linux.git
12 months agoperf: starfive: fix 64-bit only COMPILE_TEST condition
Conor Dooley [Mon, 18 Mar 2024 15:35:04 +0000 (15:35 +0000)]
perf: starfive: fix 64-bit only COMPILE_TEST condition

ARCH_STARFIVE is not restricted to 64-bit platforms, so while Will's
addition of a 64-bit only condition satisfied the build robots doing
COMPILE_TEST builds, Palmer ran into the same problems with writeq()
being undefined during regular rv32 builds.

Promote the dependency on 64-bit to its own `depends on` so that the
driver can never be included in 32-bit builds.

Reported-by: Palmer Dabbelt <palmer@rivosinc.com>
Fixes: c2b24812f7bc ("perf: starfive: Add StarLink PMU support")
Fixes: f0dbc6d0de38 ("perf: starfive: Only allow COMPILE_TEST for 64-bit architectures")
Signed-off-by: Conor Dooley <conor.dooley@microchip.com>
Acked-by: Will Deacon <will@kernel.org>
Reviewed-by: Palmer Dabbelt <palmer@rivosinc.com>
Acked-by: Palmer Dabbelt <palmer@rivosinc.com>
Acked-by: Ji Sheng Teoh <jisheng.teoh@starfivetech.com>
Acked-by: Emil Renner Berthing <emil.renner.berthing@canonical.com>
Link: https://lore.kernel.org/r/20240318-emphatic-rally-f177a4fe1bdc@spud
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agoARM64: Dynamically allocate cpumasks and increase supported CPUs to 512
Christoph Lameter (Ampere) [Thu, 7 Mar 2024 01:45:04 +0000 (17:45 -0800)]
ARM64: Dynamically allocate cpumasks and increase supported CPUs to 512

  [ a.k.a. Revert "Revert "ARM64: Dynamically allocate cpumasks and
    increase supported CPUs to 512""; originally reverted because of a
    bug in the cpufreq-dt code not using zalloc_cpumask_var() ]

Currently defconfig selects NR_CPUS=256, but some vendors (e.g. Ampere
Computing) are planning to ship systems with 512 CPUs. So that all CPUs on
these systems can be used with defconfig, we'd like to bump NR_CPUS to 512.
Therefore this patch increases the default NR_CPUS from 256 to 512.

As increasing NR_CPUS will increase the size of cpumasks, there's a fear that
this might have a significant impact on stack usage due to code which places
cpumasks on the stack. To mitigate that concern, we can select
CPUMASK_OFFSTACK. As that doesn't seem to be a problem today with
NR_CPUS=256, we only select this when NR_CPUS > 256.

CPUMASK_OFFSTACK configures the cpumasks in the kernel to be
dynamically allocated. This was used in the X86 architecture in the
past to enable support for larger CPU configurations up to 8k cpus.

With that is becomes possible to dynamically size the allocation of
the cpu bitmaps depending on the quantity of processors detected on
bootup. Memory used for cpumasks will increase if the kernel is
run on a machine with more cores.

Further increases may be needed if ARM processor vendors start
supporting more processors. Given the current inflationary trends
in core counts from multiple processor manufacturers this may occur.

There are minor regressions for hackbench. The kernel data size
for 512 cpus is smaller with offstack than with onstack.

Benchmark results using hackbench average over 10 runs of

  hackbench -s 512 -l 2000 -g 15 -f 25 -P

on Altra 80 Core

Support for 256 CPUs on stack. Baseline

  7.8564 sec

Support for 512 CUs on stack.

  7.8713 sec + 0.18%

512 CPUS offstack

  7.8916 sec + 0.44%

Kernel size comparison:

    text    data     filename Difference to onstack256 baseline
25755648 9589248     vmlinuz-6.8.0-rc4-onstack256
25755648 9607680     vmlinuz-6.8.0-rc4-onstack512 +0.19%
25755648 9603584     vmlinuz-6.8.0-rc4-offstack512 +0.14%

Tested-by: Eric Mackay <eric.mackay@oracle.com>
Reviewed-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: Christoph Lameter (Ampere) <cl@linux.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/37099a57-b655-3b3a-56d0-5f7fbd49d7db@gentwo.org
Link: https://lore.kernel.org/r/20240314125457.186678-1-m.szyprowski@samsung.com
[catalin.marinas@arm.com: use 'select' instead of duplicating 'config CPUMASK_OFFSTACK']
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agoRevert "mm: add arch hook to validate mmap() prot flags"
Catalin Marinas [Tue, 12 Mar 2024 20:00:38 +0000 (20:00 +0000)]
Revert "mm: add arch hook to validate mmap() prot flags"

This reverts commit cb1a393c40eee2f1692c995ea0cc6e45bfccde4d.

Since the arm64 WXN patch has been reverted, remove this hook as it
would not have any users.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Link: https://lore.kernel.org/r/ZfGESD3a91lxH367@arm.com
12 months agoRevert "arm64: mm: add support for WXN memory translation attribute"
Catalin Marinas [Tue, 12 Mar 2024 20:00:20 +0000 (20:00 +0000)]
Revert "arm64: mm: add support for WXN memory translation attribute"

This reverts commit 50e3ed0f93f4f62ed2aa83de5db6cb84ecdd5707.

The SCTLR_EL1.WXN control forces execute-never when a page has write
permissions. While the idea of hardening such write/exec combinations is
good, with permissions indirection enabled (FEAT_PIE) this control
becomes RES0. FEAT_PIE introduces a slightly different form of WXN which
only has an effect when the base permission is RWX and the write is
toggled by the permission overlay (FEAT_POE, not yet supported by the
arm64 kernel). Revert the patch for now.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Link: https://lore.kernel.org/r/ZfGESD3a91lxH367@arm.com
12 months agoRevert "ARM64: Dynamically allocate cpumasks and increase supported CPUs to 512"
Catalin Marinas [Mon, 11 Mar 2024 18:40:49 +0000 (18:40 +0000)]
Revert "ARM64: Dynamically allocate cpumasks and increase supported CPUs to 512"

This reverts commit 0499a78369adacec1af29340b71ff8dd375b4697.

Enabling CPUMASK_OFFSTACK on arm64 triggers a warning in the
dev_pm_opp_set_config() function followed by a failure to set the
regulators and cpufreq-dt probing error. There is no apparent reason why
this happens, so revert this commit until further investigation.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Reported-by: Marek Szyprowski <m.szyprowski@samsung.com>
Link: https://lore.kernel.org/r/c1f2902d-cefc-4122-9b86-d1d32911f590@samsung.com
12 months agoMerge branch 'for-next/stage1-lpa2' into for-next/core
Catalin Marinas [Thu, 7 Mar 2024 19:05:29 +0000 (19:05 +0000)]
Merge branch 'for-next/stage1-lpa2' into for-next/core

* for-next/stage1-lpa2: (48 commits)
  : Add support for LPA2 and WXN and stage 1
  arm64/mm: Avoid ID mapping of kpti flag if it is no longer needed
  arm64/mm: Use generic __pud_free() helper in pud_free() implementation
  arm64: gitignore: ignore relacheck
  arm64: Use Signed/Unsigned enums for TGRAN{4,16,64} and VARange
  arm64: mm: Make PUD folding check in set_pud() a runtime check
  arm64: mm: add support for WXN memory translation attribute
  mm: add arch hook to validate mmap() prot flags
  arm64: defconfig: Enable LPA2 support
  arm64: Enable 52-bit virtual addressing for 4k and 16k granule configs
  arm64: kvm: avoid CONFIG_PGTABLE_LEVELS for runtime levels
  arm64: ptdump: Deal with translation levels folded at runtime
  arm64: ptdump: Disregard unaddressable VA space
  arm64: mm: Add support for folding PUDs at runtime
  arm64: kasan: Reduce minimum shadow alignment and enable 5 level paging
  arm64: mm: Add 5 level paging support to fixmap and swapper handling
  arm64: Enable LPA2 at boot if supported by the system
  arm64: mm: add LPA2 and 5 level paging support to G-to-nG conversion
  arm64: mm: Add definitions to support 5 levels of paging
  arm64: mm: Add LPA2 support to phys<->pte conversion routines
  arm64: mm: Wire up TCR.DS bit to PTE shareability fields
  ...

12 months agoMerge branches 'for-next/reorg-va-space', 'for-next/rust-for-arm64', 'for-next/misc...
Catalin Marinas [Thu, 7 Mar 2024 19:04:55 +0000 (19:04 +0000)]
Merge branches 'for-next/reorg-va-space', 'for-next/rust-for-arm64', 'for-next/misc', 'for-next/daif-cleanup', 'for-next/kselftest', 'for-next/documentation', 'for-next/sysreg' and 'for-next/dpisa', remote-tracking branch 'arm64/for-next/perf' into for-next/core

* arm64/for-next/perf: (39 commits)
  docs: perf: Fix build warning of hisi-pcie-pmu.rst
  perf: starfive: Only allow COMPILE_TEST for 64-bit architectures
  MAINTAINERS: Add entry for StarFive StarLink PMU
  docs: perf: Add description for StarFive's StarLink PMU
  dt-bindings: perf: starfive: Add JH8100 StarLink PMU
  perf: starfive: Add StarLink PMU support
  docs: perf: Update usage for target filter of hisi-pcie-pmu
  drivers/perf: hisi_pcie: Merge find_related_event() and get_event_idx()
  drivers/perf: hisi_pcie: Relax the check on related events
  drivers/perf: hisi_pcie: Check the target filter properly
  drivers/perf: hisi_pcie: Add more events for counting TLP bandwidth
  drivers/perf: hisi_pcie: Fix incorrect counting under metric mode
  drivers/perf: hisi_pcie: Introduce hisi_pcie_pmu_get_event_ctrl_val()
  drivers/perf: hisi_pcie: Rename hisi_pcie_pmu_{config,clear}_filter()
  drivers/perf: hisi: Enable HiSilicon Erratum 162700402 quirk for HIP09
  perf/arm_cspmu: Add devicetree support
  dt-bindings/perf: Add Arm CoreSight PMU
  perf/arm_cspmu: Simplify counter reset
  perf/arm_cspmu: Simplify attribute groups
  perf/arm_cspmu: Simplify initialisation
  ...

* for-next/reorg-va-space:
  : Reorganise the arm64 kernel VA space in preparation for LPA2 support
  : (52-bit VA/PA).
  arm64: kaslr: Adjust randomization range dynamically
  arm64: mm: Reclaim unused vmemmap region for vmalloc use
  arm64: vmemmap: Avoid base2 order of struct page size to dimension region
  arm64: ptdump: Discover start of vmemmap region at runtime
  arm64: ptdump: Allow all region boundaries to be defined at boot time
  arm64: mm: Move fixmap region above vmemmap region
  arm64: mm: Move PCI I/O emulation region above the vmemmap region

* for-next/rust-for-arm64:
  : Enable Rust support for arm64
  arm64: rust: Enable Rust support for AArch64
  rust: Refactor the build target to allow the use of builtin targets

* for-next/misc:
  : Miscellaneous arm64 patches
  ARM64: Dynamically allocate cpumasks and increase supported CPUs to 512
  arm64: Remove enable_daif macro
  arm64/hw_breakpoint: Directly use ESR_ELx_WNR for an watchpoint exception
  arm64: cpufeatures: Clean up temporary variable to simplify code
  arm64: Update setup_arch() comment on interrupt masking
  arm64: remove unnecessary ifdefs around is_compat_task()
  arm64: ftrace: Don't forbid CALL_OPS+CC_OPTIMIZE_FOR_SIZE with Clang
  arm64/sme: Ensure that all fields in SMCR_EL1 are set to known values
  arm64/sve: Ensure that all fields in ZCR_EL1 are set to known values
  arm64/sve: Document that __SVE_VQ_MAX is much larger than needed
  arm64: make member of struct pt_regs and it's offset macro in the same order
  arm64: remove unneeded BUILD_BUG_ON assertion
  arm64: kretprobes: acquire the regs via a BRK exception
  arm64: io: permit offset addressing
  arm64: errata: Don't enable workarounds for "rare" errata by default

* for-next/daif-cleanup:
  : Clean up DAIF handling for EL0 returns
  arm64: Unmask Debug + SError in do_notify_resume()
  arm64: Move do_notify_resume() to entry-common.c
  arm64: Simplify do_notify_resume() DAIF masking

* for-next/kselftest:
  : Miscellaneous arm64 kselftest patches
  kselftest/arm64: Test that ptrace takes effect in the target process

* for-next/documentation:
  : arm64 documentation patches
  arm64/sme: Remove spurious 'is' in SME documentation
  arm64/fp: Clarify effect of setting an unsupported system VL
  arm64/sme: Fix cut'n'paste in ABI document
  arm64/sve: Remove bitrotted comment about syscall behaviour

* for-next/sysreg:
  : sysreg updates
  arm64/sysreg: Update ID_AA64DFR0_EL1 register
  arm64/sysreg: Update ID_DFR0_EL1 register fields
  arm64/sysreg: Add register fields for ID_AA64DFR1_EL1

* for-next/dpisa:
  : Support for 2023 dpISA extensions
  kselftest/arm64: Add 2023 DPISA hwcap test coverage
  kselftest/arm64: Add basic FPMR test
  kselftest/arm64: Handle FPMR context in generic signal frame parser
  arm64/hwcap: Define hwcaps for 2023 DPISA features
  arm64/ptrace: Expose FPMR via ptrace
  arm64/signal: Add FPMR signal handling
  arm64/fpsimd: Support FEAT_FPMR
  arm64/fpsimd: Enable host kernel access to FPMR
  arm64/cpufeature: Hook new identification registers up to cpufeature

12 months agoARM64: Dynamically allocate cpumasks and increase supported CPUs to 512
Christoph Lameter (Ampere) [Thu, 7 Mar 2024 01:45:04 +0000 (17:45 -0800)]
ARM64: Dynamically allocate cpumasks and increase supported CPUs to 512

Currently defconfig selects NR_CPUS=256, but some vendors (e.g. Ampere
Computing) are planning to ship systems with 512 CPUs. So that all CPUs on
these systems can be used with defconfig, we'd like to bump NR_CPUS to 512.
Therefore this patch increases the default NR_CPUS from 256 to 512.

As increasing NR_CPUS will increase the size of cpumasks, there's a fear that
this might have a significant impact on stack usage due to code which places
cpumasks on the stack. To mitigate that concern, we can select
CPUMASK_OFFSTACK. As that doesn't seem to be a problem today with
NR_CPUS=256, we only select this when NR_CPUS > 256.

CPUMASK_OFFSTACK configures the cpumasks in the kernel to be
dynamically allocated. This was used in the X86 architecture in the
past to enable support for larger CPU configurations up to 8k cpus.

With that is becomes possible to dynamically size the allocation of
the cpu bitmaps depending on the quantity of processors detected on
bootup. Memory used for cpumasks will increase if the kernel is
run on a machine with more cores.

Further increases may be needed if ARM processor vendors start
supporting more processors. Given the current inflationary trends
in core counts from multiple processor manufacturers this may occur.

There are minor regressions for hackbench. The kernel data size
for 512 cpus is smaller with offstack than with onstack.

Benchmark results using hackbench average over 10 runs of

  hackbench -s 512 -l 2000 -g 15 -f 25 -P

on Altra 80 Core

Support for 256 CPUs on stack. Baseline

  7.8564 sec

Support for 512 CUs on stack.

  7.8713 sec + 0.18%

512 CPUS offstack

  7.8916 sec + 0.44%

Kernel size comparison:

    text    data     filename Difference to onstack256 baseline
25755648 9589248     vmlinuz-6.8.0-rc4-onstack256
25755648 9607680     vmlinuz-6.8.0-rc4-onstack512 +0.19%
25755648 9603584     vmlinuz-6.8.0-rc4-offstack512 +0.14%

Tested-by: Eric Mackay <eric.mackay@oracle.com>
Reviewed-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: Christoph Lameter (Ampere) <cl@linux.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/37099a57-b655-3b3a-56d0-5f7fbd49d7db@gentwo.org
[catalin.marinas@arm.com: use 'select' instead of duplicating 'config CPUMASK_OFFSTACK']
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agokselftest/arm64: Add 2023 DPISA hwcap test coverage
Mark Brown [Wed, 6 Mar 2024 23:14:54 +0000 (23:14 +0000)]
kselftest/arm64: Add 2023 DPISA hwcap test coverage

Add the hwcaps added for the 2023 DPISA extensions to the hwcaps test
program.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-9-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agokselftest/arm64: Add basic FPMR test
Mark Brown [Wed, 6 Mar 2024 23:14:53 +0000 (23:14 +0000)]
kselftest/arm64: Add basic FPMR test

Verify that a FPMR frame is generated on systems that support FPMR and not
generated otherwise.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-8-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agokselftest/arm64: Handle FPMR context in generic signal frame parser
Mark Brown [Wed, 6 Mar 2024 23:14:52 +0000 (23:14 +0000)]
kselftest/arm64: Handle FPMR context in generic signal frame parser

Teach the generic signal frame parsing code about the newly added FPMR
frame, avoiding warnings every time one is generated.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-7-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agoarm64/hwcap: Define hwcaps for 2023 DPISA features
Mark Brown [Wed, 6 Mar 2024 23:14:51 +0000 (23:14 +0000)]
arm64/hwcap: Define hwcaps for 2023 DPISA features

The 2023 architecture extensions include a large number of floating point
features, most of which simply add new instructions. Add hwcaps so that
userspace can enumerate these features.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-6-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agoarm64/ptrace: Expose FPMR via ptrace
Mark Brown [Wed, 6 Mar 2024 23:14:50 +0000 (23:14 +0000)]
arm64/ptrace: Expose FPMR via ptrace

Add a new regset to expose FPMR via ptrace. It is not added to the FPSIMD
registers since that structure is exposed elsewhere without any allowance
for extension we don't add there.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-5-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agoarm64/signal: Add FPMR signal handling
Mark Brown [Wed, 6 Mar 2024 23:14:49 +0000 (23:14 +0000)]
arm64/signal: Add FPMR signal handling

Expose FPMR in the signal context on systems where it is supported. The
kernel validates the exact size of the FPSIMD registers so we can't readily
add it to fpsimd_context without disruption.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-4-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agoarm64/fpsimd: Support FEAT_FPMR
Mark Brown [Wed, 6 Mar 2024 23:14:48 +0000 (23:14 +0000)]
arm64/fpsimd: Support FEAT_FPMR

FEAT_FPMR defines a new EL0 accessible register FPMR use to configure the
FP8 related features added to the architecture at the same time. Detect
support for this register and context switch it for EL0 when present.

Due to the sharing of responsibility for saving floating point state
between the host kernel and KVM FP8 support is not yet implemented in KVM
and a stub similar to that used for SVCR is provided for FPMR in order to
avoid bisection issues. To make it easier to share host state with the
hypervisor we store FPMR as a hardened usercopy field in uw (along with
some padding).

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-3-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agoarm64/fpsimd: Enable host kernel access to FPMR
Mark Brown [Wed, 6 Mar 2024 23:14:47 +0000 (23:14 +0000)]
arm64/fpsimd: Enable host kernel access to FPMR

FEAT_FPMR provides a new generally accessible architectural register FPMR.
This is only accessible to EL0 and EL1 when HCRX_EL2.EnFPM is set to 1,
do this when the host is running. The guest part will be done along with
context switching the new register and exposing it via guest management.

Acked-by: Marc Zyngier <maz@kernel.org>
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-2-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
12 months agoarm64/cpufeature: Hook new identification registers up to cpufeature
Mark Brown [Wed, 6 Mar 2024 23:14:46 +0000 (23:14 +0000)]
arm64/cpufeature: Hook new identification registers up to cpufeature

The 2023 architecture extensions have defined several new ID registers,
hook them up to the cpufeature code so we can add feature checks and hwcaps
based on their contents.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240306-arm64-2023-dpisa-v5-1-c568edc8ed7f@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agodocs: perf: Fix build warning of hisi-pcie-pmu.rst
Yicong Yang [Tue, 5 Mar 2024 12:25:17 +0000 (20:25 +0800)]
docs: perf: Fix build warning of hisi-pcie-pmu.rst

`make htmldocs SPHINXDIRS="admin-guide"` shows below warnings:
Documentation/admin-guide/perf/hisi-pcie-pmu.rst:48: ERROR: Unexpected indentation.
Documentation/admin-guide/perf/hisi-pcie-pmu.rst:49: WARNING: Block quote ends without a blank line; unexpected unindent.

Fix this.

Closes: https://lore.kernel.org/lkml/20231011172250.5a6498e5@canb.auug.org.au/
Fixes: 89a032923d4b ("docs: perf: Update usage for target filter of hisi-pcie-pmu")
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Link: https://lore.kernel.org/r/20240305122517.12179-1-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agoperf: starfive: Only allow COMPILE_TEST for 64-bit architectures
Will Deacon [Tue, 5 Mar 2024 10:56:49 +0000 (10:56 +0000)]
perf: starfive: Only allow COMPILE_TEST for 64-bit architectures

The kbuild robot exploded while wasting its time building the Starfive
PMU driver for the 32-bit PA-RISC and Hexagon architectures.

Adjust the Kconfig dependencies so that COMPILE_TEST is only applicable
for 64-bit architectures (which implement writeq()).

Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Will Deacon <will@kernel.org>
13 months agoMAINTAINERS: Add entry for StarFive StarLink PMU
Ji Sheng Teoh [Thu, 29 Feb 2024 07:27:20 +0000 (15:27 +0800)]
MAINTAINERS: Add entry for StarFive StarLink PMU

Add maintainer entry for StarFive StarLink PMU driver, and mark it as
"Maintained"

Signed-off-by: Ji Sheng Teoh <jisheng.teoh@starfivetech.com>
Link: https://lore.kernel.org/r/20240229072720.3987876-5-jisheng.teoh@starfivetech.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodocs: perf: Add description for StarFive's StarLink PMU
Ji Sheng Teoh [Thu, 29 Feb 2024 07:27:19 +0000 (15:27 +0800)]
docs: perf: Add description for StarFive's StarLink PMU

StarFive StarLink PMU support monitoring L3 memory system PMU events.
Add documentation to describe StarFive StarLink PMU support and it's
usage.

Signed-off-by: Ji Sheng Teoh <jisheng.teoh@starfivetech.com>
Link: https://lore.kernel.org/r/20240229072720.3987876-4-jisheng.teoh@starfivetech.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodt-bindings: perf: starfive: Add JH8100 StarLink PMU
Ji Sheng Teoh [Thu, 29 Feb 2024 07:27:18 +0000 (15:27 +0800)]
dt-bindings: perf: starfive: Add JH8100 StarLink PMU

Add device tree binding for StarFive's JH8100 StarLink PMU (Performance
Monitor Unit).

Signed-off-by: Ji Sheng Teoh <jisheng.teoh@starfivetech.com>
Reviewed-by: Conor Dooley <conor.dooley@microchip.com>
Link: https://lore.kernel.org/r/20240229072720.3987876-3-jisheng.teoh@starfivetech.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agoperf: starfive: Add StarLink PMU support
Ji Sheng Teoh [Thu, 29 Feb 2024 07:27:17 +0000 (15:27 +0800)]
perf: starfive: Add StarLink PMU support

This patch adds support for StarFive's StarLink PMU (Performance
Monitor Unit). StarLink PMU integrates one or more CPU cores with
a shared L3 memory system. The PMU supports overflow interrupt,
up to 16 programmable 64bit event counters, and an independent
64bit cycle counter. StarLink PMU is accessed via MMIO.

Example Perf stat output:
[root@user]# perf stat -a -e /starfive_starlink_pmu/cycles/ \
-e /starfive_starlink_pmu/read_miss/ \
-e /starfive_starlink_pmu/read_hit/ \
-e /starfive_starlink_pmu/release_request/  \
-e /starfive_starlink_pmu/write_hit/ \
-e /starfive_starlink_pmu/write_miss/ \
-e /starfive_starlink_pmu/write_request/ \
-e /starfive_starlink_pmu/writeback/ \
-e /starfive_starlink_pmu/read_request/ \
-- openssl speed rsa2048
Doing 2048 bits private rsa's for 10s: 5 2048 bits private RSA's in
2.84s
Doing 2048 bits public rsa's for 10s: 169 2048 bits public RSA's in
2.42s
version: 3.0.11
built on: Tue Sep 19 13:02:31 2023 UTC
options: bn(64,64)
CPUINFO: N/A
                  sign    verify    sign/s verify/s
rsa 2048 bits 0.568000s 0.014320s      1.8     69.8
/////////
 Performance counter stats for 'system wide':

         649991998      starfive_starlink_pmu/cycles/
           1009690      starfive_starlink_pmu/read_miss/
           1079750      starfive_starlink_pmu/read_hit/
           2089405      starfive_starlink_pmu/release_request/
               129      starfive_starlink_pmu/write_hit/
                70      starfive_starlink_pmu/write_miss/
               194      starfive_starlink_pmu/write_request/
            150080      starfive_starlink_pmu/writeback/
           2089423      starfive_starlink_pmu/read_request/

      27.062755678 seconds time elapsed

Signed-off-by: Ji Sheng Teoh <jisheng.teoh@starfivetech.com>
Link: https://lore.kernel.org/r/20240229072720.3987876-2-jisheng.teoh@starfivetech.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodocs: perf: Update usage for target filter of hisi-pcie-pmu
Junhao He [Fri, 23 Feb 2024 10:33:59 +0000 (18:33 +0800)]
docs: perf: Update usage for target filter of hisi-pcie-pmu

One of the "port" and "bdf" target filter interface must be set, and
the related events should preferably used in the same group.
Update the usage in the documentation.

Signed-off-by: Junhao He <hejunhao3@huawei.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20240223103359.18669-9-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodrivers/perf: hisi_pcie: Merge find_related_event() and get_event_idx()
Junhao He [Fri, 23 Feb 2024 10:33:58 +0000 (18:33 +0800)]
drivers/perf: hisi_pcie: Merge find_related_event() and get_event_idx()

The function xxx_find_related_event() scan all working events to find
related events. During this process, we also can find the idle counters.
If not found related events, return the first idle counter to simplify
the code.

Signed-off-by: Junhao He <hejunhao3@huawei.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20240223103359.18669-8-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodrivers/perf: hisi_pcie: Relax the check on related events
Junhao He [Fri, 23 Feb 2024 10:33:57 +0000 (18:33 +0800)]
drivers/perf: hisi_pcie: Relax the check on related events

If we use two events with the same filter and related event type
(see the following example), the driver check whether they are related
events and are in the same group, otherwise the function
hisi_pcie_pmu_find_related_event() return -EINVAL, then the 2nd event
cannot count but the 1st event is running, although the PCIe PMU has
other idle counters.

In this case, The perf event scheduler will make the two events to
multiplex a counter, if the user use the formula
(1st event_value / 2nd event_value) to calculate the bandwidth, he/she
won't get the correct value, because they are not counting at the
same period.

This patch tries to fix this by making the related events to use
different idle counters if they are not in the same event group.

And finally, I'm going to say. The related events are best used in the
same group [1]. There are two ways to know if they are related events.
a) By event name, such as the latency events "xxx_latency, xxx_cnt" or
bandwidth events "xxx_flux, xxx_time".
b) By event type, such as "event=0xXXXX, event=0x1XXXX".

Use group to count the related events:
  [1] -e "{pmu_name/xxx_latency,port=1/,pmu_name/xxx_cnt,port=1/}"

  example:
    1st event: hisi_pcie0_core1/event=0x804,port=1
    2nd event: hisi_pcie0_core1/event=0x10804,port=1

  test cmd:
    perf stat -e hisi_pcie0_core1/event=0x804,port=1/ \
               -e hisi_pcie0_core1/event=0x10804,port=1/

  before patch:
            25,281      hisi_pcie0_core1/event=0x804,port=1/    (49.91%)
           470,598      hisi_pcie0_core1/event=0x10804,port=1/    (50.09%)

  after patch:
            24,147      hisi_pcie0_core1/event=0x804,port=1/
           474,558      hisi_pcie0_core1/event=0x10804,port=1/

Signed-off-by: Junhao He <hejunhao3@huawei.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huwei.com>
Link: https://lore.kernel.org/r/20240223103359.18669-7-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodrivers/perf: hisi_pcie: Check the target filter properly
Junhao He [Fri, 23 Feb 2024 10:33:56 +0000 (18:33 +0800)]
drivers/perf: hisi_pcie: Check the target filter properly

The PMU can monitor traffic of certain target Root Port or downstream
target Endpoint. User can specify the target filter by the "port" or
"bdf" option respectively. The PMU can only monitor the Root Port or
Endpoint on the same PCIe core so the value of "port" or "bdf" should
be valid and will be checked by the driver.

Currently at least and only one of "port" and "bdf" option must be set.
If "port" filter is not set or is set explicitly to zero (default),
driver will regard the user specifies a "bdf" option since "port" option
is a bitmask of the target Root Ports and zero is not a valid
value.

If user not explicitly set "port" or "bdf" filter, the driver uses "bdf"
default value (zero) to set target filter, but driver will skip the
check of bdf=0, although it's a valid value (meaning 0000:000:00.0).
Then the user just gets zero.

Therefore, we need to check if both "port" and "bdf" are invalid, then
return failure and report warning.

Testing:
before the patch:
                   0      hisi_pcie0_core1/rx_mrd_flux/
                   0      hisi_pcie0_core1/rx_mrd_flux,port=0/
              24,124      hisi_pcie0_core1/rx_mrd_flux,port=1/
                   0      hisi_pcie0_core1/rx_mrd_flux,bdf=0/
                   0      hisi_pcie0_core1/rx_mrd_flux,port=0x800/
     <not supported>      hisi_pcie0_core1/rx_mrd_flux,bdf=1/
              24,132      hisi_pcie0_core1/rx_mrd_flux,bdf=0x1700/
     <not supported>      hisi_pcie0_core1/rx_mrd_flux,port=0x0,bdf=0x0/
     <not supported>      hisi_pcie0_core1/rx_mrd_flux,port=0x0,bdf=0x1/
              24,138      hisi_pcie0_core1/rx_mrd_flux,port=0x0,bdf=0x1700/
              24,126      hisi_pcie0_core1/rx_mrd_flux,port=0x1,bdf=0x0/

after the patch:
     <not supported>      hisi_pcie0_core1/rx_mrd_flux/
     <not supported>      hisi_pcie0_core1/rx_mrd_flux,port=0/
              24,153      hisi_pcie0_core1/rx_mrd_flux,port=1/
                   0      hisi_pcie0_core1/rx_mrd_flux,port=0x800/
     <not supported>      hisi_pcie0_core1/rx_mrd_flux,bdf=0/
     <not supported>      hisi_pcie0_core1/rx_mrd_flux,bdf=1/
              24,117      hisi_pcie0_core1/rx_mrd_flux,bdf=0x1700/
     <not supported>      hisi_pcie0_core1/rx_mrd_flux,port=0x0,bdf=0x0/
     <not supported>      hisi_pcie0_core1/rx_mrd_flux,port=0x0,bdf=0x1/
              24,120      hisi_pcie0_core1/rx_mrd_flux,port=0x0,bdf=0x1700/
              24,123      hisi_pcie0_core1/rx_mrd_flux,port=0x1,bdf=0x0/

Signed-off-by: Junhao He <hejunhao3@huawei.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20240223103359.18669-6-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodrivers/perf: hisi_pcie: Add more events for counting TLP bandwidth
Yicong Yang [Fri, 23 Feb 2024 10:33:55 +0000 (18:33 +0800)]
drivers/perf: hisi_pcie: Add more events for counting TLP bandwidth

A typical PCIe transaction is consisted of various TLP packets in both
direction. For counting bandwidth only memory read events are exported
currently. Add memory write and completion counting events of both
direction to complete the bandwidth counting.

Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20240223103359.18669-5-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodrivers/perf: hisi_pcie: Fix incorrect counting under metric mode
Yicong Yang [Fri, 23 Feb 2024 10:33:54 +0000 (18:33 +0800)]
drivers/perf: hisi_pcie: Fix incorrect counting under metric mode

The metric counting shows incorrect results if the events in the
metric group using the same event but different filter options.
This is because we only judge the event code to decide whether
the event in the metric group should share the same hardware
counter, but ignore the settings of the filter.

For example, on a platform of 2 ports 0x1 and 0x2 but only port
0x1 has a downstream PCIe NVME device. The metric counting
shows both ports have the same counts because we misassign these
two events to one same hardware counter:
[root@localhost perf-iostat]# ./perf stat -e '{hisi_pcie0_core1/event=0x0104,port=0x2/,hisi_pcie0_core1/event=0x0104,port=0x1/}'

 Performance counter stats for 'system wide':

        7907484924      hisi_pcie0_core1/event=0x0104,port=0x2/
        7907484924      hisi_pcie0_core1/event=0x0104,port=0x1/

      10.153863691 seconds time elapsed

Fix this by using the whole config rather than the event only
to judge whether two events are the same and should share the
same hardware counter. With this patch, the metric counting in
the above case tends to be corrected:

[root@localhost perf-iostat]# ./perf stat -e '{hisi_pcie0_core1/event=0x0104,port=0x2/,hisi_pcie0_core1/event=0x0104,port=0x1/}'

 Performance counter stats for 'system wide':

                 0      hisi_pcie0_core1/event=0x0104,port=0x2/
        8123122077      hisi_pcie0_core1/event=0x0104,port=0x1/

      10.152875631 seconds time elapsed

Fixes: 8404b0fbc7fb ("drivers/perf: hisi: Add driver for HiSilicon PCIe PMU")
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20240223103359.18669-4-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodrivers/perf: hisi_pcie: Introduce hisi_pcie_pmu_get_event_ctrl_val()
Yicong Yang [Fri, 23 Feb 2024 10:33:53 +0000 (18:33 +0800)]
drivers/perf: hisi_pcie: Introduce hisi_pcie_pmu_get_event_ctrl_val()

Factor out retrieving of the register value for the
corresponding event from hisi_pcie_config_event_ctrl() into a
new function hisi_pcie_pmu_get_event_ctrl_val() allowing future
reuse.

Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20240223103359.18669-3-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodrivers/perf: hisi_pcie: Rename hisi_pcie_pmu_{config,clear}_filter()
Yicong Yang [Fri, 23 Feb 2024 10:33:52 +0000 (18:33 +0800)]
drivers/perf: hisi_pcie: Rename hisi_pcie_pmu_{config,clear}_filter()

hisi_pcie_pmu_{config,clear}_filter() are config/clear HISI_PCIE_EVENT_CTRL
register which contains not only the filter but also the event code. The
function names are bit misleading. Rename it to
hisi_pcie_pmu_{config,clear}_event_ctrl() to reflects their functions
more accurately.

Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20240223103359.18669-2-yangyicong@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agodrivers/perf: hisi: Enable HiSilicon Erratum 162700402 quirk for HIP09
Junhao He [Tue, 27 Feb 2024 12:52:31 +0000 (20:52 +0800)]
drivers/perf: hisi: Enable HiSilicon Erratum 162700402 quirk for HIP09

HiSilicon UC PMU v2 suffers the erratum 162700402 that the PMU counter
cannot be set due to the lack of clock under power saving mode. This will
lead to error or inaccurate counts. The clock can be enabled by the PMU
global enabling control.

This patch tries to fix this by set the UC PMU enable before set event
period to turn on the clock, and then restore the UC PMU configuration.
The counter register can hold its value without a clock.

Signed-off-by: Junhao He <hejunhao3@huawei.com>
Reviewed-by: Yicong Yang <yangyicong@hisilicon.com>
Link: https://lore.kernel.org/r/20240227125231.53127-1-hejunhao3@huawei.com
Signed-off-by: Will Deacon <will@kernel.org>
13 months agoarm64: Remove enable_daif macro
Jinjie Ruan [Thu, 29 Feb 2024 13:28:01 +0000 (21:28 +0800)]
arm64: Remove enable_daif macro

Since commit bb8e93a287a5 ("arm64: entry: convert SError handlers to C"),
the enable_daif assembler macro is no longer used anywhere, so remove it.

Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com>
Reviewed-by: Mark Brown <broonie@kernel.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/20240229132802.1682026-2-ruanjinjie@huawei.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/hw_breakpoint: Directly use ESR_ELx_WNR for an watchpoint exception
Anshuman Khandual [Thu, 29 Feb 2024 08:34:31 +0000 (14:04 +0530)]
arm64/hw_breakpoint: Directly use ESR_ELx_WNR for an watchpoint exception

Let's use existing ISS encoding for an watchpoint exception i.e ESR_ELx_WNR
This represents an instruction's either writing to or reading from a memory
location during an watchpoint exception. While here this drops non-standard
macro AARCH64_ESR_ACCESS_MASK.

Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: linux-arm-kernel@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Reviewed-by: Mark Brown <broonie@kernel.org>
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/20240229083431.356578-1-anshuman.khandual@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: cpufeatures: Clean up temporary variable to simplify code
Liao Chang [Thu, 29 Feb 2024 10:52:08 +0000 (10:52 +0000)]
arm64: cpufeatures: Clean up temporary variable to simplify code

Clean up one temporary variable to simplifiy code in capability
detection.

Signed-off-by: Liao Chang <liaochang1@huawei.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/20240229105208.456704-1-liaochang1@huawei.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/mm: Avoid ID mapping of kpti flag if it is no longer needed
Ard Biesheuvel [Fri, 1 Mar 2024 10:40:49 +0000 (11:40 +0100)]
arm64/mm: Avoid ID mapping of kpti flag if it is no longer needed

arm64_use_ng_mappings will be set to 'true' by the early boot code if it
decides to use non-global (nG) attributes for all kernel mappings,
typically when enabling KASLR on a system that does not implement E0PD.

In this case, the G-to-nG update routines are never called, and so there
is no reason to create the writable mapping of the associated status
flag in the ID map.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240301104046.1234309-6-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/mm: Use generic __pud_free() helper in pud_free() implementation
Ard Biesheuvel [Fri, 1 Mar 2024 10:40:48 +0000 (11:40 +0100)]
arm64/mm: Use generic __pud_free() helper in pud_free() implementation

Commit 0dd4f60a2c76 ("arm64: mm: Add support for folding PUDs at
runtime") implements specialized PUD alloc/free helpers to allow the
decision whether or not to fold PUDs to be made at runtime when the
number of paging levels is 4 or higher.

Its implementation of pud_free() is based on the generic version that
existed when the patch was first written, but in the meantime, the
freeing of a PUD has become a bit more involved, and so instead of
simply freeing the page, we should invoke the generic __pud_free() that
encapsulates whatever needs doing at this point.

This fixes a reported warning emitted by the page flags
self-diagnostics.

Reported-by: Ryan Roberts <ryan.roberts@arm.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Tested-by: Ryan Roberts <ryan.roberts@arm.com>
Link: https://lore.kernel.org/r/20240301104046.1234309-5-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Update setup_arch() comment on interrupt masking
Ryo Takakura [Wed, 28 Feb 2024 02:28:36 +0000 (11:28 +0900)]
arm64: Update setup_arch() comment on interrupt masking

DAIF_PROCCTX_NOIRQ contains the FIQ bit. Update the comment as only
asynchronous aborts are unmasked and FIQ is still masked.

Signed-off-by: Ryo Takakura <takakura@valinux.co.jp>
Link: https://lore.kernel.org/r/20240228022836.1756-1-takakura@valinux.co.jp
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: remove unnecessary ifdefs around is_compat_task()
Leonardo Bras [Tue, 9 Jan 2024 03:46:50 +0000 (00:46 -0300)]
arm64: remove unnecessary ifdefs around is_compat_task()

Currently some parts of the codebase will test for CONFIG_COMPAT before
testing is_compat_task().

is_compat_task() is a inlined function only present on CONFIG_COMPAT.
On the other hand, for !CONFIG_COMPAT, we have in linux/compat.h:

 #define is_compat_task() (0)

Since we have this define available in every usage of is_compat_task() for
!CONFIG_COMPAT, it's unnecessary to keep the ifdefs, since the compiler is
smart enough to optimize-out those snippets on CONFIG_COMPAT=n

This requires some regset code as well as a few other defines to be made
available on !CONFIG_COMPAT, so some symbols can get resolved before
getting optimized-out.

Signed-off-by: Leonardo Bras <leobras@redhat.com>
Reviewed-by: Arnd Bergmann <arnd@arndb.de>
Link: https://lore.kernel.org/r/20240109034651.478462-2-leobras@redhat.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: ftrace: Don't forbid CALL_OPS+CC_OPTIMIZE_FOR_SIZE with Clang
Stephen Boyd [Fri, 23 Feb 2024 06:40:29 +0000 (22:40 -0800)]
arm64: ftrace: Don't forbid CALL_OPS+CC_OPTIMIZE_FOR_SIZE with Clang

Per commit b3f11af9b2ce ("arm64: ftrace: forbid CALL_OPS with
CC_OPTIMIZE_FOR_SIZE"), GCC is silently ignoring `-falign-functions=N`
when passed `-Os`, causing functions to be improperly aligned. This
doesn't seem to be a problem with Clang though, where enabling CALL_OPS
with CC_OPTIMIZE_FOR_SIZE doesn't spit out any warnings at boot about
misaligned patch-sites. Only forbid CALL_OPS if GCC is used and we're
optimizing for size so that CALL_OPS can be used with clang optimizing
for size.

Cc: Jason Ling <jasonling@chromium.org>
Cc: Florian Fainelli <f.fainelli@gmail.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Nathan Chancellor <nathan@kernel.org>
Cc: Nick Desaulniers <ndesaulniers@google.com>
Cc: Bill Wendling <morbo@google.com>
Cc: Justin Stitt <justinstitt@google.com>
Cc: llvm@lists.linux.dev
Fixes: b3f11af9b2ce ("arm64: ftrace: forbid CALL_OPS with CC_OPTIMIZE_FOR_SIZE")
Signed-off-by: Stephen Boyd <swboyd@chromium.org>
Reviewed-by: Nathan Chancellor <nathan@kernel.org>
Link: https://lore.kernel.org/r/20240223064032.3463229-1-swboyd@chromium.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: gitignore: ignore relacheck
Bartosz Golaszewski [Thu, 22 Feb 2024 21:04:41 +0000 (22:04 +0100)]
arm64: gitignore: ignore relacheck

Add the generated executable for relacheck to the list of ignored files.

Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
Link: https://lore.kernel.org/r/20240222210441.33142-1-brgl@bgdev.pl
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sme: Ensure that all fields in SMCR_EL1 are set to known values
Mark Brown [Tue, 13 Feb 2024 15:32:46 +0000 (15:32 +0000)]
arm64/sme: Ensure that all fields in SMCR_EL1 are set to known values

At present nothing in our CPU initialisation code ever sets unknown fields
in SMCR_EL1 to known values, all updates to SMCR_EL1 are read/modify/write
sequences. All the unknown fields are RES0, explicitly initialise them as
such to avoid future surprises.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240213-arm64-fp-init-vec-cr-v1-2-7e7c2d584f26@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sve: Ensure that all fields in ZCR_EL1 are set to known values
Mark Brown [Tue, 13 Feb 2024 15:32:45 +0000 (15:32 +0000)]
arm64/sve: Ensure that all fields in ZCR_EL1 are set to known values

At present nothing in our CPU initialisation code ever sets unknown fields
in ZCR_EL1 to known values, all updates to ZCR_EL1 are read/modify/write
sequences for LEN. All the unknown fields are RES0, explicitly initialise
them as such to avoid future surprises.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240213-arm64-fp-init-vec-cr-v1-1-7e7c2d584f26@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sve: Document that __SVE_VQ_MAX is much larger than needed
Mark Brown [Fri, 9 Feb 2024 16:53:36 +0000 (16:53 +0000)]
arm64/sve: Document that __SVE_VQ_MAX is much larger than needed

__SVE_VQ_MAX is defined without comment as 512 but the actual
architectural maximum is 16, a substantial difference which might not
be obvious to readers especially given the several different units used
for specifying vector sizes in various contexts and the fact that it's
often used via macros.  In an effort to minimise surprises for users who
might assume the value is the architectural maximum and use it to do
things like size allocations add a comment noting the difference, and
add a note for SVE_VQ_MAX to aid discoverability.

Signed-off-by: Mark Brown <broonie@kernel.org>
Acked-by: Dave Martin <Dave.Martin@arm.com>
Link: https://lore.kernel.org/r/20240209-arm64-sve-vl-max-comment-v2-1-111b283469ee@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: make member of struct pt_regs and it's offset macro in the same order
Kemeng Shi [Tue, 30 Jan 2024 17:55:04 +0000 (01:55 +0800)]
arm64: make member of struct pt_regs and it's offset macro in the same order

In struct pt_regs, member pstate is after member pc. Move offset macro
of pstate after offset macro of pc to improve readability a little.

Signed-off-by: Kemeng Shi <shikemeng@huaweicloud.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/20240130175504.106364-1-shikemeng@huaweicloud.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: remove unneeded BUILD_BUG_ON assertion
Dawei Li [Fri, 2 Feb 2024 04:02:11 +0000 (12:02 +0800)]
arm64: remove unneeded BUILD_BUG_ON assertion

Since commit c02433dd6de3 ("arm64: split thread_info from task stack"),
CONFIG_THREAD_INFO_IN_TASK is enabled unconditionally for arm64. So
remove this always-true assertion from arch_dup_task_struct.

Signed-off-by: Dawei Li <dawei.li@shingroup.cn>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/20240202040211.3118918-1-dawei.li@shingroup.cn
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sysreg: Update ID_AA64DFR0_EL1 register
Anshuman Khandual [Tue, 20 Feb 2024 03:48:29 +0000 (09:18 +0530)]
arm64/sysreg: Update ID_AA64DFR0_EL1 register

This updates ID_AA64DFR0_EL1.PMSVer and ID_AA64DFR0_EL1.DebugVer register
fields as per the definitions based on DDI0601 2023-12.

Cc: Will Deacon <will@kernel.org>
Cc: Mark Brown <broonie@kernel.org>
Cc: linux-arm-kernel@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Reviewed-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240220034829.3098373-1-anshuman.khandual@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sysreg: Update ID_DFR0_EL1 register fields
Anshuman Khandual [Tue, 20 Feb 2024 02:53:43 +0000 (08:23 +0530)]
arm64/sysreg: Update ID_DFR0_EL1 register fields

This updates ID_DFR0_EL1.PerfMon and ID_DFR0_EL1.CopDbg register fields as
per the definitions based on DDI0601 2023-12.

Cc: Will Deacon <will@kernel.org>
Cc: Mark Brown <broonie@kernel.org>
Cc: linux-arm-kernel@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Reviewed-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240220025343.3093955-1-anshuman.khandual@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sysreg: Add register fields for ID_AA64DFR1_EL1
Anshuman Khandual [Tue, 20 Feb 2024 02:32:03 +0000 (08:02 +0530)]
arm64/sysreg: Add register fields for ID_AA64DFR1_EL1

This adds register fields for ID_AA64DFR1_EL1 as per the definitions based
on DDI0601 2023-12.

Cc: Will Deacon <will@kernel.org>
Cc: Mark Brown <broonie@kernel.org>
Cc: linux-arm-kernel@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Reviewed-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240220023203.3091229-1-anshuman.khandual@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sme: Remove spurious 'is' in SME documentation
Mark Brown [Wed, 24 Jan 2024 18:12:38 +0000 (18:12 +0000)]
arm64/sme: Remove spurious 'is' in SME documentation

Just a typographical error.

Reported-by: Edmund Grimley-Evans <edmund.grimley-evans@arm.com>
Reviewed-by: Dave Martin <Dave.Martin@arm.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240124-arm64-sve-sme-doc-v2-4-fe3964fb3c19@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/fp: Clarify effect of setting an unsupported system VL
Mark Brown [Wed, 24 Jan 2024 18:12:37 +0000 (18:12 +0000)]
arm64/fp: Clarify effect of setting an unsupported system VL

The documentation for system vector length configuration does not cover all
cases where unsupported values are written, tighten it up.

Reported-by: Edmund Grimley-Evans <edmund.grimley-evans@arm.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
Reviewed-by: Dave Martin <Dave.Martin@arm.com>
Link: https://lore.kernel.org/r/20240124-arm64-sve-sme-doc-v2-3-fe3964fb3c19@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sme: Fix cut'n'paste in ABI document
Mark Brown [Wed, 24 Jan 2024 18:12:36 +0000 (18:12 +0000)]
arm64/sme: Fix cut'n'paste in ABI document

The ABI for SME is very like that for SVE so bits of the ABI were copied
but not adequately search and replaced, fix that.

Reported-by: Edmund Grimley-Evans <edmund.grimley-evans@arm.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
Reviewed-by: Dave Martin <Dave.Martin@arm.com>
Link: https://lore.kernel.org/r/20240124-arm64-sve-sme-doc-v2-2-fe3964fb3c19@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64/sve: Remove bitrotted comment about syscall behaviour
Mark Brown [Wed, 24 Jan 2024 18:12:35 +0000 (18:12 +0000)]
arm64/sve: Remove bitrotted comment about syscall behaviour

When we documented that we always clear state not shared with FPSIMD we
didn't catch all of the places that mentioned that state might not be
cleared, remove a lingering reference.

Reported-by: Edmund Grimley-Evans <edmund.grimley-evans@arm.com>
Reviewed-by: Dave Martin <Dave.Martin@arm.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240124-arm64-sve-sme-doc-v2-1-fe3964fb3c19@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agokselftest/arm64: Test that ptrace takes effect in the target process
Mark Brown [Mon, 22 Jan 2024 21:05:03 +0000 (21:05 +0000)]
kselftest/arm64: Test that ptrace takes effect in the target process

While we have test coverage for the ptrace interface in our selftests
the current programs have a number of gaps. The testing is done per
regset so does not cover interactions and at no point do any of the
tests actually run the traced processes meaning that there is no
validation that anything we read or write corresponds to register values
the process actually sees. Let's add a new program which attempts to cover
these gaps.

Each test we do performs a single ptrace write. For each test we generate
some random initial register data in memory and then fork() and trace a
child. The child will load the generated data into the registers then
trigger a breakpoint. The parent waits for the breakpoint then reads the
entire child register state via ptrace, verifying that the values expected
were actually loaded by the child. It then does the write being tested
and resumes the child. Once resumed the child saves the register state
it sees to memory and executes another breakpoint. The parent uses
process_vm_readv() to get these values from the child and verifies that
the values were as expected before cleaning up the child.

We generate configurations with combinations of vector lengths and SVCR
values and then try every ptrace write which will implement the
transition we generated. In order to control execution time (especially
in emulation) we only cover the minimum and maximum VL for each of SVE
and SME, this will ensure we generate both increasing and decreasing
changes in vector length. In order to provide a baseline test we also
check the case where we resume the child without doing a ptrace write.

In order to simplify the generation of the test count for kselftest we
will report but skip a substantial number of tests that can't actually
be expressed via a single ptrace write, several times more than we
actually run. This is noisy and will add some overhead but is very much
simpler so is probably worth the tradeoff.

Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240122-arm64-test-ptrace-regs-v1-1-0897f822d73e@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: kretprobes: acquire the regs via a BRK exception
Mark Rutland [Thu, 8 Feb 2024 14:59:16 +0000 (14:59 +0000)]
arm64: kretprobes: acquire the regs via a BRK exception

On arm64, kprobes always take an exception and so create a struct
pt_regs through the usual exception entry logic. Similarly kretprobes
taskes and exception for function entry, but for function returns it
uses a trampoline which attempts to create a struct pt_regs without
taking an exception.

This is problematic for a few reasons, including:

1) The kretprobes trampoline neither saves nor restores all of the
   portions of PSTATE. Before invoking the handler it saves a number of
   portions of PSTATE, and after returning from the handler it restores
   NZCV before returning to the original return address provided by the
   handler.

2) The kretprobe trampoline constructs the PSTATE value piecemeal from
   special purpose registers as it cannot read all of PSTATE atomically
   without taking an exception. This is somewhat fragile, and it's not
   possible to reliably recover PSTATE information which only exists on
   some physical CPUs (e.g. when SSBS support is mismatched).

   Today the kretprobes trampoline does not record:

   - BTYPE
   - SSBS
   - ALLINT
   - SS
   - PAN
   - UAO
   - DIT
   - TCO

   ... and this will only get worse with future architecture extensions
   which add more PSTATE bits.

3) The kretprobes trampoline doesn't store portions of struct pt_regs
   (e.g. the PMR value when using pseudo-NMIs). Due to this, helpers
   which operate on a struct pt_regs, such as interrupts_enabled(), may
   not work correctly.

4) The function entry and function exit handlers run in different
   contexts. The entry handler will always be run in a debug exception
   context (which is currently treated as an NMI), but the return will
   be treated as whatever context the instrumented function was executed
   in. The differences between these contexts are liable to cause
   problems (e.g. as the two can be differently interruptible or
   preemptible, adversely affecting synchronization between the
   handlers).

5) As the kretprobes trampoline runs in the same context as the code
   being probed, it is subject to the same single-stepping context,
   which may not be desirable if this is being driven by the kprobes
   handlers.

Overall, this is fragile, painful to maintain, and gets in the way of
supporting other things (e.g. RELIABLE_STACKTRACE, FEAT_NMI).

This patch addresses these issues by replacing the kretprobes trampoline
with a `BRK` instruction, and using an exception boundary to acquire and
restore the regs, in the same way as the regular kprobes trampoline.

Ive tested this atop v6.8-rc3:

| KTAP version 1
| 1..1
|     KTAP version 1
|     # Subtest: kprobes_test
|     # module: test_kprobes
|     1..7
|     ok 1 test_kprobe
|     ok 2 test_kprobes
|     ok 3 test_kprobe_missed
|     ok 4 test_kretprobe
|     ok 5 test_kretprobes
|     ok 6 test_stacktrace_on_kretprobe
|     ok 7 test_stacktrace_on_nested_kretprobe
| # kprobes_test: pass:7 fail:0 skip:0 total:7
| # Totals: pass:7 fail:0 skip:0 total:7
| ok 1 kprobes_test

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Florent Revest <revest@chromium.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Link: https://lore.kernel.org/r/20240208145916.2004154-1-mark.rutland@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Unmask Debug + SError in do_notify_resume()
Mark Rutland [Tue, 6 Feb 2024 12:38:48 +0000 (12:38 +0000)]
arm64: Unmask Debug + SError in do_notify_resume()

When returning to a user context, the arm64 entry code masks all DAIF
exceptions before handling pending work in exit_to_user_mode_prepare()
and do_notify_resume(), where it will transiently unmask all DAIF
exceptions. This is a holdover from the old entry assembly, which
conservatively masked all DAIF exceptions, and it's only necessary to
mask interrupts at this point during the exception return path, so long
as we subsequently mask all DAIF exceptions before the actual exception
return.

While most DAIF manipulation follows a save...restore sequence, the
manipulation in do_notify_resume() is the other way around, unmasking
all DAIF exceptions before masking them again. This is unfortunate as we
unnecessarily mask Debug and SError exceptions, and it would be nice to
remove this special case to make DAIF manipulation simpler and most
consistent.

This patch changes exit_to_user_mode_prepare() and do_notify_resume() to
only mask interrupts while handling pending work, masking other DAIF
exceptions after this has completed. This removes the unusual DAIF
manipulation and allows Debug and SError exceptions to be taken for a
slightly longer window during the exception return path.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: James Morse <james.morse@arm.com>
Cc: Mark Brown <broonie@kernel.org>
Cc: Will Deacon <will@kernel.org>
Reviewed-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240206123848.1696480-4-mark.rutland@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Itaru Kitayama <itaru.kitayama@linux.dev>
13 months agoarm64: Move do_notify_resume() to entry-common.c
Mark Rutland [Tue, 6 Feb 2024 12:38:47 +0000 (12:38 +0000)]
arm64: Move do_notify_resume() to entry-common.c

Currently do_notify_resume() lives in arch/arm64/kernel/signal.c, but it would
make more sense for it to live in entry-common.c as it handles more than
signals, and is coupled with the rest of the return-to-userspace sequence (e.g.
with unusual DAIF masking that matches the exception return requirements).

Move do_notify_resume() to entry-common.c.

There should be no functional change as a result of this patch.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: James Morse <james.morse@arm.com>
Cc: Mark Brown <broonie@kernel.org>
Cc: Will Deacon <will@kernel.org>
Reviewed-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240206123848.1696480-3-mark.rutland@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Itaru Kitayama <itaru.kitayama@linux.dev>
13 months agoarm64: Simplify do_notify_resume() DAIF masking
Mark Rutland [Tue, 6 Feb 2024 12:38:46 +0000 (12:38 +0000)]
arm64: Simplify do_notify_resume() DAIF masking

In do_notify_resume, we handle _TIF_NEED_RESCHED differently from all
other flags, leaving IRQ+FIQ masked when calling into schedule(). This
masking is a historical artifact, and it is not currently necessary
to mask IRQ+FIQ when calling into schedule (as evidenced by the generic
exit_to_user_mode_loop(), which unmasks IRQs before checking
_TIF_NEED_RESCHED and calling schedule()).

This patch removes the special case for _TIF_NEED_RESCHED, moving this
check into the main loop such that schedule() will be called from a
regular process context with IRQ+FIQ unmasked. This is a minor
simplification to do_notify_resume() and brings it into line with the
generic exit_to_user_mode_loop() logic. This will also aid subsequent
rework of DAIF management.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: James Morse <james.morse@arm.com>
Cc: Mark Brown <broonie@kernel.org>
Cc: Will Deacon <will@kernel.org>
Reviewed-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240206123848.1696480-2-mark.rutland@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Itaru Kitayama <itaru.kitayama@linux.dev>
13 months agoarm64: io: permit offset addressing
Mark Rutland [Wed, 24 Jan 2024 11:12:59 +0000 (11:12 +0000)]
arm64: io: permit offset addressing

Currently our IO accessors all use register addressing without offsets,
but we could safely use offset addressing (without writeback) to
simplify and optimize the generated code.

To function correctly under a hypervisor which emulates IO accesses, we
must ensure that any faulting/trapped IO access results in an ESR_ELx
value with ESR_ELX.ISS.ISV=1 and with the tranfer register described in
ESR_ELx.ISS.SRT. This means that we can only use loads/stores of a
single general purpose register (or the zero register), and must avoid
writeback addressing modes. However, we can use immediate offset
addressing modes, as these still provide ESR_ELX.ISS.ISV=1 and a valid
ESR_ELx.ISS.SRT when those accesses fault at Stage-2.

Currently we only use register addressing without offsets. We use the
"r" constraint to place the address into a register, and manually
generate the register addressing by surrounding the resulting register
operand with square braces, e.g.

| static __always_inline void __raw_writeq(u64 val, volatile void __iomem *addr)
| {
|         asm volatile("str %x0, [%1]" : : "rZ" (val), "r" (addr));
| }

Due to this, sequences of adjacent accesses need to generate addresses
using separate instructions. For example, the following code:

| void writeq_zero_8_times(void *ptr)
| {
|        writeq_relaxed(0, ptr + 8 * 0);
|        writeq_relaxed(0, ptr + 8 * 1);
|        writeq_relaxed(0, ptr + 8 * 2);
|        writeq_relaxed(0, ptr + 8 * 3);
|        writeq_relaxed(0, ptr + 8 * 4);
|        writeq_relaxed(0, ptr + 8 * 5);
|        writeq_relaxed(0, ptr + 8 * 6);
|        writeq_relaxed(0, ptr + 8 * 7);
| }

... is compiled to:

| <writeq_zero_8_times>:
|     str     xzr, [x0]
|     add     x1, x0, #0x8
|     str     xzr, [x1]
|     add     x1, x0, #0x10
|     str     xzr, [x1]
|     add     x1, x0, #0x18
|     str     xzr, [x1]
|     add     x1, x0, #0x20
|     str     xzr, [x1]
|     add     x1, x0, #0x28
|     str     xzr, [x1]
|     add     x1, x0, #0x30
|     str     xzr, [x1]
|     add     x0, x0, #0x38
|     str     xzr, [x0]
|     ret

As described above, we could safely use immediate offset addressing,
which would allow the ADDs to be folded into the address generation for
the STRs, resulting in simpler and smaller generated assembly. We can do
this by using the "o" constraint to allow the compiler to generate
offset addressing (without writeback) for a memory operand, e.g.

| static __always_inline void __raw_writeq(u64 val, volatile void __iomem *addr)
| {
|         volatile u64 __iomem *ptr = addr;
|         asm volatile("str %x0, %1" : : "rZ" (val), "o" (*ptr));
| }

... which results in the earlier code sequence being compiled to:

| <writeq_zero_8_times>:
|     str     xzr, [x0]
|     str     xzr, [x0, #8]
|     str     xzr, [x0, #16]
|     str     xzr, [x0, #24]
|     str     xzr, [x0, #32]
|     str     xzr, [x0, #40]
|     str     xzr, [x0, #48]
|     str     xzr, [x0, #56]
|     ret

As Will notes at:

  https://lore.kernel.org/linux-arm-kernel/20240117160528.GA3398@willie-the-truck/

... some compilers struggle with a plain "o" constraint, so it's
preferable to use "Qo", where the additional "Q" constraint permits
using non-offset register addressing.

This patch modifies our IO write accessors to use "Qo" constraints,
resulting in the better code generation described above. The IO read
accessors are left as-is because ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE
requires that non-offset register addressing is used, as the LDAR
instruction does not support offset addressing.

When compiling v6.8-rc1 defconfig with GCC 13.2.0, this saves ~4KiB of
text:

| [mark@lakrids:~/src/linux]% ls -al vmlinux-*
| -rwxr-xr-x 1 mark mark 153960576 Jan 23 12:01 vmlinux-after
| -rwxr-xr-x 1 mark mark 153862192 Jan 23 11:57 vmlinux-before
|
| [mark@lakrids:~/src/linux]% size vmlinux-before vmlinux-after
|    text    data     bss     dec     hex filename
26708921        16690350         622736 44022007        29fb8f7 vmlinux-before
26704761        16690414         622736 44017911        29fa8f7 vmlinux-after

... though due to internal alignment of sections, this has no impact on
the size of the resulting Image:

| [mark@lakrids:~/src/linux]% ls -al Image-*
| -rw-r--r-- 1 mark mark 43590144 Jan 23 12:01 Image-after
| -rw-r--r-- 1 mark mark 43590144 Jan 23 11:57 Image-before

Aside from the better code generation, there should be no functional
change as a result of this patch. I have lightly tested this patch,
including booting under KVM (where some devices such as PL011 are
emulated).

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Will Deacon <will@kernel.org>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Acked-by: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20240124111259.874975-1-mark.rutland@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: errata: Don't enable workarounds for "rare" errata by default
Will Deacon [Fri, 9 Feb 2024 18:39:16 +0000 (18:39 +0000)]
arm64: errata: Don't enable workarounds for "rare" errata by default

Arm classifies some of its CPU errata as "rare", indicating that the
hardware error is unlikely to occur in practice. Given that the cost of
errata workarounds can often be significant in terms of power and
performance, don't enable workarounds for "rare" errata by default and
update our documentation to reflect that.

Cc: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will@kernel.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/r/20240209183916.25860-1-will@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Use Signed/Unsigned enums for TGRAN{4,16,64} and VARange
Marc Zyngier [Mon, 19 Feb 2024 15:13:22 +0000 (15:13 +0000)]
arm64: Use Signed/Unsigned enums for TGRAN{4,16,64} and VARange

Open-coding the feature matching parameters for LVA/LVA2 leads to
issues with upcoming changes to the cpufeature code.

By making TGRAN{4,16,64} and VARange signed/unsigned as per the
architecture, we can use the existing macros, making the feature
match robust against those changes.

Signed-off-by: Marc Zyngier <maz@kernel.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Tested-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Make PUD folding check in set_pud() a runtime check
Ard Biesheuvel [Fri, 16 Feb 2024 23:59:44 +0000 (00:59 +0100)]
arm64: mm: Make PUD folding check in set_pud() a runtime check

When set_pud() is called on a 4-level paging build config that runs with
3 levels at runtime (which happens with 16k page size builds with
support for LPA2), the updated entry is in fact a PGD in
swapper_pg_dir[], and this is mapped read-only after boot.

So in this case, the existing check needs to be performed as well, even
though __PAGETABLE_PUD_FOLDED is not #define'd. So replace the #ifdef
with a call to pgtable_l4_enabled().

Cc: Will Deacon <will@kernel.org>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240216235944.3677178-2-ardb+git@google.com
Reviewed-by: Itaru Kitayama <itaru.kitayama@fujitsu.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: add support for WXN memory translation attribute
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:29 +0000 (13:29 +0100)]
arm64: mm: add support for WXN memory translation attribute

The AArch64 virtual memory system supports a global WXN control, which
can be enabled to make all writable mappings implicitly no-exec. This is
a useful hardening feature, as it prevents mistakes in managing page
table permissions from being exploited to attack the system.

When enabled at EL1, the restrictions apply to both EL1 and EL0. EL1 is
completely under our control, and has been cleaned up to allow WXN to be
enabled from boot onwards. EL0 is not under our control, but given that
widely deployed security features such as selinux or PaX already limit
the ability of user space to create mappings that are writable and
executable at the same time, the impact of enabling this for EL0 is
expected to be limited. (For this reason, common user space libraries
that have a legitimate need for manipulating executable code already
carry fallbacks such as [0].)

If enabled at compile time, the feature can still be disabled at boot if
needed, by passing arm64.nowxn on the kernel command line.

[0] https://github.com/libffi/libffi/blob/master/src/closures.c#L440

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Kees Cook <keescook@chromium.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-88-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agomm: add arch hook to validate mmap() prot flags
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:28 +0000 (13:29 +0100)]
mm: add arch hook to validate mmap() prot flags

Add a hook to permit architectures to perform validation on the prot
flags passed to mmap(), like arch_validate_prot() does for mprotect().
This will be used by arm64 to reject PROT_WRITE+PROT_EXEC mappings on
configurations that run with WXN enabled.

Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-87-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: defconfig: Enable LPA2 support
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:27 +0000 (13:29 +0100)]
arm64: defconfig: Enable LPA2 support

We typically enable support in defconfig for all architectural features
for which we can detect at runtime if the hardware actually supports
them.

Now that we have implemented support for LPA2 based 52-bit virtual
addressing in a way that should not impact 48-bit operation on non-LPA2
CPU, we can do the same, and enable 52-bit virtual addressing by
default.

Catalin adds:

  Currently the "Virtual address space size" arch/arm64/Kconfig menu
  entry sets different defaults for each page size. However, all are
  overridden by the defconfig to 48 bits. Set the new default in
  Kconfig and remove the defconfig line.

[ardb: squash follow-up fix from Catalin]

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-86-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Enable 52-bit virtual addressing for 4k and 16k granule configs
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:26 +0000 (13:29 +0100)]
arm64: Enable 52-bit virtual addressing for 4k and 16k granule configs

Update Kconfig to permit 4k and 16k granule configurations to be built
with 52-bit virtual addressing, now that all the prerequisites are in
place.

While at it, update the feature description so it matches on the
appropriate feature bits depending on the page size. For simplicity,
let's just keep ARM64_HAS_VA52 as the feature name.

Note that LPA2 based 52-bit virtual addressing requires 52-bit physical
addressing support to be enabled as well, as programming TCR.TxSZ to
values below 16 is not allowed unless TCR.DS is set, which is what
activates the 52-bit physical addressing support.

While supporting the converse (52-bit physical addressing without 52-bit
virtual addressing) would be possible in principle, let's keep things
simple, by only allowing these features to be enabled at the same time.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-85-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: kvm: avoid CONFIG_PGTABLE_LEVELS for runtime levels
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:25 +0000 (13:29 +0100)]
arm64: kvm: avoid CONFIG_PGTABLE_LEVELS for runtime levels

get_user_mapping_size() uses vabits_actual and CONFIG_PGTABLE_LEVELS to
provide the starting point for a table walk. This is fine for LVA, as
the number of translation levels is the same regardless of whether LVA
is enabled. However, with LPA2, this will no longer be the case, so
let's derive the number of levels from the number of VA bits directly.

Acked-by: Marc Zyngier <maz@kernel.org>
Acked-by: Oliver Upton <oliver.upton@linux.dev>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-84-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: ptdump: Deal with translation levels folded at runtime
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:24 +0000 (13:29 +0100)]
arm64: ptdump: Deal with translation levels folded at runtime

Currently, the ptdump code deals with folded PMD or PUD levels at build
time, by omitting those levels when invoking note_page. IOW, note_page()
is never invoked with level == 1 if P4Ds are folded in the build
configuration.

With the introduction of LPA2 support, we will defer some of these
folding decisions to runtime, so let's take care of this by overriding
the 'level' argument when this condition triggers.

Substituting the PUD or PMD strings for "PGD" when the level in question
is folded at build time is no longer necessary, and so the conditional
expressions can be simplified. This also makes the indirection of the
'name' field unnecessary, so change that into a char[] array, and make
the whole thing __ro_after_init.

Note that the mm_p?d_folded() functions currently ignore their mm
pointer arguments, but let's wire them up correctly anyway.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-83-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: ptdump: Disregard unaddressable VA space
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:23 +0000 (13:29 +0100)]
arm64: ptdump: Disregard unaddressable VA space

Configurations built with support for 52-bit virtual addressing can also
run on CPUs that only support 48 bits of VA space, in which case only
that part of swapper_pg_dir that represents the 48-bit addressable
region is relevant, and everything else is ignored by the hardware.

Our software pagetable walker has little in the way of input address
validation, and so it will happily start a walk from an address that is
not representable by the number of paging levels that are actually
active, resulting in lots of bogus output from the page table dumper
unless we take care to start at a valid address.

So define the start address at runtime based on vabits_actual.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-82-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Add support for folding PUDs at runtime
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:22 +0000 (13:29 +0100)]
arm64: mm: Add support for folding PUDs at runtime

In order to support LPA2 on 16k pages in a way that permits non-LPA2
systems to run the same kernel image, we have to be able to fall back to
at most 48 bits of virtual addressing.

Falling back to 48 bits would result in a level 0 with only 2 entries,
which is suboptimal in terms of TLB utilization. So instead, let's fall
back to 47 bits in that case. This means we need to be able to fold PUDs
dynamically, similar to how we fold P4Ds for 48 bit virtual addressing
on LPA2 with 4k pages.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-81-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: kasan: Reduce minimum shadow alignment and enable 5 level paging
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:21 +0000 (13:29 +0100)]
arm64: kasan: Reduce minimum shadow alignment and enable 5 level paging

Allow the KASAN init code to deal with 5 levels of paging, and relax the
requirement that the shadow region is aligned to the top level pgd_t
size. This is necessary for LPA2 based 52-bit virtual addressing, where
the KASAN shadow will never be aligned to the pgd_t size. Allowing this
also enables the 16k/48-bit case for KASAN, which is a nice bonus.

This involves some hackery to manipulate the root and next level page
tables without having to distinguish all the various configurations,
including 16k/48-bits (which has a two entry pgd_t level), and LPA2
configurations running with one translation level less on non-LPA2
hardware.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-80-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Add 5 level paging support to fixmap and swapper handling
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:20 +0000 (13:29 +0100)]
arm64: mm: Add 5 level paging support to fixmap and swapper handling

Add support for using 5 levels of paging in the fixmap, as well as in
the kernel page table handling code which uses fixmaps internally.
This also handles the case where a 5 level build runs on hardware that
only supports 4 levels of paging.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-79-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Enable LPA2 at boot if supported by the system
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:19 +0000 (13:29 +0100)]
arm64: Enable LPA2 at boot if supported by the system

Update the early kernel mapping code to take 52-bit virtual addressing
into account based on the LPA2 feature. This is a bit more involved than
LVA (which is supported with 64k pages only), given that some page table
descriptor bits change meaning in this case.

To keep the handling in asm to a minimum, the initial ID map is still
created with 48-bit virtual addressing, which implies that the kernel
image must be loaded into 48-bit addressable physical memory. This is
currently required by the boot protocol, even though we happen to
support placement outside of that for LVA/64k based configurations.

Enabling LPA2 involves more than setting TCR.T1SZ to a lower value,
there is also a DS bit in TCR that needs to be set, and which changes
the meaning of bits [9:8] in all page table descriptors. Since we cannot
enable DS and every live page table descriptor at the same time, let's
pivot through another temporary mapping. This avoids the need to
reintroduce manipulations of the page tables with the MMU and caches
disabled.

To permit the LPA2 feature to be overridden on the kernel command line,
which may be necessary to work around silicon errata, or to deal with
mismatched features on heterogeneous SoC designs, test for CPU feature
overrides first, and only then enable LPA2.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-78-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: add LPA2 and 5 level paging support to G-to-nG conversion
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:18 +0000 (13:29 +0100)]
arm64: mm: add LPA2 and 5 level paging support to G-to-nG conversion

Add support for 5 level paging in the G-to-nG routine that creates its
own temporary page tables to traverse the swapper page tables. Also add
support for running the 5 level configuration with the top level folded
at runtime, to support CPUs that do not implement the LPA2 extension.

While at it, wire up the level skipping logic so it will also trigger on
4 level configurations with LPA2 enabled at build time but not active at
runtime, as we'll fall back to 3 level paging in that case.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-77-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Add definitions to support 5 levels of paging
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:17 +0000 (13:29 +0100)]
arm64: mm: Add definitions to support 5 levels of paging

Add the required types and descriptor accessors to support 5 levels of
paging in the common code. This is one of the prerequisites for
supporting 52-bit virtual addressing with 4k pages.

Note that this does not cover the code that handles kernel mappings or
the fixmap.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-76-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Add LPA2 support to phys<->pte conversion routines
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:16 +0000 (13:29 +0100)]
arm64: mm: Add LPA2 support to phys<->pte conversion routines

In preparation for enabling LPA2 support, introduce the mask values for
converting between physical addresses and their representations in a
page table descriptor.

While at it, move the pte_to_phys asm macro into its only user, so that
we can freely modify it to use its input value register as a temp
register.

For LPA2, the PTE_ADDR_MASK contains two non-adjacent sequences of zero
bits, which means it no longer fits into the immediate field of an
ordinary ALU instruction. So let's redefine it to include the bits in
between as well, and only use it when converting from physical address
to PTE representation, where the distinction does not matter. Also
update the name accordingly to emphasize this.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-75-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Wire up TCR.DS bit to PTE shareability fields
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:15 +0000 (13:29 +0100)]
arm64: mm: Wire up TCR.DS bit to PTE shareability fields

When LPA2 is enabled, bits 8 and 9 of page and block descriptors become
part of the output address instead of carrying shareability attributes
for the region in question.

So avoid setting these bits if TCR.DS == 1, which means LPA2 is enabled.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-74-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Add ESR decoding for exceptions involving translation level -1
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:14 +0000 (13:29 +0100)]
arm64: Add ESR decoding for exceptions involving translation level -1

The LPA2 feature introduces new FSC values to report abort exceptions
related to translation level -1. Define these and wire them up.

Reuse the new ESR FSC classification helpers that arrived via the KVM
arm64 tree, and update the one for translation faults to check
specifically for a translation fault at level -1. (Access flag or
permission faults cannot occur at level -1 because they alway involve a
descriptor at the superior level so changing those helpers is not
needed).

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-73-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Avoid #define'ing PTE_MAYBE_NG to 0x0 for asm use
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:13 +0000 (13:29 +0100)]
arm64: Avoid #define'ing PTE_MAYBE_NG to 0x0 for asm use

The PROT_* macros resolve to expressions that are only valid in C and
not in assembler, and so they are only usable from C code. Currently, we
make an exception for the permission indirection init code in proc.S,
which doesn't care about the bits that are conditionally set, and so we
just #define PTE_MAYBE_NG to 0x0 for any assembler file that includes
these definitions.

This is dodgy because this means that PROT_NORMAL and friends is
generally available in asm code, but defined in a way that deviates from
the definition that C code will observe, which might lead to hard to
diagnose issues down the road.

So instead, #define PTE_MAYBE_NG only in the place where the PIE
constants are evaluated, and #undef it again right after. This allows us
to drop the #define from pgtable-prot.h, and avoid the risk of deviating
definitions between asm and C.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-72-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Add feature override support for LVA
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:12 +0000 (13:29 +0100)]
arm64: mm: Add feature override support for LVA

Add support for overriding the VARange field of the MMFR2 CPU ID
register. This permits the associated LVA feature to be overridden early
enough for the boot code that creates the kernel mapping to take it into
account.

Given that LPA2 implies LVA, disabling the latter should disable the
former as well. So override the ID_AA64MMFR0.TGran field of the current
page size as well if it advertises support for 52-bit addressing.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-71-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Handle LVA support as a CPU feature
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:11 +0000 (13:29 +0100)]
arm64: mm: Handle LVA support as a CPU feature

Currently, we detect CPU support for 52-bit virtual addressing (LVA)
extremely early, before creating the kernel page tables or enabling the
MMU. We cannot override the feature this early, and so large virtual
addressing is always enabled on CPUs that implement support for it if
the software support for it was enabled at build time. It also means we
rely on non-trivial code in asm to deal with this feature.

Given that both the ID map and the TTBR1 mapping of the kernel image are
guaranteed to be 48-bit addressable, it is not actually necessary to
enable support this early, and instead, we can model it as a CPU
feature. That way, we can rely on code patching to get the correct
TCR.T1SZ values programmed on secondary boot and resume from suspend.

On the primary boot path, we simply enable the MMU with 48-bit virtual
addressing initially, and update TCR.T1SZ if LVA is supported from C
code, right before creating the kernel mapping. Given that TTBR1 still
points to reserved_pg_dir at this point, updating TCR.T1SZ should be
safe without the need for explicit TLB maintenance.

Since this gets rid of all accesses to the vabits_actual variable from
asm code that occurred before TCR.T1SZ had been programmed, we no longer
have a need for this variable, and we can replace it with a C expression
that produces the correct value directly, based on the value of TCR.T1SZ.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-70-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Revert "mm: provide idmap pointer to cpu_replace_ttbr1()"
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:10 +0000 (13:29 +0100)]
arm64: Revert "mm: provide idmap pointer to cpu_replace_ttbr1()"

This reverts commit 1682c45b920643c, which is no longer needed now that
we create the permanent kernel mapping directly during early boot.

This is a RINO (revert in name only) given that some of the code has
moved around, but the changes are straight-forward.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-69-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: omit redundant remap of kernel image
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:09 +0000 (13:29 +0100)]
arm64: mm: omit redundant remap of kernel image

Now that the early kernel mapping is created with all the right
attributes and segment boundaries, there is no longer a need to recreate
it and switch to it. This also means we no longer have to copy the kasan
shadow or some parts of the fixmap from one set of page tables to the
other.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-68-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: avoid fixmap for early swapper_pg_dir updates
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:08 +0000 (13:29 +0100)]
arm64: mm: avoid fixmap for early swapper_pg_dir updates

Early in the boot, when .rodata is still writable, we can poke
swapper_pg_dir entries directly, and there is no need to go through the
fixmap. After a future patch, we will enter the kernel with
swapper_pg_dir already active, and early swapper_pg_dir updates for
creating the fixmap page table hierarchy itself cannot go through the
fixmap for obvious reaons. So let's keep track of whether rodata is
writable, and update the descriptor directly in that case.

As the same reasoning applies to early KASAN init, make the function
noinstr as well.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-67-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: kernel: Create initial ID map from C code
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:07 +0000 (13:29 +0100)]
arm64: kernel: Create initial ID map from C code

The asm code that creates the initial ID map is rather intricate and
hard to follow. This is problematic because it makes adding support for
things like LPA2 or WXN more difficult than necessary. Also, it is
parameterized like the rest of the MM code to run with a configurable
number of levels, which is rather pointless, given that all AArch64 CPUs
implement support for 48-bit virtual addressing, and that many systems
exist with DRAM located outside of the 39-bit addressable range, which
is the only smaller VA size that is widely used, and we need additional
tricks to make things work in that combination.

So let's bite the bullet, and rip out all the asm macros, and fiddly
code, and replace it with a C implementation based on the newly added
routines for creating the early kernel VA mappings. And while at it,
create the initial ID map based on 48-bit virtual addressing as well,
regardless of the number of configured levels for the kernel proper.

Note that this code may execute with the MMU and caches disabled, and is
therefore not permitted to make unaligned accesses. This shouldn't
generally happen in any case for the algorithm as implemented, but to be
sure, let's pass -mstrict-align to the compiler just in case.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-66-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: pgtable: Decouple PGDIR size macros from PGD/PUD/PMD levels
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:06 +0000 (13:29 +0100)]
arm64: pgtable: Decouple PGDIR size macros from PGD/PUD/PMD levels

The mapping from PGD/PUD/PMD to levels and shifts is very confusing,
given that, due to folding, the shifts may be equal for different
levels, if the macros are even #define'd to begin with.

In a subsequent patch, we will modify the ID mapping code to decouple
the number of levels from the kernel's view of how these types are
folded, so prepare for this by reformulating the macros without the use
of these types.

Instead, use SWAPPER_BLOCK_SHIFT as the base quantity, and derive it
from either PAGE_SHIFT or PMD_SHIFT, which -if defined at all- are
defined unambiguously for a given page size, regardless of the number of
configured levels.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-65-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Use 48-bit virtual addressing for the permanent ID map
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:05 +0000 (13:29 +0100)]
arm64: mm: Use 48-bit virtual addressing for the permanent ID map

Even though we support loading kernels anywhere in 48-bit addressable
physical memory, we create the ID maps based on the number of levels
that we happened to configure for the kernel VA and user VA spaces.

The reason for this is that the PGD/PUD/PMD based classification of
translation levels, along with the associated folding when the number of
levels is less than 5, does not permit creating a page table hierarchy
of a set number of levels. This means that, for instance, on 39-bit VA
kernels we need to configure an additional level above PGD level on the
fly, and 36-bit VA kernels still only support 47-bit virtual addressing
with this trick applied.

Now that we have a separate helper to populate page table hierarchies
that does not define the levels in terms of PUDS/PMDS/etc at all, let's
reuse it to create the permanent ID map with a fixed VA size of 48 bits.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-64-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: head: Move early kernel mapping routines into C code
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:04 +0000 (13:29 +0100)]
arm64: head: Move early kernel mapping routines into C code

The asm version of the kernel mapping code works fine for creating a
coarse grained identity map, but for mapping the kernel down to its
exact boundaries with the right attributes, it is not suitable. This is
why we create a preliminary RWX kernel mapping first, and then rebuild
it from scratch later on.

So let's reimplement this in C, in a way that will make it unnecessary
to create the kernel page tables yet another time in paging_init().

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-63-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mmu: Make __cpu_replace_ttbr1() out of line
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:03 +0000 (13:29 +0100)]
arm64: mmu: Make __cpu_replace_ttbr1() out of line

__cpu_replace_ttbr1() is a static inline, and so it gets instantiated
wherever it is used. This is not really necessary, as it is never called
on a hot path. It also has the unfortunate side effect that the symbol
idmap_cpu_replace_ttbr1 may never be referenced from kCFI enabled C
code, and this means the type id symbol may not exist either.  This will
result in a build error once we start referring to this symbol from asm
code as well. (Note that this problem only occurs when CnP, KAsan and
suspend/resume are all disabled in the Kconfig but that is a valid
config, if unusual).

So let's just move it out of line so all callers will share the same
implementation, which will reference idmap_cpu_replace_ttbr1
unconditionally.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-62-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: mm: Make kaslr_requires_kpti() a static inline
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:02 +0000 (13:29 +0100)]
arm64: mm: Make kaslr_requires_kpti() a static inline

In preparation for moving the first assignment of arm64_use_ng_mappings
to an earlier stage in the boot, ensure that kaslr_requires_kpti() is
accessible without relying on the core kernel's view on whether or not
KASLR is enabled. So make it a static inline, and move the
kaslr_enabled() check out of it and into the callers, one of which will
disappear in a subsequent patch.

Once/when support for the obsolete ThunderX 1 platform is dropped, this
check reduces to a E0PD feature check on the local CPU.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-61-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: head: move memstart_offset_seed handling to C code
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:01 +0000 (13:29 +0100)]
arm64: head: move memstart_offset_seed handling to C code

Now that we can set BSS variables from the early code running from the
ID map, we can set memstart_offset_seed directly from the C code that
derives the value instead of passing it back and forth between C and asm
code.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-60-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: head: allocate more pages for the kernel mapping
Ard Biesheuvel [Wed, 14 Feb 2024 12:29:00 +0000 (13:29 +0100)]
arm64: head: allocate more pages for the kernel mapping

In preparation for switching to an early kernel mapping routine that
maps each segment according to its precise boundaries, and with the
correct attributes, let's allocate some extra pages for page tables for
the 4k page size configuration. This is necessary because the start and
end of each segment may not be aligned to the block size, and so we'll
need an extra page table at each segment boundary.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-59-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Add helpers to probe local CPU for PAC and BTI support
Ard Biesheuvel [Wed, 14 Feb 2024 12:28:59 +0000 (13:28 +0100)]
arm64: Add helpers to probe local CPU for PAC and BTI support

Add some helpers that will be used by the early kernel mapping code to
check feature support on the local CPU. This permits the early kernel
mapping to be created with the right attributes, removing the need for
tearing it down and recreating it.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-58-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: idreg-override: Create a pseudo feature for rodata=off
Ard Biesheuvel [Wed, 14 Feb 2024 12:28:58 +0000 (13:28 +0100)]
arm64: idreg-override: Create a pseudo feature for rodata=off

Add rodata=off to the set of kernel command line options that is parsed
early using the CPU feature override detection code, so we can easily
refer to it when creating the kernel mapping.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-57-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: kaslr: Use feature override instead of parsing the cmdline again
Ard Biesheuvel [Wed, 14 Feb 2024 12:28:57 +0000 (13:28 +0100)]
arm64: kaslr: Use feature override instead of parsing the cmdline again

The early kaslr code open codes the detection of 'nokaslr' on the kernel
command line, and this is no longer necessary now that the feature
detection code, which also looks for the same string, executes before
this code.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-56-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: cpufeature: Add helper to test for CPU feature overrides
Ard Biesheuvel [Wed, 14 Feb 2024 12:28:56 +0000 (13:28 +0100)]
arm64: cpufeature: Add helper to test for CPU feature overrides

Add some helpers to extract and apply feature overrides to the bare
idreg values. This involves inspecting the value and mask of the
specific field that we are interested in, given that an override
value/mask pair might be invalid for one field but valid for another.

Then, wire up the new helper for the hVHE test - note that we can drop
the sysreg test here, as the override will be invalid when trying to
enable hVHE on non-VHE capable hardware.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-55-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: head: move dynamic shadow call stack patching into early C runtime
Ard Biesheuvel [Wed, 14 Feb 2024 12:28:55 +0000 (13:28 +0100)]
arm64: head: move dynamic shadow call stack patching into early C runtime

Once we update the early kernel mapping code to only map the kernel once
with the right permissions, we can no longer perform code patching via
this mapping.

So move this code to an earlier stage of the boot, right after applying
the relocations.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-54-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: head: Run feature override detection before mapping the kernel
Ard Biesheuvel [Wed, 14 Feb 2024 12:28:54 +0000 (13:28 +0100)]
arm64: head: Run feature override detection before mapping the kernel

To permit the feature overrides to be taken into account before the
KASLR init code runs and the kernel mapping is created, move the
detection code to an earlier stage in the boot.

In a subsequent patch, this will be taken advantage of by merging the
preliminary and permanent mappings of the kernel text and data into a
single one that gets created and relocated before start_kernel() is
called.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-53-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: Move feature overrides into the BSS section
Ard Biesheuvel [Wed, 14 Feb 2024 12:28:53 +0000 (13:28 +0100)]
arm64: Move feature overrides into the BSS section

In order to allow the CPU feature override detection code to run even
earlier, move the feature override global variables into BSS, which is
the only part of the static kernel image that is mapped read-write in
the initial ID map.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-52-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
13 months agoarm64: head: Clear BSS and the kernel page tables in one go
Ard Biesheuvel [Wed, 14 Feb 2024 12:28:52 +0000 (13:28 +0100)]
arm64: head: Clear BSS and the kernel page tables in one go

We will move the CPU feature overrides into BSS in a subsequent patch,
and this requires that BSS is zeroed before the feature override
detection code runs. So let's map BSS read-write in the ID map, and zero
it via this mapping.

Since the kernel page tables are right next to it, and also zeroed via
the ID map, let's drop the separate clear_page_tables() function, and
just zero everything in one go.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20240214122845.2033971-51-ardb+git@google.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>