[PATCH] fs/buffer.c: make bh_lru_install() more efficient

[PATCH] fs/buffer.c: make bh_lru_install() more efficient

[Eric Biggers]

From: Eric Biggers <ebiggers@google.com>

To install a buffer_head into the cpu's LRU queue, bh_lru_install()
would construct a new copy of the queue and then memcpy it over the real
queue.  But it's easily possible to do the update in-place, which is
faster and simpler.  Some work can also be skipped if the buffer_head
was already in the queue.

As a microbenchmark I timed how long it takes to run sb_getblk()
10,000,000 times alternating between BH_LRU_SIZE + 1 blocks.
Effectively, this benchmarks looking up buffer_heads that are in the
page cache but not in the LRU:

	Before this patch: 1.758s
	After this patch: 1.653s

This patch also removes about 350 bytes of compiled code (on x86_64),
partly due to removal of the memcpy() which was being inlined+unrolled.

Signed-off-by: Eric Biggers <ebiggers@google.com>
---
 fs/buffer.c | 43 +++++++++++++++----------------------------
 1 file changed, 15 insertions(+), 28 deletions(-)

diff --git a/fs/buffer.c b/fs/buffer.c
index d21771fcf7d3..282ca52517bf 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -1273,44 +1273,31 @@ static inline void check_irqs_on(void)
 }
 
 /*
- * The LRU management algorithm is dopey-but-simple.  Sorry.
+ * Install a buffer_head into this cpu's LRU.  If not already in the LRU, it is
+ * inserted at the front, and the buffer_head at the back if any is evicted.
+ * Or, if already in the LRU it is moved to the front.
  */
 static void bh_lru_install(struct buffer_head *bh)
 {
-	struct buffer_head *evictee = NULL;
+	struct buffer_head *evictee = bh;
+	struct bh_lru *b;
+	int i;
 
 	check_irqs_on();
 	bh_lru_lock();
-	if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
-		struct buffer_head *bhs[BH_LRU_SIZE];
-		int in;
-		int out = 0;
-
-		get_bh(bh);
-		bhs[out++] = bh;
-		for (in = 0; in < BH_LRU_SIZE; in++) {
-			struct buffer_head *bh2 =
-				__this_cpu_read(bh_lrus.bhs[in]);
 
-			if (bh2 == bh) {
-				__brelse(bh2);
-			} else {
-				if (out >= BH_LRU_SIZE) {
-					BUG_ON(evictee != NULL);
-					evictee = bh2;
-				} else {
-					bhs[out++] = bh2;
-				}
-			}
+	b = this_cpu_ptr(&bh_lrus);
+	for (i = 0; i < BH_LRU_SIZE; i++) {
+		swap(evictee, b->bhs[i]);
+		if (evictee == bh) {
+			bh_lru_unlock();
+			return;
 		}
-		while (out < BH_LRU_SIZE)
-			bhs[out++] = NULL;
-		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
 	}
-	bh_lru_unlock();
 
-	if (evictee)
-		__brelse(evictee);
+	get_bh(bh);
+	bh_lru_unlock();
+	brelse(evictee);
 }
 
 /*
-- 
2.11.0

Re: [PATCH] fs/buffer.c: make bh_lru_install() more efficient

[Eric Biggers]

On Thu, Dec 29, 2016 at 01:34:45PM -0600, Eric Biggers wrote:
> From: Eric Biggers <ebiggers@google.com>
> 
> To install a buffer_head into the cpu's LRU queue, bh_lru_install()
> would construct a new copy of the queue and then memcpy it over the real
> queue.  But it's easily possible to do the update in-place, which is
> faster and simpler.  Some work can also be skipped if the buffer_head
> was already in the queue.
> 
> As a microbenchmark I timed how long it takes to run sb_getblk()
> 10,000,000 times alternating between BH_LRU_SIZE + 1 blocks.
> Effectively, this benchmarks looking up buffer_heads that are in the
> page cache but not in the LRU:
> 
> 	Before this patch: 1.758s
> 	After this patch: 1.653s
> 
> This patch also removes about 350 bytes of compiled code (on x86_64),
> partly due to removal of the memcpy() which was being inlined+unrolled.
> 
> Signed-off-by: Eric Biggers <ebiggers@google.com>

Ping?  Al, do you have any interest in taking this patch?

- Eric

Re: [PATCH] fs/buffer.c: make bh_lru_install() more efficient

[Sean Fu]

On Sat, Mar 25, 2017 at 08:34:30PM -0700, Eric Biggers wrote:
> On Thu, Dec 29, 2016 at 01:34:45PM -0600, Eric Biggers wrote:
> > From: Eric Biggers <ebiggers@google.com>
> > 
> > To install a buffer_head into the cpu's LRU queue, bh_lru_install()
> > would construct a new copy of the queue and then memcpy it over the real
> > queue.  But it's easily possible to do the update in-place, which is
> > faster and simpler.  Some work can also be skipped if the buffer_head
> > was already in the queue.
> > 
> > As a microbenchmark I timed how long it takes to run sb_getblk()
> > 10,000,000 times alternating between BH_LRU_SIZE + 1 blocks.
> > Effectively, this benchmarks looking up buffer_heads that are in the
> > page cache but not in the LRU:
> > 
> > 	Before this patch: 1.758s
> > 	After this patch: 1.653s
> > 
> > This patch also removes about 350 bytes of compiled code (on x86_64),
> > partly due to removal of the memcpy() which was being inlined+unrolled.
> > 
> > Signed-off-by: Eric Biggers <ebiggers@google.com>
> 
> Ping?  Al, do you have any interest in taking this patch?
> 
> - Eric
IMO, It is great patch.
Could you please share your microbenchmark steps?
I wanna take a deep look at it.