[PATCH] semaphore fairness patch against test11-pre6

From: David Mansfield <lkml@dm.ultramaster.com>
To: Linus Torvalds <torvalds@transmeta.com>
Cc: lkml <linux-kernel@vger.kernel.org>
Subject: [PATCH] semaphore fairness patch against test11-pre6
Date: Fri, 17 Nov 2000 20:01:41 -0500	[thread overview]
Message-ID: <3A15D4F5.B39D61BD@dm.ultramaster.com> (raw)

[-- Attachment #1: Type: text/plain, Size: 1780 bytes --]

Hi Linus et al,

I've applied your semaphore fairness patch (slightly fixed) below.  It
fixes my original bug report of vmstat, ps etc. stalls waiting for the
mmap_sem.  I can now run my memory 'hog' processes and actually see
vmstat update every second even under heavy memory pressure.  More
importantly, ps works so I can find the pid to kill.  I'm no expert in
checking for races, but I went over all (I think) the 2 process cases as
well as I could and they seem to look ok to me, but what do I know.  I
know someone else reported it didn't fix the problem, but perhaps that's
some other issue.

I ran many 'ps' (20?) in the background trying to simulate many process
contention, and everything still worked fine.  I've run some other
stress tests too (dbench, my own I/O throughput test etc) and so far
all's well (famous last words).

If you can find the time to check this out more completely, I recommend
it, because it seems like a great improvement to be able to accurately
see vmstat numbers in times of system load.  I hope the other side
effects are beneficial as well :-)

The change to the patch was that you had 'if (sleepers > 1)' when
obviously you meant 'if (sem->sleepers > 1)'...

Here's your patch again (also attached in case of mangling):

--- linux/arch/i386/kernel/semaphore.c	2000/11/16 19:58:26	1.3
+++ linux/arch/i386/kernel/semaphore.c	2000/11/17 23:12:48
@@ -64,6 +64,14 @@
 
 	spin_lock_irq(&semaphore_lock);
 	sem->sleepers++;
+
+	/*
+	 * Are there other people waiting for this?
+	 * They get to go first.
+	 */
+	if (sem->sleepers > 1)
+		goto inside;
+
 	for (;;) {
 		int sleepers = sem->sleepers;
 
@@ -76,6 +84,7 @@
 			break;
 		}
 		sem->sleepers = 1;	/* us - see -1 above */
+inside:
 		spin_unlock_irq(&semaphore_lock);
 
 		schedule();

[-- Attachment #2: sem-patch.test11-pre6 --]
[-- Type: text/plain, Size: 747 bytes --]

Index: linux/arch/i386/kernel/semaphore.c
===================================================================
RCS file: /home/kernel/cvs_master/linux/arch/i386/kernel/semaphore.c,v
retrieving revision 1.3
diff -u -r1.3 semaphore.c
--- linux/arch/i386/kernel/semaphore.c	2000/11/16 19:58:26	1.3
+++ linux/arch/i386/kernel/semaphore.c	2000/11/17 23:12:48
@@ -64,6 +64,14 @@
 
 	spin_lock_irq(&semaphore_lock);
 	sem->sleepers++;
+
+	/*
+	 * Are there other people waiting for this?
+	 * They get to go first.
+	 */
+	if (sem->sleepers > 1)
+		goto inside;
+
 	for (;;) {
 		int sleepers = sem->sleepers;
 
@@ -76,6 +84,7 @@
 			break;
 		}
 		sem->sleepers = 1;	/* us - see -1 above */
+inside:
 		spin_unlock_irq(&semaphore_lock);
 
 		schedule();
