[PATCH RFC 2/2] ext4: speed-up releasing blocks on commit

[PATCH RFC 2/2] ext4: speed-up releasing blocks on commit

[Andrey Sidorov]

Improve mb_free_blocks speed by clearing bits all at once instead of
going one by one. Rebuild buddy bitmap by going up only once per range
instead of once per block.
This algorithm relies on buddy bitmap to be correct and it can't
handle ranges with already freed blocks producing incorrect results in
these conditions.
I will add checking for error cases is to be added as soon as I get
some inputs on how to do that (and if this patch has some interest at
all). To simplify things I'd add mb_test_and_clear_bits() that will
clear given bit range, but will revert any changes made and
mb_free_blocks will go via slow one by one path in this case. But if
bitmap itself is ok, should buddies be checked?
Suggestions are welcome. Thanks!

This change was tested with AGGRESSIVE_CHECK enabled and
"mb_check_counter++ % 100" commented out in __mb_check_buddy.

Commit after unlinking 60G fallocated file (no bigalloc):
X86 before (linux 3.6-rc4):
[ 1257.114601] ext4_journal_commit_callback: enter
[ 1259.848202] ext4_journal_commit_callback: leave

X86 after:
[ 2157.358295] ext4_journal_commit_callback: enter
[ 2157.363268] ext4_journal_commit_callback: leave

MIPS before (linux 2.6.37):
<7>[240123.058000] release_blocks_on_commit: enter
<7>[240127.551000] release_blocks_on_commit: leave

MIPS after:
<7>[240313.049000] release_blocks_on_commit: enter
<7>[240313.056000] release_blocks_on_commit: leave

---
 fs/ext4/mballoc.c |  100 ++++++++++++++++++++++++++----------------------------
 1 file changed, 49 insertions(+), 51 deletions(-)

Index: linux-3.6-rc6/fs/ext4/mballoc.c
===================================================================
--- linux-3.6-rc6.orig/fs/ext4/mballoc.c	2012-09-16 17:58:51.000000000 -0400
+++ linux-3.6-rc6/fs/ext4/mballoc.c	2012-09-17 12:13:02.726293353 -0400
@@ -1254,14 +1254,27 @@ void ext4_set_bits(void *bm, int cur, in
 	}
 }

+static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
+{
+	if (mb_test_bit(*bit + side, bitmap)) {
+		mb_clear_bit(*bit, bitmap);
+		(*bit) -= side;
+		return 1;
+	}
+	else {
+		mb_set_bit(*bit += side, bitmap);
+		return -1;
+	}
+}
+
 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
 			  int first, int count)
 {
 	int block = 0;
 	int max = 0;
-	int order;
+	int order = 1;
+	int start = first, end = first + count - 1;
 	void *buddy;
-	void *buddy2;
 	struct super_block *sb = e4b->bd_sb;

 	BUG_ON(first + count > (sb->s_blocksize << 3));
@@ -1283,57 +1296,42 @@ static void mb_free_blocks(struct inode
 	else if (!block && !max)
 		e4b->bd_info->bb_fragments++;

-	/* let's maintain buddy itself */
-	while (count-- > 0) {
-		block = first++;
-		order = 0;
-
-		if (!mb_test_bit(block, e4b->bd_bitmap)) {
-			ext4_fsblk_t blocknr;
-
-			blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
-			blocknr += EXT4_C2B(EXT4_SB(sb), block);
-			ext4_grp_locked_error(sb, e4b->bd_group,
-					      inode ? inode->i_ino : 0,
-					      blocknr,
-					      "freeing already freed block "
-					      "(bit %u)", block);
+	if ((start & 1) && mb_test_bit(start - 1, e4b->bd_bitmap))
+		start++;
+	mb_clear_bits(e4b->bd_bitmap, first, count);
+	if (!(end & 1) && mb_test_bit(end + 1, e4b->bd_bitmap))
+		end--;
+
+	e4b->bd_info->bb_counters[0] += count;
+
+	if (start > end) goto done;
+
+	start >>= 1;
+	end >>= 1;
+	buddy = mb_find_buddy(e4b, order, &max);
+
+	while (buddy) {
+		void *buddy2;
+
+		if (start & 1)
+			e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&start,
buddy, -1);
+		if (!(end & 1))
+			e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&end, buddy, 1);
+		if (start > end)
+			break;
+		order++;
+
+		if (start == end || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
+			mb_clear_bits(buddy, start, end - start + 1);
+			e4b->bd_info->bb_counters[order - 1] += end - start + 1;
+			break;
 		}
-		mb_clear_bit(block, e4b->bd_bitmap);
-		e4b->bd_info->bb_counters[order]++;
-
-		/* start of the buddy */
-		buddy = mb_find_buddy(e4b, order, &max);
-
-		do {
-			block &= ~1UL;
-			if (mb_test_bit(block, buddy) ||
-					mb_test_bit(block + 1, buddy))
-				break;
-
-			/* both the buddies are free, try to coalesce them */
-			buddy2 = mb_find_buddy(e4b, order + 1, &max);
-
-			if (!buddy2)
-				break;
-
-			if (order > 0) {
-				/* for special purposes, we don't set
-				 * free bits in bitmap */
-				mb_set_bit(block, buddy);
-				mb_set_bit(block + 1, buddy);
-			}
-			e4b->bd_info->bb_counters[order]--;
-			e4b->bd_info->bb_counters[order]--;
-
-			block = block >> 1;
-			order++;
-			e4b->bd_info->bb_counters[order]++;

Re: [PATCH RFC 2/2] ext4: speed-up releasing blocks on commit

[Theodore Ts'o]

On Tue, Sep 18, 2012 at 01:00:58AM +0400, Andrey Sidorov wrote:
> Improve mb_free_blocks speed by clearing bits all at once instead of
> going one by one. Rebuild buddy bitmap by going up only once per range
> instead of once per block.
> This algorithm relies on buddy bitmap to be correct and it can't
> handle ranges with already freed blocks producing incorrect results in
> these conditions.
> I will add checking for error cases is to be added as soon as I get
> some inputs on how to do that (and if this patch has some interest at
> all). 

Thanks, this patch is interesting as well --- but yes, we do need to
make sure the changes will not make any potential inconsistency in the
buddy bitmaps any worse that they already are.

The main thing we need to consider for the buddy bitmap code is that
it's one of the more subtle and complex bits of code in the ext4 code
base.  So changing it is always scary that there is some interesting
edge case that isn't quite right.

That means the two things we need to consider are (a) adding sanity
checking code which we can compile in or out to test to make sure the
right thing happens for all of the various corner cases, and (b)
making sure we do the right thing even if the buddy bitmaps have
gotten corrupted in memory some how.

Regards,

						- Ted