8.2 ->8.3 regression on disk writes

[Michael Ross]

Hello,

using 8.2 the machine runs fine,
using 8.3 or higher, not so much.

In laymans terms,
if I do "too many" writes/time just once, the machine can't do any disk  
access for a couple of hours.

As in: What's already running stays running, no crashes or anything,
but as soon as I need to read from disk (login, start program not cached  
in memory from previous run),
I'm all out of luck.
I killed the testing ftp-transfer about 15 seconds after the transfer  
speed dropped,
now I'm waiting for 10+ minutes for ``top'' to start.


I can install ports and kernels and world fine,
but "ezjail-admin install" or transferring a few GB of files from another  
machine sends it to limbo.

The next step would likely be to go through the kernel changes between 8.2  
and 8.3 to narrow it down,
I'd appreciate pointers as to what kernel changes to look out for,
or other suggestions on what to do.

Verbose dmesg: http://gurder.ross.cx/misc/dmesg.txt

TIA,

Michael


Postscript: That's the same problem I wrote about in "Trouble with gmirror  
and device ada",
turns out using gmirror just makes the problem appear much faster as  
opposed to eventually.