sbitmap optimizations

sbitmap maintains a set of words that we use to set and clear bits, with each bit representing a tag for blk-mq. Even though we spread the bits out and maintain a hint cache, one particular bit allocated will end up being cleared in the exact same spot.

This introduces batched clearing of bits. Instead of clearing a given bit, the same bit is set in a cleared/free mask instead. If we fail allocating a bit from a given word, then we check the free mask, and batch move those cleared bits at that time. This trades 64 atomic bitops for 2 cmpxchg().

In a threaded poll test case, half the overhead of getting and clearing tags is removed with this change. On another poll test case with a single thread, performance is unchanged.

Signed-off-by Jens Axboe <[email protected]>

---
 include/linux/sbitmap.h | 31 +++++++++++++---
 lib/sbitmap.c           | 80 +++++++++++++++++++++++++++++++++++++----
 2 files changed, 100 insertions(+), 11 deletions(-)

include/linux/sbitmap.h +1 -1

index 804a50983ec5..07f117ee19dc 100644
@@ -30,14 +30,24 @@ struct seq_file;
  */
 struct sbitmap_word {
 	/**
-	 * @word: The bitmap word itself.
+	 * @depth: Number of bits being used in @word/@cleared

include/linux/sbitmap.h +22 -0

+	/**
+	 * @cleared: word holding cleared bits
+	 */
+	unsigned long cleared ____cacheline_aligned_in_smp;
+
+	/**
+	 * @swap_lock: Held while swapping word <-> cleared
+	 */
+	spinlock_t swap_lock;
 } ____cacheline_aligned_in_smp;
 
 /**

@@ -310,6 +320,19 @@ static inline void sbitmap_clear_bit(struct sbitmap *sb, unsigned int bitnr)
 	clear_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
 }
 
+/*
+ * This one is special, since it doesn't actually clear the bit, rather it
+ * sets the corresponding bit in the ->cleared mask instead. Paired with
+ * the caller doing sbitmap_batch_clear() if a given index is full, which
+ * will clear the previously freed entries in the corresponding ->word.
+ */
+static inline void sbitmap_deferred_clear_bit(struct sbitmap *sb, unsigned int bitnr)
+{
+	unsigned long *addr = &sb->map[SB_NR_TO_INDEX(sb, bitnr)].cleared;
+
+	set_bit(SB_NR_TO_BIT(sb, bitnr), addr);
+}
+
 static inline void sbitmap_clear_bit_unlock(struct sbitmap *sb,
 					    unsigned int bitnr)
 {

lib/sbitmap.c +23 -0

index 45cab6bbc1c7..f6a9553617bd 100644
@@ -59,6 +59,7 @@ int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
 	for (i = 0; i < sb->map_nr; i++) {
 		sb->map[i].depth = min(depth, bits_per_word);
 		depth -= sb->map[i].depth;
+		spin_lock_init(&sb->map[i].swap_lock);
 	}
 	return 0;
 }

@@ -111,6 +112,57 @@ static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
 	return nr;
 }
 
+/*
+ * See if we have deferred clears that we can batch move
+ */
+static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
+{
+	unsigned long mask, val;
+	bool ret = false;
+
+	spin_lock(&sb->map[index].swap_lock);
+
+	if (!sb->map[index].cleared)
+		goto out_unlock;
+
+	/*
+	 * First get a stable cleared mask, setting the old mask to 0.
+	 */
+	do {
+		mask = sb->map[index].cleared;
+	} while (cmpxchg(&sb->map[index].cleared, mask, 0) != mask);
+
+	/*
+	 * Now clear the masked bits in our free word

lib/sbitmap.c +20 -4

+static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
+				     unsigned int alloc_hint, bool round_robin)
+{
+	int nr;
+
+	do {
+		nr = __sbitmap_get_word(&sb->map[index].word,
+					sb->map[index].depth, alloc_hint,
+					!round_robin);
+		if (nr != -1)
+			break;
+		if (!sbitmap_deferred_clear(sb, index))
+			break;
+	} while (1);
+
+	return nr;
+}
+
 int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin)
 {
 	unsigned int i, index;

@@ -129,9 +181,8 @@ int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin)
 		alloc_hint = 0;
 
 	for (i = 0; i < sb->map_nr; i++) {
-		nr = __sbitmap_get_word(&sb->map[index].word,
-					sb->map[index].depth, alloc_hint,
-					!round_robin);
+		nr = sbitmap_find_bit_in_index(sb, index, alloc_hint,
+						round_robin);
 		if (nr != -1) {
 			nr += index << sb->shift;
 			break;

@@ -206,23 +257,37 @@ bool sbitmap_any_bit_clear(const struct sbitmap *sb)
 }
 EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear);
 
-unsigned int sbitmap_weight(const struct sbitmap *sb)

lib/sbitmap.c +16 -2

+		else
+			weight += bitmap_weight(&word->cleared, word->depth);
 	}
 	return weight;
 }
+
+unsigned int sbitmap_weight(const struct sbitmap *sb)
+{
+	return __sbitmap_weight(sb, true);
+}
 EXPORT_SYMBOL_GPL(sbitmap_weight);
 
+static unsigned int sbitmap_cleared(const struct sbitmap *sb)
+{
+	return __sbitmap_weight(sb, false);
+}
+
 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
 {
 	seq_printf(m, "depth=%u\n", sb->depth);
-	seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
+	seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb));
+	seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
 	seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
 	seq_printf(m, "map_nr=%u\n", sb->map_nr);
 }

@@ -514,7 +579,8 @@ EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
 			 unsigned int cpu)
 {
-	sbitmap_clear_bit_unlock(&sbq->sb, nr);
+	sbitmap_deferred_clear_bit(&sbq->sb, nr);
+
 	/*
 	 * Pairs with the memory barrier in set_current_state() to ensure the
 	 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker

Signature · 1 line

2.17.1

Quoted 35 lines · Jens Axboe

sbitmap maintains a set of words that we use to set and clear bits, with each bit representing a tag for blk-mq. Even though we spread the bits out and maintain a hint cache, one particular bit allocated will end up being cleared in the exact same spot.

This introduces batched clearing of bits. Instead of clearing a given bit, the same bit is set in a cleared/free mask instead. If we fail allocating a bit from a given word, then we check the free mask, and batch move those cleared bits at that time. This trades 64 atomic bitops for 2 cmpxchg().

In a threaded poll test case, half the overhead of getting and clearing tags is removed with this change. On another poll test case with a single thread, performance is unchanged.

Signed-off-by Jens Axboe <[email protected]>

---
 include/linux/sbitmap.h | 31 +++++++++++++---
 lib/sbitmap.c           | 80 +++++++++++++++++++++++++++++++++++++----
 2 files changed, 100 insertions(+), 11 deletions(-)

include/linux/sbitmap.h +2 -2

index 804a50983ec5..07f117ee19dc 100644
@@ -30,14 +30,24 @@ struct seq_file;
  */
 struct sbitmap_word {
 	/**
-	 * @word: The bitmap word itself.
+	 * @depth: Number of bits being used in @word/@cleared
 	 */
-	unsigned long word;
+	unsigned long depth;
 

include/linux/sbitmap.h struct seq_file; · line 40

Quoted 6 lines

include/linux/sbitmap.h +2 -2

 	/**
-	 * @depth: Number of bits being used in @word.
+	 * @word: word holding free bits
 	 */
-	unsigned long depth;
+	unsigned long word ____cacheline_aligned_in_smp;

Still splitting up word and depth in separate cachelines?

Okay, I couldn't find any holes in this one :)

Quoted 8 lines

include/linux/sbitmap.h +1 -1

-unsigned int sbitmap_weight(const struct sbitmap *sb)
+static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
 {
 	unsigned int i, weight = 0;
 
 	for (i = 0; i < sb->map_nr; i++) {
 		const struct sbitmap_word *word = &sb->map[i];

include/linux/sbitmap.h

Quoted 3 lines

include/linux/sbitmap.h +2 -1

-		weight += bitmap_weight(&word->word, word->depth);
+		if (set)
+			weight += bitmap_weight(&word->word, word->depth);

Should probably do

			weight -= bitmap_weight(&word->cleared, word->depth);

too, right?

Even if we have no waiters on any of the sbitmap_queue wait states, we still have to loop every entry to check. We do this for every IO, so the cost adds up.

Shift a bit of the cost to the slow path, when we actually have waiters. Wrap prepare_to_wait_exclusive() and finish_wait(), so we can maintain an internal count of how many are currently active. Then we can simply check this count in sbq_wake_ptr() and not have to loop if we don't have any sleepers.

Convert the two users of sbitmap with waiting, blk-mq-tag and iSCSI.

Signed-off-by Jens Axboe <[email protected]>

---
 block/blk-mq-tag.c                       | 11 ++++----
 drivers/target/iscsi/iscsi_target_util.c | 12 +++++----
 include/linux/sbitmap.h                  | 34 ++++++++++++++++++++++++
 lib/sbitmap.c                            | 28 +++++++++++++++++++
 4 files changed, 74 insertions(+), 11 deletions(-)

Quoted 36 lines · Omar Sandoval

Quoted 35 lines · Jens Axboe

sbitmap maintains a set of words that we use to set and clear bits, with each bit representing a tag for blk-mq. Even though we spread the bits out and maintain a hint cache, one particular bit allocated will end up being cleared in the exact same spot.

This introduces batched clearing of bits. Instead of clearing a given bit, the same bit is set in a cleared/free mask instead. If we fail allocating a bit from a given word, then we check the free mask, and batch move those cleared bits at that time. This trades 64 atomic bitops for 2 cmpxchg().

In a threaded poll test case, half the overhead of getting and clearing tags is removed with this change. On another poll test case with a single thread, performance is unchanged.

Signed-off-by Jens Axboe <[email protected]>

---
 include/linux/sbitmap.h | 31 +++++++++++++---
 lib/sbitmap.c           | 80 +++++++++++++++++++++++++++++++++++++----
 2 files changed, 100 insertions(+), 11 deletions(-)

include/linux/sbitmap.h +2 -2

index 804a50983ec5..07f117ee19dc 100644
@@ -30,14 +30,24 @@ struct seq_file;
  */
 struct sbitmap_word {
 	/**
-	 * @word: The bitmap word itself.
+	 * @depth: Number of bits being used in @word/@cleared
 	 */
-	unsigned long word;
+	unsigned long depth;
 

Quoted 8 lines

include/linux/sbitmap.h struct seq_file; · line 40

Quoted 6 lines

include/linux/sbitmap.h +2 -2

 	/**
-	 * @depth: Number of bits being used in @word.
+	 * @word: word holding free bits
 	 */
-	unsigned long depth;
+	unsigned long word ____cacheline_aligned_in_smp;

Still splitting up word and depth in separate cachelines?

Yeah, I mentioned that in one of the other postings, there's still a definite win to doing that.

Quoted 1 line

Okay, I couldn't find any holes in this one :)

Good to hear that :-)

Quoted 15 lines

include/linux/sbitmap.h struct seq_file;

Quoted 11 lines

include/linux/sbitmap.h +3 -2

-unsigned int sbitmap_weight(const struct sbitmap *sb)
+static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
 {
 	unsigned int i, weight = 0;
 
 	for (i = 0; i < sb->map_nr; i++) {
 		const struct sbitmap_word *word = &sb->map[i];
 
-		weight += bitmap_weight(&word->word, word->depth);
+		if (set)
+			weight += bitmap_weight(&word->word, word->depth);

Should probably do

			weight -= bitmap_weight(&word->cleared, word->depth);

Quoted 1 line

too, right?

We only use these for the debugfs stuff, how about I just make it static instead?