R10K/R12K Based O2's

Okay, this is being sent to try and spur some activity on the front of getting R10K O2's to work somewhat better. I toyed around with such a machine to see how long I could get it to stay online, and thought I'd post some results.

To start, I merged the generic pieces of Peter Fuerst's IP28 patches into an IP32 tree (piece that touch generic MIPS code, like memcpy.S, strcpy.S, etc to add R10K cache barriers) and built an R10K IP32 kernel with the gcc cache barriers patch he also has up. End result seemed to be a kernel, that while definitely not production ready, stays up longer than others.

As told to me by Ilya, using the network device too much or even remotely trying to use scsi instantly killed the machine (based on his tests a while back). My results were a bit different in that I was able to use the network for a prolonged time (over ssh too) and even managed to use the scsi disk some, downloading and unpacking a 40MB tbz2 of a Gentoo stage3 (uclibc) tarball on the system. It didn't finally lock up until an 'emerge sync', which hits the disk heavily with lots of small files (which as Ilya predicted, would kill it).

Whether this was the result of using the R10K cache barriers in both the IP28 patch and gcc, or just general kernel improvements in those drivers, is anyone's guess, but I got farther than I initially expected (I didn't hold out hope that such a system would even boot, let alone get to userland).

IP28 used protected intermediate buffers for DMA input on the sgiwd93 (scsi) driver, but not on sgiseeq (net), as this had access to bounce buffers for incoming packets. IP32's meth (net) driver was designed specifically to avoid the need of bounce buffers, though, so likely it needs some solution to work around the speculative execution bit. IP32 likely needs the same or similar implements in its scsi driver (aic7xxx) as well).

It seems when speculative execution triggers, this is seen on the primary console:

CRIME CPU error at 0x00b74fa90 status 0x00000004 CRIME CPU error at 0x00b74fa90 status 0x00000004 CRIME CPU error at 0x01005fcc0 status 0x00000004 CRIME CPU error at 0x00b74fa90 status 0x00000004 CRIME CPU error at 0x00b74fa90 status 0x00000004 CRIME CPU error at 0x00b74fa90 status 0x00000004

(once per hit; they don't happen very often unless scsi is getting a workout. The two repeating addresses are constant, it seems).

Thoughts?

For testing, I have an IP32 R10K netboot kernel available, based on 2.6.15.1, if anyone wants to try it out for themselves and report their findings.

http://dev.gentoo.org/~kumba/mips/netboot/testing/ip32/r10k/

--Kumba

Gentoo/MIPS Team Lead
Gentoo Foundation Board of Trustees

"Such is oft the course of deeds that move the wheels of the world: small hands do them because they must, while the eyes of the great are elsewhere."  --Elrond

Round 2:

Discovered I'd built my IP32 R10K kernel with:
A) An older version of Peter Fuerst's R10K Cache Barriers Patch
B) Did a code change in the above patch that generated less cache barriers instead of more

So I've re-done everything, and now have a kernel for IP32 R10K (and R12K) that uses the latest gcc cache barriers patch available on the IP28 site w/ the one code change that generates more cache barriers (See below for details).

It seems with the gcc patch and the stripped IP28 patch, a working IP32 R10K kernel can be produced that seems to handle itself very well. It occasionally spits out CRIME CPU errors, and in my case, I've triggered two Oopses (non-fatal). It seems the network adapter doesn't seem to trigger the speculative execution bug too much (which I believe is indicated by the CRIME CPU errors), however the scsi driver definitely causes them to appear every so often.

Patches and other misc. files available here:
http://dev.gentoo.org/~kumba/mips/netboot/testing/ip32/r10k/

Included there is a 2.6.15.2 kernel, the kernel patch, the gcc patch, the kernel disassembly (71MB whopper unpacked), and the cpio archive containing the netboot userland (this is actually 5-6 cpio archives slapped together with 'cat', but the kernel handles it just fine), and the md5/sha1 sums to verify against.

To boot:
bootp(): console=ttyS0,<baud> (gbefb seems to not respond on these systems)
Initramfs should take over and load a mini netboot userland. Has basic capabilities, and supports a few fs'es (xfs, ext2/3, no reiser/jfs, etc)

Stats:

# uname -a
Linux netboot-2006.0 2.6.15.2-mipsgit-20060109 #2 Sun Feb 5 19:28:36 EST 2006 mips64 R10000 V2.6  FPU V0.0 SGI O2 GNU/Linux

# uptime
21:54:14 up 1 day,  2:23, load average: 2.01, 2.02, 2.00

Currently, it's building gcc-3.4.5 as I type this, using the disk for everything but swap. There are some caveats, though, which are outlined below:

The Errors (Just a sample, and in no specific order):
CRIME CPU error at 0x00f73b850 status 0x00000004 CRIME CPU error at 0x01005fcc0 status 0x00000004 CRIME CPU error at 0x005a8b810 status 0x00000004 CRIME CPU error at 0x005a8b850 status 0x00000004 CRIME CPU error at 0x007bd7850 status 0x00000004 CRIME CPU error at 0x00802f850 status 0x00000004 CRIME CPU error at 0x003b0b850 status 0x00000004 CRIME CPU error at 0x00e33f850 status 0x00000004 CRIME CPU error at 0x00a1df850 status 0x00000004

If anyone has an idea just what those hex addresses are referencing, or what exact is the meaning of that status bit, I'd be curious to know.

The Oopses:
http://dev.gentoo.org/~kumba/mips/netboot/testing/ip32/r10k/ip32r10k-26152-oops1.txt
http://dev.gentoo.org/~kumba/mips/netboot/testing/ip32/r10k/ip32r10k-26152-oops2.txt

These seem similar to the old MC Bus Errors I hit when toying around in the early days of IP28 support. They've pretty much vanished now since Peter wrote up the new IP28 bus error handler. So whether this are the same or not is anyone's guess -- if so, I imagine there's some degree of fixups that can be done to IP32's bus error handler to take care of these. But that's just pure speculation on my part.

Read the README file there too -- it's got some small /proc tweaks that, at a first glance, help the kernel to avoid touching the disk too often by turning off swap entirely and setting two other small vm settings.

--Kumba

Gentoo/MIPS Team Lead
Gentoo Foundation Board of Trustees

"Such is oft the course of deeds that move the wheels of the world: small hands do them because they must, while the eyes of the great are elsewhere."  --Elrond