Skip to contentSkip to conversation

Hangs and reboots under high loads, oops with DEBUG_SHIRQ

DISCUSSION Attila Nagy 5 participants 7 messages Jul 30, 2007 – Aug 4, 2007

lore
Share & tools
lore
#0 lore raw

Hello,

I have four identical machines, based on Supermicro X7DBE motherboards. All of the machines have two Xeon 5130 CPUs, 16GB RAM and two add-on
cards: an Areca 1261 SATA RAID HBA and a Qlogic QLA2342 fibre channel HBA.

I would like to use these as file servers (via FC), but during the performance and reliabilty tests it turned out that the machines are very unreliable, despite that they seemed to be OK hardware-wise (memtest and the usual stuff).

During the debugging of this (seemingly) high IO load related problem, I have observed the following:
- when MSI is enabled (the first iteration), the machines sometimes "hang", but not the whole system, just the SCSI target subsystem (SCST), which makes
heavy IO on the Areca (arcmsr) and the Qlogic (patched qla2xxx) HBAs
- when MSI is disabled, I couldn't reproduce that hung up state, instead the machines sometimes throw an MCE (see below), but I couldn't find its cause
- when MSI is disabled, and CONFIG_DEBUG_SHIRQ is enabled, the machines
can't even boot normally, I get an oops instantly during the kernel initialization
- with MSI disabled sometimes the machines fail to respond, the ssh sessions terminate and on the console I can't type for very long seconds. I have nearly all debugging turned on, but can't see anything in the logs or on the console. The machine recovers from this hang automatically. The whole thing seems like when a high (eg. network) interrupt activity happens on a highly loaded machine, but I could observe this even after a fresh boot, without anything (of course minus the standard stuff, sshd, and the others) running on the machine.

The kernel is 2.6.21.5 (I've tried 2.6.18, the effects are the same), running in 64 bit mode.

The oops I get with MSI disabled and CONFIG_DEBUG_SHIRQ enabled:

[   92.681320] NET: Registered protocol family 17
[   93.491658] Unable to handle kernel NULL pointer dereference at 0000000000000000 RIP:
[   93.557402]  [<0000000000000000>]
[   93.626770] PGD 0
[   93.651106] Oops: 0010 [1] SMP
[   93.689170] CPU 1
[   93.713506] Modules linked in:
[   93.750322] Pid: 1, comm: swapper Not tainted 2.6.21.5 #1
[   93.815011] RIP: 0010:[<0000000000000000>]  [<0000000000000000>]
[   93.887187] RSP: 0018:ffff81042fc5dc68  EFLAGS: 00010002
[   93.950836] RAX: ffff81042fbe6b70 RBX: 0000000000000202 RCX: ffff81042fbe6b70
[   94.036323] RDX: ffffc20000040000 RSI: ffff81042f51cdf8 RDI: ffff81042fbe6800
[   94.121812] RBP: ffff81042fc5dd10 R08: 0000000000000000 R09: ffff81042f4c0ea8
[   94.207298] R10: 0000000000000000 R11: ffff81042fbe6800 R12: 00000000fffffff4
[   94.292788] R13: ffff81042fbe6000 R14: 0000000000000001 R15: ffffffff80399450
[   94.378275] FS:  0000000000000000(0000) GS:ffff81042fc694c8(0000) knlGS:0000000000000000
[   94.475307] CS:  0010 DS: 0018 ES: 0018 CR0: 000000008005003b
[   94.544153] CR2: 0000000000000000 CR3: 0000000000201000 CR4: 00000000000006e0
[   94.629643] Process swapper (pid: 1, threadinfo ffff81042fc5c000, task ffff81042fc58040)
[   94.726673] Stack:  ffffffff80399559 ffff81042fc5dca0 ffffffff802121b2 ffff81042f4c0ea0
[   94.823603]  ffff81042fc5dca0 ffffffff80221cd7 ffffffff802addb5 ffff81042fbe6800
[   94.913042]  ffffffff8020c9bc ffff81042fbe6b70 0000000000000246 ffff81042fbe6b70
[   95.000194] Call Trace:
[   95.031814]  [<ffffffff80399559>] e1000_intr+0x109/0x590
[   95.095461]  [<ffffffff802121b2>] poison_obj+0x42/0x60
[   95.157027]  [<ffffffff80221cd7>] dbg_redzone1+0x17/0x30
[   95.220676]  [<ffffffff802addb5>] request_irq+0x95/0x150
[   95.284324]  [<ffffffff8020c9bc>] cache_alloc_debugcheck_after+0x17c/0x1c0
[   95.366690]  [<ffffffff8020a43d>] kmem_cache_alloc+0xcd/0xf0
[   95.434500]  [<ffffffff80399450>] e1000_intr+0x0/0x590
[   95.496067]  [<ffffffff802ade00>] request_irq+0xe0/0x150
[   95.559716]  [<ffffffff8039558c>] e1000_request_irq+0x3c/0x80
[   95.628564]  [<ffffffff803985bc>] e1000_open+0x5c/0x100
[   95.691172]  [<ffffffff8041d937>] dev_open+0x37/0x80
[   95.750661]  [<ffffffff8041becd>] dev_change_flags+0x6d/0x150
[   95.819508]  [<ffffffff80616565>] ip_auto_config+0x175/0xea0
[   95.887317]  [<ffffffff80442f88>] tcp_set_default_congestion_control+0x18/0x70
[   95.973947]  [<ffffffff80442fcf>] tcp_set_default_congestion_control+0x5f/0x70
[   96.060582]  [<ffffffff80265236>] _spin_unlock+0x26/0x30
[   96.124227]  [<ffffffff805f1754>] init+0x1a4/0x2b0
[   96.181635]  [<ffffffff802a0e7b>] trace_hardirqs_on+0x14b/0x180
[   96.252563]  [<ffffffff8025ff28>] child_rip+0xa/0x12
[   96.312051]  [<ffffffff8026563b>] _spin_unlock_irq+0x2b/0x40
[   96.379859]  [<ffffffff8025f63c>] restore_args+0x0/0x30
[   96.442467]  [<ffffffff805f15b0>] init+0x0/0x2b0
[   96.497795]  [<ffffffff8025ff1e>] child_rip+0x0/0x12
[   96.557282]
[   96.575170]
[   96.575171] Code:  Bad RIP value.
[   96.633203] RIP  [<0000000000000000>]
[   96.677297]  RSP <ffff81042fc5dc68>
[   96.719105] CR2: 0000000000000000
[   96.758835] Kernel panic - not syncing: Attempted to kill init!

MCE:

[153103.918654] HARDWARE ERROR
[153103.918655] CPU 1: Machine Check Exception:                5 Bank 0: b200004010000400
[153104.066037] RIP !INEXACT! 10:<ffffffff802569e6> {mwait_idle+0x46/0x60}
[153104.145699] TSC 1167e915e93ce
[153104.183554] This is not a software problem!
[153104.234724] Run through mcelog --ascii to decode and contact your hardware vendor
[153104.325517]
[153104.325518] HARDWARE ERROR
[153104.325519] CPU 1: Machine Check Exception:                5 Bank 5: b200221024080400
[153104.472883] RIP !INEXACT! 10:<ffffffff802569e6> {mwait_idle+0x46/0x60}
[153104.552546] TSC 1167e915e9ea8
[153104.590402] This is not a software problem!
[153104.641572] Run through mcelog --ascii to decode and contact your hardware vendor
[153104.732365] Kernel panic - not syncing: Machine check

I've got exactly the same errors (only the TSC and the CPU value changing) on all four machines, could this really be a hardware error?

full dmesg: http://people.fsn.hu/~bra/linux/x7dbe-20070730/with_debug_shirq
dmesg with MSI enabled: http://people.fsn.hu/~bra/linux/x7dbe-20070730/with_msi_and_debug_shirq
kernel config: http://people.fsn.hu/~bra/linux/x7dbe-20070730/config

I've tried to disable all possible devices which consume interrupts, and placed the cards into various slots (the FC HBA is PCI-X, which the Areca is PCI-E). Currently the arcmsr and (one of, it's a dual channel HBA) qla2xxx are on a shared IRQ.

Could you please help?
Do you think this is related to the strange hang under high IO load, the occasional, complete "blackouts", where all ssh network sessions time out, but the machine recovers, and the MCEs?

I'm open for anything, have serial consoles, etc, if needed.

Please keep me on CC, I'm not subscribed.

Thanks,

#1 lore raw

O> MCE:

Quoted 8 lines
[153103.918654] HARDWARE ERROR
[153103.918655] CPU 1: Machine Check Exception:                5 Bank 0: 
b200004010000400
[153104.066037] RIP !INEXACT! 10:<ffffffff802569e6> {mwait_idle+0x46/0x60}
[153104.145699] TSC 1167e915e93ce
[153104.183554] This is not a software problem!
[153104.234724] Run through mcelog --ascii to decode and contact your 

hardware vendor

If you it through mcelog as it suggests it wil decode the meaning of the MCE data and that should give you some idea. Generally speaking MCE errors are real hardware errors but can certainly be caused by external factors (power supply glitches, heat etc)

#2 lore raw
Quoted 10 lines · Alan Cox

O> MCE:

Quoted 7 lines
[153103.918654] HARDWARE ERROR
[153103.918655] CPU 1: Machine Check Exception:                5 Bank 0: b200004010000400
[153104.066037] RIP !INEXACT! 10:<ffffffff802569e6> {mwait_idle+0x46/0x60}
[153104.145699] TSC 1167e915e93ce
[153104.183554] This is not a software problem!
[153104.234724] Run through mcelog --ascii to decode and contact your hardware vendor
Quoted 5 lines

If you it through mcelog as it suggests it wil decode the meaning of the MCE data and that should give you some idea. Generally speaking MCE errors are real hardware errors but can certainly be caused by external factors (power supply glitches, heat etc)

Sorry, of course I ran that through mcelog, but inadvertently attached the original version.

I've tried the machines with two types of power sources (different UPSes, line filtering, etc, and the chassis have redundant PSes), monitoring the temperatures (seems to be OK, the CPUs don't go over 30 °C even under load). I have the latest BIOS for the motherboard.
But I will recheck everything.

BTW, here's the output from mcelog, I see this occasionally on all four machines:

HARDWARE ERROR
HARDWARE ERROR. This is *NOT* a software problem!
Please contact your hardware vendor
CPU 1 BANK 0 TSC 1167e915e93ce
MCG status:RIPV MCIP
MCi status:
Uncorrected error
Error enabled
Processor context corrupt
MCA: Internal Timer error
STATUS b200004010000400 MCGSTATUS 5
This is not a software problem!
Run through mcelog --ascii to decode and contact your hardware vendor

HARDWARE ERROR
HARDWARE ERROR. This is *NOT* a software problem!
Please contact your hardware vendor
CPU 1 BANK 5 TSC 1167e915e9ea8
MCG status:RIPV MCIP
MCi status:
Uncorrected error
Error enabled
Processor context corrupt
MCA: Internal Timer error
STATUS b200221024080400 MCGSTATUS 5
This is not a software problem!
Run through mcelog --ascii to decode and contact your hardware vendor

Thanks,

#3 lore raw
Quoted 54 lines · Attila Nagy
Quoted 10 lines · Alan Cox

O> MCE:

Quoted 7 lines
[153103.918654] HARDWARE ERROR
[153103.918655] CPU 1: Machine Check Exception:                5 Bank 0: b200004010000400
[153104.066037] RIP !INEXACT! 10:<ffffffff802569e6> {mwait_idle+0x46/0x60}
[153104.145699] TSC 1167e915e93ce
[153104.183554] This is not a software problem!
[153104.234724] Run through mcelog --ascii to decode and contact your hardware vendor
Quoted 5 lines

If you it through mcelog as it suggests it wil decode the meaning of the MCE data and that should give you some idea. Generally speaking MCE errors are real hardware errors but can certainly be caused by external factors (power supply glitches, heat etc)

Sorry, of course I ran that through mcelog, but inadvertently attached the original version.

I've tried the machines with two types of power sources (different UPSes, line filtering, etc, and the chassis have redundant PSes), monitoring the temperatures (seems to be OK, the CPUs don't go over 30 °C even under load). I have the latest BIOS for the motherboard.
But I will recheck everything.

BTW, here's the output from mcelog, I see this occasionally on all four machines:

HARDWARE ERROR
HARDWARE ERROR. This is *NOT* a software problem!
Please contact your hardware vendor
CPU 1 BANK 0 TSC 1167e915e93ce
MCG status:RIPV MCIP
MCi status:
Uncorrected error
Error enabled
Processor context corrupt
MCA: Internal Timer error
STATUS b200004010000400 MCGSTATUS 5
This is not a software problem!
Run through mcelog --ascii to decode and contact your hardware vendor

HARDWARE ERROR
HARDWARE ERROR. This is *NOT* a software problem!
Please contact your hardware vendor
CPU 1 BANK 5 TSC 1167e915e9ea8
MCG status:RIPV MCIP
MCi status:
Uncorrected error
Error enabled
Processor context corrupt
MCA: Internal Timer error
STATUS b200221024080400 MCGSTATUS 5
This is not a software problem!
Run through mcelog --ascii to decode and contact your hardware vendor

Quoted 8 lines

Thanks,

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Attila,

We had some issues with very similar boards all of the problems seem to be around the PCIX bus area of the machine, setting the
PCIX buses to 66 mhz in the bios made things stable (but slow). Not using
the PCIX bus also seemed to make things work. We got MCE's and
other odd crashes under heavy IO loads. I believe turning things
down to 100mhz made things more stable, but things still crashed.

Supermicro reported being able to fix the issue with:
setting the PCI Configuration -> PCI-e I/O performance setting to Colasce 128B.

I am not exactly sure where to set it as we did not try it as we had already changed to a different motherboard that did not have the issue.

If this works please tell me.

Roger

#4 lore raw
Quoted 42 lines · Roger Heflin
Quoted 14 lines · Attila Nagy

HARDWARE ERROR
HARDWARE ERROR. This is *NOT* a software problem!
Please contact your hardware vendor
CPU 1 BANK 0 TSC 1167e915e93ce
MCG status:RIPV MCIP
MCi status:
Uncorrected error
Error enabled
Processor context corrupt
MCA: Internal Timer error
STATUS b200004010000400 MCGSTATUS 5
This is not a software problem!
Run through mcelog --ascii to decode and contact your hardware vendor

Quoted 13 lines

HARDWARE ERROR
HARDWARE ERROR. This is *NOT* a software problem!
Please contact your hardware vendor
CPU 1 BANK 5 TSC 1167e915e9ea8
MCG status:RIPV MCIP
MCi status:
Uncorrected error
Error enabled
Processor context corrupt
MCA: Internal Timer error
STATUS b200221024080400 MCGSTATUS 5
This is not a software problem!
Run through mcelog --ascii to decode and contact your hardware vendor

Attila,

We had some issues with very similar boards all of the problems seem to be around the PCIX bus area of the machine, setting the
PCIX buses to 66 mhz in the bios made things stable (but slow). Not using
the PCIX bus also seemed to make things work. We got MCE's and
other odd crashes under heavy IO loads. I believe turning things
down to 100mhz made things more stable, but things still crashed.

Supermicro reported being able to fix the issue with:
setting the PCI Configuration -> PCI-e I/O performance setting to Colasce 128B.

Quoted 5 lines

I am not exactly sure where to set it as we did not try it as we had already changed to a different motherboard that did not have the issue.

If this works please tell me.

Roger, you are my hero. :)
With that PCI-e setting (again, for the record, this is on a Supermicro X7DBE motherboard, and the BIOS setting is PCIe I/O performance, which has two states: Coalesce and Payload 256B) all of the four machines have survived a half day of continous bashing. Previously one, or two machines typically fell off after such amount of IO load, so it looks promising so far. I hope this won't change over the time.

BTW, this is still with 2.6.21.5, because the SCSI target stuff I use (SCST) has some -I hope temporary- problems with changed (deleted) interfaces in newer kernels.

Should the DEBUG_SHIRQ problem in e1000 affect stability (or performance)?

Thanks,

#5 lore raw

Hello Atilla ,

Quoted 55 lines · Attila Nagy
Quoted 42 lines · Roger Heflin
Quoted 14 lines · Attila Nagy

HARDWARE ERROR
HARDWARE ERROR. This is *NOT* a software problem!
Please contact your hardware vendor
CPU 1 BANK 0 TSC 1167e915e93ce
MCG status:RIPV MCIP
MCi status:
Uncorrected error
Error enabled
Processor context corrupt
MCA: Internal Timer error
STATUS b200004010000400 MCGSTATUS 5
This is not a software problem!
Run through mcelog --ascii to decode and contact your hardware vendor

Quoted 13 lines

HARDWARE ERROR
HARDWARE ERROR. This is *NOT* a software problem!
Please contact your hardware vendor
CPU 1 BANK 5 TSC 1167e915e9ea8
MCG status:RIPV MCIP
MCi status:
Uncorrected error
Error enabled
Processor context corrupt
MCA: Internal Timer error
STATUS b200221024080400 MCGSTATUS 5
This is not a software problem!
Run through mcelog --ascii to decode and contact your hardware vendor

Attila,

We had some issues with very similar boards all of the problems seem to be around the PCIX bus area of the machine, setting the
PCIX buses to 66 mhz in the bios made things stable (but slow). Not using
the PCIX bus also seemed to make things work. We got MCE's and
other odd crashes under heavy IO loads. I believe turning things
down to 100mhz made things more stable, but things still crashed.

Supermicro reported being able to fix the issue with:
setting the PCI Configuration -> PCI-e I/O performance setting to Colasce 128B.

Quoted 5 lines

I am not exactly sure where to set it as we did not try it as we had already changed to a different motherboard that did not have the issue.

If this works please tell me.

Roger, you are my hero. :)
With that PCI-e setting (again, for the record, this is on a Supermicro X7DBE motherboard, and the BIOS setting is PCIe I/O performance, which has two states: Coalesce and Payload 256B) all of the four machines have survived a half day of continous bashing. Previously one, or two machines typically fell off after such amount of IO load, so it looks promising so far. I hope this won't change over the time.

Quoted 7 lines

BTW, this is still with 2.6.21.5, because the SCSI target stuff I use (SCST) has some -I hope temporary- problems with changed (deleted) interfaces in newer kernels.

Should the DEBUG_SHIRQ problem in e1000 affect stability (or performance)?

Thanks,

I too have a SuperMicro MB , But it is a X7DB8 . Same symptoms .
Reported MCE problems here a couple of times .
I set the BIOS setting 'PCIe I/O performance', to 'Coalesce' . For everyones information , stability went way up , scsi IO is ~ half
, But if there's no stability ...

I'm going to try their 1.3b bios update & see if that helps any . iirc , Some said they'd already acquired the lastest for their MB &
that did not help them at all . What th eheck I'll give it a try anyway .
Hth , JimL

Signature · 5 lines
+-----------------------------------------------------------------+
| James   W.   Laferriere | System   Techniques | Give me VMS     |
| Network        Engineer | 663  Beaumont  Blvd |  Give me Linux  |
| [email protected] | Pacifica, CA. 94044 |   only  on  AXP |
+-----------------------------------------------------------------+
#6 lore raw
In reply to Attila Nagy
Quoted 28 lines · Attila Nagy

Hello,

I have four identical machines, based on Supermicro X7DBE motherboards. All of the machines have two Xeon 5130 CPUs, 16GB RAM and two add-on
cards: an Areca 1261 SATA RAID HBA and a Qlogic QLA2342 fibre channel HBA.

I would like to use these as file servers (via FC), but during the performance and reliabilty tests it turned out that the machines are very unreliable, despite that they seemed to be OK hardware-wise (memtest and the usual stuff).

During the debugging of this (seemingly) high IO load related problem, I have observed the following:
- when MSI is enabled (the first iteration), the machines sometimes
"hang", but
not the whole system, just the SCSI target subsystem (SCST), which makes
heavy IO on the Areca (arcmsr) and the Qlogic (patched qla2xxx) HBAs
- when MSI is disabled, I couldn't reproduce that hung up state, instead the machines sometimes throw an MCE (see below), but I couldn't find its cause
- when MSI is disabled, and CONFIG_DEBUG_SHIRQ is enabled, the machines
can't even boot normally, I get an oops instantly during the kernel initialization
- with MSI disabled sometimes the machines fail to respond, the ssh
sessions terminate
and on the console I can't type for very long seconds. I have nearly all debugging turned on, but can't see anything in the logs or on the console. The machine recovers from this hang automatically. The whole thing seems like when a high (eg. network) interrupt activity happens on a highly loaded machine, but I could observe this even after a fresh boot, without anything (of course minus the standard stuff, sshd, and the others) running on the machine.

Quoted 1 line

The kernel is 2.6.21.5 (I've tried 2.6.18, the effects are the same), running in 64 bit mode.

something is definately not happy on this system. There was a e1000 fix related to DEBUG_SHIRQ in 2.6.22, so I definately advise you to test 2.6.22.1 immediately - however:

Quoted 58 lines

The oops I get with MSI disabled and CONFIG_DEBUG_SHIRQ enabled:

[   92.681320] NET: Registered protocol family 17
[   93.491658] Unable to handle kernel NULL pointer dereference at 0000000000000000 RIP:
[   93.557402]  [<0000000000000000>]
[   93.626770] PGD 0
[   93.651106] Oops: 0010 [1] SMP
[   93.689170] CPU 1
[   93.713506] Modules linked in:
[   93.750322] Pid: 1, comm: swapper Not tainted 2.6.21.5 #1
[   93.815011] RIP: 0010:[<0000000000000000>]  [<0000000000000000>]
[   93.887187] RSP: 0018:ffff81042fc5dc68  EFLAGS: 00010002
[   93.950836] RAX: ffff81042fbe6b70 RBX: 0000000000000202 RCX: ffff81042fbe6b70
[   94.036323] RDX: ffffc20000040000 RSI: ffff81042f51cdf8 RDI: ffff81042fbe6800
[   94.121812] RBP: ffff81042fc5dd10 R08: 0000000000000000 R09: ffff81042f4c0ea8
[   94.207298] R10: 0000000000000000 R11: ffff81042fbe6800 R12: 00000000fffffff4
[   94.292788] R13: ffff81042fbe6000 R14: 0000000000000001 R15: ffffffff80399450
[   94.378275] FS:  0000000000000000(0000) GS:ffff81042fc694c8(0000) knlGS:0000000000000000
[   94.475307] CS:  0010 DS: 0018 ES: 0018 CR0: 000000008005003b
[   94.544153] CR2: 0000000000000000 CR3: 0000000000201000 CR4: 00000000000006e0
[   94.629643] Process swapper (pid: 1, threadinfo ffff81042fc5c000, task ffff81042fc58040)
[   94.726673] Stack:  ffffffff80399559 ffff81042fc5dca0 ffffffff802121b2 ffff81042f4c0ea0
[   94.823603]  ffff81042fc5dca0 ffffffff80221cd7 ffffffff802addb5 ffff81042fbe6800
[   94.913042]  ffffffff8020c9bc ffff81042fbe6b70 0000000000000246 ffff81042fbe6b70
[   95.000194] Call Trace:
[   95.031814]  [<ffffffff80399559>] e1000_intr+0x109/0x590
[   95.095461]  [<ffffffff802121b2>] poison_obj+0x42/0x60
[   95.157027]  [<ffffffff80221cd7>] dbg_redzone1+0x17/0x30
[   95.220676]  [<ffffffff802addb5>] request_irq+0x95/0x150
[   95.284324]  [<ffffffff8020c9bc>] cache_alloc_debugcheck_after+0x17c/0x1c0
[   95.366690]  [<ffffffff8020a43d>] kmem_cache_alloc+0xcd/0xf0
[   95.434500]  [<ffffffff80399450>] e1000_intr+0x0/0x590
[   95.496067]  [<ffffffff802ade00>] request_irq+0xe0/0x150
[   95.559716]  [<ffffffff8039558c>] e1000_request_irq+0x3c/0x80
[   95.628564]  [<ffffffff803985bc>] e1000_open+0x5c/0x100
[   95.691172]  [<ffffffff8041d937>] dev_open+0x37/0x80
[   95.750661]  [<ffffffff8041becd>] dev_change_flags+0x6d/0x150
[   95.819508]  [<ffffffff80616565>] ip_auto_config+0x175/0xea0
[   95.887317]  [<ffffffff80442f88>] tcp_set_default_congestion_control+0x18/0x70
[   95.973947]  [<ffffffff80442fcf>] tcp_set_default_congestion_control+0x5f/0x70
[   96.060582]  [<ffffffff80265236>] _spin_unlock+0x26/0x30
[   96.124227]  [<ffffffff805f1754>] init+0x1a4/0x2b0
[   96.181635]  [<ffffffff802a0e7b>] trace_hardirqs_on+0x14b/0x180
[   96.252563]  [<ffffffff8025ff28>] child_rip+0xa/0x12
[   96.312051]  [<ffffffff8026563b>] _spin_unlock_irq+0x2b/0x40
[   96.379859]  [<ffffffff8025f63c>] restore_args+0x0/0x30
[   96.442467]  [<ffffffff805f15b0>] init+0x0/0x2b0
[   96.497795]  [<ffffffff8025ff1e>] child_rip+0x0/0x12
[   96.557282]
[   96.575170]
[   96.575171] Code:  Bad RIP value.
[   96.633203] RIP  [<0000000000000000>]
[   96.677297]  RSP <ffff81042fc5dc68>
[   96.719105] CR2: 0000000000000000
[   96.758835] Kernel panic - not syncing: Attempted to kill init!

MCE:

Quoted 14 lines
[153103.918654] HARDWARE ERROR
[153103.918655] CPU 1: Machine Check Exception:                5 Bank 0: b200004010000400
[153104.066037] RIP !INEXACT! 10:<ffffffff802569e6> {mwait_idle+0x46/0x60}
[153104.145699] TSC 1167e915e93ce
[153104.183554] This is not a software problem!
[153104.234724] Run through mcelog --ascii to decode and contact your hardware vendor
[153104.325517]
[153104.325518] HARDWARE ERROR
[153104.325519] CPU 1: Machine Check Exception:                5 Bank 5: b200221024080400
[153104.472883] RIP !INEXACT! 10:<ffffffff802569e6> {mwait_idle+0x46/0x60}
[153104.552546] TSC 1167e915e9ea8
[153104.590402] This is not a software problem!
[153104.641572] Run through mcelog --ascii to decode and contact your hardware vendor
[153104.732365] Kernel panic - not syncing: Machine check

this is serious problems that might not be resolved and be a symptom of a true hardware issue. Looking at the time it seems an issue on itself and unrelated to the e1000 debug_shirq fix.

Quoted 2 lines

I've got exactly the same errors (only the TSC and the CPU value changing) on all four machines, could this really be a hardware error?

yes

Quoted 9 lines

full dmesg: http://people.fsn.hu/~bra/linux/x7dbe-20070730/with_debug_shirq
dmesg with MSI enabled: http://people.fsn.hu/~bra/linux/x7dbe-20070730/with_msi_and_debug_shirq
kernel config: http://people.fsn.hu/~bra/linux/x7dbe-20070730/config

I've tried to disable all possible devices which consume interrupts, and placed the cards into various slots (the FC HBA is PCI-X, which the Areca is PCI-E). Currently the arcmsr and (one of, it's a dual channel HBA) qla2xxx are on a shared IRQ.

Quoted 4 lines

Could you please help?
Do you think this is related to the strange hang under high IO load, the occasional, complete "blackouts", where all ssh network sessions time out, but the machine recovers, and the MCEs?

Like I said, please try 2.6.22 which should have the e1000 issue fixed. The MCE looks real and a different problem

Auke