Hi,
Currently APEI depends on x86 architecture. It is because of many x86 specific features like "IA-32 Architecture Corrected Machine Check " error source or NMI hardware error notification. However, many other features like "PCI Express Device AER Structure" or GHES via external interrupt can be still used perfectly by other architectures. So my idea is to move x86 dependency away form Kconfig to APEI areas where it really applies to.
I have started refactoring ghes.c driver in that direction. And here comes my confusion, how should we treat x86 related parts, as fixed profile? (which means we could use ACPI_REDUCED_HARDWARE or CONFIG_ACPI_REDUCED_HARDWARE_ONLY define). I would like to ask for your opinion.
Regards, Tomasz
On Wednesday, March 26, 2014 01:08:10 PM Tomasz Nowicki wrote:
Hi,
Hi,
This is a question for Tony, Boris and Mauro (CCed now).
Currently APEI depends on x86 architecture. It is because of many x86 specific features like "IA-32 Architecture Corrected Machine Check " error source or NMI hardware error notification. However, many other features like "PCI Express Device AER Structure" or GHES via external interrupt can be still used perfectly by other architectures. So my idea is to move x86 dependency away form Kconfig to APEI areas where it really applies to.
I have started refactoring ghes.c driver in that direction. And here comes my confusion, how should we treat x86 related parts, as fixed profile? (which means we could use ACPI_REDUCED_HARDWARE or CONFIG_ACPI_REDUCED_HARDWARE_ONLY define). I would like to ask for your opinion.
Thanks!
Em Wed, 26 Mar 2014 15:55:07 +0100 "Rafael J. Wysocki" rjw@rjwysocki.net escreveu:
On Wednesday, March 26, 2014 01:08:10 PM Tomasz Nowicki wrote:
Hi,
Hi,
This is a question for Tony, Boris and Mauro (CCed now).
Currently APEI depends on x86 architecture. It is because of many x86 specific features like "IA-32 Architecture Corrected Machine Check " error source or NMI hardware error notification. However, many other features like "PCI Express Device AER Structure" or GHES via external interrupt can be still used perfectly by other architectures. So my idea is to move x86 dependency away form Kconfig to APEI areas where it really applies to.
I have started refactoring ghes.c driver in that direction. And here comes my confusion, how should we treat x86 related parts, as fixed profile? (which means we could use ACPI_REDUCED_HARDWARE or CONFIG_ACPI_REDUCED_HARDWARE_ONLY define). I would like to ask for your opinion.
That's a good question, and probably depends on how are you mapping the ACPI changes. For example, are you moving acpi out of /arch?
As I answered to a similar questioning, IMHO, the better would be to have the hardware error report mechanisms on /drivers/ras, and have there some Kconfig items that would depend on X86 to enable certain drivers.
Also, I don't like to have something like ACPI_REDUCED_foo. IMHO, the better would be to do the reverse: to have Kconfig symbols enabling the extra X86-specific functionality, and have them mapped into separate files/drivers, with proper KConfig names, like ACPI_X86 or ACPI_X86_NMI.
Yet, it would be better if you could be a little more specific about what are your plans and what are the common/not-common features that you're mapping.
Regards, Mauro
On Wed, Mar 26, 2014 at 12:10:47PM -0300, Mauro Carvalho Chehab wrote:
Yet, it would be better if you could be a little more specific about what are your plans and what are the common/not-common features that you're mapping.
Well, I don't see anything x86-specific in ghes.c on a quick scan - just the GHES gunk itself, which is the spec. Which begs the question, how much of the APEI spec is ARM going to implement and if it is a subset, this definitely needs to be handled cleanly. Judging by your text, it seems like you want to ignore the correctable errors part...?
So yes, Tomasz, you want to be much more specific here :-)
Thanks.
On 03/26/14 21:36, Borislav Petkov wrote:
On Wed, Mar 26, 2014 at 12:10:47PM -0300, Mauro Carvalho Chehab wrote:
Yet, it would be better if you could be a little more specific about what are your plans and what are the common/not-common features that you're mapping.
Well, I don't see anything x86-specific in ghes.c on a quick scan - just the GHES gunk itself, which is the spec. Which begs the question, how much of the APEI spec is ARM going to implement and if it is a subset, this definitely needs to be handled cleanly. Judging by your text, it seems like you want to ignore the correctable errors part...?
So yes, Tomasz, you want to be much more specific here :-)
Thanks for your comment. Please refer to my email to Mauro. I have put more explanation there.
Correctable errors are fine for ARM. But ARM is not going to have NMI ever. So as responded to Mauro, I am focusing on making NMI as the feature for ghes.c.
Thanks, Tomasz
On 03/26/14 16:10, Mauro Carvalho Chehab wrote:
Em Wed, 26 Mar 2014 15:55:07 +0100 "Rafael J. Wysocki" rjw@rjwysocki.net escreveu:
On Wednesday, March 26, 2014 01:08:10 PM Tomasz Nowicki wrote:
Hi,
Hi,
This is a question for Tony, Boris and Mauro (CCed now).
Currently APEI depends on x86 architecture. It is because of many x86 specific features like "IA-32 Architecture Corrected Machine Check " error source or NMI hardware error notification. However, many other features like "PCI Express Device AER Structure" or GHES via external interrupt can be still used perfectly by other architectures. So my idea is to move x86 dependency away form Kconfig to APEI areas where it really applies to.
I have started refactoring ghes.c driver in that direction. And here comes my confusion, how should we treat x86 related parts, as fixed profile? (which means we could use ACPI_REDUCED_HARDWARE or CONFIG_ACPI_REDUCED_HARDWARE_ONLY define). I would like to ask for your opinion.
That's a good question, and probably depends on how are you mapping the ACPI changes. For example, are you moving acpi out of /arch?
As I answered to a similar questioning, IMHO, the better would be to have the hardware error report mechanisms on /drivers/ras, and have there some Kconfig items that would depend on X86 to enable certain drivers.
Also, I don't like to have something like ACPI_REDUCED_foo. IMHO, the better would be to do the reverse: to have Kconfig symbols enabling the extra X86-specific functionality, and have them mapped into separate files/drivers, with proper KConfig names, like ACPI_X86 or ACPI_X86_NMI.
Yet, it would be better if you could be a little more specific about what are your plans and what are the common/not-common features that you're mapping.
Yes and sorry, I should have been more specific here.
After scanning APEI code it seems like NMI notification of GHES implies APEI x86 dependency for Kconfig, so I am targeting ghes.c.
I agree that ACPI_REDUCED_foo is not suitable for that purpose. However, ACPI_X86_NMI sounds good to me. I also have been thinking of moving NMI code (from ghes.c) to separate file but NMI and IRQ context are tightly coupled. That convinced me to leave it in ghes.c for now but I need to look at it closer.
Thanks, Tomasz
-
Borislav Petkov -
Mauro Carvalho Chehab -
Rafael J. Wysocki -
Tomasz Nowicki