Skip to contentSkip to conversation

x86, NMI, Add support to notify hardware error with unknown NMI

RFC Andi Kleen 4 participants 5 messages Sep 14, 2010 – Sep 15, 2010

lore
Share & tools
lore

Andi Kleen

author
#0 lore raw
Quoted 4 lines

The proper approach would be not to add hacks to the NMI code but to implement southbridge drivers - which would also have NMI callbacks.

BTW southbridges do less and less regarding PCI.

Quoted 2 lines

These are unchartered waters, but variance in that space is reducing systematically so it would be worth a shot.

You don't really need special drivers for AER -- it's fully standardized and works generically. I think the old PCI-X error bits Don was interested in were also all architectural.

The driver for the first is already there, just right now the information is not dumped in the right places.

There are a few platform specific error signals, but they tend to be rather obscure stuff. The "meat" is all in the standard.

-Andi

#1 lore raw
Quoted 14 lines · Andi Kleen
Quoted 4 lines

The proper approach would be not to add hacks to the NMI code but to implement southbridge drivers - which would also have NMI callbacks.

BTW southbridges do less and less regarding PCI.

Quoted 2 lines

These are unchartered waters, but variance in that space is reducing systematically so it would be worth a shot.

You don't really need special drivers for AER -- it's fully standardized and works generically. I think the old PCI-X error bits Don was interested in were also all architectural.

Quoted 7 lines

The driver for the first is already there, just right now the information is not dumped in the right places.

There are a few platform specific error signals, but they tend to be rather obscure stuff. The "meat" is all in the standard.

From what someone explained to me, the only platform that implements AER

correctly is Nehalem based ones. So for AMD and Intel Core arches, AER is not expected to work. Is that incorrect?

Cheers,
Don

author
#2 lore raw
Quoted 5 lines

From what someone explained to me, the only platform that implements AER correctly is Nehalem based ones. So for AMD and Intel Core arches, AER is not expected to work. Is that incorrect?

At least for Intel it's definitely not true. AER goes back a lot of generations, pretty much all that support PCI Express.

There were some bugs in it of course and some workarounds needed, but no show stoppers in any particular part to my knowledge.

Haven't kept fully track of AMD systems recently, but would surprise me if it was the case on AMD either.

This applies to servers. On clients AER is more spotty and various simply don't have it.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only.
#3 lore raw
In reply to Andi Kleen
Quoted 5 lines · Some One
Quoted 3 lines

The proper approach would be not to add hacks to the NMI code but to implement southbridge drivers - which would also have NMI callbacks.

BTW southbridges do less and less regarding PCI.

Except WICTCR.

Quoted 4 lines
Quoted 2 lines

These are unchartered waters, but variance in that space is reducing systematically so it would be worth a shot.

You don't really need special drivers for AER [...]

On the contrary, we need proper driverization for _everything_ new. Embedded x86 is here to stay, so we are abstracting away each and every bit of the platform. See struct x86_ops for a highlevel platform driver
- but a more specific, southbridge-encompassing driver framework can be created too.

Thanks,

Ingo

#4 lore raw
Quoted 12 lines · Ingo Molnar
Quoted 5 lines · Some One
Quoted 3 lines

The proper approach would be not to add hacks to the NMI code but to implement southbridge drivers - which would also have NMI callbacks.

BTW southbridges do less and less regarding PCI.

Except WICTCR.

Quoted 3 lines
Quoted 2 lines

These are unchartered waters, but variance in that space is reducing systematically so it would be worth a shot.

Quoted 7 lines
Quoted 1 line

You don't really need special drivers for AER [...]

On the contrary, we need proper driverization for _everything_ new. Embedded x86 is here to stay, so we are abstracting away each and every bit of the platform. See struct x86_ops for a highlevel platform driver
- but a more specific, southbridge-encompassing driver framework can be created too.

All in all, we can have proper drivers to printk/write to NVRAM the error information collected for unknown NMI, regardless they are AER or Southbridge driver or both. And, the drivers will register handlers for unknown NMI. Do you agree?

Best Regards,
Huang Ying