On Thu, Sep 7, 2023 at 8:56 AM Simon Glass <sjg@xxxxxxxxxxxx> wrote: > > Hi Ard, > > On Thu, 7 Sept 2023 at 07:31, Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > > > On Wed, 6 Sept 2023 at 18:50, Simon Glass <sjg@xxxxxxxxxxxx> wrote: > > > > > > Hi Ard, > > > > > > On Wed, Sep 6, 2023, 10:09 Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > >> > > >> On Wed, 6 Sept 2023 at 16:54, Simon Glass <sjg@xxxxxxxxxxxx> wrote: > > >> > > > >> > Hi Rob, Ard, > > >> > > > >> > On Wed, 6 Sept 2023 at 08:34, Rob Herring <robh@xxxxxxxxxx> wrote: > > >> > > > > >> > > On Tue, Sep 5, 2023 at 4:44 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > >> > > > > > >> > > > On Thu, 31 Aug 2023 at 01:18, Simon Glass <sjg@xxxxxxxxxxxx> wrote: > > >> > > > > > > >> > > > > The Devicetree specification skips over handling of a logical view of > > >> > > > > the memory map, pointing users to the UEFI specification. > > >> > > > > > > >> > > > > It is common to split firmware into 'Platform Init', which does the > > >> > > > > initial hardware setup and a "Payload" which selects the OS to be booted. > > >> > > > > Thus an handover interface is required between these two pieces. > > >> > > > > > > >> > > > > Where UEFI boot-time services are not available, but UEFI firmware is > > >> > > > > present on either side of this interface, information about memory usage > > >> > > > > and attributes must be presented to the "Payload" in some form. > > >> > > > > > > >> > > > > > >> > > > I don't think the UEFI references are needed or helpful here. > > >> > > > > > >> > > > > This aims to provide an small schema addition for this mapping. > > >> > > > > > > >> > > > > For now, no attempt is made to create an exhaustive binding, so there are > > >> > > > > some example types listed. More can be added later. > > >> > > > > > > >> > > > > The compatible string is not included, since the node name is enough to > > >> > > > > indicate the purpose of a node, as per the existing reserved-memory > > >> > > > > schema. > > >> > > > > >> > > Node names reflect the 'class', but not what's specifically in the > > >> > > node. So really, all reserved-memory nodes should have the same name, > > >> > > but that ship already sailed for existing users. 'compatible' is the > > >> > > right thing here. As to what the node name should be, well, we haven't > > >> > > defined that. I think we just used 'memory' on some platforms. > > >> > > > >> > OK > > >> > > > >> > > > > >> > > > > This binding does not include a binding for the memory 'attribute' > > >> > > > > property, defined by EFI_BOOT_SERVICES.GetMemoryMap(). It may be useful > > >> > > > > to have that as well, but perhaps not as a bit mask. > > >> > > > > > > >> > > > > Signed-off-by: Simon Glass <sjg@xxxxxxxxxxxx> > > >> > > > > --- > > >> > > > > > > >> > > > > Changes in v5: > > >> > > > > - Drop the memory-map node (should have done that in v4) > > >> > > > > - Tidy up schema a bit > > >> > > > > > > >> > > > > Changes in v4: > > >> > > > > - Make use of the reserved-memory node instead of creating a new one > > >> > > > > > > >> > > > > Changes in v3: > > >> > > > > - Reword commit message again > > >> > > > > - cc a lot more people, from the FFI patch > > >> > > > > - Split out the attributes into the /memory nodes > > >> > > > > > > >> > > > > Changes in v2: > > >> > > > > - Reword commit message > > >> > > > > > > >> > > > > .../reserved-memory/common-reserved.yaml | 53 +++++++++++++++++++ > > >> > > > > 1 file changed, 53 insertions(+) > > >> > > > > create mode 100644 dtschema/schemas/reserved-memory/common-reserved.yaml > > >> > > > > > > >> > > > > diff --git a/dtschema/schemas/reserved-memory/common-reserved.yaml b/dtschema/schemas/reserved-memory/common-reserved.yaml > > >> > > > > new file mode 100644 > > >> > > > > index 0000000..d1b466b > > >> > > > > --- /dev/null > > >> > > > > +++ b/dtschema/schemas/reserved-memory/common-reserved.yaml > > >> > > > > @@ -0,0 +1,53 @@ > > >> > > > > +# SPDX-License-Identifier: GPL-2.0-only OR BSD-2-Clause > > >> > > > > +%YAML 1.2 > > >> > > > > +--- > > >> > > > > +$id: http://devicetree.org/schemas/reserved-memory/common-reserved.yaml# > > >> > > > > +$schema: http://devicetree.org/meta-schemas/core.yaml# > > >> > > > > + > > >> > > > > +title: Common memory reservations > > >> > > > > + > > >> > > > > +description: | > > >> > > > > + Specifies that the reserved memory region can be used for the purpose > > >> > > > > + indicated by its node name. > > >> > > > > + > > >> > > > > + Clients may reuse this reserved memory if they understand what it is for. > > >> > > > > + > > >> > > > > +maintainers: > > >> > > > > + - Simon Glass <sjg@xxxxxxxxxxxx> > > >> > > > > + > > >> > > > > +allOf: > > >> > > > > + - $ref: reserved-memory.yaml > > >> > > > > + > > >> > > > > +properties: > > >> > > > > + $nodename: > > >> > > > > + enum: > > >> > > > > + - acpi-reclaim > > >> > > > > + - acpi-nvs > > >> > > > > + - boot-code > > >> > > > > + - boot-data > > >> > > > > + - runtime-code > > >> > > > > + - runtime-data > > >> > > > > + > > >> > > > > > >> > > > These types are used by firmware to describe the nature of certain > > >> > > > memory regions to the OS. Boot code and data can be discarded, as well > > >> > > > as ACPI reclaim after its contents have been consumed. Runtime code > > >> > > > and data need to be mapped for runtime features to work. > > >> > > > > > >> > > > When one firmware phase communicates the purpose of a certain memory > > >> > > > reservation to another, it is typically not limited to whether its > > >> > > > needs to be preserved and when it needs to be mapped (and with which > > >> > > > attributes). I'd expect a memory reservation appearing under this node > > >> > > > to have a clearly defined purpose, and the subsequent phases need to > > >> > > > be able to discover this information. > > >> > > > > > >> > > > For example, a communication buffer for secure<->non-secure > > >> > > > communication or a page with spin tables used by PSCI. None of the > > >> > > > proposed labels are appropriate for this, and I'd much rather have a > > >> > > > compatible string or some other property that clarifies the nature in > > >> > > > a more suitable way. Note that 'no-map' already exists to indicate > > >> > > > that the CPU should not map this memory unless it does so for the > > >> > > > specific purpose that the reservation was made for. > > >> > > > > >> > > I agree. I think compatible is the better approach. Some property like > > >> > > 'discard' may not be sufficient information if the OS needs to consume > > >> > > the region first and then discard it. Better to state exactly what's > > >> > > there and then the OS can imply the rest. > > >> > > > >> > OK, so what sort of compatible strings? > > >> > > > >> > How about: > > >> > "acpi-reclaim" - holds ACPI tables; memory can be reclaimed once the > > >> > tables are read and no-longer needed > > >> > > >> ACPI reclaim is a policy, not a purpose. This memory could contain > > >> many different things. > > >> > > >> > "boot-code" - holds boot code; memory can be reclaimed once the boot > > >> > phase is complete > > >> > "runtime-code" - holds runtime code; memory can be reclaimed only if > > >> > this code will not be used from that point > > >> > > > >> > > >> These are also policies. They can be inferred from the purpose. > > >> > > >> > etc. We can then have more specific compatibles, like: > > >> > > > >> > "psci-spin-table" - holds PSCI spin tables > > >> > > > >> > so you could do: > > >> > > > >> > compatible = "runtime-code", "psci-spin-table"; > > >> > > > >> > > >> I understand that this binding targets firmware<->firmware rather than > > >> firmware<->OS, which makes it much more difficult to keep it both > > >> generic and sufficiently descriptive. > > >> > > >> However, I still feel that all the overlap with UEFI memory types is > > >> not what we want here. UEFI knows how to manage its own memory map, > > >> what it needs to know is what memory is already in use and for which > > >> exact purpose. Whether or not that implies that the memory can be > > >> freed at some point or can be mapped or not should follow from that. > > > > > > > > > Can you please make a suggestion? I am unsure what you are looking for. > > > > > > > I'm happy to help flesh this out, but you still have not provided us > > with an actual use case, so I can only draw from my own experience > > putting together firmware for virtual and physical ARM machines. > > I did explain that this is needed when Tianocore is on both sides of > the interface, since Platform Init places some things in memory and > the Payload needs to preserve them there, and/or know where they are. > > I think the problem might be that you don't agree with that, but it > seems to be a fact, so I am not sure how I can alter it. > > Please can you clearly explain which part of the use case you are missing. > > > > > All the ACPI and UEFI lingo needs to be dropped. This is relevant only > > to the OS, and only if the prior stage exposes UEFI interfaces, in > > which case it does not need this binding. > > OK I can drop those from the commit message. > > > > > So on one side, there is the requirement for each memory reservation > > to be described with sufficient detail so that a subsequent boot stage > > (firmware or OS) can use it for its intended purpose, provided that > > this boot stage is aware of its purpose (i.e., it has a driver that > > matches on the compatible string in question, and actually maps/uses > > the memory) > > > > On the other side, we need to describe how a memory reservation should > > be treated if the boot stage doesn't know its purpose, has no interest > > in using it or has consumed the contents and has no longer a need for > > the region. We already have no-map to describe that the memory should > > never be mapped (and this may be disregarded by an actual driver for > > the region). I imagine we might add 'discardable' as a boolean DT > > property, meaning that stage N may use the memory whichever way it > > wants if it is not going to use it for its intended purpose, provided > > that it deletes the node from the DT before passing it on to stage > > N+1. > > OK. For now I think that everything is discardable, so long as the > Payload knows the purpose and that it not needed. That is what Rob > seemed to be saying. If we add 'discardable', does that mean that > things default to non-discardable? Would that not be a change of > behaviour for existing users? I believe I said "discardable" should be implied from the compatible, not be a discreet property. It is possible that some region is passed thru to multiple stages and is only discardable by the last stage (or next to last or...). So discardable depends on the consumer. There has to be some logic of "I know what this foo-bar region is for and know it isn't needed". If the consumer doesn't know what the region is, then it must default to leaving it reserved. > > One thing that needs to be clarified is how this binding interacts > > with /memory nodes. I assume that currently, /reserved-memory is > > independent, i.e., it could describe mappable memory that is not > > covered by /memory at all. If this is the case, we have to decide > > whether or not discardable regions can be treated in the same way, or > > whether we should require that discardable regions are covered by > > /memory. > > I would expect all memory to be described in /memory nodes. What is > the use case for omitting it? That is the expectation though nothing prevents a platform from omitting areas. Most likely only the beginning or end of memory. System partitioning usecases likely just describe each partition's portion of memory. I don't recall ever seeing a bunch of memory regions rather than 1-2 with reserved-memory nodes. Also, I don't think we prevent a reserved region outside the memory region(s) though I imagine (at least Linux) just ignores them. > Are you thinking of SRAM, etc? We have mmio-sram binding for that which has its own way to carve up the regions. Rob
