[pipe] b8597fdc23: -8% regression of hackbench throughput

[fengguang . wu]

Hi Dmitry,

We measured some 8% drop of hackbench throughput with your small commit
b8597fdc23 ("fs/pipe.c: skip file_update_time on frozen fs"). Comparing to its
parent commit, we also see increased context switches:

5b6e04beb19abc9  b8597fdc23f8f6e26a531e9a7  
---------------  -------------------------  
    231874 ~ 1%      -7.3%     214834 ~ 2%  lkp-snb01/micro/hackbench/1600%-process-pipe
     87774 ~ 3%     -10.7%      78419 ~ 4%  xps2/micro/hackbench/1600%-process-pipe
    319649           -8.3%     293253       TOTAL hackbench.throughput

5b6e04beb19abc9  b8597fdc23f8f6e26a531e9a7  
---------------  -------------------------  
   3673577 ~ 1%     +17.4%    4312376 ~ 2%  lkp-snb01/micro/hackbench/1600%-process-pipe
    195669 ~ 3%    +156.6%     502173 ~ 4%  xps2/micro/hackbench/1600%-process-pipe
   3869246          +24.4%    4814550       TOTAL vmstat.system.cs

5b6e04beb19abc9  b8597fdc23f8f6e26a531e9a7  
---------------  -------------------------  
 6.311e+08 ~ 2%     +33.7%  8.437e+08 ~ 1%  lkp-snb01/micro/hackbench/1600%-process-pipe
  21485632 ~ 3%    +411.9%    1.1e+08 ~ 8%  xps2/micro/hackbench/1600%-process-pipe
 6.526e+08          +46.1%  9.537e+08       TOTAL time.involuntary_context_switches

5b6e04beb19abc9  b8597fdc23f8f6e26a531e9a7  
---------------  -------------------------  
 1.563e+09 ~ 2%     +12.8%  1.763e+09 ~ 3%  lkp-snb01/micro/hackbench/1600%-process-pipe
  96703947 ~ 3%    +102.7%   1.96e+08 ~ 3%  xps2/micro/hackbench/1600%-process-pipe
  1.66e+09          +18.0%  1.959e+09       TOTAL time.voluntary_context_switches

The bisect looks stable:

                           time.involuntary_context_switches

     9e+08 ++---------------------------------------------------------------+
           |                                  O                             |
   8.5e+08 O+O O O O                   O    O     O O O                     |
           |         O O O O O O O   O   O      O                           |
     8e+08 ++                      O                                        |
           |                                                                |
   7.5e+08 ++                                                               |
           |                                                                |
     7e+08 ++                                                               |
           |.*   *.*.*.*.*   *.                  .*   *.                    |
   6.5e+08 *+ + +         + +  *.*.*.*       .*.*  + +  *.    *.   .*   *.  |
           |   *           *          + .*..*       *     *. +  *.*  + +  *.*
     6e+08 ++                          *                    *         *     |
           |                                                                |
   5.5e+08 ++---------------------------------------------------------------+

Thanks,
Fengguang
>From b8597fdc23f8f6e26a531e9a7da3c5fe7b99752c Mon Sep 17 00:00:00 2001
From: Dmitry Monakhov <dmonakhov@xxxxxxxxxx>
Date: Tue, 17 Dec 2013 10:45:35 +1100
Subject: [PATCH] fs/pipe.c: skip file_update_time on frozen fs

Pipe has no data associated with fs so it is not good idea to block
pipe_write() if FS is frozen, but we can not update file's time on such
filesystem.  Let's use same idea as we use in touch_time().

Addresses https://bugzilla.kernel.org/show_bug.cgi?id=65701

Signed-off-by: Dmitry Monakhov <dmonakhov@xxxxxxxxxx>
Reviewed-by: Jan Kara <jack@xxxxxxx>
Cc: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
---
 fs/pipe.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/pipe.c b/fs/pipe.c
index 0e0752e..78fd0d0 100644
--- a/fs/pipe.c
+++ b/fs/pipe.c
@@ -663,10 +663,11 @@ out:
 		wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 	}
-	if (ret > 0) {
+	if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 		int err = file_update_time(filp);
 		if (err)
 			ret = err;
+		sb_end_write(file_inode(filp)->i_sb);
 	}
 	return ret;
 }
-- 
1.8.5.2