readahead: fix sequential read cache miss detection
authorDamien Ramonda <damien.ramonda@intel.com>
Tue, 12 Nov 2013 23:08:16 +0000 (15:08 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 13 Nov 2013 03:09:09 +0000 (12:09 +0900)
The kernel's readahead algorithm sometimes interprets random read
accesses as sequential and triggers unnecessary data prefecthing from
storage device (impacting random read average latency).

In order to identify sequential cache read misses, the readahead
algorithm intends to check whether offset - previous offset == 1
(trivial sequential reads) or offset - previous offset == 0 (sequential
reads not aligned on page boundary):

  if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)

The current offset is stored in the "offset" variable of type "pgoff_t"
(unsigned long), while previous offset is stored in "ra->prev_pos" of
type "loff_t" (long long).  Therefore, operands of the if statement are
implicitly converted to type long long.  Consequently, when previous
offset > current offset (which happens on random pattern), the if
condition is true and access is wrongly interpeted as sequential.  An
unnecessary data prefetching is triggered, impacting the average random
read latency.

Storing the previous offset value in a "pgoff_t" variable (unsigned
long) fixes the sequential read detection logic.

Signed-off-by: Damien Ramonda <damien.ramonda@intel.com>
Reviewed-by: Fengguang Wu <fengguang.wu@intel.com>
Acked-by: Pierre Tardy <pierre.tardy@intel.com>
Acked-by: David Cohen <david.a.cohen@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/readahead.c

index 50241836fe82d8d986e458172f13d8d8f2bb2f62..7cdbb44aa90bd99bef05f43ee27cc13b2121ed7a 100644 (file)
@@ -401,6 +401,7 @@ ondemand_readahead(struct address_space *mapping,
                   unsigned long req_size)
 {
        unsigned long max = max_sane_readahead(ra->ra_pages);
+       pgoff_t prev_offset;
 
        /*
         * start of file
@@ -452,8 +453,11 @@ ondemand_readahead(struct address_space *mapping,
 
        /*
         * sequential cache miss
+        * trivial case: (offset - prev_offset) == 1
+        * unaligned reads: (offset - prev_offset) == 0
         */
-       if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
+       prev_offset = (unsigned long long)ra->prev_pos >> PAGE_CACHE_SHIFT;
+       if (offset - prev_offset <= 1UL)
                goto initial_readahead;
 
        /*