Re: [Patch v7 13/24] perf: Add sampling support for SIMD registers

[Mi, Dapeng]

On 3/24/2026 8:41 AM, Dapeng Mi wrote:
> From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>
> Users may be interested in sampling SIMD registers during profiling.
> The current sample_regs_* structure does not have sufficient space
> for all SIMD registers.
>
> To address this, new attribute fields sample_simd_{pred,vec}_reg_* are
> added to struct perf_event_attr to represent the SIMD registers that are
> expected to be sampled.
>
> Currently, the perf/x86 code supports XMM registers in sample_regs_*.
> To unify the configuration of SIMD registers and ensure a consistent
> method for configuring XMM and other SIMD registers, a new event
> attribute field, sample_simd_regs_enabled, is introduced. When
> sample_simd_regs_enabled is set, it indicates that all SIMD registers,
> including XMM, will be represented by the newly introduced
> sample_simd_{pred|vec}_reg_* fields. The original XMM space in
> sample_regs_* is reserved for future uses.
>
> Since SIMD registers are wider than 64 bits, a new output format is
> introduced. The number and width of SIMD registers are dumped first,
> followed by the register values. The number and width are based on the
> user's configuration. If they differ (e.g., on ARM), an ARCH-specific
> perf_output_sample_simd_regs function can be implemented separately.
>
> A new ABI, PERF_SAMPLE_REGS_ABI_SIMD, is added to indicate the new format.
> The enum perf_sample_regs_abi is now a bitmap. This change should not
> impact existing tools, as the version and bitmap remain the same for
> values 1 and 2.
>
> Additionally, two new __weak functions are introduced:
> - perf_simd_reg_value(): Retrieves the value of the requested SIMD
>   register.
> - perf_simd_reg_validate(): Validates the configuration of the SIMD
>   registers.
>
> A new flag, PERF_PMU_CAP_SIMD_REGS, is added to indicate that the PMU
> supports SIMD register dumping. An error is generated if
> sample_simd_{pred|vec}_reg_* is mistakenly set for a PMU that does not
> support this capability.
>
> Suggested-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> Co-developed-by: Dapeng Mi <dapeng1.mi@xxxxxxxxxxxxxxx>
> Signed-off-by: Dapeng Mi <dapeng1.mi@xxxxxxxxxxxxxxx>
> ---
>
> V7: Add macro word_for_each_set_bit() to simplify u64 set-bit iteration.
>
>  include/linux/perf_event.h      |   8 +++
>  include/linux/perf_regs.h       |   4 ++
>  include/uapi/linux/perf_event.h |  50 ++++++++++++++--
>  kernel/events/core.c            | 102 +++++++++++++++++++++++++++++---
>  tools/perf/util/header.c        |   3 +-
>  5 files changed, 153 insertions(+), 14 deletions(-)
>
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index e8b0d8e2d2af..137d6e4a3403 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -306,6 +306,7 @@ struct perf_event_pmu_context;
>  #define PERF_PMU_CAP_AUX_PAUSE		0x0200
>  #define PERF_PMU_CAP_AUX_PREFER_LARGE	0x0400
>  #define PERF_PMU_CAP_MEDIATED_VPMU	0x0800
> +#define PERF_PMU_CAP_SIMD_REGS		0x1000
>  
>  /**
>   * pmu::scope
> @@ -1534,6 +1535,13 @@ perf_event__output_id_sample(struct perf_event *event,
>  extern void
>  perf_log_lost_samples(struct perf_event *event, u64 lost);
>  
> +static inline bool event_has_simd_regs(struct perf_event *event)
> +{
> +	struct perf_event_attr *attr = &event->attr;
> +
> +	return attr->sample_simd_regs_enabled != 0;
> +}
> +
>  static inline bool event_has_extended_regs(struct perf_event *event)
>  {
>  	struct perf_event_attr *attr = &event->attr;
> diff --git a/include/linux/perf_regs.h b/include/linux/perf_regs.h
> index 144bcc3ff19f..518f28c6a7d4 100644
> --- a/include/linux/perf_regs.h
> +++ b/include/linux/perf_regs.h
> @@ -14,6 +14,10 @@ int perf_reg_validate(u64 mask);
>  u64 perf_reg_abi(struct task_struct *task);
>  void perf_get_regs_user(struct perf_regs *regs_user,
>  			struct pt_regs *regs);
> +int perf_simd_reg_validate(u16 vec_qwords, u64 vec_mask,
> +			   u16 pred_qwords, u32 pred_mask);
> +u64 perf_simd_reg_value(struct pt_regs *regs, int idx,
> +			u16 qwords_idx, bool pred);
>  
>  #ifdef CONFIG_HAVE_PERF_REGS
>  #include <asm/perf_regs.h>
> diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
> index fd10aa8d697f..b8c8953928f8 100644
> --- a/include/uapi/linux/perf_event.h
> +++ b/include/uapi/linux/perf_event.h
> @@ -314,8 +314,9 @@ enum {
>   */
>  enum perf_sample_regs_abi {
>  	PERF_SAMPLE_REGS_ABI_NONE		= 0,
> -	PERF_SAMPLE_REGS_ABI_32			= 1,
> -	PERF_SAMPLE_REGS_ABI_64			= 2,
> +	PERF_SAMPLE_REGS_ABI_32			= (1 << 0),
> +	PERF_SAMPLE_REGS_ABI_64			= (1 << 1),
> +	PERF_SAMPLE_REGS_ABI_SIMD		= (1 << 2),
>  };
>  
>  /*
> @@ -383,6 +384,7 @@ enum perf_event_read_format {
>  #define PERF_ATTR_SIZE_VER7			128	/* Add: sig_data */
>  #define PERF_ATTR_SIZE_VER8			136	/* Add: config3 */
>  #define PERF_ATTR_SIZE_VER9			144	/* add: config4 */
> +#define PERF_ATTR_SIZE_VER10			176	/* Add: sample_simd_{pred,vec}_reg_* */
>  
>  /*
>   * 'struct perf_event_attr' contains various attributes that define
> @@ -547,6 +549,30 @@ struct perf_event_attr {
>  
>  	__u64	config3; /* extension of config2 */
>  	__u64	config4; /* extension of config3 */
> +
> +	/*
> +	 * Defines the sampling SIMD/PRED registers bitmap and qwords
> +	 * (8 bytes) length.
> +	 *
> +	 * sample_simd_regs_enabled != 0 indicates there are SIMD/PRED registers
> +	 * to be sampled, the SIMD/PRED registers bitmap and qwords length are
> +	 * represented in sample_{simd|pred}_pred_reg_{intr|user} and
> +	 * sample_simd_{vec|pred}_reg_qwords fields.
> +	 *
> +	 * sample_simd_regs_enabled == 0 indicates no SIMD/PRED registers are
> +	 * sampled.
> +	 */
> +	union {
> +		__u16 sample_simd_regs_enabled;
> +		__u16 sample_simd_pred_reg_qwords;
> +	};
> +	__u16	sample_simd_vec_reg_qwords;
> +	__u32	__reserved_4;

Sashiko comments

"

Since __reserved_4 is newly introduced to maintain alignment, should there be
validation in perf_copy_attr() in kernel/events/core.c to ensure it is
strictly zeroed out?
Without validation to reject non-zero values, userspace applications might
pass uninitialized memory here, potentially preventing the kernel from safely
repurposing __reserved_4 in the future without breaking existing binaries.

"

It's correct. Just updating the perf_attr_check(), but forget the
perf_copy_attr().


> +
> +	__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 +1046,15 @@ 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 +1081,15 @@ 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
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 7558bc5b1e73..de42575f517b 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -7753,22 +7753,60 @@ void __weak perf_get_regs_user(struct perf_regs *regs_user,
>  	regs_user->abi = perf_reg_abi(current);
>  }
>  
> +#define word_for_each_set_bit(bit, val)			\
> +	for (unsigned long long __v = (val);		\
> +	     __v && ((bit = __builtin_ctzll(__v)), 1);	\
> +	     __v &= __v - 1)
> +
>  static void
>  perf_output_sample_regs(struct perf_output_handle *handle,
>  			struct pt_regs *regs, u64 mask)
>  {
>  	int bit;
> -	DECLARE_BITMAP(_mask, 64);
> -
> -	bitmap_from_u64(_mask, mask);
> -	for_each_set_bit(bit, _mask, sizeof(mask) * BITS_PER_BYTE) {
> -		u64 val;
>  
> -		val = perf_reg_value(regs, bit);
> +	word_for_each_set_bit(bit, mask) {
> +		u64 val = perf_reg_value(regs, bit);
>  		perf_output_put(handle, val);
>  	}
>  }
>  
> +static void
> +perf_output_sample_simd_regs(struct perf_output_handle *handle,
> +			     struct perf_event *event,
> +			     struct pt_regs *regs,
> +			     u64 mask, u32 pred_mask)

Sashiko comments

"

Does the static declaration of perf_output_sample_simd_regs() prevent
architectures from overriding it as described in the commit message?


Additionally, can the generic perf_prepare_sample() function reserve enough
space for this new payload?

"

Yes,  the commit message needs to be enhanced and remove the description
"If they differ (e.g., on ARM), an ARCH-specific
perf_output_sample_simd_regs function can be implemented separately."

Currently the space for the new payload is reserved in
x86_pmu_setup_regs_data(), but it's indeed not a good place, would move it
into perf_prepare_sample(). Thanks.


> +{
> +	u16 pred_qwords = event->attr.sample_simd_pred_reg_qwords;
> +	u16 vec_qwords = event->attr.sample_simd_vec_reg_qwords;
> +	u16 nr_vectors = hweight64(mask);
> +	u16 nr_pred = hweight32(pred_mask);
> +	int bit;
> +
> +	perf_output_put(handle, nr_vectors);
> +	perf_output_put(handle, vec_qwords);
> +	perf_output_put(handle, nr_pred);
> +	perf_output_put(handle, pred_qwords);
> +
> +	if (nr_vectors) {
> +		word_for_each_set_bit(bit, mask) {
> +			for (int i = 0; i < vec_qwords; i++) {
> +				u64 val = perf_simd_reg_value(regs, bit,
> +							      i, false);
> +				perf_output_put(handle, val);
> +			}
> +		}
> +	}
> +	if (nr_pred) {
> +		word_for_each_set_bit(bit, pred_mask) {
> +			for (int i = 0; i < pred_qwords; i++) {
> +				u64 val = perf_simd_reg_value(regs, bit,
> +							      i, true);
> +				perf_output_put(handle, val);
> +			}
> +		}
> +	}
> +}
> +
>  static void perf_sample_regs_user(struct perf_regs *regs_user,
>  				  struct pt_regs *regs)
>  {
> @@ -7790,6 +7828,17 @@ static void perf_sample_regs_intr(struct perf_regs *regs_intr,
>  	regs_intr->abi  = perf_reg_abi(current);
>  }
>  
> +int __weak perf_simd_reg_validate(u16 vec_qwords, u64 vec_mask,
> +				  u16 pred_qwords, u32 pred_mask)
> +{
> +	return vec_qwords || vec_mask || pred_qwords || pred_mask ? -ENOSYS : 0;
> +}
> +
> +u64 __weak perf_simd_reg_value(struct pt_regs *regs, int idx,
> +			       u16 qwords_idx, bool pred)
> +{
> +	return 0;
> +}
>  
>  /*
>   * Get remaining task size from user stack pointer.
> @@ -8320,10 +8369,17 @@ void perf_output_sample(struct perf_output_handle *handle,
>  		perf_output_put(handle, abi);
>  
>  		if (abi) {
> -			u64 mask = event->attr.sample_regs_user;
> +			struct perf_event_attr *attr = &event->attr;
> +			u64 mask = attr->sample_regs_user;
>  			perf_output_sample_regs(handle,
>  						data->regs_user.regs,
>  						mask);
> +			if (abi & PERF_SAMPLE_REGS_ABI_SIMD) {
> +				perf_output_sample_simd_regs(handle, event,
> +							     data->regs_user.regs,
> +							     attr->sample_simd_vec_reg_user,
> +							     attr->sample_simd_pred_reg_user);
> +			}
>  		}
>  	}
>  
> @@ -8351,11 +8407,18 @@ void perf_output_sample(struct perf_output_handle *handle,
>  		perf_output_put(handle, abi);
>  
>  		if (abi) {
> -			u64 mask = event->attr.sample_regs_intr;
> +			struct perf_event_attr *attr = &event->attr;
> +			u64 mask = attr->sample_regs_intr;
>  
>  			perf_output_sample_regs(handle,
>  						data->regs_intr.regs,
>  						mask);
> +			if (abi & PERF_SAMPLE_REGS_ABI_SIMD) {
> +				perf_output_sample_simd_regs(handle, event,
> +							     data->regs_intr.regs,
> +							     attr->sample_simd_vec_reg_intr,
> +							     attr->sample_simd_pred_reg_intr);
> +			}
>  		}
>  	}
>  
> @@ -13011,6 +13074,12 @@ static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)
>  	if (ret)
>  		goto err_pmu;
>  
> +	if (!(pmu->capabilities & PERF_PMU_CAP_SIMD_REGS) &&
> +	    event_has_simd_regs(event)) {
> +		ret = -EOPNOTSUPP;
> +		goto err_destroy;
> +	}
> +
>  	if (!(pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS) &&
>  	    event_has_extended_regs(event)) {
>  		ret = -EOPNOTSUPP;
> @@ -13556,6 +13625,12 @@ static int perf_copy_attr(struct perf_event_attr __user *uattr,
>  		ret = perf_reg_validate(attr->sample_regs_user);
>  		if (ret)
>  			return ret;
> +		ret = perf_simd_reg_validate(attr->sample_simd_vec_reg_qwords,
> +					     attr->sample_simd_vec_reg_user,
> +					     attr->sample_simd_pred_reg_qwords,
> +					     attr->sample_simd_pred_reg_user);
> +		if (ret)
> +			return ret;
>  	}
>  
>  	if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
> @@ -13576,8 +13651,17 @@ static int perf_copy_attr(struct perf_event_attr __user *uattr,
>  	if (!attr->sample_max_stack)
>  		attr->sample_max_stack = sysctl_perf_event_max_stack;
>  
> -	if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
> +	if (attr->sample_type & PERF_SAMPLE_REGS_INTR) {
>  		ret = perf_reg_validate(attr->sample_regs_intr);
> +		if (ret)
> +			return ret;
> +		ret = perf_simd_reg_validate(attr->sample_simd_vec_reg_qwords,
> +					     attr->sample_simd_vec_reg_intr,
> +					     attr->sample_simd_pred_reg_qwords,
> +					     attr->sample_simd_pred_reg_intr);
> +		if (ret)
> +			return ret;
> +	}
>  
>  #ifndef CONFIG_CGROUP_PERF
>  	if (attr->sample_type & PERF_SAMPLE_CGROUP)
> diff --git a/tools/perf/util/header.c b/tools/perf/util/header.c
> index 9142a8ba4019..f84200b9dd57 100644
> --- a/tools/perf/util/header.c
> +++ b/tools/perf/util/header.c
> @@ -2051,7 +2051,8 @@ static void free_event_desc(struct evsel *events)
>  
>  static bool perf_attr_check(struct perf_event_attr *attr)
>  {
> -	if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3) {
> +	if (attr->__reserved_1 || attr->__reserved_2 ||
> +	    attr->__reserved_3 || attr->__reserved_4) {
>  		pr_warning("Reserved bits are set unexpectedly. "
>  			   "Please update perf tool.\n");
>  		return false;