system freeze with xfs

When I create an xfs filesystem and run
'stress' (http://weather.ou.edu/~apw/projects/stress/stress.html) I can induce a system freeze (hang, no panic). There are no issues if I create a jfs filesystem (on the same disk) or run 'stress' on a different disk (ext3)

I'd like to know how I could go about debugging this, otherwise I will have to switch filesystem types.

Details:
% stress -d 2 --hdd-noclean --hdd-bytes 3G --verbose

Fedora Core 5
2.6.20-1.2312.fc5 #1 Tue Apr 10 15:09:44 EDT 2007 i686 athlon i386 GNU/Linux

# /sbin/mkfs.xfs -f  /dev/VolTest/test 
meta-data=/dev/VolTest/test      isize=256    agcount=16, agsize=198656
blks
         =                       sectsz=512  
data     =                       bsize=4096   blocks=3178496, imaxpct=25
         =                       sunit=0      swidth=0 blks, unwritten=1
naming   =version 2              bsize=4096  
log      =internal log           bsize=4096   blocks=2560, version=1
         =                       sectsz=512   sunit=0 blks
realtime =none                   extsz=65536  blocks=0, rtextents=0

# /sbin/hdparm /dev/hdk 

/dev/hdk:

 multcount    = 16 (on)
 IO_support   =  0 (default 16-bit)
 unmaskirq    =  0 (off)
 using_dma    =  1 (on)
 keepsettings =  0 (off)
 readonly     =  0 (off)
 readahead    = 256 (on)
 geometry     = 25228/16/63, sectors = 25429824, start = 0

is that a threat? ;-)

Try enabling sysrq, and do sysrq-t when it freezes, as a first pass, to see where all the threads are (stuck) at.

You might also try this on a stock upstream kernel, and see if you have the same problem with 8k stacks vs. 4k (as is in FC6)

You could also set up kdump and/or netdump (whatever FC5 supports...) and get a system dump at the time it freezes up, via sysrq-c.

Thanks,

-Eric

Smells like 4k stack on i386 blowing up; try w/o the 4k stacks option and see how that hold up