[Devel] Re: [Q] PCI Express and ide (native) leads to irq storm?

[Vasily Averin]

Tejun Heo wrote:
> Vasily Averin wrote:
>> Alan Cox wrote:
>>> Ar Gwe, 2006-10-27 am 17:17 +0400, ysgrifennodd Vasily Averin:
>>>> Could somebody please help me to troubleshoot this issue? I've seen this issue
>>>> on the customer nodes and would like to know how I can work-around this issue
>>>> without any changes inside motherboard BIOS.
>>> If its an IRQ routing triggered problem you probably can't, at least not
>>> the IDE error. The oops wants debugging further because it shouldn't
>>> have oopsed on that error merely given up.
>> Alan,
>> I've reproduced this issue on linux 2.6.19-rc5 kernel.
>>
>> As far as I see if IDE controller is switched into native mode it shares irq
>> together with one of PCI Express Ports. It seems for me the last device is
>> guilty in this issue, becuase of it shares IDE irq on all the checked nodes.
>> and I do not know the ways to change their irq number or disable this device at all.
>>
>> I means the following devices:
>>
>> on Intel 915G-based nodes
>> 0000:00:1c.2 Class 0604: 8086:2664 (rev 03)
>> 0000:00:1c.2 PCI bridge: Intel Corporation 82801FB/FBM/FR/FW/FRW (ICH6 Family)
>> PCI Express Port 3 (rev 03)
>>
>> on Intel E7520 node:
>> 00:04.0 0604: 8086:3597 (rev 0a)
>> 00:05.0 0604: 8086:3598 (rev 0a)
>> 00:04.0 PCI bridge: Intel Corporation E7525/E7520 PCI Express Port B (rev 0a)
>> 00:05.0 PCI bridge: Intel Corporation E7520 PCI Express Port B1 (rev 0a)
>>
>> I've checked Intel chipset spec updates but do not found any related issues.
>>
>> Please see http://bugzilla.kernel.org/show_bug.cgi?id=7518 for details
> 
> Okay, I tracked this one down.  It's pretty interesting.
> 
> In short, some piix controllers including ICH7, when put into enhanced 
> mode (PCI native mode), uses BMDMA Interrupt bit as interrupt 
> pending/clear bit for *all* commands.  ie. Reading STATUS does NOT clear 
> IRQ even for PIO commands.  1 should be written to BMDMA Interrupt bit 
> to clear IRQ.  That's what's causing IRQ storm.  IDE driver does what 
> it's supposed to do but IRQ is just stuck at low active.
> 
> Fortunately, libata is immune to the problem because it does 
> ap->ops->irq_clear(ap) in ata_host_intr() regardless of command type in 
> flight.  So, not loading IDE piix and using libata to drive all piix 
> ports solves the problem.

I've disabled IDE support in the config and recompiled the kernel.
It seems you are right, problem go away, new kernel was booted without any
problems and works well.

> I guess this behavior is unique to some piixs in enhanced mode 
> considering wide use of IDE driver.  Fixing this in IDE driver is pain 
> in the ass because IRQ handler is scattered all over the place.  I'm 
> thinking about adding big warning message saying "IRQ storm can occur 
> and you better switch to libata if that happens".  But if anyone else is 
> up for the job of fixing IDE, please don't hesitate.

I'm very happy that we have found the cause of this issue, however it seems for
me you do not understand fully its severity for linux end-users.

At the present moment this issue is present in all vendor kernels, and they
cannot be installed on the huge number of end-user nodes. Moreover, end-user
nodes can have installed old Linux distribution where initscripts do not loads
all the detected modules at the boot-time. Linux may be installed  and the
following situation is possible: kernel was booted and works well until some
user will going to access the CDROM.

>From end-users point of view this issue looks mystic and very dump: is the linux
stable? is it ready for desktop? $%^&#! It crashes when I accessing the CDROM! :(

As a linux support engeneer I've seen this issue several times on the user-nodes
and it was very hard to understand what's happened and how to prevent this issue
in the future. First question is resolved now but from support point of view it
is very important to find some workaround against this issue on existing
distributions. Right now I see only one way: if this issue is detected on the
user node, we can add something like "ide=disable" into kernel commandline.

Probably the better solution exists?

thank you,
	Vasily Averin