linux-kernel - Re: [PATCH -v6 0/2] Fixing the issue with memory-mapped file times

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date:	Fri, 18 Jan 2008 22:48:38 +0300
From:	"Anton Salikhmetov" <salikhmetov@...il.com>
To:	"Miklos Szeredi" <miklos@...redi.hu>
Cc:	linux-mm@...ck.org, jakob@...hought.net,
	linux-kernel@...r.kernel.org, valdis.kletnieks@...edu,
	riel@...hat.com, ksm@...dk, staubach@...hat.com,
	jesper.juhl@...il.com, torvalds@...ux-foundation.org,
	a.p.zijlstra@...llo.nl, akpm@...ux-foundation.org,
	protasnb@...il.com, r.e.wolff@...wizard.nl,
	hidave.darkstar@...il.com, hch@...radead.org
Subject: Re: [PATCH -v6 0/2] Fixing the issue with memory-mapped file times

2008/1/18, Miklos Szeredi <miklos@...redi.hu>:
> > 4. Performance test was done using the program available from the
> > following link:
> >
> > http://bugzilla.kernel.org/attachment.cgi?id=14493
> >
> > Result: the impact of the changes was negligible for files of a few
> > hundred megabytes.
>
> Could you also test with ext4 and post some numbers?  Afaik, ext4 uses
> nanosecond timestamps, so the time updating code would be exercised
> more during the page faults.
>
> What about performance impact on msync(MS_ASYNC)?  Could you please do
> some measurment of that as well?

Did a quick test on an ext4 partition. This is how it looks like:

debian:~/miklos# ./miklos_test /mnt/file
begin   1200662360      1200662360      1200662353
write   1200662361      1200662361      1200662353
mmap    1200662361      1200662361      1200662362
b       1200662363      1200662363      1200662362
msync b 1200662363      1200662363      1200662362
c       1200662365      1200662365      1200662362
msync c 1200662365      1200662365      1200662362
d       1200662367      1200662367      1200662362
munmap  1200662367      1200662367      1200662362
close   1200662367      1200662367      1200662362
sync    1200662367      1200662367      1200662362
debian:~/miklos# mount | grep /mnt
/root/image.ext4 on /mnt type ext4dev (rw,loop=/dev/loop0)

> What about performance impact on msync(MS_ASYNC)?  Could you please do
> some measurment of that as well?

Following are the results of the measurements. Here are the relevant
portions of the test program:

>>>

#define FILE_SIZE (1024 * 1024 * 512)

p = mmap(0, FILE_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);

/* Bring the pages in */
for (i = 0; i < FILE_SIZE; i += 4096)
        tmp = p[i];

/* Dirty the pages */
for (i = 0; i < FILE_SIZE; i += 4096)
        p[i] = i;

/* How long did we spend in msync(MS_ASYNC)? */
gettimeofday(&tv_start, NULL);
msync(p, FILE_SIZE, MS_ASYNC);
gettimeofday(&tv_stop, NULL);

<<<

For reference tests, the following platforms were used:

1. HP-UX B.11.31, PA-RISC 8800 processor (800 MHz, 64 MB), Memory: 4 GB.

2. HP-UX B.11.31, 2 Intel(R) Itanium 2 9000 series processors (1.59 GHz, 18 MB),
   Memory: 15.98 GB.

3. FreeBSD 6.2-RELEASE, Intel(R) Pentium(R) III CPU family 1400MHz, 2 CPUs.
   Memory: 4G.

The tests of my solution were performed using the following platform:

A KVM x86_64 guest OS, current Git kernel. Host system: Intel(R) Core(TM)2
Duo CPU T7300 @ 2.00GHz. Further referred to as "the first case".

The following tables give the time difference between the two calls
to gettimeofday(). The test program was run three times in a raw
with a delay of one second between consecutive runs. On Linux
systems, the following commands were issued prior to running the
tests:

echo 80 >/proc/sys/vm/dirty_ratio
echo 80 >/proc/sys/vm/dirty_background_ratio
echo 30000 >/proc/sys/vm/dirty_expire_centisecs
sync

Table 1. Reference platforms.

------------------------------------------------------------
|             | HP-UX/PA-RISC | HP-UX/Itanium | FreeBSD    |
------------------------------------------------------------
| First run   | 263405 usec   | 202283 usec   | 90 SECONDS |
------------------------------------------------------------
| Second run  | 262253 usec   | 172837 usec   | 90 SECONDS |
------------------------------------------------------------
| Third run   | 238465 usec   | 238465 usec   | 90 SECONDS |
------------------------------------------------------------

It looks like FreeBSD is a clear outsider here. Note that FreeBSD
showed an almost liner depencence of the time spent in the
msync(MS_ASYNC) call on the file size.

Table 2. The Qemu system. File size is 512M.

---------------------------------------------------
|            | Before the patch | After the patch |
---------------------------------------------------
| First run  |     35 usec      |   5852 usec     |
---------------------------------------------------
| Second run |     35 usec      |   4444 usec     |
---------------------------------------------------
| Third run  |     35 usec      |   6330 usec     |
---------------------------------------------------

I think that the data above prove the viability of the solution I
presented. Also, I guess that this bug fix is most probably ready
for getting upstream.

Please apply the sixth version of my solution.

>
> Thanks,
> Miklos
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/