[dm-devel] [PATCH 2/3] dm-bufio: Add DM_BUFIO_GET_CANT_SLEEP

[Nathan Huckleberry]

Add an optional flag that ensures dm_bufio_get does not sleep.  This
allows the dm-bufio cache to be queried from interrupt context.

To ensure that dm-bufio does not sleep, dm-bufio must use a spinlock
instead of a mutex.  Additionally, to avoid deadlocks, special care must
be taken so that dm-bufio does not sleep while holding the spinlock.

DM_BUFIO_GET_CANT_SLEEP is useful in some contexts, such as dm-verity,
so that we can query the dm-bufio cache in a tasklet.  If the required
data is cached, processing can be handled immediately in the tasklet
instead of waiting for a work-queue job to be scheduled.  This can
reduce latency when there is high CPU load and memory pressure.

Signed-off-by: Nathan Huckleberry <nhuck at google.com>
---
 drivers/md/dm-bufio.c    | 26 ++++++++++++++++++++++----
 include/linux/dm-bufio.h |  5 +++++
 2 files changed, 27 insertions(+), 4 deletions(-)

diff --git a/drivers/md/dm-bufio.c b/drivers/md/dm-bufio.c
index ad5603eb12e3..3edeca7cfca6 100644
--- a/drivers/md/dm-bufio.c
+++ b/drivers/md/dm-bufio.c
@@ -81,6 +81,8 @@
  */
 struct dm_bufio_client {
 	struct mutex lock;
+	spinlock_t spinlock;
+	unsigned long spinlock_flags;
 
 	struct list_head lru[LIST_SIZE];
 	unsigned long n_buffers[LIST_SIZE];
@@ -90,6 +92,7 @@ struct dm_bufio_client {
 	s8 sectors_per_block_bits;
 	void (*alloc_callback)(struct dm_buffer *);
 	void (*write_callback)(struct dm_buffer *);
+	bool may_sleep;
 
 	struct kmem_cache *slab_buffer;
 	struct kmem_cache *slab_cache;
@@ -167,17 +170,26 @@ struct dm_buffer {
 
 static void dm_bufio_lock(struct dm_bufio_client *c)
 {
-	mutex_lock_nested(&c->lock, dm_bufio_in_request());
+	if (c->may_sleep)
+		mutex_lock_nested(&c->lock, dm_bufio_in_request());
+	else
+		spin_lock_irqsave_nested(&c->spinlock, c->spinlock_flags, dm_bufio_in_request());
 }
 
 static int dm_bufio_trylock(struct dm_bufio_client *c)
 {
-	return mutex_trylock(&c->lock);
+	if (c->may_sleep)
+		return mutex_trylock(&c->lock);
+	else
+		return spin_trylock_irqsave(&c->spinlock, c->spinlock_flags);
 }
 
 static void dm_bufio_unlock(struct dm_bufio_client *c)
 {
-	mutex_unlock(&c->lock);
+	if (c->may_sleep)
+		mutex_unlock(&c->lock);
+	else
+		spin_unlock_irqrestore(&c->spinlock, c->spinlock_flags);
 }
 
 /*----------------------------------------------------------------*/
@@ -878,7 +890,7 @@ static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client
 	 * be allocated.
 	 */
 	while (1) {
-		if (dm_bufio_cache_size_latch != 1) {
+		if (dm_bufio_cache_size_latch != 1 && c->may_sleep) {
 			b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 			if (b)
 				return b;
@@ -1041,6 +1053,7 @@ static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
 	if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
 		return NULL;
 
+
 	b->hold_count++;
 	__relink_lru(b, test_bit(B_DIRTY, &b->state) ||
 		     test_bit(B_WRITING, &b->state));
@@ -1748,12 +1761,17 @@ struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsign
 	c->alloc_callback = alloc_callback;
 	c->write_callback = write_callback;
 
+	c->may_sleep = true;
+	if (flags & DM_BUFIO_GET_CANT_SLEEP)
+		c->may_sleep = false;
+
 	for (i = 0; i < LIST_SIZE; i++) {
 		INIT_LIST_HEAD(&c->lru[i]);
 		c->n_buffers[i] = 0;
 	}
 
 	mutex_init(&c->lock);
+	spin_lock_init(&c->spinlock);
 	INIT_LIST_HEAD(&c->reserved_buffers);
 	c->need_reserved_buffers = reserved_buffers;
 
diff --git a/include/linux/dm-bufio.h b/include/linux/dm-bufio.h
index e21480715255..2a78f0cb8e71 100644
--- a/include/linux/dm-bufio.h
+++ b/include/linux/dm-bufio.h
@@ -17,6 +17,11 @@
 struct dm_bufio_client;
 struct dm_buffer;
 
+/*
+ * Flags for dm_bufio_client_create
+ */
+#define DM_BUFIO_GET_CANT_SLEEP 0x1
+
 /*
  * Create a buffered IO cache on a given device
  */
-- 
2.37.1.359.gd136c6c3e2-goog