mbox series

[v7,0/6] arm64: Permit EFI boot with MMU and caches on

Message ID 20230111102236.1430401-1-ardb@kernel.org (mailing list archive)
Headers show
Series arm64: Permit EFI boot with MMU and caches on | expand

Message

Ard Biesheuvel Jan. 11, 2023, 10:22 a.m. UTC
The purpose of this series is to remove any explicit cache maintenance
for coherency during early boot. Software managed coherency is error
prone and tedious, and running with the MMU off is generally bad for
performance, and it becomes unnecessary if we simply retain the
cacheable 1:1 mapping of all of system RAM provided by EFI, and use it
to populate the initial ID map page tables. After setting up this
preliminary ID map, we disable the MMU, drop to EL1, reprogram the MAIR,
TCR and SCTLR registers as before, and proceed as usual, avoiding the
need for any manipulations of memory while the MMU and caches are off.

The only properties of the firmware provided 1:1 map we rely on is that
it does not require any explicit cache maintenance for coherency, and
that it covers the entire memory footprint of the image, including the
BSS and padding at the end - all else is under control of the kernel
itself, as before.

The final patch updates the EFI stub code so that it no longer disables
the MMU and caches or cleans the entire image to the PoC. Note that
some cache maintenace for I/D coherence may still be needed, in the
zboot case (which decompresses and boots a compressed kernel image) or
in cases where the image is moved in memory.

Changes since v6:
- drop the 64k alignment patch, which is not strictly a prerequisite,
  and will be revisited later if needed
- add back EFI stub changes now that all dependencies are in mainline
- panic() the kernel later in the boot if we detected a non-EFI boot
  occurring with the MMU and caches enabled

Changes since v5:
- add a special entry point into the boot sequence that is to be used by
  EFI only, and only permit booting with the MMU enabled when using that
  boot path;
- omit the final patch that would need to go via the EFI tree in any
  case - adding the new entrypoint specific for EFI makes it conflict
  even more badly, and I'll try to revisit this during the merge window
  or simply defer the final piece for the next release;

Changes since v4:
- add patch to align the callers of finalise_el2()
- also clean HYP text to the PoC when booting at EL2 with the MMU on
- add a warning and a taint when doing non-EFI boot with the MMU and
  caches enabled
- rebase onto zboot changes in efi/next - this means that patches #6 and
  #7 will not apply onto arm64/for-next so a shared stable branch will
  be needed if we want to queue this up for v6.2

Changes since v3:
- drop EFI_LOADER_CODE memory type patch that has been queued in the
  mean time
- rebased onto [partial] series that moves efi-entry.S into the libstub/
  source directory
- fixed a correctness issue in patch #2

Cc: Will Deacon <will@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>

Ard Biesheuvel (6):
  arm64: head: Move all finalise_el2 calls to after __enable_mmu
  arm64: kernel: move identity map out of .text mapping
  arm64: head: record the MMU state at primary entry
  arm64: head: avoid cache invalidation when entering with the MMU on
  arm64: head: Clean the ID map and the HYP text to the PoC if needed
  efi: arm64: enter with MMU and caches enabled

 arch/arm64/include/asm/efi.h               |  2 +
 arch/arm64/kernel/head.S                   | 89 +++++++++++++++-----
 arch/arm64/kernel/image-vars.h             |  5 +-
 arch/arm64/kernel/setup.c                  | 17 +++-
 arch/arm64/kernel/sleep.S                  |  6 +-
 arch/arm64/kernel/vmlinux.lds.S            |  2 +-
 arch/arm64/mm/cache.S                      |  1 +
 arch/arm64/mm/proc.S                       |  2 -
 drivers/firmware/efi/libstub/Makefile      |  4 +-
 drivers/firmware/efi/libstub/arm64-entry.S | 67 ---------------
 drivers/firmware/efi/libstub/arm64-stub.c  | 26 ++++--
 drivers/firmware/efi/libstub/arm64.c       | 41 +++++++--
 12 files changed, 151 insertions(+), 111 deletions(-)
 delete mode 100644 drivers/firmware/efi/libstub/arm64-entry.S

Comments

Ard Biesheuvel Jan. 11, 2023, 10:26 a.m. UTC | #1
On Wed, 11 Jan 2023 at 11:23, Ard Biesheuvel <ardb@kernel.org> wrote:
>
> The purpose of this series is to remove any explicit cache maintenance
> for coherency during early boot. Software managed coherency is error
> prone and tedious, and running with the MMU off is generally bad for
> performance, and it becomes unnecessary if we simply retain the
> cacheable 1:1 mapping of all of system RAM provided by EFI, and use it
> to populate the initial ID map page tables. After setting up this
> preliminary ID map, we disable the MMU, drop to EL1, reprogram the MAIR,
> TCR and SCTLR registers as before, and proceed as usual, avoiding the
> need for any manipulations of memory while the MMU and caches are off.
>

Forgot to mention: as it happens, patches #1 and #2 in this series
also work around a problem that was reported the other day, where the
allyesconfig build fails to link [0]

I bisected this to

commit 5e5ff73c2e5863f93fc5fd78d178cd8f2af12464
Author: Sai Prakash Ranjan <quic_saipraka@quicinc.com>
Date:   Mon Oct 17 20:04:50 2022 +0530

    asm-generic/io: Add _RET_IP_ to MMIO trace for more accurate debug info

which seems entirely unrelated, but looks like it may be causing the
number of direct calls (and therefore the number of trampolines) to
increase, causing the ID map to blow up like it does.

[0] https://lore.kernel.org/all/CAMj1kXGAf7ikEU5jLoik0xrOde0xBg0yJkOo5=PtEtNXoUxMXA@mail.gmail.com/



> The only properties of the firmware provided 1:1 map we rely on is that
> it does not require any explicit cache maintenance for coherency, and
> that it covers the entire memory footprint of the image, including the
> BSS and padding at the end - all else is under control of the kernel
> itself, as before.
>
> The final patch updates the EFI stub code so that it no longer disables
> the MMU and caches or cleans the entire image to the PoC. Note that
> some cache maintenace for I/D coherence may still be needed, in the
> zboot case (which decompresses and boots a compressed kernel image) or
> in cases where the image is moved in memory.
>
> Changes since v6:
> - drop the 64k alignment patch, which is not strictly a prerequisite,
>   and will be revisited later if needed
> - add back EFI stub changes now that all dependencies are in mainline
> - panic() the kernel later in the boot if we detected a non-EFI boot
>   occurring with the MMU and caches enabled
>
> Changes since v5:
> - add a special entry point into the boot sequence that is to be used by
>   EFI only, and only permit booting with the MMU enabled when using that
>   boot path;
> - omit the final patch that would need to go via the EFI tree in any
>   case - adding the new entrypoint specific for EFI makes it conflict
>   even more badly, and I'll try to revisit this during the merge window
>   or simply defer the final piece for the next release;
>
> Changes since v4:
> - add patch to align the callers of finalise_el2()
> - also clean HYP text to the PoC when booting at EL2 with the MMU on
> - add a warning and a taint when doing non-EFI boot with the MMU and
>   caches enabled
> - rebase onto zboot changes in efi/next - this means that patches #6 and
>   #7 will not apply onto arm64/for-next so a shared stable branch will
>   be needed if we want to queue this up for v6.2
>
> Changes since v3:
> - drop EFI_LOADER_CODE memory type patch that has been queued in the
>   mean time
> - rebased onto [partial] series that moves efi-entry.S into the libstub/
>   source directory
> - fixed a correctness issue in patch #2
>
> Cc: Will Deacon <will@kernel.org>
> Cc: Catalin Marinas <catalin.marinas@arm.com>
> Cc: Marc Zyngier <maz@kernel.org>
> Cc: Mark Rutland <mark.rutland@arm.com>
>
> Ard Biesheuvel (6):
>   arm64: head: Move all finalise_el2 calls to after __enable_mmu
>   arm64: kernel: move identity map out of .text mapping
>   arm64: head: record the MMU state at primary entry
>   arm64: head: avoid cache invalidation when entering with the MMU on
>   arm64: head: Clean the ID map and the HYP text to the PoC if needed
>   efi: arm64: enter with MMU and caches enabled
>
>  arch/arm64/include/asm/efi.h               |  2 +
>  arch/arm64/kernel/head.S                   | 89 +++++++++++++++-----
>  arch/arm64/kernel/image-vars.h             |  5 +-
>  arch/arm64/kernel/setup.c                  | 17 +++-
>  arch/arm64/kernel/sleep.S                  |  6 +-
>  arch/arm64/kernel/vmlinux.lds.S            |  2 +-
>  arch/arm64/mm/cache.S                      |  1 +
>  arch/arm64/mm/proc.S                       |  2 -
>  drivers/firmware/efi/libstub/Makefile      |  4 +-
>  drivers/firmware/efi/libstub/arm64-entry.S | 67 ---------------
>  drivers/firmware/efi/libstub/arm64-stub.c  | 26 ++++--
>  drivers/firmware/efi/libstub/arm64.c       | 41 +++++++--
>  12 files changed, 151 insertions(+), 111 deletions(-)
>  delete mode 100644 drivers/firmware/efi/libstub/arm64-entry.S
>
> --
> 2.39.0
>
Catalin Marinas Jan. 24, 2023, 12:10 p.m. UTC | #2
On Wed, 11 Jan 2023 11:22:30 +0100, Ard Biesheuvel wrote:
> The purpose of this series is to remove any explicit cache maintenance
> for coherency during early boot. Software managed coherency is error
> prone and tedious, and running with the MMU off is generally bad for
> performance, and it becomes unnecessary if we simply retain the
> cacheable 1:1 mapping of all of system RAM provided by EFI, and use it
> to populate the initial ID map page tables. After setting up this
> preliminary ID map, we disable the MMU, drop to EL1, reprogram the MAIR,
> TCR and SCTLR registers as before, and proceed as usual, avoiding the
> need for any manipulations of memory while the MMU and caches are off.
> 
> [...]

Applied to arm64 (for-next/efi-boot-mmu-on), thanks!

[1/6] arm64: head: Move all finalise_el2 calls to after __enable_mmu
      https://git.kernel.org/arm64/c/82e4958800c0
[2/6] arm64: kernel: move identity map out of .text mapping
      https://git.kernel.org/arm64/c/af7249b317e4
[3/6] arm64: head: record the MMU state at primary entry
      https://git.kernel.org/arm64/c/9d7c13e5dde3
[4/6] arm64: head: avoid cache invalidation when entering with the MMU on
      https://git.kernel.org/arm64/c/32b135a7fafe
[5/6] arm64: head: Clean the ID map and the HYP text to the PoC if needed
      https://git.kernel.org/arm64/c/3dcf60bbfd28
[6/6] efi: arm64: enter with MMU and caches enabled
      https://git.kernel.org/arm64/c/617861703830