Crash with Z77 chipset

Hello,
(probably a bit long mail, but i will try to describe what i did or tried)
using ASRock Z77 Pro3 motherboard with Z77 chipset, 4xSATA WDC WD1002FAEX-00Z3A0 drives, Debian Linux (basic installation, no X or other services).
Problem: any intense I/O to disk causes system to crash. Easiest method (for me) to reproduce (100% so far) problem - just do mkfs.ext2 /dev/sdb3 (any failsystem will work, the same goes for dd if=/dev/zero of=/dev/sdb bs=1M, just a bit slower). Before crash inode creation slows down, for ~10 seconds, then stops at all (and crash immediately). What i tried:
- first i noticed that system will crash with default debian kernel
(2.6.32-5-amd64). This is only one kernel which writes something to message log, and crashes when writing inodes at count ~3250/7464. It writes info to /var/log/messages and console, system becomes unresponsive (kernel.panic from sysctl does not reboot system, same goes for software watchdog - you need to "manually" reboot system)
- i recompiled current stable kernel (3.6.10) with
CONFIG_DETECT_HUNG_TASK=y and CONFIG_BOOTPARAM_HUNG_TASK_PANIC=y and re-tested. System hangs when writing inodes at ~3450/7464, no info on screen or syslog. System could be rebooted with echo b > /proc/sysrq-trigger on another console, console is responsive, but any disk access will hung console. Sometimes (rarely) system becomes unresponsive, and reboots after timeout
- i recompiled todays git kernel, recompiled with the same parameters.
It hangs ~6400/7464 (note - goes much further than previous versions), but completely - does not reboot itself, does not respond to ping, only poweroff helps. Nothing in syslog, photo from screen will be attached with logs in next post (can't be scrolled up/down - so no info what happened earlier)

Observations:
- system could be "alive" and working with low disk activity for long
time (at least, more than week). But enough to do some disk I/O - crash (for example, copying bzip'ed kernel image from one place to another is enough to trigger crash)
- disk type does not matter. I tried to attach Hitachi HDS722020ALA330
disk instead of WD - the same (i would say, it crashed even earlier, but didn't measured exactly)
- SATA cables are replaced, system could run prime95 torture test for
several hours - so i could say that RAM/CPU isn't a problem here
- could be crashed with activity on any disk. I tried to make RAID10,
LVM on top - crash; disassembled md array, tested with disk activity to _all_ disks separately - any disk activity could crash system
- tested all "quick" solutions i could find on internet, including
module params "acpi=off noapic", "libata.noacpi=1", "libata.force=1.5Gbps", some other woodoo magic like disabling write cache or disabling NCQ - no difference (probably tested something more, like 'norst', i forgot already) Attached zip'ed logs - one from 2.6 kernel (with trace), another from today's git kernel (entire log from boot to crash, next line in log starts again with rsyslog...).
Also, screen images from "dead" system (nothing in logs, and i can't scroll up):
- todays git kernel: http://i49.tinypic.com/js0xl2.jpg
- 3.6.10 on shutdown (crashed): http://i47.tinypic.com/2exv4fr.jpg

Because this problem is easily reproducible - i could try to get as much information as i can, if you ask. Minor problem - i do not have physical access to system, so if tests should be done with latest kernel (which hangs completely and needs access to system for restart) - i can do tests only at day, when others could access and reboot system.

Thanks.

My first thought would be that a power problem is a possibility. These kinds of setups with multiple HDs in a RAID setup are known to cause these issues in some cases if the PSU isn't adequate. It tends to show up in situations like this where all hard drives are maxed out with disk activity and they all pull their peak power at the same time - if the voltage dips too low you can get problems with the SATA link dropping, etc.

You might want to try running with only one or two disks powered up, or try moving disks to different power cables, etc. to see if that affects the problem.

I do not think PSU is a problem, because:
1) All hard disks combined draw less energy than loaded CPU, even at heavy load (from HDD datasheet: "Read/Write: 6.80 Watts; Idle 6.10 Watts" - difference is 0.7W per HDD, so < 3W combined, CPU draws ~40W when loaded, compared to idle). Loading CPU/RAM to max does not crash system at all
2) I'm planning power supplies at 2x needed power (you know, all those "Chinese Watt" system is unreliable). Anyway, should be more than enough for whole system (and CPU is almost at idle when creating filesystem, so load on PSU is very low - should be < 70W - that's almost nothing on 560W PSU, even counting "Chinese Watt" coefficient)
3) If PSU is fault - why it fails at exact the same place? Most of hardware failures have some "random" factor - you get segfaults at random places from faulty RAM, crashes from dying PSU when doing random tasks... But now it fails at exactly the same place (when using the same kernel)
4) Let's say PSU is faulty. Then how comes, that with 3.6.10 kernel i still have control over system (when it crashes) - so only disk subsystem fails? Because it has only one 12V rail - you cannot disconnect disks from system, without killing motherboard power too. But after crash i still can do ssh root@deadhost 'echo b > /proc/sysrq-trigger' - so system is alive and working well (just disks are dead)

I could imagine that motherboard itself is faulty (well, interesting anyway - why it fails only on heavy I/O load), so i will try to get Windows Server installed to check if that will work.

The CPU has a voltage regulator in front of it which can compensate for dips in the input voltage. The disks don't. The wattage figures don't necessarily account for short-duration power draw peaks. And depending on how the drives are hooked up, especially if they are all on one cable, they can potentially see a problematic voltage drop.