[PATCH for-next 0/2] optimise sbitmap deferred clear

[PATCH for-next 0/2] optimise sbitmap deferred clear

[Pavel Begunkov]

In short, sbitmap_deferred_clear() lights up pretty much in my profiler,
so these optimisations gave me ~1% more t-put. All the heavy stuff is in
[2/2]. Would love someones eye to check it.

I also want to replace cmpxchg() in that function with a single
atomic and, that's slightly lighter and also transfers it from
lock-free to wait-free, that's pretty neat.
The problem is that apparently there is a non-atomic_t atomic and,
and atomic_t is unsigned int but all the bitmap do unsigned long.
Advice is welcome.

Pavel Begunkov (2):
  sbitmap: optimise sbitmap_deferred_clear()
  sbitmap: remove swap_lock

 include/linux/sbitmap.h |  5 -----
 lib/sbitmap.c           | 21 +++++++--------------
 2 files changed, 7 insertions(+), 19 deletions(-)

-- 
2.24.0

[PATCH 1/2] sbitmap: optimise sbitmap_deferred_clear()

[Pavel Begunkov]

Because of spinlocks and atomics sbitmap_deferred_clear() have to reload
&sb->map[index] on each access even though the map address won't change.
Hint it by explicitly caching it in a variable.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 lib/sbitmap.c | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff --git a/lib/sbitmap.c b/lib/sbitmap.c
index 267aa7709416..49afb34e8340 100644
--- a/lib/sbitmap.c
+++ b/lib/sbitmap.c
@@ -14,30 +14,31 @@
  */
 static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
 {
+	struct sbitmap_word *map = &sb->map[index];
 	unsigned long mask, val;
 	bool ret = false;
 	unsigned long flags;
 
-	spin_lock_irqsave(&sb->map[index].swap_lock, flags);
+	spin_lock_irqsave(&map->swap_lock, flags);
 
-	if (!sb->map[index].cleared)
+	if (!map->cleared)
 		goto out_unlock;
 
 	/*
 	 * First get a stable cleared mask, setting the old mask to 0.
 	 */
-	mask = xchg(&sb->map[index].cleared, 0);
+	mask = xchg(&map->cleared, 0);
 
 	/*
 	 * Now clear the masked bits in our free word
 	 */
 	do {
-		val = sb->map[index].word;
-	} while (cmpxchg(&sb->map[index].word, val, val & ~mask) != val);
+		val = map->word;
+	} while (cmpxchg(&map->word, val, val & ~mask) != val);
 
 	ret = true;
 out_unlock:
-	spin_unlock_irqrestore(&sb->map[index].swap_lock, flags);
+	spin_unlock_irqrestore(&map->swap_lock, flags);
 	return ret;
 }
 
-- 
2.24.0

[PATCH 2/2] sbitmap: remove swap_lock

[Pavel Begunkov]

map->swap_lock serialises concurrent calls to sbitmap_deferred_clear(),
however that function is already works in atomic fashion and guarantees
to not loose bits while applying map->cleared bitmask.

Remove spinlocking in sbitmap_deferred_clear(). For a one-threaded
tag allocation heavy test on top of nullblk it yields ~1.0-1.5% t-put
increase, and according to perf 3% -> 1.5% cycle reduction of
sbitmap_get().

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 include/linux/sbitmap.h |  5 -----
 lib/sbitmap.c           | 14 +++-----------
 2 files changed, 3 insertions(+), 16 deletions(-)

diff --git a/include/linux/sbitmap.h b/include/linux/sbitmap.h
index e40d019c3d9d..74cc6384715e 100644
--- a/include/linux/sbitmap.h
+++ b/include/linux/sbitmap.h
@@ -32,11 +32,6 @@ struct sbitmap_word {
 	 * @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;
 
 /**
diff --git a/lib/sbitmap.c b/lib/sbitmap.c
index 49afb34e8340..238d9849f24b 100644
--- a/lib/sbitmap.c
+++ b/lib/sbitmap.c
@@ -16,13 +16,9 @@ static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
 {
 	struct sbitmap_word *map = &sb->map[index];
 	unsigned long mask, val;
-	bool ret = false;
-	unsigned long flags;
 
-	spin_lock_irqsave(&map->swap_lock, flags);
-
-	if (!map->cleared)
-		goto out_unlock;
+	if (!READ_ONCE(map->cleared))
+		return false;
 
 	/*
 	 * First get a stable cleared mask, setting the old mask to 0.
@@ -36,10 +32,7 @@ static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
 		val = map->word;
 	} while (cmpxchg(&map->word, val, val & ~mask) != val);
 
-	ret = true;
-out_unlock:
-	spin_unlock_irqrestore(&map->swap_lock, flags);
-	return ret;
+	return true;
 }
 
 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
@@ -81,7 +74,6 @@ 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;
 }
-- 
2.24.0