arm64 UEFI early FDT handling

This is a followup to the "arm64: update/clarify/relax Image and FDT placement rules" series I sent a while ago:
(http://article.gmane.org/gmane.linux.ports.arm.kernel/407148)

This has now been split in two series: this first series deals with the early FDT handling, primarily in the context of UEFI, but not exclusively.

A number of minor issues exist in the early UEFI/FDT handling path, such as:
- when booting via UEFI, memreserve entries are removed from the device tree but the /reserved-memory node is not

- memory nodes are removed from the device tree in a way that is not officially supported by the libfdt API (i.e., you cannot delete nodes while traversing the tree)
- removal of memory nodes may discard annotations (such as NUMA topology) that should ideally be retained, or may corrupt the tree by discarding nodes referenced by phandles.

Patches #1 and #2 introduce an arm64 specific version of early_init_dt_add_memory_arch() so that we can modify it later to ignore DT memory nodes if booting via UEFI.

Patch #3 moves some UEFI+FDT init code around before making changes to it.

Patch #4 moves the UEFI initialization to before the early FDT scanning so we know at that point whether we are booting via UEFI or not.

Patch #5 changes the UEFI init code so that memory nodes are simply ignored, so that they don't have to be removed by the stub anymore.

Patch #6 does the same as #5, but for memreserves and the /reserved-memory node.

Changes since v2:
- instead of copying the generic implementation, turn early_init_dt_add_memory_arch() into a weak alias so that it is still accessible to overrides

Changes since v1:
- dropped first two patches, they have been merged into v4.2-rc1
- dropped last patch regarding FDT placement by the stub, this is not entirely relevant to the primary issue targeted by this series
- rebased onto for-next/core (arm64) as of today

Ard Biesheuvel (6):
of/fdt: make generic early_init_dt_add_memory_arch() a weak alias
arm64: override generic version of early_init_dt_add_memory_arch()
efi: move FDT handling to separate object file
arm64/efi: move EFI /chosen node parsing before early FDT processing arm64/efi: ignore DT memory nodes instead of removing them arm64/efi: ignore DT memreserve entries instead of removing them

 arch/arm64/include/asm/efi.h       |  2 +
 arch/arm64/kernel/efi.c            | 26 ++++--
 arch/arm64/kernel/setup.c          |  3 +
 arch/arm64/mm/init.c               | 12 ++-
 drivers/firmware/efi/Makefile      |  1 +
 drivers/firmware/efi/efi-fdt.c     | 73 +++++++++++++++++
 drivers/firmware/efi/efi.c         | 84 --------------------
 drivers/firmware/efi/libstub/fdt.c | 33 +-------
 drivers/of/fdt.c                   |  7 +-
 include/linux/efi.h                |  2 +-
 include/linux/of_fdt.h             |  1 +
 11 files changed, 116 insertions(+), 128 deletions(-)
 create mode 100644 drivers/firmware/efi/efi-fdt.c

--
1.9.1

(+ Grant)

Quoted 10 lines · Ard Biesheuvel

This is a followup to the "arm64: update/clarify/relax Image and FDT placement rules" series I sent a while ago:
(http://article.gmane.org/gmane.linux.ports.arm.kernel/407148)

This has now been split in two series: this first series deals with the early FDT handling, primarily in the context of UEFI, but not exclusively.

A number of minor issues exist in the early UEFI/FDT handling path, such as:
- when booting via UEFI, memreserve entries are removed from the device tree but the /reserved-memory node is not

After reading Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt again, I think simply ignoring the reserved-memory node is not the way to go. The reason is that it may contain dynamic allocations that are referenced by other nodes in the DT, and there is no good technical reason IMO to disallow those. OTOH, static allocations may conflict with the UEFI memory map, so those need to be dropped or at least checked against the memory map. The problem here is that static nodes may also be referenced by phandle, so we need to handle the referring node in some way as well.

So I think we have a number of options:
- ignore /memreserve/s and reject static allocations in /reserved-memory (*) but honor dynamic ones
- ignore /memreserve/s and honor all of /reserved-memory after checking that static allocations don't conflict
- honor all /memreserve/s and /reserved-memory nodes and check all for conflicts
- ...

(*) static allocations for regions that the UEFI memory map does not describe should be OK, though

I personally prefer the first one, since a dynamic allocation implicitly conveys that the region does not contain anything special when coming out of boot, and there is very little we need to do other than perform the actual reservation. Static allocations are ambiguous in the sense that there is no annotation that explains the choice of address.

Thoughts, please?

Hi Ard,

Quoted 42 lines · Ard Biesheuvel

(+ Grant)

Quoted 10 lines · Ard Biesheuvel

This is a followup to the "arm64: update/clarify/relax Image and FDT placement rules" series I sent a while ago:
(http://article.gmane.org/gmane.linux.ports.arm.kernel/407148)

This has now been split in two series: this first series deals with the early FDT handling, primarily in the context of UEFI, but not exclusively.

A number of minor issues exist in the early UEFI/FDT handling path, such as:
- when booting via UEFI, memreserve entries are removed from the device tree but the /reserved-memory node is not

After reading Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt again, I think simply ignoring the reserved-memory node is not the way to go. The reason is that it may contain dynamic allocations that are referenced by other nodes in the DT, and there is no good technical reason IMO to disallow those. OTOH, static allocations may conflict with the UEFI memory map, so those need to be dropped or at least checked against the memory map. The problem here is that static nodes may also be referenced by phandle, so we need to handle the referring node in some way as well.

So I think we have a number of options:
- ignore /memreserve/s and reject static allocations in /reserved-memory (*) but honor dynamic ones
- ignore /memreserve/s and honor all of /reserved-memory after checking that static allocations don't conflict
- honor all /memreserve/s and /reserved-memory nodes and check all for conflicts
- ...

(*) static allocations for regions that the UEFI memory map does not describe should be OK, though

I personally prefer the first one, since a dynamic allocation implicitly conveys that the region does not contain anything special when coming out of boot, and there is very little we need to do other than perform the actual reservation. Static allocations are ambiguous in the sense that there is no annotation that explains the choice of address.

Quoted 1 line

Thoughts, please?

What's the status of this series? It was on my "list of patches to watch" that I'm just refreshing for 4.5, but I can't see any comments on-list about it.

Quoted 40 lines · Ard Biesheuvel

Quoted 10 lines · Ard Biesheuvel

This is a followup to the "arm64: update/clarify/relax Image and FDT placement rules" series I sent a while ago:
(http://article.gmane.org/gmane.linux.ports.arm.kernel/407148)

This has now been split in two series: this first series deals with the early FDT handling, primarily in the context of UEFI, but not exclusively.

A number of minor issues exist in the early UEFI/FDT handling path, such as:
- when booting via UEFI, memreserve entries are removed from the device tree but the /reserved-memory node is not

After reading Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt again, I think simply ignoring the reserved-memory node is not the way to go. The reason is that it may contain dynamic allocations that are referenced by other nodes in the DT, and there is no good technical reason IMO to disallow those. OTOH, static allocations may conflict with the UEFI memory map, so those need to be dropped or at least checked against the memory map. The problem here is that static nodes may also be referenced by phandle, so we need to handle the referring node in some way as well.

So I think we have a number of options:
- ignore /memreserve/s and reject static allocations in /reserved-memory (*) but honor dynamic ones
- ignore /memreserve/s and honor all of /reserved-memory after checking that static allocations don't conflict
- honor all /memreserve/s and /reserved-memory nodes and check all for conflicts
- ...

(*) static allocations for regions that the UEFI memory map does not describe should be OK, though

I personally prefer the first one, since a dynamic allocation implicitly conveys that the region does not contain anything special when coming out of boot, and there is very little we need to do other than perform the actual reservation. Static allocations are ambiguous in the sense that there is no annotation that explains the choice of address.

Quoted 1 line

Thoughts, please?

What's the status of this series? It was on my "list of patches to watch" that I'm just refreshing for 4.5, but I can't see any comments on-list about it.

Quoted 43 lines · Will Deacon

Quoted 41 lines · Ard Biesheuvel

Quoted 10 lines · Ard Biesheuvel

This is a followup to the "arm64: update/clarify/relax Image and FDT placement rules" series I sent a while ago:
(http://article.gmane.org/gmane.linux.ports.arm.kernel/407148)

This has now been split in two series: this first series deals with the early FDT handling, primarily in the context of UEFI, but not exclusively.

A number of minor issues exist in the early UEFI/FDT handling path, such as:
- when booting via UEFI, memreserve entries are removed from the device tree but the /reserved-memory node is not

After reading Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt again, I think simply ignoring the reserved-memory node is not the way to go. The reason is that it may contain dynamic allocations that are referenced by other nodes in the DT, and there is no good technical reason IMO to disallow those. OTOH, static allocations may conflict with the UEFI memory map, so those need to be dropped or at least checked against the memory map. The problem here is that static nodes may also be referenced by phandle, so we need to handle the referring node in some way as well.

So I think we have a number of options:
- ignore /memreserve/s and reject static allocations in /reserved-memory (*) but honor dynamic ones
- ignore /memreserve/s and honor all of /reserved-memory after checking that static allocations don't conflict
- honor all /memreserve/s and /reserved-memory nodes and check all for conflicts
- ...

(*) static allocations for regions that the UEFI memory map does not describe should be OK, though

I personally prefer the first one, since a dynamic allocation implicitly conveys that the region does not contain anything special when coming out of boot, and there is very little we need to do other than perform the actual reservation. Static allocations are ambiguous in the sense that there is no annotation that explains the choice of address.

Thoughts, please?

Quoted 3 lines

What's the status of this series? It was on my "list of patches to watch" that I'm just refreshing for 4.5, but I can't see any comments on-list about it.

I thought it's being taken over by this series:

http://lkml.kernel.org/r/1446126059-25336-1-git-send-email-ard.biesheuvel at linaro.org