Re: [PATCH v1 01/25] perf sample: Document struct perf_sample

[Namhyung Kim <namhyung@kernel.org>]

On Mon, Feb 09, 2026 at 09:40:08AM -0800, Ian Rogers wrote:
> Add kernel-doc for struct perf_sample capturing the somewhat unusual
> population of fields and lifetime relationships.

Thanks for doing this.

> 
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/util/sample.h | 90 ++++++++++++++++++++++++++++++++++++++--
>  1 file changed, 87 insertions(+), 3 deletions(-)
> 
> diff --git a/tools/perf/util/sample.h b/tools/perf/util/sample.h
> index 3cce8dd202aa..3fd2a5e01308 100644
> --- a/tools/perf/util/sample.h
> +++ b/tools/perf/util/sample.h
> @@ -81,47 +81,131 @@ struct simd_flags {
>  #define SIMD_OP_FLAGS_PRED_PARTIAL	0x01	/* partial predicate */
>  #define SIMD_OP_FLAGS_PRED_EMPTY	0x02	/* empty predicate */
>  
> +/**
> + * struct perf_sample
> + *
> + * A sample is generally filled in by evlist__parse_sample/evsel__parse_sample
> + * which fills in the variables from a "union perf_event *event" which is data
> + * from a perf ring buffer or perf.data file. The "event" sample is variable in
> + * length as determined by the perf_event_attr (in the evsel) and details within
> + * the sample event itself. A struct perf_sample avoids needing to care about
> + * the variable length nature of the original event.
> + *
> + * To avoid being excessively large parts of the struct perf_sample are pointers
> + * into the original sample event. In general the lifetime of a struct
> + * perf_sample needs to be less than the "union perf_event *event" it was
> + * derived from.
> + *
> + * The struct regs_dump user_regs and intr_regs are lazily allocated again for
> + * size reasons, due to them holding a cache of looked up registers. The
> + * function pair of perf_sample__init and perf_sample__exit correctly initialize
> + * and clean up these values.
> + */
>  struct perf_sample {
> +	/** @ip: The sample event PERF_SAMPLE_IP value. */
>  	u64 ip;
> -	u32 pid, tid;
> +	/** @pid: The sample event PERF_SAMPLE_TID pid value. */
> +	u32 pid;
> +	/** @tid: The sample event PERF_SAMPLE_TID tid value. */
> +	u32 tid;
> +	/** @time: The sample event PERF_SAMPLE_TIME value. */
>  	u64 time;
> +	/** @addr: The sample event PERF_SAMPLE_ADDR value. */
>  	u64 addr;
> +	/** @id: The sample event PERF_SAMPLE_ID value. */
>  	u64 id;
> +	/** @stream_id: The sample event PERF_SAMPLE_STREAM_ID value. */
>  	u64 stream_id;
> +	/** @period: The sample event PERF_SAMPLE_PERIOD value. */
>  	u64 period;
> +	/** @weight: Data determined by PERF_SAMPLE_WEIGHT or PERF_SAMPLE_WEIGHT_STRUCT. */
>  	u64 weight;
> +	/** @transaction: The sample event PERF_SAMPLE_TRANSACTION value. */
>  	u64 transaction;
> +	/** @insn_cnt: Filled in and used by intel-pt. */
>  	u64 insn_cnt;
> +	/** @cyc_cnt: Filled in and used by intel-pt. */
>  	u64 cyc_cnt;
> +	/** @cpu: The sample event PERF_SAMPLE_CPU value. */
>  	u32 cpu;
> +	/**
> +	 * @raw_size: The size in bytes of raw data from PERF_SAMPLE_RAW. For
> +	 *            alignment reasons this should always be a multiple of
> +	 *            sizeof(u64) + sizeof(u32).
> +	 */
>  	u32 raw_size;
> +	/** @data_src: The sample event PERF_SAMPLE_DATA_SRC value. */
>  	u64 data_src;
> +	/** @phys_addr: The sample event PERF_SAMPLE_PHYS_ADDR value. */
>  	u64 phys_addr;
> +	/** @data_page_size: The sample event PERF_SAMPLE_DATA_PAGE_SIZE value. */
>  	u64 data_page_size;
> +	/** @code_page_size: The sample event PERF_SAMPLE_CODE_PAGE_SIZE value. */
>  	u64 code_page_size;
> +	/** @cgroup: The sample event PERF_SAMPLE_CGROUP value. */
>  	u64 cgroup;
> +	/** @flags: Extra flag data from auxiliary events like intel-pt. */
>  	u32 flags;
> +	/** @machine_pid: The guest machine pid derived from the sample id. */
>  	u32 machine_pid;
> +	/** @vcpu: The guest machine vcpu derived from the sample id. */
>  	u32 vcpu;
> +	/** @insn_len: Instruction length from auxiliary events like intel-pt. */
>  	u16 insn_len;

Does it control the insn array later?


> +	/**
> +	 * @cpumode: The cpumode from struct perf_event_header misc variable
> +	 *           masked with CPUMODE_MASK. Gives user, kernel and hypervisor
> +	 *           information.
> +	 */
>  	u8  cpumode;
> +	/** @misc: The entire struct perf_event_header misc variable. */
>  	u16 misc;
> +	/** @ins_lat: Instruction latency information from auxiliary events like intel-pt. */
>  	u16 ins_lat;

I think this is weight2 coming from PERF_SAMPLE_WEIGHT_STRUCT.


>  	/** @weight3: On x86 holds retire_lat, on powerpc holds p_stage_cyc. */
>  	u16 weight3;

You may also want to mention it's from WEIGHT_STRUCT.


> -	bool no_hw_idx;		/* No hw_idx collected in branch_stack */
> -	bool deferred_callchain;	/* Has deferred user callchains */
> +	/**
> +	 * @no_hw_idx: For PERF_SAMPLE_BRANCH_STACK, true when
> +	 *             PERF_SAMPLE_BRANCH_HW_INDEX isn't set.
> +	 */
> +	bool no_hw_idx;
> +	/**
> +	 * @deferred_callchain: When processing PERF_SAMPLE_CALLCHAIN a deferred
> +	 *                      user callchain marker was encountered.
> +	 */
> +	bool deferred_callchain;

If this is set, then callchain entry is allocated for deferred
callchains.


> +	/**
> +	 * @deferred_cookie: Identifier of the deferred callchain in the later
> +	 *                   PERF_RECORD_CALLCHAIN_DEFERRED event.
> +	 */
>  	u64 deferred_cookie;
> +	/** @insn: A copy of the sampled instruction filled in by perf_sample__fetch_insn. */
>  	char insn[MAX_INSN];
> +	/** @raw_data: Byte aligned pointer into the original event for PERF_SAMPLE_RAW data. */

I think it's 32-bit aligned (or shifted) as it comes right after the
raw_size (32-bit).  As other fields in the event are 64-bit aligned this
caused some troubles like in perf trace when you access the pointer
directly and assume it's naturally aligned.


>  	void *raw_data;
> +	/** @callchain: Pointer into the original event for PERF_SAMPLE_CALLCHAIN data. */
>  	struct ip_callchain *callchain;

If deferred_callchain is not set, it just points to data in the mmap
buffer and it should not be freed.


> +	/** @branch_stack: Pointer into the original event for PERF_SAMPLE_BRANCH_STACK data. */
>  	struct branch_stack *branch_stack;
> +	/**
> +	 * @branch_stack_cntr: Pointer into the original event for
> +	 *                     PERF_SAMPLE_BRANCH_COUNTERS data.
> +	 */
>  	u64 *branch_stack_cntr;
> +	/** @user_regs: Values and pointers into the sample for PERF_SAMPLE_REGS_USER. */
>  	struct regs_dump  *user_regs;
> +	/** @intr_regs: Values and pointers into the sample for PERF_SAMPLE_REGS_INTR. */
>  	struct regs_dump  *intr_regs;
> +	/** @user_stack: Size and pointer into the sample for PERF_SAMPLE_STACK_USER. */
>  	struct stack_dump user_stack;
> +	/** @read: The sample event PERF_SAMPLE_READ counter values. */

The actual format depends on read_format in the event attribute.

Thanks,
Namhyung


>  	struct sample_read read;
> +	/**
> +	 * @aux_sample: Similar to raw data but with a 64-bit size and
> +	 *              alignment, PERF_SAMPLE_AUX data.
> +	 */
>  	struct aux_sample aux_sample;
> +	/** @simd_flags: SIMD flag information from ARM SPE auxiliary events. */
>  	struct simd_flags simd_flags;
>  };
>  
> -- 
> 2.53.0.239.g8d8fc8a987-goog
>