block: Added in stricter no merge semantics for block I/O

Updated 'nomerges' tunable to accept a value of '2' - indicating that _no_
merges at all are to be attempted (not even the simple one-hit cache).

The following table illustrates the additional benefit - 5 minute runs of
a random I/O load were applied to a dozen devices on a 16-way x86_64 system.

nomerges        Throughput      %System         Improvement (tput / %sys)
--------        ------------    -----------     -------------------------
0               12.45 MB/sec    0.669365609
1               12.50 MB/sec    0.641519199     0.40% / 2.71%
2               12.52 MB/sec    0.639849750     0.56% / 2.96%

Signed-off-by: Alan D. Brunelle <alan.brunelle@hp.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

5 files changed, 39 insertions, 10 deletions

diff --git a/Documentation/ABI/testing/sysfs-block b/Documentation/ABI/testing/sysfs-block
index d2f90334bb93..4873c759d535 100644
--- a/Documentation/ABI/testing/sysfs-block
+++ b/Documentation/ABI/testing/sysfs-block
@@ -128,3 +128,17 @@ Description:
                 preferred request size for workloads where sustained
                 throughput is desired.  If no optimal I/O size is
                 reported this file contains 0.
+
+What:           /sys/block/<disk>/queue/nomerges
+Date:           January 2010
+Contact:
+Description:
+                Standard I/O elevator operations include attempts to
+                merge contiguous I/Os. For known random I/O loads these
+                attempts will always fail and result in extra cycles
+                being spent in the kernel. This allows one to turn off
+                this behavior on one of two ways: When set to 1, complex
+                merge checks are disabled, but the simple one-shot merges
+                with the previous I/O request are enabled. When set to 2,
+                all merge tries are disabled. The default value is 0 -
+                which enables all types of merge tries.
diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index e164403f60e1..f65274081c8d 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -25,11 +25,11 @@ size allowed by the hardware.
 
 nomerges (RW)
 -------------
-This enables the user to disable the lookup logic involved with IO merging
-requests in the block layer. Merging may still occur through a direct
-1-hit cache, since that comes for (almost) free. The IO scheduler will not
-waste cycles doing tree/hash lookups for merges if nomerges is 1. Defaults
-to 0, enabling all merges.
+This enables the user to disable the lookup logic involved with IO
+merging requests in the block layer. By default (0) all merges are
+enabled. When set to 1 only simple one-hit merges will be tried. When
+set to 2 no merge algorithms will be tried (including one-hit or more
+complex tree/hash lookups).
 
 nr_requests (RW)
 ----------------
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 8606c9543fdd..e85442415db3 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -189,7 +189,8 @@ static ssize_t queue_nonrot_store(struct request_queue *q, const char *page,
 
 static ssize_t queue_nomerges_show(struct request_queue *q, char *page)
 {
-        return queue_var_show(blk_queue_nomerges(q), page);
+        return queue_var_show((blk_queue_nomerges(q) << 1) |
+                               blk_queue_noxmerges(q), page);
 }
 
 static ssize_t queue_nomerges_store(struct request_queue *q, const char *page,
@@ -199,10 +200,12 @@ static ssize_t queue_nomerges_store(struct request_queue *q, const char *page,
         ssize_t ret = queue_var_store(&nm, page, count);
 
         spin_lock_irq(q->queue_lock);
-        if (nm)
+        queue_flag_clear(QUEUE_FLAG_NOMERGES, q);
+        queue_flag_clear(QUEUE_FLAG_NOXMERGES, q);
+        if (nm == 2)
                 queue_flag_set(QUEUE_FLAG_NOMERGES, q);
-        else
-                queue_flag_clear(QUEUE_FLAG_NOMERGES, q);
+        else if (nm)
+                queue_flag_set(QUEUE_FLAG_NOXMERGES, q);
         spin_unlock_irq(q->queue_lock);
 
         return ret;
diff --git a/block/elevator.c b/block/elevator.c
index 9ad5ccc4c5ee..ee3a883840f2 100644
--- a/block/elevator.c
+++ b/block/elevator.c
@@ -474,6 +474,15 @@ int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
         int ret;
 
         /*
+         * Levels of merges:
+         *      nomerges:  No merges at all attempted
+         *      noxmerges: Only simple one-hit cache try
+         *      merges:    All merge tries attempted
+         */
+        if (blk_queue_nomerges(q))
+                return ELEVATOR_NO_MERGE;
+
+        /*
          * First try one-hit cache.
          */
         if (q->last_merge) {
@@ -484,7 +493,7 @@ int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
                 }
         }
 
-        if (blk_queue_nomerges(q))
+        if (blk_queue_noxmerges(q))
                 return ELEVATOR_NO_MERGE;
 
         /*
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index ffb13ad35716..f71f5c58620c 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -463,6 +463,7 @@ struct request_queue
 #define QUEUE_FLAG_IO_STAT     15       /* do IO stats */
 #define QUEUE_FLAG_CQ          16       /* hardware does queuing */
 #define QUEUE_FLAG_DISCARD     17       /* supports DISCARD */
+#define QUEUE_FLAG_NOXMERGES   18       /* No extended merges */
 
 #define QUEUE_FLAG_DEFAULT      ((1 << QUEUE_FLAG_IO_STAT) |            \
                                  (1 << QUEUE_FLAG_CLUSTER) |            \
@@ -589,6 +590,8 @@ enum {
 #define blk_queue_queuing(q)    test_bit(QUEUE_FLAG_CQ, &(q)->queue_flags)
 #define blk_queue_stopped(q)    test_bit(QUEUE_FLAG_STOPPED, &(q)->queue_flags)
 #define blk_queue_nomerges(q)   test_bit(QUEUE_FLAG_NOMERGES, &(q)->queue_flags)
+#define blk_queue_noxmerges(q)  \
+        test_bit(QUEUE_FLAG_NOXMERGES, &(q)->queue_flags)
 #define blk_queue_nonrot(q)     test_bit(QUEUE_FLAG_NONROT, &(q)->queue_flags)
 #define blk_queue_io_stat(q)    test_bit(QUEUE_FLAG_IO_STAT, &(q)->queue_flags)
 #define blk_queue_flushing(q)   ((q)->ordseq)