bpf: add two helpers to read perf event enabled/running time
Subsystems (10)
v7 cover bpf: add two helpers to read perf event enabled/running time
Hardware pmu counters are limited resources. When there are more pmu based perf events opened than available counters, kernel will multiplex these events so each event gets certain percentage (but not 100%) of the pmu time. In case that multiplexing happens, the number of samples or counter value will not reflect the case compared to no multiplexing. This makes comparison between different runs difficult.
Typically, the number of samples or counter value should be normalized before comparing to other experiments. The typical normalization is done like:
normalized_num_samples = num_samples * time_enabled / time_running normalized_counter_value = counter_value * time_enabled / time_running
where time_enabled is the time enabled for event and time_running is the time running for event since last normalization.
This patch set implements two helper functions.
The helper bpf_perf_event_read_value reads counter/time_enabled/time_running for perf event array map. The helper bpf_perf_prog_read_value read counter/time_enabled/time_running for bpf prog with type BPF_PROG_TYPE_PERF_EVENT.
[Dave, Peter,
Patch set has been restructured such that all perf (non-bpf) related changes are in the first commit. The whole patch set should be merged into net-next and the first commit cherry-picked to tip to avoid conflicts during the merge window.
Thanks!
]
Changelogs:v6->v7:
. restructure the patch set so that all perf related changes are
in the first one to make cherry-pick easier,
. address Daniel's comments to clear user buffer in all error cases
for both helpers.v5->v6:
. rebase on top of net-next.v4->v5:
. fix some coding style issues,
. memset the input buffer in case of error for ARG_PTR_TO_UNINIT_MEM
type of argument.v3->v4:
. fix a build failure.v2->v3:
. counters should be read in order to read enabled/running time. This is to
prevent that counters and enabled/running time may be read separately.v1->v2:
. reading enabled/running time should be together with reading counters
which contains the logic to ensure the result is valid.
Yonghong Song (5):
bpf: perf event change needed for subsequent bpf helpers
bpf: add helper bpf_perf_event_read_value for perf event array map
bpf: add a test case for helper bpf_perf_event_read_value
bpf: add helper bpf_perf_prog_read_value
bpf: add a test case for helper bpf_perf_prog_read_value
include/linux/perf_event.h | 7 ++- include/uapi/linux/bpf.h | 29 +++++++++++- kernel/bpf/arraymap.c | 2 +- kernel/bpf/verifier.c | 4 +- kernel/events/core.c | 15 ++++++- kernel/trace/bpf_trace.c | 73 +++++++++++++++++++++++++++++-- samples/bpf/trace_event_kern.c | 10 +++++ samples/bpf/trace_event_user.c | 13 +++--- samples/bpf/tracex6_kern.c | 26 +++++++++++ samples/bpf/tracex6_user.c | 13 +++++- tools/include/uapi/linux/bpf.h | 4 +- tools/testing/selftests/bpf/bpf_helpers.h | 6 +++ 12 files changed, 183 insertions(+), 19 deletions(-)
Signature · 1 line
2.9.5
PATCH v7 1/5 bpf: perf event change needed for subsequent bpf helpers
This patch does not impact existing functionalities. It contains the changes in perf event area needed for subsequent bpf_perf_event_read_value and bpf_perf_prog_read_value helpers.
--- include/linux/perf_event.h | 7 +++++-- kernel/bpf/arraymap.c | 2 +- kernel/events/core.c | 15 +++++++++++++-- kernel/trace/bpf_trace.c | 2 +- 4 files changed, 20 insertions(+), 6 deletions(-)
include/linux/perf_event.h +5 -2
@@ -806,6 +806,7 @@ struct perf_output_handle {struct bpf_perf_event_data_kern {struct pt_regs *regs;struct perf_sample_data *data; +struct perf_event *event; };#ifdef CONFIG_CGROUP_PERF
@@ -884,7 +885,8 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, void *context);extern void perf_pmu_migrate_context (struct pmu *pmu,int src_cpu,int dst_cpu); -int perf_event_read_local (struct perf_event *event,u64 *value); +int perf_event_read_local (struct perf_event *event,u64 *value, +u64 *enabled,u64 *running);extern u64 perf_event_read_value (struct perf_event *event,u64 *enabled,u64 *running);
@@ -1286,7 +1288,8 @@ static inline const struct perf_event_attr *perf_event_attrs(struct perf_event * {return ERR_PTR (-EINVAL ); } -static inline int perf_event_read_local (struct perf_event *event,u64 *value) +static inline int perf_event_read_local (struct perf_event *event,u64 *value, +u64 *enabled,u64 *running) {return -EINVAL ; }
kernel/bpf/arraymap.c +1 -1
@@ -492,7 +492,7 @@ static void *perf_event_fd_array_get_ptr(struct bpf_map *map, ee =ERR_PTR (-EOPNOTSUPP ); event = perf_file->private_data; -if (perf_event_read_local (event, &value) == -EOPNOTSUPP ) +if (perf_event_read_local (event, &value,NULL ,NULL ) == -EOPNOTSUPP ) goto err_out; ee =bpf_event_entry_gen (perf_file, map_file);
kernel/events/core.c +13 -2
@@ -3684,10 +3684,12 @@ static inline u64 perf_event_count(struct perf_event *event)* will not be local and we cannot read them atomically * - must not have a pmu::count method */ -int perf_event_read_local (struct perf_event *event,u64 *value) +int perf_event_read_local (struct perf_event *event,u64 *value, +u64 *enabled,u64 *running) {unsigned long flags;int ret =0 ; +u64 now;/* * Disabling interrupts avoids all counter scheduling (context
@@ -3718,13 +3720,21 @@ int perf_event_read_local(struct perf_event *event, u64 *value) goto out; } + now = event->shadow_ctx_time +perf_clock (); +if (enabled) + *enabled = now - event->tstamp_enabled;/* * If the event is currently on this CPU, its either a per-task event, * or local to this CPU. Furthermore it means its ACTIVE (otherwise * oncpu == -1). */ -if (event->oncpu ==smp_processor_id ()) +if (event->oncpu ==smp_processor_id ()) { event->pmu->read (event); +if (running) + *running = now - event->tstamp_running; + }else if (running) { + *running = event->total_time_running; + } *value =local64_read (&event->count); out:
@@ -8072,6 +8082,7 @@ static void bpf_overflow_handler(struct perf_event *event,struct bpf_perf_event_data_kern ctx = { .data = data, .regs = regs, + .event = event, };int ret =0 ;
kernel/trace/bpf_trace.c +1 -1
@@ -275,7 +275,7 @@ BPF_CALL_2(bpf_perf_event_read, struct bpf_map *, map, u64, flags)if (!ee)return -ENOENT ; - err =perf_event_read_local (ee->event, &value); + err =perf_event_read_local (ee->event, &value,NULL ,NULL );/* * this api is ugly since we miss [-22..-2] range of valid * counter values, but that's uapi
Signature · 1 line
2.9.5
Re: [PATCH net-next v7 1/5] bpf: perf event change needed for subsequent bpf helpers
Quoted 6 lines · Yonghong Song
This patch does not impact existing functionalities. It contains the changes in perf event area needed for subsequent
bpf_perf_event_read_valueandbpf_perf_prog_read_valuehelpers.Signed-off-by Yonghong Song <[email protected]>
And as discussed, I'll take this patch into my dev tree while Dave will take all of them into the network tree.
tip-bot for Yonghong Song
merge bot[tip:perf/core] perf/bpf: Extend the perf_event_read_local() interface, a.k.a. "bpf: perf event change needed for subsequent bpf helpers"
In reply to Yonghong Songperf/bpf: Extend the perf_event_read_local() interface, a.k.a. "bpf: perf event change needed for subsequent bpf helpers"
Commit-ID: 7d9285e82db5defca4d9674ba089429eeca0c697 Gitweb: https://git.kernel.org/tip/7d9285e82db5defca4d9674ba089429eeca0c697 Author: Yonghong Song <[email protected]> AuthorDate: Thu, 5 Oct 2017 09:19:19 -0700 Committer: Ingo Molnar <[email protected]> CommitDate: Fri, 27 Oct 2017 10:31:56 +0200
perf/bpf: Extend the perf_event_read_local() interface, a.k.a. "bpf: perf event change needed for subsequent bpf helpers"
eBPF programs would like access to the (perf) event enabled and running times along with the event value, such that they can deal with event multiplexing (among other things).
This patch extends the interface; a future eBPF patch will utilize the new functionality.
[ Note, there's a same-content commit with a poor changelog and a meaningless
title in the networking tree as well - but we need this change for subsequent perf work, so apply it here as well, with a proper changelog. Hopefully Git will be able to sort out this somewhat messy workflow, if there are no other, conflicting changes to these files. ]
--- include/linux/perf_event.h | 7 +++++-- kernel/bpf/arraymap.c | 2 +- kernel/events/core.c | 15 +++++++++++++-- kernel/trace/bpf_trace.c | 2 +- 4 files changed, 20 insertions(+), 6 deletions(-)
include/linux/perf_event.h +5 -2
@@ -806,6 +806,7 @@ struct perf_output_handle {struct bpf_perf_event_data_kern {struct pt_regs *regs;struct perf_sample_data *data; +struct perf_event *event; };#ifdef CONFIG_CGROUP_PERF
@@ -884,7 +885,8 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, void *context);extern void perf_pmu_migrate_context (struct pmu *pmu,int src_cpu,int dst_cpu); -int perf_event_read_local (struct perf_event *event,u64 *value); +int perf_event_read_local (struct perf_event *event,u64 *value, +u64 *enabled,u64 *running);extern u64 perf_event_read_value (struct perf_event *event,u64 *enabled,u64 *running);
@@ -1286,7 +1288,8 @@ static inline const struct perf_event_attr *perf_event_attrs(struct perf_event * {return ERR_PTR (-EINVAL ); } -static inline int perf_event_read_local (struct perf_event *event,u64 *value) +static inline int perf_event_read_local (struct perf_event *event,u64 *value, +u64 *enabled,u64 *running) {return -EINVAL ; }
kernel/bpf/arraymap.c +1 -1
@@ -492,7 +492,7 @@ static void *perf_event_fd_array_get_ptr(struct bpf_map *map, ee =ERR_PTR (-EOPNOTSUPP ); event = perf_file->private_data; -if (perf_event_read_local (event, &value) == -EOPNOTSUPP ) +if (perf_event_read_local (event, &value,NULL ,NULL ) == -EOPNOTSUPP ) goto err_out; ee =bpf_event_entry_gen (perf_file, map_file);
kernel/events/core.c +13 -2
@@ -3684,10 +3684,12 @@ static inline u64 perf_event_count(struct perf_event *event)* will not be local and we cannot read them atomically * - must not have a pmu::count method */ -int perf_event_read_local (struct perf_event *event,u64 *value) +int perf_event_read_local (struct perf_event *event,u64 *value, +u64 *enabled,u64 *running) {unsigned long flags;int ret =0 ; +u64 now;/* * Disabling interrupts avoids all counter scheduling (context
@@ -3718,13 +3720,21 @@ int perf_event_read_local(struct perf_event *event, u64 *value) goto out; } + now = event->shadow_ctx_time +perf_clock (); +if (enabled) + *enabled = now - event->tstamp_enabled;/* * If the event is currently on this CPU, its either a per-task event, * or local to this CPU. Furthermore it means its ACTIVE (otherwise * oncpu == -1). */ -if (event->oncpu ==smp_processor_id ()) +if (event->oncpu ==smp_processor_id ()) { event->pmu->read (event); +if (running) + *running = now - event->tstamp_running; + }else if (running) { + *running = event->total_time_running; + } *value =local64_read (&event->count); out:
@@ -8072,6 +8082,7 @@ static void bpf_overflow_handler(struct perf_event *event,struct bpf_perf_event_data_kern ctx = { .data = data, .regs = regs, + .event = event, };int ret =0 ;
kernel/trace/bpf_trace.c +1 -1
@@ -275,7 +275,7 @@ BPF_CALL_2(bpf_perf_event_read, struct bpf_map *, map, u64, flags)if (!ee)return -ENOENT ; - err =perf_event_read_local (ee->event, &value); + err =perf_event_read_local (ee->event, &value,NULL ,NULL );/* * this api is ugly since we miss [-22..-2] range of valid * counter values, but that's uapi
PATCH v7 2/5 bpf: add helper bpf_perf_event_read_value for perf event array map
In reply to Yonghong SongHardware pmu counters are limited resources. When there are more pmu based perf events opened than available counters, kernel will multiplex these events so each event gets certain percentage (but not 100%) of the pmu time. In case that multiplexing happens, the number of samples or counter value will not reflect the case compared to no multiplexing. This makes comparison between different runs difficult.
Typically, the number of samples or counter value should be normalized before comparing to other experiments. The typical normalization is done like:
normalized_num_samples = num_samples * time_enabled / time_running normalized_counter_value = counter_value * time_enabled / time_running
where time_enabled is the time enabled for event and time_running is the time running for event since last normalization.
This patch adds helper bpf_perf_event_read_value for kprobed based perf event array map, to read perf counter and enabled/running time. The enabled/running time is accumulated since the perf event open. To achieve scaling factor between two bpf invocations, users can can use cpu_id as the key (which is typical for perf array usage model) to remember the previous value and do the calculation inside the bpf program.
--- include/uapi/linux/bpf.h | 21 +++++++++++++++++++-- kernel/bpf/verifier.c | 4 +++- kernel/trace/bpf_trace.c | 45 +++++++++++++++++++++++++++++++++++++++++---- 3 files changed, 63 insertions(+), 7 deletions(-)
include/uapi/linux/bpf.h +19 -2
@@ -641,6 +641,14 @@ union bpf_attr {* @xdp_md: pointer to xdp_md * @delta: An positive/negative integer to be added to xdp_md.data_meta * Return: 0 on success or negative on error +* +* int bpf_perf_event_read_value(map, flags, buf, buf_size) +* read perf event counter value and perf event enabled/running time +* @map: pointer to perf_event_array map +* @flags: index of event in the map or bitmask flags +* @buf: buf to fill +* @buf_size: size of the buf +* Return: 0 on success or negative error code */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \
@@ -697,7 +705,8 @@ union bpf_attr {FN (redirect_map), \FN (sk_redirect_map), \FN (sock_map_update), \ -FN (xdp_adjust_meta), +FN (xdp_adjust_meta), \ +FN (perf_event_read_value),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call
@@ -741,7 +750,9 @@ enum bpf_func_id {#define BPF_F_ZERO_CSUM_TX (1ULL << 1)#define BPF_F_DONT_FRAGMENT (1ULL << 2) -/* BPF_FUNC_perf_event_output and BPF_FUNC_perf_event_read flags. */ +/* BPF_FUNC_perf_event_output, BPF_FUNC_perf_event_read and +* BPF_FUNC_perf_event_read_value flags. +*/ #define BPF_F_INDEX_MASK 0xffffffffULL#define BPF_F_CURRENT_CPU BPF_F_INDEX_MASK/* BPF_FUNC_perf_event_output for sk_buff input context. */
@@ -934,4 +945,10 @@ enum {#define TCP_BPF_IW 1001/* Set TCP initial congestion window */ #define TCP_BPF_SNDCWND_CLAMP 1002/* Set sndcwnd_clamp */ +struct bpf_perf_event_value { +__u64 counter; +__u64 enabled; +__u64 running; +}; +#endif /* _UAPI__LINUX_BPF_H__ */
kernel/bpf/verifier.c +3 -1
@@ -1552,7 +1552,8 @@ static int check_map_func_compatibility(struct bpf_map *map, int func_id)break ;case BPF_MAP_TYPE_PERF_EVENT_ARRAY :if (func_id != BPF_FUNC_perf_event_read && - func_id != BPF_FUNC_perf_event_output) + func_id != BPF_FUNC_perf_event_output && + func_id != BPF_FUNC_perf_event_read_value) goto error;break ;case BPF_MAP_TYPE_STACK_TRACE :
@@ -1595,6 +1596,7 @@ static int check_map_func_compatibility(struct bpf_map *map, int func_id)break ;case BPF_FUNC_perf_event_read:case BPF_FUNC_perf_event_output: +case BPF_FUNC_perf_event_read_value:if (map->map_type !=BPF_MAP_TYPE_PERF_EVENT_ARRAY ) goto error;break ;
kernel/trace/bpf_trace.c +41 -4
74 lines not on this page · raw message
Signature · 1 line
2.9.5
PATCH v7 3/5 bpf: add a test case for helper bpf_perf_event_read_value
In reply to Yonghong SongThe bpf sample program tracex6 is enhanced to use the new helper to read enabled/running time as well.
--- samples/bpf/tracex6_kern.c | 26 ++++++++++++++++++++++++++ samples/bpf/tracex6_user.c | 13 ++++++++++++- tools/include/uapi/linux/bpf.h | 3 ++- tools/testing/selftests/bpf/bpf_helpers.h | 3 +++ 4 files changed, 43 insertions(+), 2 deletions(-)
samples/bpf/tracex6_kern.c +26 -0
@@ -15,6 +15,12 @@ struct bpf_map_def SEC("maps") values = { .value_size = sizeof(u64), .max_entries = 64, }; +struct bpf_map_def SEC("maps") values2 = { + .type = BPF_MAP_TYPE_HASH, + .key_size = sizeof(int), + .value_size = sizeof(struct bpf_perf_event_value), + .max_entries = 64, +}; SEC("kprobe/htab_map_get_next_key")int bpf_prog1 (struct pt_regs *ctx)
@@ -37,5 +43,25 @@ int bpf_prog1(struct pt_regs *ctx)return 0 ; } +SEC("kprobe/htab_map_lookup_elem") +int bpf_prog2 (struct pt_regs *ctx) +{ +u32 key =bpf_get_smp_processor_id (); +struct bpf_perf_event_value *val, buf; +int error; + + error =bpf_perf_event_read_value (&counters, key, &buf,sizeof (buf)); +if (error) +return 0 ; + + val =bpf_map_lookup_elem (&values2, &key); +if (val) + *val = buf; +else +bpf_map_update_elem (&values2, &key, &buf,BPF_NOEXIST ); + +return 0 ; +} + char _license[] SEC("license") = "GPL"; u32 _version SEC("version") = LINUX_VERSION_CODE;
samples/bpf/tracex6_user.c +12 -1
@@ -22,6 +22,7 @@static void check_on_cpu (int cpu,struct perf_event_attr *attr) { +struct bpf_perf_event_value value2;int pmu_fd, error =0 ;cpu_set_t set;__u64 value;
@@ -46,8 +47,18 @@ static void check_on_cpu(int cpu, struct perf_event_attr *attr)fprintf (stderr,"Value missing for CPU %d\n" , cpu); error =1 ; goto on_exit; + }else { +fprintf (stderr,"CPU %d: %llu\n" , cpu, value); + } +/* The above bpf_map_lookup_elem should trigger the second kprobe */ +if (bpf_map_lookup_elem (map_fd[2 ], &cpu, &value2)) { +fprintf (stderr,"Value2 missing for CPU %d\n" , cpu); + error =1 ; + goto on_exit; + }else { +fprintf (stderr,"CPU %d: counter: %llu, enabled: %llu, running: %llu\n" , cpu, + value2.counter, value2.enabled, value2.running); } -fprintf (stderr,"CPU %d: %llu\n" , cpu, value); on_exit:assert (bpf_map_delete_elem (map_fd[0 ], &cpu) ==0 || error);
tools/include/uapi/linux/bpf.h +2 -1
@@ -697,7 +697,8 @@ union bpf_attr {FN (redirect_map), \FN (sk_redirect_map), \FN (sock_map_update), \ -FN (xdp_adjust_meta), +FN (xdp_adjust_meta), \ +FN (perf_event_read_value),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call
tools/testing/selftests/bpf/bpf_helpers.h +3 -0
@@ -72,6 +72,9 @@ static int (*bpf_sk_redirect_map)(void *map, int key, int flags) =static int (*bpf_sock_map_update)(void *map,void *key,void *value,unsigned long long flags) = (void *) BPF_FUNC_sock_map_update; +static int (*bpf_perf_event_read_value)(void *map,unsigned long long flags, +void *buf,unsigned int buf_size) = + (void *) BPF_FUNC_perf_event_read_value;/* llvm builtin functions that eBPF C program may use to
Signature · 1 line
2.9.5
PATCH v7 4/5 bpf: add helper bpf_perf_prog_read_value
In reply to Yonghong SongThis patch adds helper bpf_perf_prog_read_cvalue for perf event based bpf programs, to read event counter and enabled/running time.
The enabled/running time is accumulated since the perf event open.
The typical use case for perf event based bpf program is to attach itself to a single event. In such cases, if it is desirable to get scaling factor between two bpf invocations, users can can save the time values in a map, and use the value from the map and the current value to calculate the scaling factor.
--- include/uapi/linux/bpf.h | 10 +++++++++- kernel/trace/bpf_trace.c | 28 ++++++++++++++++++++++++++++ 2 files changed, 37 insertions(+), 1 deletion(-)
include/uapi/linux/bpf.h +9 -1
@@ -649,6 +649,13 @@ union bpf_attr {* @buf: buf to fill * @buf_size: size of the buf * Return: 0 on success or negative error code +* +* int bpf_perf_prog_read_value(ctx, buf, buf_size) +* read perf prog attached perf event counter and enabled/running time +* @ctx: pointer to ctx +* @buf: buf to fill +* @buf_size: size of the buf +* Return : 0 on success or negative error code */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \
@@ -706,7 +713,8 @@ union bpf_attr {FN (sk_redirect_map), \FN (sock_map_update), \FN (xdp_adjust_meta), \ -FN (perf_event_read_value), +FN (perf_event_read_value), \ +FN (perf_prog_read_value),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call
kernel/trace/bpf_trace.c +28 -0
@@ -613,6 +613,32 @@ static const struct bpf_func_proto bpf_get_stackid_proto_tp = { .arg3_type = ARG_ANYTHING, }; +BPF_CALL_3(bpf_perf_prog_read_value_tp, struct bpf_perf_event_data_kern *, ctx, + struct bpf_perf_event_value *, buf, u32, size) +{ +int err = -EINVAL ; + +if (unlikely (size !=sizeof (struct bpf_perf_event_value ))) + goto clear; + err =perf_event_read_local (ctx->event, &buf->counter, &buf->enabled, + &buf->running); +if (unlikely (err)) + goto clear; +return 0 ; +clear: +memset (buf,0 , size); +return err; +} + +static const struct bpf_func_proto bpf_perf_prog_read_value_proto_tp = { + .func = bpf_perf_prog_read_value_tp, + .gpl_only = true, + .ret_type =RET_INTEGER , + .arg1_type =ARG_PTR_TO_CTX , + .arg2_type =ARG_PTR_TO_UNINIT_MEM , + .arg3_type =ARG_CONST_SIZE , +}; +static const struct bpf_func_proto *tp_prog_func_proto (enum bpf_func_id func_id) {switch (func_id) {
@@ -620,6 +646,8 @@ static const struct bpf_func_proto *tp_prog_func_proto(enum bpf_func_id func_id)return &bpf_perf_event_output_proto_tp;case BPF_FUNC_get_stackid:return &bpf_get_stackid_proto_tp; +case BPF_FUNC_perf_prog_read_value: +return &bpf_perf_prog_read_value_proto_tp;default :return tracing_func_proto (func_id); }
Signature · 1 line
2.9.5
PATCH v7 5/5 bpf: add a test case for helper bpf_perf_prog_read_value
In reply to Yonghong SongThe bpf sample program trace_event is enhanced to use the new helper to print out enabled/running time.
--- samples/bpf/trace_event_kern.c | 10 ++++++++++ samples/bpf/trace_event_user.c | 13 ++++++++----- tools/include/uapi/linux/bpf.h | 3 ++- tools/testing/selftests/bpf/bpf_helpers.h | 3 +++ 4 files changed, 23 insertions(+), 6 deletions(-)
samples/bpf/trace_event_kern.c +10 -0
@@ -37,10 +37,14 @@ struct bpf_map_def SEC("maps") stackmap = { SEC("perf_event")int bpf_prog1 (struct bpf_perf_event_data *ctx) { +char time_fmt1[] ="Time Enabled: %llu, Time Running: %llu" ; +char time_fmt2[] ="Get Time Failed, ErrCode: %d" ;char fmt[] ="CPU-%d period %lld ip %llx" ;u32 cpu =bpf_get_smp_processor_id (); +struct bpf_perf_event_value value_buf;struct key_t key;u64 *val, one =1 ; +int ret;if (ctx->sample_period <10000 )/* ignore warmup */
@@ -54,6 +58,12 @@ int bpf_prog1(struct bpf_perf_event_data *ctx)return 0 ; } + ret =bpf_perf_prog_read_value (ctx, (void *)&value_buf,sizeof (struct bpf_perf_event_value )); +if (!ret) +bpf_trace_printk (time_fmt1,sizeof (time_fmt1), value_buf.enabled, value_buf.running); +else +bpf_trace_printk (time_fmt2,sizeof (time_fmt2), ret); + val =bpf_map_lookup_elem (&counts, &key);if (val) (*val)++;
samples/bpf/trace_event_user.c +8 -5
@@ -127,6 +127,9 @@ static void test_perf_event_all_cpu(struct perf_event_attr *attr)int *pmu_fd =malloc (nr_cpus *sizeof (int ));int i, error =0 ; +/* system wide perf event, no need to inherit */ + attr->inherit =0 ; +/* open perf_event on all cpus */ for (i =0 ; i < nr_cpus; i++) { pmu_fd[i] =sys_perf_event_open (attr,-1 , i,-1 ,0 );
@@ -154,6 +157,11 @@ static void test_perf_event_task(struct perf_event_attr *attr) {int pmu_fd; +/* per task perf event, enable inherit so the "dd ..." command can be traced properly. +* Enabling inherit will cause bpf_perf_prog_read_time helper failure. +*/ + attr->inherit =1 ; +/* open task bound event */ pmu_fd =sys_perf_event_open (attr,0 ,-1 ,-1 ,0 );if (pmu_fd <0 ) {
@@ -175,14 +183,12 @@ static void test_bpf_perf_event(void) .freq = 1, .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES, - .inherit = 1, };struct perf_event_attr attr_type_sw = { .sample_freq =SAMPLE_FREQ , .freq =1 , .type =PERF_TYPE_SOFTWARE , .config =PERF_COUNT_SW_CPU_CLOCK , - .inherit =1 , };struct perf_event_attr attr_hw_cache_l1d = { .sample_freq =SAMPLE_FREQ ,
@@ -192,7 +198,6 @@ static void test_bpf_perf_event(void)PERF_COUNT_HW_CACHE_L1D | (PERF_COUNT_HW_CACHE_OP_READ <<8 ) | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16), - .inherit = 1, };struct perf_event_attr attr_hw_cache_branch_miss = { .sample_freq = SAMPLE_FREQ,
@@ -202,7 +207,6 @@ static void test_bpf_perf_event(void)PERF_COUNT_HW_CACHE_BPU | (PERF_COUNT_HW_CACHE_OP_READ <<8 ) | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16), - .inherit = 1, };struct perf_event_attr attr_type_raw = { .sample_freq = SAMPLE_FREQ,
@@ -210,7 +214,6 @@ static void test_bpf_perf_event(void) .type = PERF_TYPE_RAW, /* Intel Instruction Retired */ .config = 0xc0, - .inherit = 1, };printf ("Test HW_CPU_CYCLES\n" );
tools/include/uapi/linux/bpf.h +2 -1
@@ -698,7 +698,8 @@ union bpf_attr {FN (sk_redirect_map), \FN (sock_map_update), \FN (xdp_adjust_meta), \ -FN (perf_event_read_value), +FN (perf_event_read_value), \ +FN (perf_prog_read_value),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call
tools/testing/selftests/bpf/bpf_helpers.h +3 -0
@@ -75,6 +75,9 @@ static int (*bpf_sock_map_update)(void *map, void *key, void *value,static int (*bpf_perf_event_read_value)(void *map,unsigned long long flags,void *buf,unsigned int buf_size) = (void *) BPF_FUNC_perf_event_read_value; +static int (*bpf_perf_prog_read_value)(void *ctx,void *buf, +unsigned int buf_size) = + (void *) BPF_FUNC_perf_prog_read_value;/* llvm builtin functions that eBPF C program may use to
Signature · 1 line
2.9.5
David Miller
maintainerFrom: Yonghong Song <[email protected]>
Date: Thu, 5 Oct 2017 09:19:18 -0700
Quoted 16 lines
Hardware pmu counters are limited resources. When there are more pmu based perf events opened than available counters, kernel will multiplex these events so each event gets certain percentage (but not 100%) of the pmu time. In case that multiplexing happens, the number of samples or counter value will not reflect the case compared to no multiplexing. This makes comparison between different runs difficult.
Typically, the number of samples or counter value should be normalized before comparing to other experiments. The typical normalization is done like:
normalized_num_samples=num_samples*time_enabled/time_runningnormalized_counter_value=counter_value*time_enabled/time_running
wheretime_enabledis the time enabled for event andtime_runningis the time running for event since last normalization.
Quoted 4 lines
This patch set implements two helper functions.
The helperbpf_perf_event_read_valuereads counter/time_enabled/time_running for perf event array map. The helperbpf_perf_prog_read_valueread counter/time_enabled/time_running for bpf prog with typeBPF_PROG_TYPE_PERF_EVENT.
Series applied to net-next, thanks.