
 2 files changed, 37 insertions(+), 7 deletions(-)

Index: linux-dm/drivers/md/dm-bufio.c
===================================================================
--- linux-dm.orig/drivers/md/dm-bufio.c	2022-09-05 13:29:52.000000000 +0200
+++ linux-dm/drivers/md/dm-bufio.c	2022-09-27 15:43:18.000000000 +0200
@@ -169,29 +169,32 @@ static DEFINE_STATIC_KEY_FALSE(no_sleep_
 
 #define dm_bufio_in_request()	(!!current->bio_list)
 
-static void dm_bufio_lock(struct dm_bufio_client *c)
+void dm_bufio_lock(struct dm_bufio_client *c)
 {
 	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
 		spin_lock_bh(&c->spinlock);
 	else
 		mutex_lock_nested(&c->lock, dm_bufio_in_request());
 }
+EXPORT_SYMBOL_GPL(dm_bufio_lock);
 
-static int dm_bufio_trylock(struct dm_bufio_client *c)
+int dm_bufio_trylock(struct dm_bufio_client *c)
 {
 	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
 		return spin_trylock_bh(&c->spinlock);
 	else
 		return mutex_trylock(&c->lock);
 }
+EXPORT_SYMBOL_GPL(dm_bufio_trylock);
 
-static void dm_bufio_unlock(struct dm_bufio_client *c)
+void dm_bufio_unlock(struct dm_bufio_client *c)
 {
 	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
 		spin_unlock_bh(&c->spinlock);
 	else
 		mutex_unlock(&c->lock);
 }
+EXPORT_SYMBOL_GPL(dm_bufio_unlock);
 
 /*----------------------------------------------------------------*/
 
@@ -870,7 +873,8 @@ enum new_flag {
 	NF_FRESH = 0,
 	NF_READ = 1,
 	NF_GET = 2,
-	NF_PREFETCH = 3
+	NF_GET_UNLOCKED = 3,
+	NF_PREFETCH = 4
 };
 
 /*
@@ -1013,7 +1017,7 @@ static struct dm_buffer *__bufio_new(str
 	if (b)
 		goto found_buffer;
 
-	if (nf == NF_GET)
+	if (nf == NF_GET || nf == NF_GET_UNLOCKED)
 		return NULL;
 
 	new_b = __alloc_buffer_wait(c, nf);
@@ -1058,10 +1062,11 @@ found_buffer:
 	 * If the user called both dm_bufio_prefetch and dm_bufio_get on
 	 * the same buffer, it would deadlock if we waited.
 	 */
-	if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
+	if ((nf == NF_GET || nf == NF_GET_UNLOCKED) && unlikely(test_bit(B_READING, &b->state)))
 		return NULL;
 
-	b->hold_count++;
+	b->hold_count += nf != NF_GET_UNLOCKED;
+
 	__relink_lru(b, test_bit(B_DIRTY, &b->state) ||
 		     test_bit(B_WRITING, &b->state));
 	return b;
@@ -1154,6 +1159,16 @@ void *dm_bufio_new(struct dm_bufio_clien
 }
 EXPORT_SYMBOL_GPL(dm_bufio_new);
 
+void *dm_bufio_get_unlocked(struct dm_bufio_client *c, sector_t block)
+{
+	int need_submit;
+	struct dm_buffer *b = __bufio_new(c, block, NF_GET_UNLOCKED, &need_submit, NULL);
+	if (b)
+		return b->data;
+	return NULL;
+}
+EXPORT_SYMBOL_GPL(dm_bufio_get_unlocked);
+
 void dm_bufio_prefetch(struct dm_bufio_client *c,
 		       sector_t block, unsigned n_blocks)
 {
Index: linux-dm/include/linux/dm-bufio.h
===================================================================
--- linux-dm.orig/include/linux/dm-bufio.h	2022-09-05 13:29:56.000000000 +0200
+++ linux-dm/include/linux/dm-bufio.h	2022-09-27 15:36:49.000000000 +0200
@@ -38,6 +38,14 @@ dm_bufio_client_create(struct block_devi
 void dm_bufio_client_destroy(struct dm_bufio_client *c);
 
 /*
+ * Lock and unlock the bufio client - this is needed if we want to call
+ * dm_bufio_get_unlocked.
+ */
+void dm_bufio_lock(struct dm_bufio_client *c);
+int dm_bufio_trylock(struct dm_bufio_client *c);
+void dm_bufio_unlock(struct dm_bufio_client *c);
+
+/*
  * Set the sector range.
  * When this function is called, there must be no I/O in progress on the bufio
  * client.
@@ -76,6 +84,13 @@ void *dm_bufio_new(struct dm_bufio_clien
 		   struct dm_buffer **bp);
 
 /*
+ * Like dm_bufio_get, but assume that the client is already locked with
+ * dm_bufio_lock/dm_bufio_unlock. dm_bufio_release should not be called;
+ * the caller should call dm_bufio_unlock to release the buffer.
+ */
+void *dm_bufio_get_unlocked(struct dm_bufio_client *c, sector_t block);
+
+/*
  * Prefetch the specified blocks to the cache.
  * The function starts to read the blocks and returns without waiting for
  * I/O to finish.