[plug] 2.4.24 mremap root exploit

[Craig Ringer]

On Thu, 2004-02-19 at 08:16, Tony Breeds wrote:

> > Something is clearly not right here. Whether it's LVM2/device-mapper (if
> > so, LVM1 had the same problem), the kernel I/O scheduler, the 3ware
> > driver, the 3ware card firmware, or something else I do not know.
> > Suggestions appreciated.
> 
> Well I don't run LVM at all and I'm not seeing any IO  starvation.  Can
> you run the same test on the same hardware w/o LVM/DM available?

Not at present, but I'll be able to do some no-LVM testing when I take
the next system snapshot.

BTW, some more research along a line I didn't think of before has turned
up the possibility that it is the 3ware driver. It looks like the driver
may keep a very deep queue (256 entries) by default, and this is
apparently not at all good for write-heavy workloads without TCQ-capable
disks. I'm dropping the queue depth to 16, and next time I get a chance
to reboot the server I'll find out if it helps.

The driver exports the queue depth in sysfs, but you can't alter it
there. Grr.

Craig Ringer