Re: Direct I/O bug in kernel

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]


----- Original Message -----

> From: Hillf Danton <dhillf@xxxxxxxxx>
> To: Victor Meyerson <calculuspenguin@xxxxxxxxx>
> Cc: "linux-mips@xxxxxxxxxxxxxx" <linux-mips@xxxxxxxxxxxxxx>; Ralf Baechle <ralf@xxxxxxxxxxxxxx>; LKML <linux-kernel@xxxxxxxxxxxxxxx>
> Sent: Tuesday, July 24, 2012 6:04 AM
> Subject: Re: Direct I/O bug in kernel
> 
> On Sun, Jul 22, 2012 at 10:05 AM, Victor Meyerson
> <calculuspenguin@xxxxxxxxx> wrote:
>>  Hi,
>> 
>>  I recently found a bug related to direct io in post 3.3 linux kernels.  
> Fortunately, my hardware (a Cobalt Qube2) is supported by the vanilla kernel so 
> I did not need additional patch sets to get the machine to boot.  I ran git 
> bisect on the main tree[1] and tested the various bisect results until git 
> reported the first bad commit.  After several bisects and many reboots, git 
> reported that [2] was the first bad commit.
>> 
>>  In testing this I came up with a repeatable process.  Unfortunately, I do 
> not have any other MIPS hardware to test this on and I believe that based on the 
> commit in question that it is MIPS related.  My procedure is as follows:
>> 
>>  1) Create a random file to be used on the two kernels (one before the 
> commit, and one that includes the commit)
>>  $ dd if=/dev/urandom of=random-file bs=512 count=30720
>>  30720+0 records in
>>  30720+0 records out
>>  15728640 bytes (16 MB) copied, 60.7035 s, 259 kB/s
>>  $ chmod -w random-file
>> 
>>  2) Reboot to the kernel before the commit and run dd with direct io.  
> Repeat.
>>  $ uname -a
>>  Linux horadric 3.2.0-dirty #2 Fri Jul 13 06:20:22 PDT 2012 mips64 Nevada 
> V10.0 FPU V10.0 Cobalt Qube2 GNU/Linux
>>  $ dd if=random-file of=portion-of-random-3.2.0 bs=512 count=20480 
> iflag=direct
>>  20480+0 records in
>>  20480+0 records out
>>  10485760 bytes (10 MB) copied, 42.3636 s, 248 kB/s
>>  $ reboot
>>  $ dd if=random-file of=portion-of-random-3.2.0-2 bs=512 count=20480 
> iflag=direct
>>  20480+0 records in
>>  20480+0 records out
>>  10485760 bytes (10 MB) copied, 42.5252 s, 247 kB/s
>> 
>>  3) Reboot to the kernel with the commit and run dd with direct io.  Repeat.
>>  $ uname -a
>>  Linux horadric 3.2.0-rc4-00003-gb1c10be-dirty #15 Fri Jul 20 15:05:13 PDT 
> 2012 mips64 Nevada V10.0 FPU V10.0 Cobalt Qube2 GNU/Linux
>>  $ dd if=random-file of=portion-of-random-3.2.0-rc4 bs=512 count=20480 
> iflag=direct
>>  20480+0 records in
>>  20480+0 records out
>>  10485760 bytes (10 MB) copied, 40.6226 s, 258 kB/s
>>  $ reboot
>>  $ dd if=random-file of=portion-of-random-3.2.0-rc4-2 bs=512 count=20480 
> iflag=direct
>>  20480+0 records in
>>  20480+0 records out
>>  10485760 bytes (10 MB) copied, 40.8856 s, 256 kB/s
>> 
> Hi Victor,
> 
> Create files with
> 
>     dd if=random-file of=portion-of-random-3.2.0-rc4    bs=8k
> count=1280 iflag=direct
>     dd if=random-file of=portion-of-random-3.2.0-rc4-2 bs=8k
> count=1280 iflag=direct
> 
> without reboot(why reboot needed?), then see the changes in checksums.
> 
> Thanks
> Hillf
> 

Hi Hillf,

I rebooted in an attempt to make sure nothing was cached between runs.  In any case, here are the results without a reboot:

$ dd if=random-file of=portion-of-random-3.2.0-rc4 bs=8k count=1280 iflag=direct
1280+0 records in
1280+0 records out
10485760 bytes (10 MB) copied, 6.00599 s, 1.7 MB/s
$ dd if=random-file of=portion-of-random-3.2.0-rc4-2 bs=8k count=1280 iflag=direct
1280+0 records in
1280+0 records out
10485760 bytes (10 MB) copied, 5.25964 s, 2.0 MB/s
$ sha256sum portion-of-random-3.2.0-rc4*
4c56820030ce22e6cc96127a53f6025d11a78f2fd3b0c1dec44f6d6746f70bbd  portion-of-random-3.2.0-rc4
05c41d626a67b9bcddb0e7b905533c63a0866092b819bf01cdb2a80f29c2b162  portion-of-random-3.2.0-rc4-2

Still different checksums and I used the same random-file from my first test.

Victor

>>  4) Compare checksums of the resulting files.
>>  $ sha256sum portion-of-random-3.2.0*
>>  c98a6e949b36448842a21f68e7c6a5daff1f161e1eb3e3529176cf56bf5af89e  
> portion-of-random-3.2.0
>>  c98a6e949b36448842a21f68e7c6a5daff1f161e1eb3e3529176cf56bf5af89e  
> portion-of-random-3.2.0-2
>>  dca27da87a78580b8a34bbff2790ae80d3aa880d5d00fc2126f109d6fff9e056  
> portion-of-random-3.2.0-rc4
>>  703cf02d4fa90679d4a75900e7e5a3b8c3000a65bfc475610b10f17bb88bedbc  
> portion-of-random-3.2.0-rc4-2
>> 
>>  Notice how the last two files have different checksums between themselves 
> and even different from the first two files.  This lead me to believe that there 
> is a problem with direct io.  All the files are the same size and should include 
> the same portion of the random file created in step 1).
>> 
>>  My configuration is the Cobalt Qube2 with a 64-bit kernel and an n32 
> userspace.  Hopefully someone with a much more deeper understanding of the 
> kernel can confirm and provide a fix for this (assuming one has not been created 
> yet).
>> 
>>  Thanks.  Let me know if there is any additional information that may help 
> with the investigation.
>> 
>>  Victor
>> 
>> 
>>  [1] http://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
>>  [2] 
> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux.git;a=commitdiff;h=b1c10bea620f79109b5cc9935267bea4f6f29ac6
> 



[Linux MIPS Home]     [LKML Archive]     [Linux ARM]     [Linux]     [Git]     [Photo]     [Yosemite News]     [Linux SCSI]     [Linux Hams]

Add to Google Powered by Linux