readahead: sequential mmap readahead
authorWu Fengguang <fengguang.wu@intel.com>
Tue, 16 Jun 2009 22:31:28 +0000 (15:31 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 17 Jun 2009 02:47:29 +0000 (19:47 -0700)
Auto-detect sequential mmap reads and do readahead for them.

The sequential mmap readahead will be triggered when
- sync readahead: it's a major fault and (prev_offset == offset-1);
- async readahead: minor fault on PG_readahead page with valid readahead state.

The benefits of doing readahead instead of read-around:
- less I/O wait thanks to async readahead
- double real I/O size and no more cache hits

The single stream case is improved a little.
For 100,000 sequential mmap reads:

                                    user       system    cpu        total
(1-1)  plain -mm, 128KB readaround: 3.224      2.554     48.40%     11.838
(1-2)  plain -mm, 256KB readaround: 3.170      2.392     46.20%     11.976
(2)  patched -mm, 128KB readahead:  3.117      2.448     47.33%     11.607

The patched (2) has smallest total time, since it has no cache hit overheads
and less I/O block time(thanks to async readahead). Here the I/O size
makes no much difference, since there's only one single stream.

Note that (1-1)'s real I/O size is 64KB and (1-2)'s real I/O size is 128KB,
since the half of the read-around pages will be readahead cache hits.

This is going to make _real_ differences for _concurrent_ IO streams.

Cc: Nick Piggin <npiggin@suse.de>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/filemap.c

index 827536485599c5e1bfdd637697902252682123b5..99977f0a94e4719f9a3644ff018aa35ea3cfbc6f 100644 (file)
@@ -1472,7 +1472,8 @@ static void do_sync_mmap_readahead(struct vm_area_struct *vma,
        if (VM_RandomReadHint(vma))
                return;
 
-       if (VM_SequentialReadHint(vma)) {
+       if (VM_SequentialReadHint(vma) ||
+                       offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
                page_cache_sync_readahead(mapping, ra, file, offset, 1);
                return;
        }