Bug 300 - Lookup while "mkisofs"ing a DVD-Image
: Lookup while "mkisofs"ing a DVD-Image
Status: RESOLVED FIXED
Product: XFS
Classification: Unclassified
Component: XFS kernel code
: Current
: Linux
: major
: ---
Assigned To: Eric Sandeen
:
:
:
Depends on:
Blocks:
  Show dependency treegraph
 
Reported: 2003-12-31 11:44 CST by Matthias Schniedermeyer
Modified: 2004-01-14 10:46 CST (History)
0 users

See Also:


Attachments
Patch to fix large extents on large allocation groups (7.60 KB, patch)
2004-01-14 08:45 CST, Eric Sandeen
Details | Diff

Note You need to log in before you can comment on or make changes to this bug.
Description Matthias Schniedermeyer 2003-12-31 11:44:15 CST
When i try to create a DVD-Image with mkisofs (4.2 GB in size) i get a (most
times) reproducable lookup at about 95% to 100% completition (6 out of 8 tries).

This happens both with 2.4.24-pre3 (vanilla) and 2.4.23 + XFS-Patch.
Nothing appears in the log-files and nothing appears on the screen (The 8 tries
where with a freshly booted system on the console (no X11)).
Compilers i tried: gcc-3.3.3(Debian Prerelease 20031229) and gcc-2.95.4

System is a Dual-P3-933Mhz, 3GB-RAM.
The involved HDDs (both 100GB Western Digital, 7200RPM) are connected to a
HighPoint RapidRAID 1540.
The system is 3 years old and worked flawlessly with reiserfs for months.
Comment 1 Eric Sandeen 2003-12-31 12:57:31 CST
Please strace the process to see where it's stopping,
and if possible can you test it on a non-xfs filesystem as well?

You could also get the kernel tree from oss.sgi.com cvs, this will include
kdb and you could get a kernel backtrace to see where things are stuck.

As it is, I don't think there's enough info to debug anything.

Ah, one further thought...

please include xfs_info output for the mount point you're using, this may be
the large AG problem again.
Comment 2 Eric Sandeen 2003-12-31 12:58:49 CST
Actually please include the xfs_info output first, this is simplest
and will quickly rule out or confirm a specific problem, thanks.
Comment 3 Matthias Schniedermeyer 2003-12-31 13:31:49 CST
Here is the xfs_info
I formated the partition without further option to mkfs.xfs

# xfs_info /dev/hdk1
meta-data=/x4                    isize=256    agcount=16, agsize=1526339 blks
         =                       sectsz=512
data     =                       bsize=4096   blocks=24421424, imaxpct=25
         =                       sunit=0      swidth=0 blks, unwritten=1
naming   =version 2              bsize=4096
log      =internal               bsize=4096   blocks=11924, version=1
         =                       sectsz=512   sunit=0 blks
realtime =none                   extsz=65536  blocks=0, rtextents=0
Comment 4 Matthias Schniedermeyer 2003-12-31 13:35:38 CST
Testing with non-xfs partition.

Before today i used reiserfs for month, worked flawlessly. (The last
hardware-change is more than 3 month ago). Kernel 2.4.23 ran since about 1 week
after it was released without problems.

The first lockup was even with reiserfs as the source-partition after the first
lookup i "converted" the source-partition to xfs. (Btw. "cp -av"ing 40GB of data
went through without problems).

Later i will try the other suggestions.
Comment 5 Eric Sandeen 2003-12-31 15:02:01 CST
Ok, your allocation groups are > 4G (they are 5G or so)
which means that file extents can grow past 4G, and this overflows
the 32 bits that contain the length in an iomap structure.

I'm working on the fix for this; if you want to re-mkfs and test with mkfs options
to constrain the AGs to 4G or less, that should fix your problem in the meantime.

(the -d agsize=4g option to mkfs)
Comment 6 Matthias Schniedermeyer 2003-12-31 16:10:29 CST
95GB/16 = 5,9GB. Sure this is more than 4GB. :-(

OK. I've recreated the filesystem with doubled agcount.

meta-data=/x4                    isize=256    agcount=32, agsize=763170 blks
         =                       sectsz=512
data     =                       bsize=4096   blocks=24421438, imaxpct=25
         =                       sunit=0      swidth=0 blks, unwritten=1
naming   =version 2              bsize=4096
log      =internal               bsize=4096   blocks=11924, version=1
         =                       sectsz=512   sunit=0 blks
realtime =none                   extsz=65536  blocks=0, rtextents=0

This also explains why "cp -av"ing the data around ("dump & restore") wasn't a
problem because the biggest file copied was only(tm) 3,4 GB in size.

<Recreating & filling the filesystem finished>

Creating the DVD-Image hadn't crashed the machine (3 tries).
I hope this was my first and last problem with XFS (knocking on wood). :-)
Comment 7 Eric Sandeen 2004-01-14 08:45:43 CST
Created attachment 97 [details]
Patch to fix large extents on large allocation groups

The attached patch should resolve this bug
Comment 8 Eric Sandeen 2004-01-14 08:46:07 CST
The attached patch should resolve this bug.