freebsd stable - Oct 2006 - Promise SATA timeout problems (PR 103435)

I'm experiencing a problem that seems to be the one reported in:

   http://www.freebsd.org/cgi/query-pr.cgi?pr=103435&cat=i386

I've got a Promise SATA300 TX4 card (note, this is =not= a RAID
card) and two Seagate 300GB SATA drives, in an old (800-MHz
original "Slot A" Athlon) experimental box currently running
6.1-RELEASE-p9.

If I generate a heavy I/O load on the disks, I start getting a
bunch of timeout messages -- various "taskqueue timeout"s, READ_DMA
and READ_DMA48 retries, and even an occasional UDMA ICRC error.

I can produce this stuff pretty reliably by starting up two "dd"
commands -- copying each drive to /dev/null -- and running them
concurrently.

As best I can tell from searching on the net, this problem has been
around for quite some time, in both FreeBSD and Linux systems, and
no one seems to have managed to figure out what's causing it.

Needless to say, I'd love to see a fix.  Any ideas?

Rich Wales
Palo Alto, CA, USA
richw@richw.org
http://www.richw.org

Several days ago, I wrote (in -stable, -current, and -hardware)
about problems I had been having with a Promise SATA300 TX4 card
(PDC40718 chip).

Per a suggestion from S?ren, I rechecked my hardware and my BIOS
settings.  I eventually found that I was able to make the timeout
problems disappear by disabling PCI master burst mode in the BIOS
setup of my system.  (For the record, I've been working with an
800-MHz original "Slot A" Athlon on a VIA VAM-0070 motherboard --
a board which appears to be a rebranded EPoX EP-7KXA.)

Not surprisingly, turning off burst mode had a significant impact
on the disk I/O speed of the system, but at least it's stable now,
with no more timeouts or failures.

I'm not sure if the primary culprit here is the Promise card or
the motherboard, but it suggests to me that some systems with iffy
PCI bus tolerances might require suboptimal BIOS settings in order
to work properly with some Promise controllers.  At least, this is
something to try if you encounter problems with a Promise card.

As for whether any effective steps can be taken in the driver to
make it more robust against this problem, I'll have to leave that
to people who are more familiar with the hardware and the code.

I've updated PR 103435 with this info.

Rich Wales
Palo Alto, CA, USA
richw@richw.org
http://www.richw.org