Re: [PATCH v7 2/2] schemas: Add some common reserved-memory usages

[Ard Biesheuvel]

On Thu, 4 Jan 2024 at 01:25, Chiu, Chasel <chasel.chiu@xxxxxxxxx> wrote:
>
>
>
> > -----Original Message-----
> > From: Ard Biesheuvel <ardb@xxxxxxxxxx>
> > Sent: Wednesday, January 3, 2024 7:22 AM
> > To: Chiu, Chasel <chasel.chiu@xxxxxxxxx>
> > Cc: Simon Glass <sjg@xxxxxxxxxxxx>; devicetree@xxxxxxxxxxxxxxx; Mark Rutland
> > <mark.rutland@xxxxxxx>; Rob Herring <robh@xxxxxxxxxx>; Tan, Lean Sheng
> > <sheng.tan@xxxxxxxxxxxxx>; lkml <linux-kernel@xxxxxxxxxxxxxxx>; Dhaval
> > Sharma <dhaval@xxxxxxxxxxxx>; Brune, Maximilian
> > <maximilian.brune@xxxxxxxxxxxxx>; Yunhui Cui <cuiyunhui@xxxxxxxxxxxxx>;
> > Dong, Guo <guo.dong@xxxxxxxxx>; Tom Rini <trini@xxxxxxxxxxxx>; ron minnich
> > <rminnich@xxxxxxxxx>; Guo, Gua <gua.guo@xxxxxxxxx>; linux-
> > acpi@xxxxxxxxxxxxxxx; U-Boot Mailing List <u-boot@xxxxxxxxxxxxx>
> > Subject: Re: [PATCH v7 2/2] schemas: Add some common reserved-memory
> > usages
> >
> > On Fri, 22 Dec 2023 at 20:52, Chiu, Chasel <chasel.chiu@xxxxxxxxx> wrote:
> > >
> > >
> > > Please see my reply below inline.
> > >
> > > Thanks,
> > > Chasel
> > >
> > ...
> > > > > > The gEfiMemoryTypeInformationGuid HOB typically carries platform
> > > > > > defaults, and the actual memory type information is kept in a
> > > > > > non-volatile EFI variable, which gets updated when the memory
> > > > > > usage changes. Is this different for UefiPayloadPkg?
> > > > > >
> > > > > > (For those among the cc'ees less versed in EFI/EDK2: when you
> > > > > > get the 'config changed -rebooting' message from the boot
> > > > > > firmware, it typically means that this memory type table has
> > > > > > changed, and a reboot is necessary.)
> > > > > >
> > > > > > So the platform init needs to read this variable, or get the
> > > > > > information in a different way. I assume it is the payload, not
> > > > > > the platform init that updates the variable when necessary. This
> > > > > > means the information flows from payload(n) to platform
> > > > > > init(n+1), where n is a monotonic index tracking consecutive boots of the
> > system.
> > > > > >
> > > > > > Can you explain how the DT fits into this? How are the
> > > > > > runtime-code and runtime-data memory reservation nodes under
> > > > > > /reserved-memory used to implement this information exchange
> > > > > > between platform init and payload? And how do the HOB and the EFI
> > variable fit into this picture?
> > > > >
> > > > >
> > > > > 1. With some offline discussion, we would move
> > > > > gEfiMemoryTypeInformationGuid usage to FDT->upl-custom node. This
> > > > > is because it is edk2 implementation choice and non-edk2
> > > > > PlatformInit or Payload may not have such memory optimization
> > > > > implementation. (not a generic usage/requirement for PlatformInit
> > > > > and Payload)
> > > > >
> > > > > The edk2 example flow will be like below:
> > > > >
> > > > > PlatformInit to GetVariable of gEfiMemoryTypeInformationGuid and
> > > > > create Hob-
> > > > >
> > > > >   PlatformInit to initialize FDT->upl-custom node to report
> > > > gEfiMemoryTypeInformationGuid HOB information ->
> > > > >     UefiPayload entry to re-create gEfiMemoryTypeInformationGuid
> > > > > HOB basing
> > > > on FDT input (instead of the default MemoryType inside UefiPayload)
> > > > ->
> > > > >       UefiPayload DxeMain/Gcd will consume
> > > > > gEfiMemoryTypeInformationGuid
> > > > Hob for memory type information ->
> > > > >         UefiPayload to initialize UEFI environment (mainly DXE dispatcher) ->
> > > > >           (additional FV binary appended to common UefiPayload
> > > > > binary)
> > > > PlatformPayload to provide VariableService which is platform
> > > > specific ->
> > > > >             UefiPayload UefiBootManager will SetVariable if memory
> > > > > type change
> > > > needed and request a warm reset ->
> > > > >               Back to PlatformInit ...
> > > > >
> > > >
> > > > OK so the upl-custom node can do whatever it needs to. I imagine
> > > > these will include the memory descriptor attribute field, and other
> > > > parts that may be missing from the /reserved-memory DT node specification?
> > >
> > >
> > > Yes, if needed by edk2 specific implementation, not generic enough, we may
> > consider to use upl-custom node to pass those data.
> > >
> > >
> > > >
> > > > >
> > > > > 2. Now the proposed reserved-memory node usages will be for
> > > > > PlatformInit to
> > > > provide data which may be used by Payload or OS. This is not edk2
> > > > specific and any PlatformInit/Payload could have same support.
> > > > > Note: all of below are optional and PlatformInit may choose to
> > > > > implement some
> > > > of them or not.
> > > > >
> > > > >       - acpi
> > > > > If PlatformInit created some ACPI tables, this will report a
> > > > > memory region which
> > > > contains all the tables to Payload and Payload may base on this to
> > > > add some more tables if required.
> > > > >
> > > > >       - acpi-nvs
> > > > > If PlatformInit has created some ACPI tables which having ACPI NVS
> > > > > memory
> > > > dependency, this will be that nvs region.
> > > > >
> > > >
> > > > These make sense.
> > > >
> > > > >       - boot-code
> > > > > When PlatformInit having some FW boot phase code that could be
> > > > > freed for OS to use when payload transferring control to UEFI OS
> > > > >
> > > > >       - boot-data
> > > > > When PlatformInit having some FW boot phase data that could be
> > > > > freed for OS
> > > > to use when payload transferring control to UEFI OS.
> > > > >
> > > > >       - runtime-code
> > > > > PlatformInit may provide some services code that can be used for
> > > > > Payload to
> > > > initialize UEFI Runtime Services for supporting UEFI OS.
> > > > >
> > > > >       - runtime-data
> > > > > PlatformInit may provide some services data that can be used for
> > > > > Payload to
> > > > Initialize UEFI Runtime Services for supporting UEFI OS.
> > > > >
> > > >
> > > > A UEFI OS must consume this information from the UEFI memory map,
> > > > not from the /reserved-memory nodes. So these nodes must either not
> > > > be visible to the OS at all, or carry an annotation that the OS must ignore
> > them.
> > > >
> > > > Would it be possible to include a restriction in the DT schema that
> > > > these are only valid in the firmware boot phase?
> > >
> > >
> > > https://uefi.org/specs/UEFI/2.10/07_Services_Boot_Services.html#efi-bo
> > > ot-services-exitbootservices Per UEFI specification, UEFI OS will
> > > always call UEFI GetMemoryMap function to retrieve memory map, so FDT
> > node present or not does not matter to UEFI OS. We probably could have
> > annotation in UPL specification to emphasize this.
> > > I'm not familiar with Linux FDT boot, but if non-UEFI OS does not call UEFI
> > GetMemoryMap() and does not know what is runtime-code/data, boot-
> > code/data, it might just treat such reserved-memory nodes as 'regular' reserved
> > memory nodes, and that's still ok because non-UEFI OS will not call to any
> > runtime service or re-purpose boot-code/data memory regions.
> > >
> >
> > You are saying the same thing but in a different way. A UEFI OS must only rely on
> > GetMemoryMap(), and not on the /reserved-memory node to obtain this
> > information. But this requirement needs to be stated
> > somewhere: the UEFI spec does not reason about other sources of EFI memory
> > information at all, and this DT schema does not mention any of this either.
> >
> > > Would you provide a real OS case which will be impacted by this reserved-
> > memory schema so we can discuss basing on real case?
> > >
> >
> > Funny, that is what I have been trying to get from you :-)
> >
> > The problem I am anticipating here is that the information in /reserved-memory
> > may be out of sync with the EFI memory map. It needs to be made clear that the
> > EFI memory map is the only source of truth when the OS is involved, and this
> > /reserved-memory mechanism should only be used by other firmware stages. But
> > the schema does not mention this at all. The schema also does not mention that
> > the information in /reserved-memory is not actually sufficient to reconstruct the
> > EFI memory map that the firmware payload expects (which is why the upl-
> > custom-node exists too)
>
>
>
> Does below solve your concerns if we mention those in schema description? (please feel free to add more if you have)
> . boot-code/boot-data and runtime-code/runtime-data usages are following UEFI specification
>   . before ExitBootServices: https://uefi.org/specs/UEFI/2.10/07_Services_Boot_Services.html#memory-type-usage-before-exitbootservices
>   . after ExitBootServices: https://uefi.org/specs/UEFI/2.10/07_Services_Boot_Services.html#memory-type-usage-after-exitbootservices
> . These usages do not intend to construct full UEFI memory map, it is only for PlatformInit to pass pre-installed tables or services to Payload for supporting UEFI OS boot.
> . These usages are optional
> . Typically UEFI OS boot will always call GetMemoryMap() to retrieve memory map following UEFI spec, no matter DT nodes present or not (https://uefi.org/specs/UEFI/2.10/07_Services_Boot_Services.html#efi-boot-services-exitbootservices)
> . Typically Non-UEFI OS boot will treat those  boot* or runtime* reserved-memory as 'regular' reserved memory if present.
>

This already helps quite a lot, thanks.

But why should a non-UEFI OS be required to keep boot* or runtime*
regions reserved? The firmware stage that boots the OS knows whether
it is performing an UEFI boot or a non-UEFI boot, and it should only
present the information that goes along with that. The OS should never
have to worry about reconciling two sources of truth.

And to Rob's point about boot / runtime being ill-defined: I would
argue that 'runtime' quite clearly implies 'under the OS', and so UEFI
runtime* reservations are assumed to always be relevant to UEFI OSes.

I think there is a fundamental difference of opinion here, where the
position of the firmware developers is that the DT should be the same
across all boot stages, while my position reasoning from the OS side
is that the OS should be able to observe only the abstractions that
are part of the contract between firmware and OS.