Currently fp-stress only covers userspace use of floating point, it does
not cover any kernel mode uses. Since currently kernel mode floating
point usage can't be preempted and there are explicit preemption points in
the existing implementations this isn't so important for fp-stress but
when we readd preemption it will be good to try to exercise it.
When the arm64 accelerated crypto operations are implemented we can
relatively straightforwardly trigger kernel mode floating point usage by
using the crypto userspace API to hash data, using the splice() support
in an effort to minimise copying. We use /proc/crypto to check which
accelerated implementations are available, picking the first symmetric
hash we find. We run the kernel mode test unconditionally, replacing the
second copy of the FPSIMD testcase for systems with FPSIMD only. If we
don't think there are any suitable kernel mode implementations we fall back
to running another copy of fpsimd-stress.
There are a number issues with this approach, we don't actually verify
that we are using an accelerated (or even CPU) implementation of the
algorithm being tested and even with attempting to use splice() to
minimise copying there are sizing limits on how much data gets spliced
at once.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20240521-arm64-fp-stress-kernel-v1-1-e38f107baad4@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
* arm64/for-next/perf:
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
drivers/perf: fsl_imx8_ddr_perf: Remove set-but-not-used variable
perf: arm_spe: Support new SPEv1.2/v8.7 'not taken' event
perf: arm_spe: Use new PMSIDR_EL1 register enums
perf: arm_spe: Drop BIT() and use FIELD_GET/PREP accessors
arm64/sysreg: Convert SPE registers to automatic generation
arm64: Drop SYS_ from SPE register defines
perf: arm_spe: Use feature numbering for PMSEVFR_EL1 defines
perf/marvell: Add ACPI support to TAD uncore driver
perf/marvell: Add ACPI support to DDR uncore driver
perf/arm-cmn: Reset DTM_PMU_CONFIG at probe
drivers/perf: hisi: Extract initialization of "cpa_pmu->pmu"
drivers/perf: hisi: Simplify the parameters of hisi_pmu_init()
drivers/perf: hisi: Advertise the PERF_PMU_CAP_NO_EXCLUDE capability
* for-next/sysreg:
: arm64 sysreg and cpufeature fixes/updates
KVM: arm64: Use symbolic definition for ISR_EL1.A
arm64/sysreg: Add definition of ISR_EL1
arm64/sysreg: Add definition for ICC_NMIAR1_EL1
arm64/cpufeature: Remove 4 bit assumption in ARM64_FEATURE_MASK()
arm64/sysreg: Fix errors in 32 bit enumeration values
arm64/cpufeature: Fix field sign for DIT hwcap detection
* for-next/sme:
: SME-related updates
arm64/sme: Optimise SME exit on syscall entry
arm64/sme: Don't use streaming mode to probe the maximum SME VL
arm64/ptrace: Use system_supports_tpidr2() to check for TPIDR2 support
* for-next/kselftest: (23 commits)
: arm64 kselftest fixes and improvements
kselftest/arm64: Don't require FA64 for streaming SVE+ZA tests
kselftest/arm64: Copy whole EXTRA context
kselftest/arm64: Fix enumeration of systems without 128 bit SME for SSVE+ZA
kselftest/arm64: Fix enumeration of systems without 128 bit SME
kselftest/arm64: Don't require FA64 for streaming SVE tests
kselftest/arm64: Limit the maximum VL we try to set via ptrace
kselftest/arm64: Correct buffer size for SME ZA storage
kselftest/arm64: Remove the local NUM_VL definition
kselftest/arm64: Verify simultaneous SSVE and ZA context generation
kselftest/arm64: Verify that SSVE signal context has SVE_SIG_FLAG_SM set
kselftest/arm64: Remove spurious comment from MTE test Makefile
kselftest/arm64: Support build of MTE tests with clang
kselftest/arm64: Initialise current at build time in signal tests
kselftest/arm64: Don't pass headers to the compiler as source
kselftest/arm64: Remove redundant _start labels from FP tests
kselftest/arm64: Fix .pushsection for strings in FP tests
kselftest/arm64: Run BTI selftests on systems without BTI
kselftest/arm64: Fix test numbering when skipping tests
kselftest/arm64: Skip non-power of 2 SVE vector lengths in fp-stress
kselftest/arm64: Only enumerate power of two VLs in syscall-abi
...
* for-next/misc:
: Miscellaneous arm64 updates
arm64/mm: Intercept pfn changes in set_pte_at()
Documentation: arm64: correct spelling
arm64: traps: attempt to dump all instructions
arm64: Apply dynamic shadow call stack patching in two passes
arm64: el2_setup.h: fix spelling typo in comments
arm64: Kconfig: fix spelling
arm64: cpufeature: Use kstrtobool() instead of strtobool()
arm64: Avoid repeated AA64MMFR1_EL1 register read on pagefault path
arm64: make ARCH_FORCE_MAX_ORDER selectable
* for-next/sme2: (23 commits)
: Support for arm64 SME 2 and 2.1
arm64/sme: Fix __finalise_el2 SMEver check
kselftest/arm64: Remove redundant _start labels from zt-test
kselftest/arm64: Add coverage of SME 2 and 2.1 hwcaps
kselftest/arm64: Add coverage of the ZT ptrace regset
kselftest/arm64: Add SME2 coverage to syscall-abi
kselftest/arm64: Add test coverage for ZT register signal frames
kselftest/arm64: Teach the generic signal context validation about ZT
kselftest/arm64: Enumerate SME2 in the signal test utility code
kselftest/arm64: Cover ZT in the FP stress test
kselftest/arm64: Add a stress test program for ZT0
arm64/sme: Add hwcaps for SME 2 and 2.1 features
arm64/sme: Implement ZT0 ptrace support
arm64/sme: Implement signal handling for ZT
arm64/sme: Implement context switching for ZT0
arm64/sme: Provide storage for ZT0
arm64/sme: Add basic enumeration for SME2
arm64/sme: Enable host kernel to access ZT0
arm64/sme: Manually encode ZT0 load and store instructions
arm64/esr: Document ISS for ZT0 being disabled
arm64/sme: Document SME 2 and SME 2.1 ABI
...
* for-next/tpidr2:
: Include TPIDR2 in the signal context
kselftest/arm64: Add test case for TPIDR2 signal frame records
kselftest/arm64: Add TPIDR2 to the set of known signal context records
arm64/signal: Include TPIDR2 in the signal context
arm64/sme: Document ABI for TPIDR2 signal information
* for-next/scs:
: arm64: harden shadow call stack pointer handling
arm64: Stash shadow stack pointer in the task struct on interrupt
arm64: Always load shadow stack pointer directly from the task struct
* for-next/compat-hwcap:
: arm64: Expose compat ARMv8 AArch32 features (HWCAPs)
arm64: Add compat hwcap SSBS
arm64: Add compat hwcap SB
arm64: Add compat hwcap I8MM
arm64: Add compat hwcap ASIMDBF16
arm64: Add compat hwcap ASIMDFHM
arm64: Add compat hwcap ASIMDDP
arm64: Add compat hwcap FPHP and ASIMDHP
* for-next/ftrace:
: Add arm64 support for DYNAMICE_FTRACE_WITH_CALL_OPS
arm64: avoid executing padding bytes during kexec / hibernation
arm64: Implement HAVE_DYNAMIC_FTRACE_WITH_CALL_OPS
arm64: ftrace: Update stale comment
arm64: patching: Add aarch64_insn_write_literal_u64()
arm64: insn: Add helpers for BTI
arm64: Extend support for CONFIG_FUNCTION_ALIGNMENT
ACPI: Don't build ACPICA with '-Os'
Compiler attributes: GCC cold function alignment workarounds
ftrace: Add DYNAMIC_FTRACE_WITH_CALL_OPS
* for-next/efi-boot-mmu-on:
: Permit arm64 EFI boot with MMU and caches on
arm64: kprobes: Drop ID map text from kprobes blacklist
arm64: head: Switch endianness before populating the ID map
efi: arm64: enter with MMU and caches enabled
arm64: head: Clean the ID map and the HYP text to the PoC if needed
arm64: head: avoid cache invalidation when entering with the MMU on
arm64: head: record the MMU state at primary entry
arm64: kernel: move identity map out of .text mapping
arm64: head: Move all finalise_el2 calls to after __enable_mmu
* for-next/ptrauth:
: arm64 pointer authentication cleanup
arm64: pauth: don't sign leaf functions
arm64: unify asm-arch manipulation
* for-next/pseudo-nmi:
: Pseudo-NMI code generation optimisations
arm64: irqflags: use alternative branches for pseudo-NMI logic
arm64: add ARM64_HAS_GIC_PRIO_RELAXED_SYNC cpucap
arm64: make ARM64_HAS_GIC_PRIO_MASKING depend on ARM64_HAS_GIC_CPUIF_SYSREGS
arm64: rename ARM64_HAS_IRQ_PRIO_MASKING to ARM64_HAS_GIC_PRIO_MASKING
arm64: rename ARM64_HAS_SYSREG_GIC_CPUIF to ARM64_HAS_GIC_CPUIF_SYSREGS
As documented in issue C215 in the known issues list for DDI0487I.a [1] Arm
will be making a retroactive change to SVE to remove the possibility of
selecting non power of two vector lengths. This has no impact on existing
physical implementations but most virtual implementations have implemented
the full range of permissible vector lengths. Given how demanding fp-stress
is for these implementations update to only attempt to enumerate the power
of two vector lengths, reducing the load created on existing virtual
implementations and only exercising the functionality that will be seen in
physical implementations.
[1] https://developer.arm.com/documentation/102105/ia-00/
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20221220-arm64-fp-stress-pow2-v1-1-d0ce756b57af@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
When everything is starting up we are likely to have a lot of child
processes producing output at once. This means that we can reduce
overhead a bit by allowing epoll_wait() to return more than one
descriptor at once, it cuts down on the number of system calls we need
to do which on virtual platforms where the syscall overhead is a bit
more noticable and we're likely to have a lot more children active can
make a small but noticable difference.
On physical platforms the relatively small number of processes being run
and vastly improved speeds push the effects of this change into the
noise.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20221129215926.442895-4-broonie@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
Now we hold execution of the stress test programs until all children are
started there is no need to drain output while that is happening.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20221129215926.442895-3-broonie@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
At present fp-stress has a bit of a thundering herd problem since the
children it spawns start running immediately, meaning that they can start
starving the parent process of CPU before it has even started all the
children. This is much more severe on virtual platforms since they tend to
support far more SVE and SME vector lengths, be slower in general and for
some have issues with performance when simulating multiple CPUs.
We can mitigate this problem by having all the child processes block before
starting the test program, meaning that we at least have all the child
processes started before we start heavily using CPU. We still have the same
load issues while waiting for the actual stress test programs to start up
and produce output but they're at least all ready to go before that kicks
in, resulting in substantial reductions in overall runtime on some of the
severely affected systems. One test was showing about 20% improvement.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20221129215926.442895-2-broonie@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
Since we now flush output immediately on starting children we should ensure
that the child name is set beforehand so that any output that does get
flushed from the newly created child has the name of the child attached.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20221124120722.150988-1-broonie@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
Especially when the test is configured to run for a longer time it can be
reassuring to users to see that the supervising program is running OK so
provide a message every second when the output timer expires.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20221017144553.773176-3-broonie@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
Currently we don't have an explicit check that when it's been a second
since we have seen output produced from the test programs starting up that
means all of them are running and we should start both sending signals and
timing out. This is not reliable, especially on very heavily loaded systems
where the test programs might take longer than a second to run.
We do skip sending signals to children that have not produced output yet
so we won't cause them to exit unexpectedly by sending a signal but this
can create confusion when interpreting output, for example appearing to
show the tests running for less time than expected or appearing to show
missed signal deliveries. Avoid issues by explicitly checking that we have
seen output from all the child processes before we start sending signals
or counting test run time.
This is especially likely on virtual platforms with large numbers of vector
lengths supported since the platforms are slow and there will be a lot of
tasks per CPU.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20221017144553.773176-2-broonie@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
Currently we treat any error when reading from the child as a failure and
don't read any more output from that child as a result. This ignores the
fact that it is valid for read() to return EINTR as the error code if there
is a signal pending so we could stop handling the output of children,
especially during exit when we will get some SIGCHLD signals delivered to
us. Fix this by pulling the read handling out into a separate function
which returns a flag if reads should be continued and wrapping it in a
loop.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20220921181345.618085-4-broonie@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Once we have started exiting the termination handler will have the same
effect as what we're already running so set the termination flag at that
point.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20220921181345.618085-3-broonie@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
When fp-stress gets a termination signal it sets a flag telling itself to
exit and sends a termination signal to all the children. If the flag is set
then don't bother repeating this process, it isn't going to accomplish
anything other than consume CPU time which can be an issue when running in
emulation.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20220921181345.618085-2-broonie@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Currently the stress test programs for floating point context switching are
run by hand, there are extremely simplistic harnesses which run some copies
of each test individually but they are not integrated into kselftest and
with SVE and SME they only run with whatever vector length the process has
by default. This is hassle when running the tests and means that they're
not being run at all by CI systems picking up kselftest.
In order to improve our coverage and provide a more convenient interface
provide a harness program which starts enough stress test programs up to
cause context switching and runs them for a set period. If only FPSIMD is
available in the system we start two copies of the FPSIMD stress test per
CPU, otherwise we start one copy of the FPSIMD and then start the SVE,
streaming SVE and ZA tests once per CPU for each available VL they have
to run on. We then run for a set period monitoring for any errors
reported by the test programs before cleanly terminating them.
In order to provide additional coverage of signal handling and some extra
noise in the scheduling we send a SIGUSR2 to the stress tests once a
second, the tests will count the number of signals they get.
Since kselftest is generally expected to run quickly we by default only run
for ten seconds. This is enough to show if there is anything cripplingly
wrong but not exactly a thorough soak test, for interactive and more
focused use a command line option -t N is provided which overrides the
length of time to run for (specified in seconds) and if 0 is specified then
there is no timeout and the test must be manually terminated. The timeout
is counted in seconds with no output, this is done to account for the
potentially slow startup time for the test programs on virtual platforms
which tend to struggle during startup as they are both slow and tend to
support a wide range of vector lengths.
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20220829154452.824870-5-broonie@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
