fs: fix performance regression in clean_bdev_aliases()
authorJan Kara <jack@suse.cz>
Wed, 6 Sep 2017 23:21:24 +0000 (16:21 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 7 Sep 2017 00:27:26 +0000 (17:27 -0700)
Commit e64855c6cfaa ("fs: Add helper to clean bdev aliases under a bh
and use it") added a wrapper for clean_bdev_aliases() that invalidates
bdev aliases underlying a single buffer head.

However this has caused a performance regression for bonnie++ benchmark
on ext4 filesystem when delayed allocation is turned off (ext3 mode) -
average of 3 runs:

  Hmean SeqOut Char  164787.55 (  0.00%) 107189.06 (-34.95%)
  Hmean SeqOut Block 219883.89 (  0.00%) 168870.32 (-23.20%)

The reason for this regression is that clean_bdev_aliases() is slower
when called for a single block because pagevec_lookup() it uses will end
up iterating through the radix tree until it finds a page (which may
take a while) but we are only interested whether there's a page at a
particular index.

Fix the problem by using pagevec_lookup_range() instead which avoids the
needless iteration.

Fixes: e64855c6cfaa ("fs: Add helper to clean bdev aliases under a bh and use it")
Link: http://lkml.kernel.org/r/20170726114704.7626-5-jack@suse.cz
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Jens Axboe <axboe@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/buffer.c

index 5b20893708e29bb4004436dec756ce8f6397a8d4..7e531bb356bd3ad58757f550fdec459106644d44 100644 (file)
@@ -1627,19 +1627,18 @@ void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
        struct pagevec pvec;
        pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
        pgoff_t end;
        struct pagevec pvec;
        pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
        pgoff_t end;
-       int i;
+       int i, count;
        struct buffer_head *bh;
        struct buffer_head *head;
 
        end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
        pagevec_init(&pvec, 0);
        struct buffer_head *bh;
        struct buffer_head *head;
 
        end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
        pagevec_init(&pvec, 0);
-       while (index <= end && pagevec_lookup(&pvec, bd_mapping, &index,
-                       min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
-               for (i = 0; i < pagevec_count(&pvec); i++) {
+       while (pagevec_lookup_range(&pvec, bd_mapping, &index, end,
+                                   PAGEVEC_SIZE)) {
+               count = pagevec_count(&pvec);
+               for (i = 0; i < count; i++) {
                        struct page *page = pvec.pages[i];
 
                        struct page *page = pvec.pages[i];
 
-                       if (page->index > end)
-                               break;
                        if (!page_has_buffers(page))
                                continue;
                        /*
                        if (!page_has_buffers(page))
                                continue;
                        /*
@@ -1669,6 +1668,9 @@ unlock_page:
                }
                pagevec_release(&pvec);
                cond_resched();
                }
                pagevec_release(&pvec);
                cond_resched();
+               /* End of range already reached? */
+               if (index > end || !index)
+                       break;
        }
 }
 EXPORT_SYMBOL(clean_bdev_aliases);
        }
 }
 EXPORT_SYMBOL(clean_bdev_aliases);