I tried force enabling IOMMU for ASUS AM1I-A by changing " #define BLDCFG_IOMMU_SUPPORT FALSE " to TRUE at ./src/mainboard/asus/am1i-a/buildOpts.c and adding " device pci 0.2 on end # IOMMU " line to ./src/mainboard/asus/am1i-a/devicetree.cb : device domain 0 on subsystemid 0x1043 0x8623 inherit chip northbridge/amd/agesa/family16kb device pci 0.0 on end # Root Complex device pci 0.2 on end # IOMMU - because that's usually where AMD IOMMU PCI device resides.
However, even after doing that, I got " PCI: Static device PCI: 00:00.2 not found, disabling it. " Googling for other people's lspci did not give any results for IOMMU at 16h Jaguar platforms. Then I checked the datasheets at https://en.wikipedia.org/wiki/Puma_(microarchitecture)#External_links - and discovered that Puma's datasheet contains a lot of IOMMU references, while Jaguar's - only a few.
So: after 15h - which had IOMMU - this IOMMU has been temporarily removed from early 16h Jaguar (except from a few related bits as accidental leftovers), before being returned with late 16h Puma - together with the appearance of AMD PSP! My wild guess is that AMD needed some time to remake IOMMU for 16h to function together with AMD PSP (Platform Security Processor), and maybe they intended to implement PSP at early 16h as well - but it wasn't ready in time for early 16h so they removed IOMMU from it too.
In any case, it seems that IOMMU is impossible at AM1I-A because it is 16h Jaguar. And I tested this on the latest available CPU for AM1 socket, ultra rare Athlon 5370 (slightly faster than Athlon 5350 but no major differences). Hope this report helps to save some of your time.
Best regards, Mike Banon
Hi Mike,
I'm nearly certain that APU (Family 16h Models 00-0Fh aka Kabini) didn't have a usable IOMMU. It was fused away.
So: after 15h - which had IOMMU...
FWIW, not all Family 15h did either. That's the type of feature that could be specified or removed for different SKUs. For example in Trinity, IIRC, consumer products in mobile packages had their IOMMUs fused off because of the power they'd sip. IOMMU wasn't considered a compelling feature for mobile products at that time. Any Trinity in a desktop package allowed them to be enabled. And, unfortunately reading the BKDG doesn't give you that level of granularity -- it explains the highest level of functionality available, and sometimes more.
My wild guess is that AMD
needed some time to remake IOMMU for 16h to function together with AMD PSP (Platform Security Processor), and maybe they intended to implement PSP at early 16h as well - but it wasn't ready in time for early 16h so they removed IOMMU from it too.
Without getting into how features are developed or considered for POR, let me just end with that's definitely not how I remember the discussions for either of those two features.
Thanks, Marshall
On Thu, Jun 20, 2019 at 11:21 AM Mike Banon mikebdp2@gmail.com wrote:
I tried force enabling IOMMU for ASUS AM1I-A by changing " #define BLDCFG_IOMMU_SUPPORT FALSE " to TRUE at ./src/mainboard/asus/am1i-a/buildOpts.c and adding " device pci 0.2 on end # IOMMU " line to ./src/mainboard/asus/am1i-a/devicetree.cb : device domain 0 on subsystemid 0x1043 0x8623 inherit chip northbridge/amd/agesa/family16kb device pci 0.0 on end # Root Complex device pci 0.2 on end # IOMMU
- because that's usually where AMD IOMMU PCI device resides.
However, even after doing that, I got " PCI: Static device PCI: 00:00.2 not found, disabling it. " Googling for other people's lspci did not give any results for IOMMU at 16h Jaguar platforms. Then I checked the datasheets at https://en.wikipedia.org/wiki/Puma_(microarchitecture)#External_links
- and discovered that Puma's datasheet contains a lot of IOMMU
references, while Jaguar's - only a few.
So: after 15h - which had IOMMU - this IOMMU has been temporarily removed from early 16h Jaguar (except from a few related bits as accidental leftovers), before being returned with late 16h Puma - together with the appearance of AMD PSP! My wild guess is that AMD needed some time to remake IOMMU for 16h to function together with AMD PSP (Platform Security Processor), and maybe they intended to implement PSP at early 16h as well - but it wasn't ready in time for early 16h so they removed IOMMU from it too.
In any case, it seems that IOMMU is impossible at AM1I-A because it is 16h Jaguar. And I tested this on the latest available CPU for AM1 socket, ultra rare Athlon 5370 (slightly faster than Athlon 5350 but no major differences). Hope this report helps to save some of your time.
Best regards, Mike Banon _______________________________________________ coreboot mailing list -- coreboot@coreboot.org To unsubscribe send an email to coreboot-leave@coreboot.org
Hello,
On Thu, Jun 20, 2019 at 9:05 PM Marshall Dawson marshalldawson3rd@gmail.com wrote:
My wild guess is that AMD needed some time to remake IOMMU for 16h to function together with AMD PSP (Platform Security Processor), and maybe they intended to implement PSP at early 16h as well - but it wasn't ready in time for early 16h so they removed IOMMU from it too.
Without getting into how features are developed or considered for POR, let me just end with that's definitely not how I remember the discussions for either of those two features.
Mike, I would appreciate if you would please refrain from posting baseless speculations on this mailing list. I believe you have been asked to not do so on earlier occasions, so it should not come to you as a surprise. I would not want unfounded guesses to damage this list's credibility, and I believe others would not want that either.
Regards,
Angel Pons
Hi Marshall,
For example in Trinity, IIRC, consumer products in mobile packages had their IOMMUs fused off because of the power they'd sip. IOMMU wasn't considered a compelling feature for mobile products at that time.
Luckily mobile A10-5750M (although this is 15h Richland, slightly newer than Trinity) found in G505S - has a functional IOMMU. H owever it wasn't working at proprietary UEFI - only with coreboot. Thank you very much for your explanation, and especially
let me just end with that's definitely not how I remember the discussions for either of those two features
note.
please refrain from posting baseless speculations on this mailing list
Well, I just wanted someone to confirm or deny it. No harm from a "baseless speculation" if it is disproved at the next letter. Thanks to Marshall's letter, now I know that the return of IOMMU to 16h Puma together with the introduction of AMD PSP - is just a pure coincidence.
Best regards, Mike Banon
-
Angel Pons -
Marshall Dawson -
Mike Banon