Commit ea86ea2c authored by Jens Axboe's avatar Jens Axboe

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: default avatarOmar Sandoval <osandov@fb.com>
Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
parent 531724ab
......@@ -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.
......
......@@ -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
*/
do {
val = sb->map[index].word;
} while (cmpxchg(&sb->map[index].word, val, val & ~mask) != val);
ret = true;
out_unlock:
spin_unlock(&sb->map[index].swap_lock);
return ret;
}
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 +257,36 @@ bool sbitmap_any_bit_clear(const struct sbitmap *sb)
}
EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear);
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);
else
weight += bitmap_weight(&word->cleared, word->depth);
}
return weight;
}
EXPORT_SYMBOL_GPL(sbitmap_weight);
static unsigned int sbitmap_weight(const struct sbitmap *sb)
{
return __sbitmap_weight(sb, true);
}
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 +578,8 @@ 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
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment