Has anyone else seen any form of in memory or on disk corruption?

[John Baldwin]

On Friday 04 July 2008 12:58:07 pm gnn at freebsd.org wrote:
> Hi,
> 
> I've been working on the following brain teasing (breaking?) problem
> for about a week now.  What I'm seeing is that on large memory
> machines, those with more than 4G of RAM, the ungzipping/untarring of
> files fails due to gzip thinking the file is corrupt.  The way to
> reproduce this is:
> 
> 1) Create a bunch of gzip/tar balls in the 1-20MB range.
> 2) Reboot FreeBSD 7.0 release
> 3) Run gzip -t over all the files.
> 
> I have hundreds of these files to run this over, and a full check
> takes about 3 hours, but I usually see some form of corruption within
> the first 20 minutes.
> 
> Other important factors:
> 
> 1) This is on very modern, 2P/4Core (8 cores total) hardware
> 2) The disks are 1TB SATA set up in JBOD.
> 3) The machines have 16G of RAM.
> 4) Corruption is seen only after a reboot, if the machines continue to
> run corruption is never seen again, until another reboot.
> 5) The systems are all Xeon running amd64
> 6) The disk controller is an AMCC 9650, but we do see this very rarely
> with the on board controlller.

If this is one of the ATA controllers where it tries to use 63k transfers (126 
* DEV_BSIZE) instead of 64k, then change it to 32k (64 * DEV_BSIZE).  W/o 
this fix I see massive data corruption (couldn't even build a kernel with the 
fix, had to reinstall the box) on HT1000 ATA chipsets.  Crashdumps also don't 
seem to work reliably w/o changing that.

-- 
John Baldwin