Previously, wp_config0/2 registers were used for primary match group and
wp_config1/3 registers for secondary match group. In order to support
tertiary match group, this patch decouples the registers and the groups.
Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Link: https://lore.kernel.org/r/20240618005056.3092866-2-ilkka@os.amperecomputing.com
Signed-off-by: Will Deacon <will@kernel.org>
Now that perf supports giving the PMU device a parent, we can use our
platform device to make the relationship between CMN instances and PMU
IDs trivially discoverable, from either nominal direction:
root@crazy-taxi:~# ls /sys/devices/platform/ARMHC600:00 | grep cmn
arm_cmn_0
root@crazy-taxi:~# realpath /sys/bus/event_source/devices/arm_cmn_0/..
/sys/devices/platform/ARMHC600:00
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/25d4428df1ddad966c74a3ed60171cd3ca6c8b66.1712682917.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
In general it's preferable to avoid placing cpumasks on the stack, as
for large values of NR_CPUS these can consume significant amounts of
stack space and make stack overflows more likely.
Use cpumask_any_and_but() to avoid the need for a temporary cpumask on
the stack.
Suggested-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Dawei Li <dawei.li@shingroup.cn>
Link: https://lore.kernel.org/r/20240403155950.2068109-4-dawei.li@shingroup.cn
Signed-off-by: Will Deacon <will@kernel.org>
* Reorganise the arm64 kernel VA space and add support for LPA2 (at
stage 1, KVM stage 2 was merged earlier) - 52-bit VA/PA address range
with 4KB and 16KB pages
* Enable Rust on arm64
* Support for the 2023 dpISA extensions (data processing ISA), host only
* arm64 perf updates:
- StarFive's StarLink (integrates one or more CPU cores with a shared
L3 memory system) PMU support
- Enable HiSilicon Erratum 162700402 quirk for HIP09
- Several updates for the HiSilicon PCIe PMU driver
- Arm CoreSight PMU support
- Convert all drivers under drivers/perf/ to use .remove_new()
* Miscellaneous:
- Don't enable workarounds for "rare" errata by default
- Clean up the DAIF flags handling for EL0 returns (in preparation for
NMI support)
- Kselftest update for ptrace()
- Update some of the sysreg field definitions
- Slight improvement in the code generation for inline asm I/O
accessors to permit offset addressing
- kretprobes: acquire regs via a BRK exception (previously done via a
trampoline handler)
- SVE/SME cleanups, comment updates
- Allow CALL_OPS+CC_OPTIMIZE_FOR_SIZE with clang (previously disabled
due to gcc silently ignoring -falign-functions=N)
-----BEGIN PGP SIGNATURE-----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=KTD/
-----END PGP SIGNATURE-----
Merge tag 'arm64-upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux
Pull arm64 updates from Catalin Marinas:
"The major features are support for LPA2 (52-bit VA/PA with 4K and 16K
pages), the dpISA extension and Rust enabled on arm64. The changes are
mostly contained within the usual arch/arm64/, drivers/perf, the arm64
Documentation and kselftests. The exception is the Rust support which
touches some generic build files.
Summary:
- Reorganise the arm64 kernel VA space and add support for LPA2 (at
stage 1, KVM stage 2 was merged earlier) - 52-bit VA/PA address
range with 4KB and 16KB pages
- Enable Rust on arm64
- Support for the 2023 dpISA extensions (data processing ISA), host
only
- arm64 perf updates:
- StarFive's StarLink (integrates one or more CPU cores with a
shared L3 memory system) PMU support
- Enable HiSilicon Erratum 162700402 quirk for HIP09
- Several updates for the HiSilicon PCIe PMU driver
- Arm CoreSight PMU support
- Convert all drivers under drivers/perf/ to use .remove_new()
- Miscellaneous:
- Don't enable workarounds for "rare" errata by default
- Clean up the DAIF flags handling for EL0 returns (in preparation
for NMI support)
- Kselftest update for ptrace()
- Update some of the sysreg field definitions
- Slight improvement in the code generation for inline asm I/O
accessors to permit offset addressing
- kretprobes: acquire regs via a BRK exception (previously done
via a trampoline handler)
- SVE/SME cleanups, comment updates
- Allow CALL_OPS+CC_OPTIMIZE_FOR_SIZE with clang (previously
disabled due to gcc silently ignoring -falign-functions=N)"
* tag 'arm64-upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux: (134 commits)
Revert "mm: add arch hook to validate mmap() prot flags"
Revert "arm64: mm: add support for WXN memory translation attribute"
Revert "ARM64: Dynamically allocate cpumasks and increase supported CPUs to 512"
ARM64: Dynamically allocate cpumasks and increase supported CPUs to 512
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
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
...
AmpereOneX mesh implementation has a bug in HN-P nodes that makes them
report incorrect child count. The failing crosspoints report 8 children
while they only have two.
When the driver tries to access the inexistent child nodes, it believes it
has reached an invalid node type and probing fails. The workaround is to
ignore those incorrect child nodes and continue normally.
Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
[ rm: rewrote simpler generalised version ]
Tested-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/ce4b1442135fe03d0de41859b04b268c88c854a3.1707498577.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
The debugfs pretty-printer was written for the CMN-600 assumptions of a
maximum 8x8 mesh, but CMN-700 now allows coordinates and ID values up to
12 and 128 respectively, which can overflow the format strings, mess up
the alignment of the table and hurt overall readability. This table does
prove useful for double-checking that the driver is picking up the
topology of new systems correctly and for verifying user expectations,
so tweak the formatting to stay nice and readable with wider values.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/1d1517eadd1bac5992fab679c9dc531b381944da.1702484646.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
The .remove() callback for a platform driver returns an int which makes
many driver authors wrongly assume it's possible to do error handling by
returning an error code. However the value returned is ignored (apart
from emitting a warning) and this typically results in resource leaks.
To improve here there is a quest to make the remove callback return
void. In the first step of this quest all drivers are converted to
.remove_new(), which already returns void. Eventually after all drivers
are converted, .remove_new() will be renamed to .remove().
Trivially convert this driver from always returning zero in the remove
callback to the void returning variant.
Signed-off-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>
Link: https://lore.kernel.org/r/8698ca612e17292f8a8bbb2d1c0f6be4b2053da7.1702648125.git.u.kleine-koenig@pengutronix.de
Signed-off-by: Will Deacon <will@kernel.org>
Merge in arm64 fixes queued for 6.7 so that kpti_install_ng_mappings()
can be updated to use arm64_kernel_unmapped_at_el0() instead of checking
the ARM64_UNMAP_KERNEL_AT_EL0 CPU capability directly.
* for-next/fixes:
arm64: mm: Always make sw-dirty PTEs hw-dirty in pte_modify
perf/arm-cmn: Fail DTC counter allocation correctly
arm64: Avoid enabling KPTI unnecessarily
Calling arm_cmn_event_clear() before all DTC indices are allocated is
wrong, and can lead to arm_cmn_event_add() erroneously clearing live
counters from full DTCs where allocation fails. Since the DTC counters
are only updated by arm_cmn_init_counter() after all DTC and DTM
allocations succeed, nothing actually needs cleaning up in this case
anyway, and it should just return directly as it did before.
Fixes: 7633ec2c26 ("perf/arm-cmn: Rework DTC counters (again)")
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Reviewed-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Acked-by: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/ed589c0d8e4130dc68b8ad1625226d28bdc185d4.1702322847.git.robin.murphy@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
A subtle copy-paste error managed to slip through the reorganisation
of these patches in development, and not only give some HN-F events
the wrong type, but use that wrong type before the subsequent patch
defined it. Too late to fix history, but we can at least fix the bug.
Fixes: b1b7dc38e4 ("perf/arm-cmn: Refactor HN-F event selector macros")
Reported-by: Jing Zhang <renyu.zj@linux.alibaba.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/5a22439de84ff188ef76674798052448eb03a3e1.1700740693.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
* Major refactoring of the CPU capability detection logic resulting in
the removal of the cpus_have_const_cap() function and migrating the
code to "alternative" branches where possible
* Backtrace/kgdb: use IPIs and pseudo-NMI
* Perf and PMU:
- Add support for Ampere SoC PMUs
- Multi-DTC improvements for larger CMN configurations with multiple
Debug & Trace Controllers
- Rework the Arm CoreSight PMU driver to allow separate registration of
vendor backend modules
- Fixes: add missing MODULE_DEVICE_TABLE to the amlogic perf
driver; use device_get_match_data() in the xgene driver; fix NULL
pointer dereference in the hisi driver caused by calling
cpuhp_state_remove_instance(); use-after-free in the hisi driver
* HWCAP updates:
- FEAT_SVE_B16B16 (BFloat16)
- FEAT_LRCPC3 (release consistency model)
- FEAT_LSE128 (128-bit atomic instructions)
* SVE: remove a couple of pseudo registers from the cpufeature code.
There is logic in place already to detect mismatched SVE features
* Miscellaneous:
- Reduce the default swiotlb size (currently 64MB) if no ZONE_DMA
bouncing is needed. The buffer is still required for small kmalloc()
buffers
- Fix module PLT counting with !RANDOMIZE_BASE
- Restrict CPU_BIG_ENDIAN to LLVM IAS 15.x or newer move
synchronisation code out of the set_ptes() loop
- More compact cpufeature displaying enabled cores
- Kselftest updates for the new CPU features
-----BEGIN PGP SIGNATURE-----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=QaEo
-----END PGP SIGNATURE-----
Merge tag 'arm64-upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux
Pull arm64 updates from Catalin Marinas:
"No major architecture features this time around, just some new HWCAP
definitions, support for the Ampere SoC PMUs and a few fixes/cleanups.
The bulk of the changes is reworking of the CPU capability checking
code (cpus_have_cap() etc).
- Major refactoring of the CPU capability detection logic resulting
in the removal of the cpus_have_const_cap() function and migrating
the code to "alternative" branches where possible
- Backtrace/kgdb: use IPIs and pseudo-NMI
- Perf and PMU:
- Add support for Ampere SoC PMUs
- Multi-DTC improvements for larger CMN configurations with
multiple Debug & Trace Controllers
- Rework the Arm CoreSight PMU driver to allow separate
registration of vendor backend modules
- Fixes: add missing MODULE_DEVICE_TABLE to the amlogic perf
driver; use device_get_match_data() in the xgene driver; fix
NULL pointer dereference in the hisi driver caused by calling
cpuhp_state_remove_instance(); use-after-free in the hisi driver
- HWCAP updates:
- FEAT_SVE_B16B16 (BFloat16)
- FEAT_LRCPC3 (release consistency model)
- FEAT_LSE128 (128-bit atomic instructions)
- SVE: remove a couple of pseudo registers from the cpufeature code.
There is logic in place already to detect mismatched SVE features
- Miscellaneous:
- Reduce the default swiotlb size (currently 64MB) if no ZONE_DMA
bouncing is needed. The buffer is still required for small
kmalloc() buffers
- Fix module PLT counting with !RANDOMIZE_BASE
- Restrict CPU_BIG_ENDIAN to LLVM IAS 15.x or newer move
synchronisation code out of the set_ptes() loop
- More compact cpufeature displaying enabled cores
- Kselftest updates for the new CPU features"
* tag 'arm64-upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux: (83 commits)
arm64: Restrict CPU_BIG_ENDIAN to GNU as or LLVM IAS 15.x or newer
arm64: module: Fix PLT counting when CONFIG_RANDOMIZE_BASE=n
arm64, irqchip/gic-v3, ACPI: Move MADT GICC enabled check into a helper
perf: hisi: Fix use-after-free when register pmu fails
drivers/perf: hisi_pcie: Initialize event->cpu only on success
drivers/perf: hisi_pcie: Check the type first in pmu::event_init()
arm64: cpufeature: Change DBM to display enabled cores
arm64: cpufeature: Display the set of cores with a feature
perf/arm-cmn: Enable per-DTC counter allocation
perf/arm-cmn: Rework DTC counters (again)
perf/arm-cmn: Fix DTC domain detection
drivers: perf: arm_pmuv3: Drop some unused arguments from armv8_pmu_init()
drivers: perf: arm_pmuv3: Read PMMIR_EL1 unconditionally
drivers/perf: hisi: use cpuhp_state_remove_instance_nocalls() for hisi_hns3_pmu uninit process
clocksource/drivers/arm_arch_timer: limit XGene-1 workaround
arm64: Remove system_uses_lse_atomics()
arm64: Mark the 'addr' argument to set_ptes() and __set_pte_at() as unused
drivers/perf: xgene: Use device_get_match_data()
perf/amlogic: add missing MODULE_DEVICE_TABLE
arm64/mm: Hoist synchronization out of set_ptes() loop
...
Finally enable independent per-DTC-domain counter allocation, except on
CMN-600 where we still need to cope with not knowing the domain topology
and thus keep counter indices sychronised across domains. This allows
users to simultaneously count up to 8 targeted events per domain, rather
than 8 globally, for up to 4x wider coverage on maximum configurations.
Even though this now looks deceptively simple, I stand by my previous
assertion that it was a flippin' nightmare to implement; all the real
head-scratchers are hidden in the foundations in the previous patch...
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/849f65566582cb102c6d0843d0f26e231180f8ac.1697824215.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
The bitmap-based scheme for tracking DTC counter usage turns out to be a
complete dead-end for its imagined purpose, since by the time we have to
keep track of a per-DTC counter index anyway, we already have enough
information to make the bitmap itself redundant. Revert the remains of
it back to almost the original scheme, but now expanded to track per-DTC
indices, in preparation for making use of them in anger.
Note that since cycle count events always use a dedicated counter on a
single DTC, we reuse the field to encode their DTC index directly.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Reviewed-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Link: https://lore.kernel.org/r/5f6ade76b47f033836d7a36c03555da896dfb4a3.1697824215.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
It transpires that dtm_unit_info is another register which got shuffled
in CMN-700 without me noticing. Fix that in a way which also proactively
fixes the fragile laziness of its consumer, just in case any further
fields ever get added alongside dtc_domain.
Fixes: 23760a0144 ("perf/arm-cmn: Add CMN-700 support")
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Reviewed-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Link: https://lore.kernel.org/r/3076ee83d0554f6939fbb6ee49ab2bdb28d8c7ee.1697824215.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
The register por_dt_pmovsr Bits[7:0] indicates overflow from counters 7
to 0. But in arm_cmn_handle_irq(), only handled the overflow status of
Bits[3:0] which results in unhandled overflow status of counters 4 to 7.
So let the overflow status of DTC counters 4 to 7 to be handled.
Fixes: 0ba64770a2 ("perf: Add Arm CMN-600 PMU driver")
Signed-off-by: Jing Zhang <renyu.zj@linux.alibaba.com>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/1695612152-123633-1-git-send-email-renyu.zj@linux.alibaba.com
Signed-off-by: Will Deacon <will@kernel.org>
CMN-700 r3 has a special configuration option for a so-called "Super
Home Node", which is a superset of the standard HN-F that also manages
remote-chip coherency for multi-chip setups. As such it has a similar
but expanded set of PMU events compared to HN-F, with some additional
filtering options to boot.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/49153b72253f6af0e625cb55b9e1b825b110c49c.1688746690.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
As the name suggests, the "partial DAT flit" event is only counted for
the DAT channel, and furthermore is only applicable to device ports, not
mesh links (strictly it's only device ports with CHI-A requesters
connected, but detecting that degree of detail is more bother than it's
worth). Stop generating spurious event aliases for other combinations
which aren't meaningful.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/b01a58e3ff05c322547fbfd015f6dbfedf555ed3.1688746690.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Expose a sysfs identifier encapsulating the CMN part number and revision
so that jevents can narrow down a fundamental set of possible events for
calculating metrics. Configuration-dependent aspects - such as whether a
given node type is present, and/or a given node ID is valid - are still
not covered, and in general it's hard to see how userspace could handle
them, so we won't be removing any data or validation logic from the
driver any time soon, but at least it's a step in a useful direction.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Reviewed-and-tested-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Tested-by: Jing Zhang <renyu.zj@linux.alibaba.com>
Link: https://lore.kernel.org/r/b8a14c14fcdf028939ebf57849863e8ae01743de.1686588640.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
CMN implements a set of CoreSight-format peripheral ID registers which
in principle we should be able to use to identify the hardware. However
so far we have avoided trying to use the part number field since the
TRMs have all described it as "configuration dependent". It turns out,
though, that this is a quirk of the documentation generation process,
and in fact the part number should always be a stable well-defined field
which we can trust.
To that end, revamp our model detection to rely less on ACPI/DT, and
pave the way towards further using the hardware information as an
identifier for userspace jevent metrics. This includes renaming the
revision constants to maximise readability.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Reviewed-and-tested-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Link: https://lore.kernel.org/r/3c791eaae814b0126f9adbd5419bfb4a600dade7.1686588640.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
It turns out that my naive DTC reset logic fails to work as intended,
since, after checking with the hardware designers, the PMU actually
needs to be fully enabled in order to correctly clear any pending
overflows. Therefore, invert the sequence to start with turning on both
enables so that we can reliably get the DTCs into a known state, then
moving to our normal counters-stopped state from there. Since all the
DTM counters have already been unpaired during the initial discovery
pass, we just need to additionally reset the cycle counters to ensure
that no other unexpected overflows occur during this period.
Fixes: 0ba64770a2 ("perf: Add Arm CMN-600 PMU driver")
Reported-by: Geoff Blake <blakgeof@amazon.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/0ea4559261ea394f827c9aee5168c77a60aaee03.1684946389.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
When the "extra device ports" configuration was first added, the
additional mxp_device_port_connect_info registers were added around the
existing mxp_mesh_port_connect_info registers. What I missed about
CMN-700 is that it shuffled them around to remove this discontinuity.
As such, tweak the definitions and factor out a helper for reading these
registers so we can deal with this discrepancy easily, which does at
least allow nicely tidying up the callsites. With this we can then also
do the nice thing and skip accesses completely rather than relying on
RES0 behaviour where we know the extra registers aren't defined.
Fixes: 23760a0144 ("perf/arm-cmn: Add CMN-700 support")
Reported-by: Jing Zhang <renyu.zj@linux.alibaba.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/71d129241d4d7923cde72a0e5b4c8d2f6084525f.1681295193.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
DTC cycle count events don't have anything to validate or initialise in
themselves, but we should not forget to still validate their whole group
context. Otherwise, we may fail to correctly reject a contrived group
containing an impossible number of cycles events.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/3124e8c276a1f513c1a415dc839ca4181b3c8bc8.1680522545.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
As eventid field was expanded to support new mesh versions, it started to
overlap with wp_combine field. Move wp_combine to fix the issue.
Fixes: 23760a0144 ("perf/arm-cmn: Add CMN-700 support")
Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Link: https://lore.kernel.org/r/20230301175540.19891-1-ilkka@os.amperecomputing.com
Signed-off-by: Will Deacon <will@kernel.org>
- Support for arm64 SME 2 and 2.1. SME2 introduces a new 512-bit
architectural register (ZT0, for the look-up table feature) that Linux
needs to save/restore.
- Include TPIDR2 in the signal context and add the corresponding
kselftests.
- Perf updates: Arm SPEv1.2 support, HiSilicon uncore PMU updates, ACPI
support to the Marvell DDR and TAD PMU drivers, reset DTM_PMU_CONFIG
(ARM CMN) at probe time.
- Support for DYNAMIC_FTRACE_WITH_CALL_OPS on arm64.
- Permit EFI boot with MMU and caches on. Instead of cleaning the entire
loaded kernel image to the PoC and disabling the MMU and caches before
branching to the kernel bare metal entry point, leave the MMU and
caches enabled and rely on EFI's cacheable 1:1 mapping of all of
system RAM to populate the initial page tables.
- Expose the AArch32 (compat) ELF_HWCAP features to user in an arm64
kernel (the arm32 kernel only defines the values).
- Harden the arm64 shadow call stack pointer handling: stash the shadow
stack pointer in the task struct on interrupt, load it directly from
this structure.
- Signal handling cleanups to remove redundant validation of size
information and avoid reading the same data from userspace twice.
- Refactor the hwcap macros to make use of the automatically generated
ID registers. It should make new hwcaps writing less error prone.
- Further arm64 sysreg conversion and some fixes.
- arm64 kselftest fixes and improvements.
- Pointer authentication cleanups: don't sign leaf functions, unify
asm-arch manipulation.
- Pseudo-NMI code generation optimisations.
- Minor fixes for SME and TPIDR2 handling.
- Miscellaneous updates: ARCH_FORCE_MAX_ORDER is now selectable, replace
strtobool() to kstrtobool() in the cpufeature.c code, apply dynamic
shadow call stack in two passes, intercept pfn changes in set_pte_at()
without the required break-before-make sequence, attempt to dump all
instructions on unhandled kernel faults.
-----BEGIN PGP SIGNATURE-----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=VmXL
-----END PGP SIGNATURE-----
Merge tag 'arm64-upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux
Pull arm64 updates from Catalin Marinas:
- Support for arm64 SME 2 and 2.1. SME2 introduces a new 512-bit
architectural register (ZT0, for the look-up table feature) that
Linux needs to save/restore
- Include TPIDR2 in the signal context and add the corresponding
kselftests
- Perf updates: Arm SPEv1.2 support, HiSilicon uncore PMU updates, ACPI
support to the Marvell DDR and TAD PMU drivers, reset DTM_PMU_CONFIG
(ARM CMN) at probe time
- Support for DYNAMIC_FTRACE_WITH_CALL_OPS on arm64
- Permit EFI boot with MMU and caches on. Instead of cleaning the
entire loaded kernel image to the PoC and disabling the MMU and
caches before branching to the kernel bare metal entry point, leave
the MMU and caches enabled and rely on EFI's cacheable 1:1 mapping of
all of system RAM to populate the initial page tables
- Expose the AArch32 (compat) ELF_HWCAP features to user in an arm64
kernel (the arm32 kernel only defines the values)
- Harden the arm64 shadow call stack pointer handling: stash the shadow
stack pointer in the task struct on interrupt, load it directly from
this structure
- Signal handling cleanups to remove redundant validation of size
information and avoid reading the same data from userspace twice
- Refactor the hwcap macros to make use of the automatically generated
ID registers. It should make new hwcaps writing less error prone
- Further arm64 sysreg conversion and some fixes
- arm64 kselftest fixes and improvements
- Pointer authentication cleanups: don't sign leaf functions, unify
asm-arch manipulation
- Pseudo-NMI code generation optimisations
- Minor fixes for SME and TPIDR2 handling
- Miscellaneous updates: ARCH_FORCE_MAX_ORDER is now selectable,
replace strtobool() to kstrtobool() in the cpufeature.c code, apply
dynamic shadow call stack in two passes, intercept pfn changes in
set_pte_at() without the required break-before-make sequence, attempt
to dump all instructions on unhandled kernel faults
* tag 'arm64-upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux: (130 commits)
arm64: fix .idmap.text assertion for large kernels
kselftest/arm64: Don't require FA64 for streaming SVE+ZA tests
kselftest/arm64: Copy whole EXTRA context
arm64: kprobes: Drop ID map text from kprobes blacklist
perf: arm_spe: Print the version of SPE detected
perf: arm_spe: Add support for SPEv1.2 inverted event filtering
perf: Add perf_event_attr::config3
arm64/sme: Fix __finalise_el2 SMEver check
drivers/perf: fsl_imx8_ddr_perf: Remove set-but-not-used variable
arm64/signal: Only read new data when parsing the ZT context
arm64/signal: Only read new data when parsing the ZA context
arm64/signal: Only read new data when parsing the SVE context
arm64/signal: Avoid rereading context frame sizes
arm64/signal: Make interface for restore_fpsimd_context() consistent
arm64/signal: Remove redundant size validation from parse_user_sigframe()
arm64/signal: Don't redundantly verify FPSIMD magic
arm64/cpufeature: Use helper macros to specify hwcaps
arm64/cpufeature: Always use symbolic name for feature value in hwcaps
arm64/sysreg: Initial unsigned annotations for ID registers
arm64/sysreg: Initial annotation of signed ID registers
...
It turns out the optimisation implemented by commit 4f2c3872dd is
totally broken, since all the places that consume hw->dtcs_used for
events other than cycle count are still not expecting it to be sparsely
populated, and fail to read all the relevant DTC counters correctly if
so.
If implemented correctly, the optimisation potentially saves up to 3
register reads per event update, which is reasonably significant for
events targeting a single node, but still not worth a massive amount of
additional code complexity overall. Getting it right within the current
design looks a fair bit more involved than it was ever intended to be,
so let's just make a functional revert which restores the old behaviour
while still backporting easily.
Fixes: 4f2c3872dd ("perf/arm-cmn: Optimise DTC counter accesses")
Reported-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/b41bb4ed7283c3d8400ce5cf5e6ec94915e6750f.1674498637.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Although we treat the DTM counters as free-running such that we're not
too concerned about the initial DTM state, it's possible for a previous
user to have left DTM counters enabled and paired with DTC counters.
Thus if the first events are scheduled using some, but not all, DTMs,
the as-yet-unused ones could end up adding spurious increments to the
event counts at the DTC. Make sure we sync our initial DTM_PMU_CONFIG
state to all the DTMs at probe time to avoid that possibility.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/ba5f38b3dc733cd06bfb5e659b697e76d18c2183.1670269572.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
CMN-600 uses bits [27:0] for child node address offset while bits [30:28]
are required to be zero.
For CMN-650, the child node address offset field has been increased
to include bits [29:0] while leaving only bit 30 set to zero.
Let's include the missing two bits and assume older implementations
comply with the spec and set bits [29:28] to 0.
Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Fixes: 60d1504070 ("perf/arm-cmn: Support new IP features")
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/20220808195455.79277-1-ilkka@os.amperecomputing.com
Signed-off-by: Will Deacon <will@kernel.org>
The debugfs code is lazy, and since it only keeps the bottom byte of
each connect_info register to save space, it also treats the whole thing
as the device_type since the other bits were reserved anyway. Upon
closer inspection, though, this is no longer true on newer IP versions,
so let's be good and decode the exact field properly. This should help
it not get confused when a Component Aggregation Layer is present (which
is already implied if Node IDs are found for both device addresses
represented by the next two lines of the table).
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/6a13a6128a28cfe2eec6d09cf372a167ec9c3b65.1652274773.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Carefully considering the bounds of an array is all well and good,
until you forget that that array also contains a NULL sentinel at
the end and dereference it. So close...
Reported-by: Qian Cai <quic_qiancai@quicinc.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/bebba768156aa3c0757140457bdd0fec10819388.1652217788.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Add the identifiers, events, and subtleties for CMN-700. Highlights
include yet more options for doubling up CHI channels, which finally
grows event IDs beyond 8 bits for XPs, and a new set of CML gateway
nodes adding support for CXL as well as CCIX, where the Link Agent is
now internal to the CMN mesh so we gain regular PMU events for that too.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Tested-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Link: https://lore.kernel.org/r/cf892baa0d0258ea6cd6544b15171be0069a083a.1650320598.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
So far, DNs and HN-Fs have each had one event ralated to occupancy
trackers which are filtered by a separate field. CMN-700 raises the
stakes by introducing two more sets of HN-F events with corresponding
additional filter fields. Prepare for this by refactoring our filter
selection and tracking logic to account for multiple filter types
coexisting on the same node. This need not affect the uAPI, which can
just continue to encode any per-event filter setting in the "occupid"
config field, even if it's technically not the most accurate name for
some of them.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Tested-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Link: https://lore.kernel.org/r/1aa47ba0455b144c416537f6b0e58dc93b467a00.1650320598.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Add the identifiers and events for CMN-650, which slots into its
evolutionary position between CMN-600 and the 700-series products.
Imagine CMN-600 made bigger, and with most of the rough edges smoothed
off, but that then balanced out by some bonkers PMU functionality for
the new HN-P enhancement in CMN-650r2.
Most of the CXG events are actually common to newer revisions of CMN-600
too, so they're arguably a little late; oh well.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Tested-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Link: https://lore.kernel.org/r/b0adc5824db53f71a2b561c293e2120390106536.1650320598.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
From CMN-650 onwards, some of the fields in the watchpoint config
registers moved subtly enough to easily overlook. Watchpoint events are
still only partially supported on newer IPs - which in itself deserves
noting - but were not intended to become any *less* functional than on
CMN-600.
Fixes: 60d1504070 ("perf/arm-cmn: Support new IP features")
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/e1ce4c2f1e4f73ab1c60c3a85e4037cd62dd6352.1645727871.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
In general, detailed performance analysis will require knoweldge of the
the SoC beyond the CMN itself - e.g. which actual CPUs/peripherals/etc.
are connected to each node. However for certain development and bringup
tasks it can be useful to have a quick overview of the CMN internal
topology to hand too. Add a debugfs file to map this out.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/159fd4d7e19fb3c8801a8cb64ee73ec50f55903c.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
The second generation of CMN IPs add new node types and significantly
expand the configuration space with options for extra device ports on
edge XPs, either plumbed into the regular DTM or with extra dedicated
DTMs to monitor them, plus larger (and smaller) mesh sizes. Add basic
support for pulling this new information out of the hardware, piping
it around as necessary, and handling (most of) the new choices.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/e58b495bcc7deec3882be4bac910ed0bf6979674.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
In preparation for supporting newer CMN products, let's introduce a
means to differentiate the features and events which are specific to a
particular IP from those which remain common to the whole family. The
newer designs have also smoothed off some of the rough edges in terms
of discoverability, so separate out the parts of the flow which have
effectively now become CMN-600 quirks.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/9f6368cdca4c821d801138939508a5bba54ccabb.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
With the value of CMN_MAX_DTMS increasing significantly, our validation
data structure is set to get quite big. Technically we could pack it at
least twice as densely, since we only need around 19 bits of information
per DTM, but that makes the code even more mind-bogglingly impenetrable,
and even half of "quite big" may still be uncomfortably large for a
stack frame (~1KB). Just move it to an off-stack allocation instead.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/0cabff2e5839ddc0979e757c55515966f65359e4.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
In cases where we do know which DTC domain a node belongs to, we can
skip initialising or reading the global count in DTCs where we know
it won't change. The machinery to achieve that is mostly in place
already, so finish hooking it up by converting the vestigial domain
tracking to propagate suitable bitmaps all the way through to events.
Note that this does not allow allocating such an unused counter to a
different event on that DTC, because that is a flippin' nightmare.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/51d930fd945ef51c81f5889ccca055c302b0a1d0.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
When multiple nodes of the same type are connected to the same XP
(particularly in CAL configurations), it seems that they are likely
to be consecutive in logical ID. Therefore, we're likely to gain a
small benefit from an easy tweak to optimise out consecutive reads
of the same set of DTM counters for an aggregated event.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/7777d77c2df17693cd3dabb6e268906e15238d82.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Untangle DTMs from XPs into a dedicated abstraction. This helps make
things a little more obvious and robust, but primarily paves the way
for further development where new IPs can grow extra DTMs per XP.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/9cca18b1b98f482df7f1aaf3d3213e7f39500423.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Refactor the places where we scan through the set of nodes to switch
from explicit array indexing to pointer-based iteration. This leads to
slightly simpler object code, but also makes the source less dense and
more pleasant for further development. It also unearths an almost-bug
in arm_cmn_event_init() where we've been depending on the "array index"
of NULL relative to cmn->dns being a sufficiently large number, yuck.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/ee0c9eda9a643f46001ac43aadf3f0b1fd5660dd.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Add a bit more abstraction for the places where we decompose node IDs.
This will help keep things nice and manageable when we come to add yet
more variables which affect the node ID format. Also use the opportunity
to move the rest of the low-level node management helpers back up to the
logical place they were meant to be - how they ended up buried right in
the middle of the event-related definitions is somewhat of a mystery...
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/a2242a8c3c96056c13a04ae87bf2047e5e64d2d9.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Although CMN is currently (and overwhelmingly likely to remain) deployed
in arm64-only (modulo userspace) systems, the 64-bit "dependency" for
compile-testing was just laziness due to heavy reliance on readq/writeq
accessors. Since we only need one extra include for robustness in that
regard, let's pull that in, widen the compile-test coverage, and fix up
the smattering of type laziness that that brings to light.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/baee9ee0d0bdad8aaeb70f5a4b98d8fd4b1f5786.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
On a system with multiple CMN meshes, ideally we'd want to access each
PMU from within its own mesh, rather than with a long CML round-trip,
wherever feasible. Since such a system is likely to be presented as
multiple NUMA nodes, let's also hope a proximity domain is specified
for each CMN programming interface, and use that to guide our choice
of IRQ affinity to favour a node-local CPU where possible.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/32438b0d016e0649d882d47d30ac2000484287b9.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
Attempting to migrate the PMU context after we've unregistered the PMU
device, or especially if we never successfully registered it in the
first place, is a woefully bad idea. It's also fundamentally pointless
anyway. Make sure to unregister an instance from the hotplug handler
*without* invoking the teardown callback.
Fixes: 0ba64770a2 ("perf: Add Arm CMN-600 PMU driver")
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Link: https://lore.kernel.org/r/2c221d745544774e4b07583b65b5d4d94f7e0fe4.1638530442.git.robin.murphy@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
