sbitmap: ammortize cost of clearing bits

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.

Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

1 files changed, 27 insertions, 6 deletions

diff --git a/include/linux/sbitmap.h b/include/linux/sbitmap.h
index 804a50983ec5..81359d45751e 100644
--- a/include/linux/sbitmap.h
+++ b/include/linux/sbitmap.h
@@ -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;
 
         /**
-         * @depth: Number of bits being used in @word.
+         * @word: word holding free bits
          */
-        unsigned long depth;
+        unsigned long word ____cacheline_aligned_in_smp;
+
+        /**
+         * @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)
 {
@@ -321,8 +344,6 @@ static inline int sbitmap_test_bit(struct sbitmap *sb, unsigned int bitnr)
         return test_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
 }
 
-unsigned int sbitmap_weight(const struct sbitmap *sb);
-
 /**
  * sbitmap_show() - Dump &struct sbitmap information to a &struct seq_file.
  * @sb: Bitmap to show.