Re: [PATCH Part1 RFC v4 24/36] x86/compressed/acpi: move EFI config table access to common code

[Michael Roth]

On Thu, Aug 19, 2021 at 07:09:42PM +0200, Borislav Petkov wrote:
> On Thu, Aug 19, 2021 at 09:58:31AM -0500, Michael Roth wrote:
> > Not sure what you mean here. All the interfaces introduced here are used
> > by acpi.c. There is another helper added later (efi_bp_find_vendor_table())
> > in "enable SEV-SNP-validated CPUID in #VC handler", since it's not used
> > here by acpi.c.
> 
> Maybe I got confused by the amount of changes in a single patch. I'll
> try harder with your v5. :)
> 
> > There is the aforementioned efi_bp_find_vendor_table() that does the
> > simple iteration, but I wasn't sure how to build the "find one of these,
> > but this one is preferred" logic into it in a reasonable way.
> 
> Instead of efi_foreach_conf_entry() you simply do a bog-down simple
> loop and each time you stop at a table, you examine it and overwrite
> pointers, if you've found something better.
> 
> With "overwrite pointers" I mean you cache the pointers to those conf
> tables you iterate over and dig out so that you don't have to do it a
> second time. That is, *if* you need them a second time. I believe you
> call at least efi_bp_get_conf_table() twice... you get the idea.

Sorry I'm still a little confused on how to determine "something better",
since it's acpi.c that decides which GUID is preferred, whereas
efi_find_vendor_table() is a library function with no outside knowledge
other than its arguments, so to return the preferred pointer it would need
both/multiple GUIDs passed in as arguments, wouldn't it? (or a callback)

Another alternative is something like what
drivers/firmware/efi/efi.c:common_tables does, where interesting table
GUIDs are each associated with a pointer, and all the pointers can then
be initialized with to the corresponding table address with a single pass.
But would need to be careful to re-initialize those pointers when BSS gets
cleared, or declare them in __section(".data"). Is that closer to what
you were thinking?

> 
> > I could just call it once for each of these GUIDs though. I was
> > hesitant to do so since it's less efficient than existing code, but if
> > it's worth it for the simplification then I'm all for it.
> 
> Yeah, this is executed once during boot so I don't think you can make it
> more efficient than a single iteration over the config table blobs.

In v5, I've simplified things to just call efi_find_vendor_table() once
for ACPI_20_TABLE_GUID, then once for ACPI_TABLE_GUID if that's not
available. So definitely doesn't sound like what you are suggesting here,
but does at least simplify code and gets rid of the efi_foreach* stuff. But
happy to rework things if you had something else in mind.

> 
> I hope that makes more sense.
> 
> -- 
> Regards/Gruss,
>     Boris.
> 
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpeople.kernel.org%2Ftglx%2Fnotes-about-netiquette&data=04%7C01%7Cmichael.roth%40amd.com%7C2a4304e70b5b4f2137f808d963340eec%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637649897546039774%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=GKDogD%2BOCN0swhmT4RZ2%2B3JmURF3e4qq%2FzgrxxFqJt0%3D&reserved=0