dm kcopyd: avoid queue shuffle

Write throughput to LVM snapshot origin volume is an order
of magnitude slower than those to LV without snapshots or
snapshot target volumes, especially in the case of sequential
writes with O_SYNC on.

The following patch originally written by Kevin Jamieson and
Jan Blunck and slightly modified for the current RCs by myself
tries to improve the performance by modifying the behaviour
of kcopyd, so that it pushes back an I/O job to the head of
the job queue instead of the tail as process_jobs() currently
does when it has to wait for free pages. This way, write
requests aren't shuffled to cause extra seeks.

I tested the patch against 2.6.27-rc5 and got the following results.
The test is a dd command writing to snapshot origin followed by fsync
to the file just created/updated.  A couple of filesystem benchmarks
gave me similar results in case of sequential writes, while random
writes didn't suffer much.

dd if=/dev/zero of=<somewhere on snapshot origin> bs=4096 count=...
   [conv=notrunc when updating]

1) linux 2.6.27-rc5 without the patch, write to snapshot origin,
average throughput (MB/s)
                     10M     100M    1000M
create,dd         511.46   610.72    11.81
create,dd+fsync     7.10     6.77     8.13
update,dd         431.63   917.41    12.75
update,dd+fsync     7.79     7.43     8.12

compared with write throughput to LV without any snapshots,
all dd+fsync and 1000 MiB writes perform very poorly.

                     10M     100M    1000M
create,dd         555.03   608.98   123.29
create,dd+fsync   114.27    72.78    76.65
update,dd         152.34  1267.27   124.04
update,dd+fsync   130.56    77.81    77.84

2) linux 2.6.27-rc5 with the patch, write to snapshot origin,
average throughput (MB/s)

                     10M     100M    1000M
create,dd         537.06   589.44    46.21
create,dd+fsync    31.63    29.19    29.23
update,dd         487.59   897.65    37.76
update,dd+fsync    34.12    30.07    26.85

Although still not on par with plain LV performance -
cannot be avoided because it's copy on write anyway -
this simple patch successfully improves throughtput
of dd+fsync while not affecting the rest.

Signed-off-by: Jan Blunck <jblunck@suse.de>
Signed-off-by: Kazuo Ito <ito.kazuo@oss.ntt.co.jp>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Cc: stable@kernel.org

diff --git a/drivers/md/dm-kcopyd.c b/drivers/md/dm-kcopyd.c
index 996802b8a4522e50126365d070f21458f9b68be2..8f153530d6dfac3377c3159867b6242e4b614a3f 100644 (file)
--- a/drivers/md/dm-kcopyd.c
+++ b/drivers/md/dm-kcopyd.c
@@ -268,6 +268,17 @@ static void push(struct list_head *jobs, struct kcopyd_job *job)
        spin_unlock_irqrestore(&kc->job_lock, flags);
 }
 
+
+static void push_head(struct list_head *jobs, struct kcopyd_job *job)
+{
+       unsigned long flags;
+       struct dm_kcopyd_client *kc = job->kc;
+
+       spin_lock_irqsave(&kc->job_lock, flags);
+       list_add(&job->list, jobs);
+       spin_unlock_irqrestore(&kc->job_lock, flags);
+}
+
 /*
  * These three functions process 1 item from the corresponding
  * job list.
@@ -398,7 +409,7 @@ static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
                         * We couldn't service this job ATM, so
                         * push this job back onto the list.
                         */
-                       push(jobs, job);
+                       push_head(jobs, job);
                        break;
                }