The series can be found at [1]. The series adds support for the KVM_ARM_VCPU_PMU_V3_SET_PMU PMU attribute, which allows userspace to set a PMU for a VCPU. This PMU is used by KVM when creating perf events to emulate the guest PMU. Without settings this attribute, the PMU used when creating events is the first one that successfully probed when booting, but this is unreliable as the probe order can change (if the order of the PMUs is changed in the DTB or if asynchronous driver probing is enabled on the host's command line), and furthermore it requires the user to have intimate knowledge of how the PMU was chosen in order to pin the VM on the correct physical CPUs. With KVM_ARM_VCPU_PMU_V3_SET_PMU, the user is still expected to pin the VCPUs on a particular set of CPUs, but now it can be any CPUs as long as they share the same PMU. The set does not depend anymore on the driver probe order and all that is necessary for the user to know is which CPUs share the same PMU, which I believe is a more reasonable expectation for the average user. Pinning the VCPU threads on the desired CPUs can be done using two methods: 1. By pinning the kvmtool process using taskset. 2. By pinning only the VCPU threads using the kvmtool command line argument --vcpu-affinity, which this series introduces. Regarding method #1, the series has the potential to break users of kvmtool on heterogeneous systems. For example, launching kvmtool like this: ./lkvm run -c2 -m512 -k <kernel> -d <disk> --pmu will exit with an error when Linux migrates a VCPU thread to a physical CPU which has a different PMU than the PMU associated with the main thread when the emulated PMU has been set for the entire VM. I believe this is an improvement over the current behaviour for two reasons: * PMU events would have stopped working anyway when the thread migration happens, but now KVM kills the VM instead of events silently stopping. This can save someone a lot of time and pain debugging the VM. * Using taskset to pin kvmtool to a set of CPUs which have a particular PMU now works. However, I understand that breaking existing scripts/use cases might not be desirable, and if requested I can change the series to set the PMU only when --vcpu-affinity is specified and preserve the current behaviour. Patches #1-#3 are fixes and can be taken independently of this series. Patches #4-#7 move the PMU code to aarch64, where it belongs, because the PMU has never been supported on KVM for arm. This also paves the way for pulling in the KVM_ARM_VCPU_PMU_V3_SET_PMU attribute, which was not defined for KVM for arm (when KVM supported arm). This also can be merged right now, independently of the other patches. Patch #8 adds the cpumask_* functions which are necessary for subsequent patches. Patch #10 adds basic support for KVM_ARM_VCPU_PMU_V3_SET_PMU; the user is still expected to use taskset to pin the entire VM to the correct CPUs. Patch #11 expands on #10 by adding the --vcpu-affinity command line argument to pin VCPUs to the correct CPUs without pinning the rest of the kvmtool threads. Changes from v2: * Patch #3 (" arm/arm64: pmu.h: Add missing header guards") is new. * Use __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ as per C preprocessor common predefined macros [2]. * Do not die if kvmtool is not able to find a PMU associated with the CPUs specified by the user with --vcpu-affinity. There are valid reasons for someone to do this (like testing what happens with the default PMU, or what happens to a guest when PMU events are not working, etc). * Updated --pmu help text to explain how the PMU is chosen. * Minor cosmetic fixes. Changes from v1: * Patch #2 ("bitops.h: Include wordsize.h to provide the __WORDSIZE define") is new. * Added for_each_cpu(), cpumask_and() and cpumask_subset() functions and all the cpumask_* functions are added in one patch. * Bumped NR_CPUS fro arm64 to 4096 to match the Linux Kconfig option. * Reworked the way kvmtool specific header files were included to use quotes to clearly differentiate them from the system level headers and to keep the style consistent with the current code (for example, #include <linux/bitops.h> is now "linux/bitops.h"). * Patch #10 ("arm64: Add --vcpu-affinity command line argument") is new. [1] https://gitlab.arm.com/linux-arm/kvmtool-ae/-/tree/pmu-big-little-fix-v3 [2] https://gcc.gnu.org/onlinedocs/cpp/Common-Predefined-Macros.html Alexandru Elisei (11): linux/err.h: Add missing stdbool.h include linux/bitops.h: Include wordsize.h to provide the __WORDSIZE define arm/arm64: pmu.h: Add missing header guards arm: Move arch specific VCPU features to the arch specific function arm: Get rid of the ARM_VCPU_FEATURE_FLAGS() macro arm: Make the PMUv3 emulation code arm64 specific arm64: Rework set_pmu_attr() Add cpumask functions update_headers.sh: Sync ABI headers with Linux v5.18-rc2 arm64: Add support for KVM_ARM_VCPU_PMU_V3_SET_PMU arm64: Add --vcpu-affinity command line argument Makefile | 8 +- arm/aarch32/include/asm/kernel.h | 8 + arm/aarch32/include/kvm/kvm-cpu-arch.h | 4 - arm/aarch64/arm-cpu.c | 3 +- arm/aarch64/include/asm/kernel.h | 8 + arm/aarch64/include/asm/kvm.h | 11 + .../arm-common => aarch64/include/asm}/pmu.h | 4 + arm/aarch64/include/kvm/kvm-config-arch.h | 9 +- arm/aarch64/include/kvm/kvm-cpu-arch.h | 6 - arm/aarch64/kvm-cpu.c | 21 ++ arm/aarch64/kvm.c | 32 +++ arm/aarch64/pmu.c | 237 ++++++++++++++++ arm/include/arm-common/kvm-arch.h | 7 + arm/include/arm-common/kvm-config-arch.h | 1 + arm/kvm-cpu.c | 14 +- arm/pmu.c | 76 ------ include/linux/bitmap.h | 71 +++++ include/linux/bitops.h | 4 + include/linux/bits.h | 8 + include/linux/cpumask.h | 67 +++++ include/linux/err.h | 2 + include/linux/find.h | 30 ++ include/linux/kernel.h | 6 + include/linux/kvm.h | 15 +- mips/include/asm/kernel.h | 8 + powerpc/include/asm/kernel.h | 8 + util/bitmap.c | 256 ++++++++++++++++++ util/find.c | 40 +++ x86/include/asm/kernel.h | 8 + 29 files changed, 870 insertions(+), 102 deletions(-) create mode 100644 arm/aarch32/include/asm/kernel.h create mode 100644 arm/aarch64/include/asm/kernel.h rename arm/{include/arm-common => aarch64/include/asm}/pmu.h (55%) create mode 100644 arm/aarch64/pmu.c delete mode 100644 arm/pmu.c create mode 100644 include/linux/bitmap.h create mode 100644 include/linux/bits.h create mode 100644 include/linux/cpumask.h create mode 100644 include/linux/find.h create mode 100644 mips/include/asm/kernel.h create mode 100644 powerpc/include/asm/kernel.h create mode 100644 util/bitmap.c create mode 100644 util/find.c create mode 100644 x86/include/asm/kernel.h -- 2.25.1 _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm