Re: [Patch v7 00/24] Support SIMD/eGPRs/SSP registers sampling for perf
From: Mi, Dapeng
Date: Wed Mar 25 2026 - 05:46:42 EST
All comments from Sashiko
(https://sashiko.dev/#/patchset/20260324004118.3772171-1-dapeng1.mi%40linux.intel.com)
have been reviewed.
About 70% comments make sense and are posted the individual patches. About
10% comments are correct but have nothing to do with this patch series,
like "intel_pmu_init_hybrid() fails to allocate x86_pmu.hybrid_pmu, but
current code doesn't check if x86_pmu.hybrid_pmu is null before accessing
it". The left ~20% comments are not correct and would be ignored.
All the correct comments would be addressed in the next version. Thanks.
On 3/24/2026 8:40 AM, Dapeng Mi wrote:
> Changes since V6:
> - Fix potential overwritten issue in hybrid PMU structure (patch 01/24)
> - Restrict PEBS events work on GP counters if no PEBS baseline suggested
> (patch 02/24)
> - Use per-cpu x86_intr_regs for perf_event_nmi_handler() instead of
> temporary variable (patch 06/24)
> - Add helper update_fpu_state_and_flag() to ensure TIF_NEED_FPU_LOAD is
> set after save_fpregs_to_fpstate() call (patch 09/24)
> - Optimize and simplify x86_pmu_sample_xregs(), etc. (patch 11/24)
> - Add macro word_for_each_set_bit() to simplify u64 set-bit iteration
> (patch 13/24)
> - Add sanity check for PEBS fragment size (patch 24/24)
>
> Changes since V5:
> - Introduce 3 commits to fix newly found PEBS issues (Patch 01~03/19)
> - Address Peter comments, including,
> * Fully support user-regs sampling of the SIMD/eGPRs/SSP registers
> * Adjust newly added fields in perf_event_attr to avoid holes
> * Fix the endian issue introduced by for_each_set_bit() in
> event/core.c
> * Remove some unnecessary macros from UAPI header perf_regs.h
> * Enhance b2b NMI detection for all PEBS handlers to ensure identical
> behaviors of all PEBS handlers
> - Split perf-tools patches which would be posted in a separate patchset
> later
>
> Changes since V4:
> - Rewrite some functions comments and commit messages (Dave)
> - Add arch-PEBS based SIMD/eGPRs/SSP sampling support (Patch 15/19)
> - Fix "suspecious NMI" warnning observed on PTL/NVL P-core and DMR by
> activating back-to-back NMI detection mechanism (Patch 16/19)
> - Fix some minor issues on perf-tool patches (Patch 18/19)
>
> Changes since V3:
> - Drop the SIMD registers if an NMI hits kernel mode for REGS_USER.
> - Only dump the available regs, rather than zero and dump the
> unavailable regs. It's possible that the dumped registers are a subset
> of the requested registers.
> - Some minor updates to address Dapeng's comments in V3.
>
> Changes since V2:
> - Use the FPU format for the x86_pmu.ext_regs_mask as well
> - Add a check before invoking xsaves_nmi()
> - Add perf_simd_reg_check() to retrieve the number of available
> registers. If the kernel fails to get the requested registers, e.g.,
> XSAVES fails, nothing dumps to the userspace (the V2 dumps all 0s).
> - Add POC perf tool patches
>
> Changes since V1:
> - Apply the new interfaces to configure and dump the SIMD registers
> - Utilize the existing FPU functions, e.g., xstate_calculate_size,
> get_xsave_addr().
>
> Starting from Intel Ice Lake, XMM registers can be collected in a PEBS
> record. Future Architecture PEBS will include additional registers such
> as YMM, ZMM, OPMASK, SSP and APX eGPRs, contingent on hardware support.
>
> This patch set introduces a software solution to mitigate the hardware
> requirement by utilizing the XSAVES command to retrieve the requested
> registers in the overflow handler. This feature is no longer limited to
> PEBS events or specific platforms. While the hardware solution remains
> preferable due to its lower overhead and higher accuracy, this software
> approach provides a viable alternative.
>
> The solution is theoretically compatible with all x86 platforms but is
> currently enabled on newer platforms, including Sapphire Rapids and
> later P-core server platforms, Sierra Forest and later E-core server
> platforms and recent Client platforms, like Arrow Lake, Panther Lake and
> Nova Lake.
>
> Newly supported registers include YMM, ZMM, OPMASK, SSP, and APX eGPRs.
> Due to space constraints in sample_regs_user/intr, new fields have been
> introduced in the perf_event_attr structure to accommodate these
> registers.
>
> After a long discussion in V1,
> https://lore.kernel.org/lkml/3f1c9a9e-cb63-47ff-a5e9-06555fa6cc9a@xxxxxxxxxxxxxxx/
> The below new fields are introduced.
>
> @@ -547,6 +549,25 @@ struct perf_event_attr {
>
> __u64 config3; /* extension of config2 */
> __u64 config4; /* extension of config3 */
> +
> + /*
> + * Defines set of SIMD registers to dump on samples.
> + * The sample_simd_regs_enabled !=0 implies the
> + * set of SIMD registers is used to config all SIMD registers.
> + * If !sample_simd_regs_enabled, sample_regs_XXX may be used to
> + * config some SIMD registers on X86.
> + */
> + union {
> + __u16 sample_simd_regs_enabled;
> + __u16 sample_simd_pred_reg_qwords;
> + };
> + __u16 sample_simd_vec_reg_qwords;
> + __u32 __reserved_4;
> +
> + __u32 sample_simd_pred_reg_intr;
> + __u32 sample_simd_pred_reg_user;
> + __u64 sample_simd_vec_reg_intr;
> + __u64 sample_simd_vec_reg_user;
> };
>
> /*
> @@ -1020,7 +1041,15 @@ enum perf_event_type {
> * } && PERF_SAMPLE_BRANCH_STACK
> *
> * { u64 abi; # enum perf_sample_regs_abi
> - * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
> + * u64 regs[weight(mask)];
> + * struct {
> + * u16 nr_vectors; # 0 ... weight(sample_simd_vec_reg_user)
> + * u16 vector_qwords; # 0 ... sample_simd_vec_reg_qwords
> + * u16 nr_pred; # 0 ... weight(sample_simd_pred_reg_user)
> + * u16 pred_qwords; # 0 ... sample_simd_pred_reg_qwords
> + * u64 data[nr_vectors * vector_qwords + nr_pred * pred_qwords];
> + * } && (abi & PERF_SAMPLE_REGS_ABI_SIMD)
> + * } && PERF_SAMPLE_REGS_USER
> *
> * { u64 size;
> * char data[size];
> @@ -1047,7 +1076,15 @@ enum perf_event_type {
> * { u64 data_src; } && PERF_SAMPLE_DATA_SRC
> * { u64 transaction; } && PERF_SAMPLE_TRANSACTION
> * { u64 abi; # enum perf_sample_regs_abi
> - * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
> + * u64 regs[weight(mask)];
> + * struct {
> + * u16 nr_vectors; # 0 ... weight(sample_simd_vec_reg_intr)
> + * u16 vector_qwords; # 0 ... sample_simd_vec_reg_qwords
> + * u16 nr_pred; # 0 ... weight(sample_simd_pred_reg_intr)
> + * u16 pred_qwords; # 0 ... sample_simd_pred_reg_qwords
> + * u64 data[nr_vectors * vector_qwords + nr_pred * pred_qwords];
> + * } && (abi & PERF_SAMPLE_REGS_ABI_SIMD)
> + * } && PERF_SAMPLE_REGS_INTR
> * { u64 phys_addr;} && PERF_SAMPLE_PHYS_ADDR
> * { u64 cgroup;} && PERF_SAMPLE_CGROUP
> * { u64 data_page_size;} && PERF_SAMPLE_DATA_PAGE_SIZE
>
>
> To maintain simplicity, a single field, sample_{simd|pred}_vec_reg_qwords,
> is introduced to indicate register width. For example:
> - sample_simd_vec_reg_qwords = 2 for XMM registers (128 bits) on x86
> - sample_simd_vec_reg_qwords = 4 for YMM registers (256 bits) on x86
>
> Four additional fields, sample_{simd|pred}_vec_reg_{intr|user}, represent
> the bitmap of sampling registers. For instance, the bitmap for x86
> XMM registers is 0xffff (16 XMM registers). Although users can
> theoretically sample a subset of registers, the current perf-tool
> implementation supports sampling all registers of each type to avoid
> complexity.
>
> A new ABI, PERF_SAMPLE_REGS_ABI_SIMD, is introduced to signal user space
> tools about the presence of SIMD registers in sampling records. When this
> flag is detected, tools should recognize that extra SIMD register data
> follows the general register data. The layout of the extra SIMD register
> data is displayed as follow.
>
> u16 nr_vectors;
> u16 vector_qwords;
> u16 nr_pred;
> u16 pred_qwords;
> u64 data[nr_vectors * vector_qwords + nr_pred * pred_qwords];
>
> With this patch set, sampling for the aforementioned registers is
> supported on the Intel Nova Lake platform.
>
> Examples:
> $perf record -I?
> available registers: AX BX CX DX SI DI BP SP IP FLAGS CS SS R8 R9 R10
> R11 R12 R13 R14 R15 R16 R17 R18 R19 R20 R21 R22 R23 R24 R25 R26 R27 R28
> R29 R30 R31 SSP XMM0-15 YMM0-15 ZMM0-31 OPMASK0-7
>
> $perf record --user-regs=?
> available registers: AX BX CX DX SI DI BP SP IP FLAGS CS SS R8 R9 R10
> R11 R12 R13 R14 R15 R16 R17 R18 R19 R20 R21 R22 R23 R24 R25 R26 R27 R28
> R29 R30 R31 SSP XMM0-15 YMM0-15 ZMM0-31 OPMASK0-7
>
> $perf record -e branches:p -Iax,bx,r8,r16,r31,ssp,xmm,ymm,zmm,opmask -c 100000 ./test
> $perf report -D
>
> ... ...
> 14027761992115 0xcf30 [0x8a8]: PERF_RECORD_SAMPLE(IP, 0x1): 29964/29964:
> 0xffffffff9f085e24 period: 100000 addr: 0
> ... intr regs: mask 0x18001010003 ABI 64-bit
> .... AX 0xdffffc0000000000
> .... BX 0xffff8882297685e8
> .... R8 0x0000000000000000
> .... R16 0x0000000000000000
> .... R31 0x0000000000000000
> .... SSP 0x0000000000000000
> ... SIMD ABI nr_vectors 32 vector_qwords 8 nr_pred 8 pred_qwords 1
> .... ZMM [0] 0xffffffffffffffff
> .... ZMM [0] 0x0000000000000001
> .... ZMM [0] 0x0000000000000000
> .... ZMM [0] 0x0000000000000000
> .... ZMM [0] 0x0000000000000000
> .... ZMM [0] 0x0000000000000000
> .... ZMM [0] 0x0000000000000000
> .... ZMM [0] 0x0000000000000000
> .... ZMM [1] 0x003a6b6165506d56
> ... ...
> .... ZMM [31] 0x0000000000000000
> .... ZMM [31] 0x0000000000000000
> .... ZMM [31] 0x0000000000000000
> .... ZMM [31] 0x0000000000000000
> .... ZMM [31] 0x0000000000000000
> .... ZMM [31] 0x0000000000000000
> .... ZMM [31] 0x0000000000000000
> .... ZMM [31] 0x0000000000000000
> .... OPMASK[0] 0x00000000fffffe00
> .... OPMASK[1] 0x0000000000ffffff
> .... OPMASK[2] 0x000000000000007f
> .... OPMASK[3] 0x0000000000000000
> .... OPMASK[4] 0x0000000000010080
> .... OPMASK[5] 0x0000000000000000
> .... OPMASK[6] 0x0000400004000000
> .... OPMASK[7] 0x0000000000000000
> ... ...
>
>
> History:
> v6: https://lore.kernel.org/all/20260209072047.2180332-1-dapeng1.mi@xxxxxxxxxxxxxxx/
> v5: https://lore.kernel.org/all/20251203065500.2597594-1-dapeng1.mi@xxxxxxxxxxxxxxx/
> v4: https://lore.kernel.org/all/20250925061213.178796-1-dapeng1.mi@xxxxxxxxxxxxxxx/
> v3: https://lore.kernel.org/lkml/20250815213435.1702022-1-kan.liang@xxxxxxxxxxxxxxx/
> v2: https://lore.kernel.org/lkml/20250626195610.405379-1-kan.liang@xxxxxxxxxxxxxxx/
> v1: https://lore.kernel.org/lkml/20250613134943.3186517-1-kan.liang@xxxxxxxxxxxxxxx/
>
> Dapeng Mi (12):
> perf/x86: Move hybrid PMU initialization before x86_pmu_starting_cpu()
> perf/x86/intel: Avoid PEBS event on fixed counters without extended
> PEBS
> perf/x86/intel: Enable large PEBS sampling for XMMs
> perf/x86/intel: Convert x86_perf_regs to per-cpu variables
> perf: Eliminate duplicate arch-specific functions definations
> x86/fpu: Ensure TIF_NEED_FPU_LOAD is set after saving FPU state
> perf/x86: Enable XMM Register Sampling for Non-PEBS Events
> perf/x86: Enable XMM register sampling for REGS_USER case
> perf: Enhance perf_reg_validate() with simd_enabled argument
> perf/x86/intel: Enable arch-PEBS based SIMD/eGPRs/SSP sampling
> perf/x86: Activate back-to-back NMI detection for arch-PEBS induced
> NMIs
> perf/x86/intel: Add sanity check for PEBS fragment size
>
> Kan Liang (12):
> perf/x86: Use x86_perf_regs in the x86 nmi handler
> perf/x86: Introduce x86-specific x86_pmu_setup_regs_data()
> x86/fpu/xstate: Add xsaves_nmi() helper
> perf: Move and rename has_extended_regs() for ARCH-specific use
> perf: Add sampling support for SIMD registers
> perf/x86: Enable XMM sampling using sample_simd_vec_reg_* fields
> perf/x86: Enable YMM sampling using sample_simd_vec_reg_* fields
> perf/x86: Enable ZMM sampling using sample_simd_vec_reg_* fields
> perf/x86: Enable OPMASK sampling using sample_simd_pred_reg_* fields
> perf/x86: Enable eGPRs sampling using sample_regs_* fields
> perf/x86: Enable SSP sampling using sample_regs_* fields
> perf/x86/intel: Enable PERF_PMU_CAP_SIMD_REGS capability
>
> arch/arm/kernel/perf_regs.c | 8 +-
> arch/arm64/kernel/perf_regs.c | 8 +-
> arch/csky/kernel/perf_regs.c | 8 +-
> arch/loongarch/kernel/perf_regs.c | 8 +-
> arch/mips/kernel/perf_regs.c | 8 +-
> arch/parisc/kernel/perf_regs.c | 8 +-
> arch/powerpc/perf/perf_regs.c | 2 +-
> arch/riscv/kernel/perf_regs.c | 8 +-
> arch/s390/kernel/perf_regs.c | 2 +-
> arch/x86/events/core.c | 392 +++++++++++++++++++++++++-
> arch/x86/events/intel/core.c | 127 ++++++++-
> arch/x86/events/intel/ds.c | 195 ++++++++++---
> arch/x86/events/perf_event.h | 85 +++++-
> arch/x86/include/asm/fpu/sched.h | 5 +-
> arch/x86/include/asm/fpu/xstate.h | 3 +
> arch/x86/include/asm/msr-index.h | 7 +
> arch/x86/include/asm/perf_event.h | 38 ++-
> arch/x86/include/uapi/asm/perf_regs.h | 51 ++++
> arch/x86/kernel/fpu/core.c | 27 +-
> arch/x86/kernel/fpu/xstate.c | 25 +-
> arch/x86/kernel/perf_regs.c | 134 +++++++--
> include/linux/perf_event.h | 16 ++
> include/linux/perf_regs.h | 36 +--
> include/uapi/linux/perf_event.h | 50 +++-
> kernel/events/core.c | 138 +++++++--
> tools/perf/util/header.c | 3 +-
> 26 files changed, 1193 insertions(+), 199 deletions(-)
>