perf lock contention: Support task/addr aggregation mode (v1)
Subsystems (2)
Namhyung Kim
maintainercover perf lock contention: Support task/addr aggregation mode (v1)
Hello,
This patchset adds two more aggregation modes for perf lock contention.
The first one is the per-task mode with -t/--threads option. The option was there already but it remained broken with BPF for a while. Now it supports the output with and without BPF.
# perf lock contention -a -b -t -- sleep 1 contended total wait max wait avg wait pid comm
11 11.85 us 2.23 us 1.08 us 0 swapper
2 11.13 us 10.22 us 5.56 us 749053 ThreadPoolForeg
1 8.15 us 8.15 us 8.15 us 321353 Chrome_ChildIOT
2 2.73 us 1.77 us 1.37 us 320761 Chrome_ChildIOT
1 1.40 us 1.40 us 1.40 us 320502 chrome
1 379 ns 379 ns 379 ns 321227 chromeThe other one is the per-lock-instance mode with -l/--lock-addr option. If the lock has a symbol, it will be displayed as well.
# perf lock contention -a -b -l -- sleep 1 contended total wait max wait avg wait address symbol
3 4.79 us 2.33 us 1.60 us ffffffffbaed50c0 rcu_state
4 4.19 us 1.62 us 1.05 us ffffffffbae07a40 jiffies_lock
1 1.94 us 1.94 us 1.94 us ffff9262277861e0
1 387 ns 387 ns 387 ns ffff9260bfda4f60It's based on the current acme/tmp.perf/core branch. You can find the code in the 'perf/lock-con-aggr-v1' branch in
git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
Thanks,
Namhyung
Namhyung Kim (4):
perf lock contention: Add lock_data.h for common data perf lock contention: Implement -t/--threads option for BPF perf lock contention: Add -l/--lock-addr option perf test: Update perf lock contention test
tools/perf/Documentation/perf-lock.txt | 4 + tools/perf/builtin-lock.c | 95 ++++++++++++++----- tools/perf/tests/shell/lock_contention.sh | 48 ++++++++++ tools/perf/util/bpf_lock_contention.c | 72 ++++++++++---- .../perf/util/bpf_skel/lock_contention.bpf.c | 67 +++++++++---- tools/perf/util/bpf_skel/lock_data.h | 30 ++++++ tools/perf/util/lock-contention.h | 1 + 7 files changed, 255 insertions(+), 62 deletions(-) create mode 100644 tools/perf/util/bpf_skel/lock_data.h
base-commit: b22802e295a80ec16e355d7208d2fbbd7bbc1b7a
Signature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Namhyung Kim
maintainerPATCH 1/4 perf lock contention: Add lock_data.h for common data
Accessing BPF maps should use the same data types. Add bpf_skel/lock_data.h to define the common data structures. No functional changes.
--- tools/perf/util/bpf_lock_contention.c | 19 ++++-------- .../perf/util/bpf_skel/lock_contention.bpf.c | 17 ++--------- tools/perf/util/bpf_skel/lock_data.h | 30 +++++++++++++++++++ 3 files changed, 38 insertions(+), 28 deletions(-) create mode 100644 tools/perf/util/bpf_skel/lock_data.h
tools/perf/util/bpf_lock_contention.c +6 -13
@@ -12,17 +12,10 @@#include <bpf/bpf.h> #include "bpf_skel/lock_contention.skel.h" +#include "bpf_skel/lock_data.h" static struct lock_contention_bpf *skel; -struct lock_contention_data { -u64 total_time; -u64 min_time; -u64 max_time; -u32 count; -u32 flags; -}; -int lock_contention_prepare (struct lock_contention *con) {int i, fd;
@@ -110,8 +103,8 @@ int lock_contention_stop(void)int lock_contention_read (struct lock_contention *con) {int fd, stack, err =0 ; -s32 prev_key, key; -struct lock_contention_data data = {}; +struct contention_key *prev_key, key; +struct contention_data data = {};struct lock_stat *st =NULL ;struct machine *machine = con->machine;u64 *stack_trace;
@@ -126,8 +119,8 @@ int lock_contention_read(struct lock_contention *con)if (stack_trace ==NULL )return -1 ; - prev_key =0 ; -while (!bpf_map_get_next_key (fd, &prev_key, &key)) { + prev_key =NULL ; +while (!bpf_map_get_next_key (fd, prev_key, &key)) {struct map *kmap;struct symbol *sym;int idx =0 ;
@@ -184,7 +177,7 @@ int lock_contention_read(struct lock_contention *con) }hlist_add_head (&st->hash_entry, con->result); - prev_key = key; + prev_key = &key;/* we're fine now, reset the values */ st =NULL ;
tools/perf/util/bpf_skel/lock_contention.bpf.c +2 -15
@@ -5,24 +5,11 @@#include <bpf/bpf_tracing.h> #include <bpf/bpf_core_read.h> -/* maximum stack trace depth */ -#define MAX_STACKS 8 +#include "lock_data.h" /* default buffer size */ #define MAX_ENTRIES 10240 -struct contention_key { -__s32 stack_id; -}; - -struct contention_data { -__u64 total_time; -__u64 min_time; -__u64 max_time; -__u32 count; -__u32 flags; -}; -struct tstamp_data {__u64 timestamp;__u64 lock;
@@ -34,7 +21,7 @@ struct tstamp_data {struct { __uint(type, BPF_MAP_TYPE_STACK_TRACE); __uint(key_size, sizeof(__u32)); - __uint(value_size, MAX_STACKS * sizeof(__u64)); + __uint(value_size, sizeof(__u64)); __uint(max_entries, MAX_ENTRIES); } stacks SEC(".maps");
tools/perf/util/bpf_skel/lock_data.h new file +30 -0
@@ -0,0 +1,30 @@ +// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) +/* Data structures shared between BPF and tools. */ +#ifndef UTIL_BPF_SKEL_LOCK_DATA_H +#define UTIL_BPF_SKEL_LOCK_DATA_H + +struct contention_key { +s32 stack_or_task_id; +}; + +#define TASK_COMM_LEN 16 + +struct contention_task_data { +char comm[TASK_COMM_LEN ]; +}; + +struct contention_data { +u64 total_time; +u64 min_time; +u64 max_time; +u32 count; +u32 flags; +}; + +enum lock_aggr_mode { +LOCK_AGGR_ADDR =0 , +LOCK_AGGR_TASK , +LOCK_AGGR_CALLER , +}; + +#endif /* UTIL_BPF_SKEL_LOCK_DATA_H */
Signature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 1/4] perf lock contention: Add lock_data.h for common data
Em Fri, Dec 09, 2022 at 11:07:24AM -0800, Namhyung Kim escreveu:
Quoted 2 lines
Accessing BPF maps should use the same data types. Add
bpf_skel/lock_data.hto define the common data structures. No functional changes.
You forgot to update one of the stack_id users, that field got renamed:
util/bpf_skel/lock_contention.bpf.c:144:6: error: no member named 'stack_id' in 'struct contention_key'
key.stack_id = pelem->stack_id;
~~~ ^1 error generated.
make[2]: *** [Makefile.perf:1075: /tmp/build/perf/util/bpf_skel/.tmp/lock_contention.bpf.o] Error 1 make[1]: *** [Makefile.perf:236: sub-make] Error 2
make: *** [Makefile:113: install-bin] Error 2
make: Leaving directory '/var/home/acme/git/perf/tools/perf'
Performance counter stats for 'make -k NO_LIBTRACEEVENT=1 BUILD_BPF_SKEL=1 CORESIGHT=1 O=/tmp/build/perf -C tools/perf install-bin':
7,005,216,342 cycles:u
11,851,225,594 instructions:u # 1.69 insn per cycle
3.168945139 seconds time elapsed
1.730964000 seconds user
1.578932000 seconds sys⬢[acme@toolbox perf]$ git log --oneline -4f6e7a5f1db49dc8e (HEAD) perf lock contention: Add lock_data.h for common data 5d9b55713c5c037f perf python: Account for multiple words in CCd9078bf3f3320457 perf off_cpu: Fix a typo in BTF tracepoint name, it should be 'btf_trace_sched_switch' 3b7ea76f0f7844f5 perf test: Update event group check for support of uncore event ⬢[acme@toolbox perf]$
After some point it builds.
I'm fixing this to keep it bisectable.
- Arnaldo
Quoted 139 lines
Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/util/bpf_lock_contention.c | 19 ++++-------- .../perf/util/bpf_skel/lock_contention.bpf.c | 17 ++--------- tools/perf/util/bpf_skel/lock_data.h | 30 +++++++++++++++++++ 3 files changed, 38 insertions(+), 28 deletions(-) create mode 100644 tools/perf/util/bpf_skel/lock_data.htools/perf/util/bpf_lock_contention.c +6
-13@@ -12,17 +12,10 @@#include <bpf/bpf.h> #include "bpf_skel/lock_contention.skel.h" +#include "bpf_skel/lock_data.h" static struct lock_contention_bpf *skel; -struct lock_contention_data { -u64 total_time; -u64 min_time; -u64 max_time; -u32 count; -u32 flags; -}; -int lock_contention_prepare (struct lock_contention *con) {int i, fd;@@ -110,8 +103,8 @@ int lock_contention_stop(void)int lock_contention_read (struct lock_contention *con) {int fd, stack, err =0 ; -s32 prev_key, key; -struct lock_contention_data data = {}; +struct contention_key *prev_key, key; +struct contention_data data = {};struct lock_stat *st =NULL ;struct machine *machine = con->machine;u64 *stack_trace;@@ -126,8 +119,8 @@ int lock_contention_read(struct lock_contention *con)if (stack_trace ==NULL )return -1 ; - prev_key =0 ; -while (!bpf_map_get_next_key (fd, &prev_key, &key)) { + prev_key =NULL ; +while (!bpf_map_get_next_key (fd, prev_key, &key)) {struct map *kmap;struct symbol *sym;int idx =0 ;@@ -184,7 +177,7 @@ int lock_contention_read(struct lock_contention *con) }hlist_add_head (&st->hash_entry, con->result); - prev_key = key; + prev_key = &key;/* we're fine now, reset the values */ st =NULL ;tools/perf/util/bpf_skel/lock_contention.bpf.c +2
-15@@ -5,24 +5,11 @@#include <bpf/bpf_tracing.h> #include <bpf/bpf_core_read.h> -/* maximum stack trace depth */ -#define MAX_STACKS 8 +#include "lock_data.h" /* default buffer size */ #define MAX_ENTRIES 10240 -struct contention_key { -__s32 stack_id; -}; - -struct contention_data { -__u64 total_time; -__u64 min_time; -__u64 max_time; -__u32 count; -__u32 flags; -}; -struct tstamp_data {__u64 timestamp;__u64 lock;@@ -34,7 +21,7 @@ struct tstamp_data {struct { __uint(type, BPF_MAP_TYPE_STACK_TRACE); __uint(key_size, sizeof(__u32)); - __uint(value_size, MAX_STACKS * sizeof(__u64)); + __uint(value_size, sizeof(__u64)); __uint(max_entries, MAX_ENTRIES); } stacks SEC(".maps");tools/perf/util/bpf_skel/lock_data.h new file +30
-0@@ -0,0 +1,30 @@ +// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) +/* Data structures shared between BPF and tools. */ +#ifndef UTIL_BPF_SKEL_LOCK_DATA_H +#define UTIL_BPF_SKEL_LOCK_DATA_H + +struct contention_key { +s32 stack_or_task_id; +}; + +#define TASK_COMM_LEN 16 + +struct contention_task_data { +char comm[TASK_COMM_LEN ]; +}; + +struct contention_data { +u64 total_time; +u64 min_time; +u64 max_time; +u32 count; +u32 flags; +}; + +enum lock_aggr_mode { +LOCK_AGGR_ADDR =0 , +LOCK_AGGR_TASK , +LOCK_AGGR_CALLER , +}; + +#endif /* UTIL_BPF_SKEL_LOCK_DATA_H */ Signature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Signature · 1 line
- Arnaldo
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 1/4] perf lock contention: Add lock_data.h for common data
Em Mon, Dec 12, 2022 at 04:42:30PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 33 lines
Em Fri, Dec 09, 2022 at 11:07:24AM -0800, Namhyung Kim escreveu:
Quoted 2 lines
Accessing BPF maps should use the same data types. Add
bpf_skel/lock_data.hto define the common data structures. No functional changes.You forgot to update one of the
stack_idusers, that field got renamed:
util/bpf_skel/lock_contention.bpf.c:144:6: error: no member named 'stack_id' in 'structcontention_key'key.stack_id = pelem->stack_id; ~~~ ^1 error generated.
make[2]: *** [Makefile.perf:1075:/tmp/build/perf/util/bpf_skel/.tmp/lock_contention.bpf.o] Error 1 make[1]: *** [Makefile.perf:236: sub-make] Error 2
make: *** [Makefile:113: install-bin] Error 2
make: Leaving directory '/var/home/acme/git/perf/tools/perf'Performance counter stats for 'make -k NO_LIBTRACEEVENT=1 BUILD_BPF_SKEL=1 CORESIGHT=1 O=/tmp/build/perf -C
tools/perfinstall-bin':7,005,216,342 cycles:u 11,851,225,594 instructions:u # 1.69 insn per cycle 3.168945139 seconds time elapsed 1.730964000 seconds user 1.578932000 seconds sys⬢[acme@toolbox perf]$ git log --oneline -4
f6e7a5f1db49dc8e(HEAD) perf lock contention: Addlock_data.hfor common data5d9b55713c5c037fperf python: Account for multiple words in CCd9078bf3f3320457perfoff_cpu: Fix a typo in BTF tracepoint name, it should be 'btf_trace_sched_switch'3b7ea76f0f7844f5perf test: Update event group check for support of uncore event ⬢[acme@toolbox perf]$
Quoted 3 lines
After some point it builds.
I'm fixing this to keep it bisectable.
I folded this:
tools/perf/util/bpf_skel/lock_contention.bpf.c +1 -1
@@ -141,7 +141,7 @@ int contention_end(u64 *ctx) duration =bpf_ktime_get_ns () - pelem->timestamp; - key.stack_id = pelem->stack_id; + key.stack_or_task_id = pelem->stack_id; data =bpf_map_lookup_elem (&lock_stat, &key);if (!data) {struct contention_data first = {
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 1/4] perf lock contention: Add lock_data.h for common data
Em Mon, Dec 12, 2022 at 04:43:43PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 45 lines
Em Mon, Dec 12, 2022 at 04:42:30PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 33 lines
Em Fri, Dec 09, 2022 at 11:07:24AM -0800, Namhyung Kim escreveu:
Quoted 2 lines
Accessing BPF maps should use the same data types. Add
bpf_skel/lock_data.hto define the common data structures. No functional changes.You forgot to update one of the
stack_idusers, that field got renamed:
util/bpf_skel/lock_contention.bpf.c:144:6: error: no member named 'stack_id' in 'structcontention_key'key.stack_id = pelem->stack_id; ~~~ ^1 error generated.
make[2]: *** [Makefile.perf:1075:/tmp/build/perf/util/bpf_skel/.tmp/lock_contention.bpf.o] Error 1 make[1]: *** [Makefile.perf:236: sub-make] Error 2
make: *** [Makefile:113: install-bin] Error 2
make: Leaving directory '/var/home/acme/git/perf/tools/perf'Performance counter stats for 'make -k NO_LIBTRACEEVENT=1 BUILD_BPF_SKEL=1 CORESIGHT=1 O=/tmp/build/perf -C
tools/perfinstall-bin':7,005,216,342 cycles:u 11,851,225,594 instructions:u # 1.69 insn per cycle 3.168945139 seconds time elapsed 1.730964000 seconds user 1.578932000 seconds sys⬢[acme@toolbox perf]$ git log --oneline -4
f6e7a5f1db49dc8e(HEAD) perf lock contention: Addlock_data.hfor common data5d9b55713c5c037fperf python: Account for multiple words in CCd9078bf3f3320457perfoff_cpu: Fix a typo in BTF tracepoint name, it should be 'btf_trace_sched_switch'3b7ea76f0f7844f5perf test: Update event group check for support of uncore event ⬢[acme@toolbox perf]$Quoted 3 lines
After some point it builds.
I'm fixing this to keep it bisectable.
I folded this:
tools/perf/util/bpf_skel/lock_contention.bpf.c +0
-0
tools/perf/util/bpf_skel/lock_contention.bpf.c int contention_end(u64 *ctx) · line 147
Quoted 8 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +1
-1@@ -141,7 +141,7 @@ int contention_end(u64 *ctx) duration =bpf_ktime_get_ns () - pelem->timestamp; - key.stack_id = pelem->stack_id; + key.stack_or_task_id = pelem->stack_id; data =bpf_map_lookup_elem (&lock_stat, &key);if (!data) {struct contention_data first = {
And then fixed up this:
Could not apply 3d4947c7bd10beba... perf lock contention: Implement -t/--threads option for BPF ⬢[acme@toolbox perf]$
⬢[acme@toolbox perf]$
⬢[acme@toolbox perf]$ git diff
tools/perf/util/bpf_skel/lock_contention.bpf.c +0 -0
@@@ -141,7 -168,13 +168,17 @@@ int contention_end(u64 *ctx
duration = bpf_ktime_get_ns() - pelem->timestamp;
++<<<<<<< HEAD
+ key.stack_or_task_id = pelem->stack_id;
++=======
+ if (aggr_mode == LOCK_AGGR_CALLER) {
+ key.stack_or_task_id = pelem->stack_id;
+ } else {
+ key.stack_or_task_id = pid;
+ update_task_data(pid);
+ }
+
++>>>>>>> 3d4947c7bd10beba (perf lock contention: Implement -t/--threads option for BPF)
Namhyung Kim
maintainerRe: [PATCH 1/4] perf lock contention: Add lock_data.h for common data
Hi Arnaldo,
Quoted 17 lines · Arnaldo Carvalho de Melo
Em Mon, Dec 12, 2022 at 04:43:43PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 15 lines
Em Mon, Dec 12, 2022 at 04:42:30PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 14 lines
Em Fri, Dec 09, 2022 at 11:07:24AM -0800, Namhyung Kim escreveu:
Quoted 2 lines
Accessing BPF maps should use the same data types. Add
bpf_skel/lock_data.hto define the common data structures. No functional changes.You forgot to update one of the
stack_idusers, that field got renamed:
util/bpf_skel/lock_contention.bpf.c:144:6: error: no member named 'stack_id' in 'structcontention_key'key.stack_id = pelem->stack_id; ~~~ ^1 error generated.
make[2]: *** [Makefile.perf:1075:/tmp/build/perf/util/bpf_skel/.tmp/lock_contention.bpf.o] Error 1 make[1]: *** [Makefile.perf:236: sub-make] Error 2
make: *** [Makefile:113: install-bin] Error 2
make: Leaving directory '/var/home/acme/git/perf/tools/perf'
Oops, right.
Quoted 31 lines
Quoted 31 lines
Quoted 19 lines
Performance counter stats for 'make -k NO_LIBTRACEEVENT=1 BUILD_BPF_SKEL=1 CORESIGHT=1 O=/tmp/build/perf -C
tools/perfinstall-bin':7,005,216,342 cycles:u 11,851,225,594 instructions:u # 1.69 insn per cycle 3.168945139 seconds time elapsed 1.730964000 seconds user 1.578932000 seconds sys⬢[acme@toolbox perf]$ git log --oneline -4
f6e7a5f1db49dc8e(HEAD) perf lock contention: Addlock_data.hfor common data5d9b55713c5c037fperf python: Account for multiple words in CCd9078bf3f3320457perfoff_cpu: Fix a typo in BTF tracepoint name, it should be 'btf_trace_sched_switch'3b7ea76f0f7844f5perf test: Update event group check for support of uncore event ⬢[acme@toolbox perf]$Quoted 3 lines
After some point it builds.
I'm fixing this to keep it bisectable.
I folded this:
tools/perf/util/bpf_skel/lock_contention.bpf.c +0
-0@@ -141,7 +141,7 @@ int contention_end(u64 *ctx)
Quoted 7 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c int contention_end(u64 *ctx) · line 147
Quoted 7 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +1
-1duration =bpf_ktime_get_ns () - pelem->timestamp; - key.stack_id = pelem->stack_id; + key.stack_or_task_id = pelem->stack_id; data =bpf_map_lookup_elem (&lock_stat, &key);if (!data) {struct contention_data first = {
Thanks for fixing this.
Quoted 17 lines
And then fixed up this:
Could not apply
3d4947c7bd10beba... perf lock contention: Implement -t/--threads option for BPF ⬢[acme@toolbox perf]$
⬢[acme@toolbox perf]$
⬢[acme@toolbox perf]$ git difftools/perf/util/bpf_skel/lock_contention.bpf.c +1
-0@@@ -141,7 -168,13 +168,17 @@@ int contention_end(u64 *ctx duration =bpf_ktime_get_ns () - pelem->timestamp; ++<<<<<<< HEAD
tools/perf/util/bpf_skel/lock_contention.bpf.c
Quoted 10 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +9
-0+ key.stack_or_task_id = pelem->stack_id; ++======= +if (aggr_mode ==LOCK_AGGR_CALLER ) { + key.stack_or_task_id = pelem->stack_id; + }else { + key.stack_or_task_id = pid; +update_task_data (pid); + } + ++>>>>>>> 3d4947c7bd10beba (perf lock contention: Implement -t/--threads option for BPF)
Sure, it looks good to me.
Thanks,
Namhyung
Namhyung Kim
maintainerPATCH 2/4 perf lock contention: Implement -t/--threads option for BPF
In reply to Namhyung KimThe BPF didn't show the per-thread stat properly. Use task's thread id (PID) as a key instead of stack_id and add a task_data map to save task comm names.
$ sudo ./perf lock con -abt -E 5 sleep 1 contended total wait max wait avg wait pid comm
1 740.66 ms 740.66 ms 740.66 ms 1950 nv_queue
3 305.50 ms 298.19 ms 101.83 ms 1884 nvidia-modeset/
1 25.14 us 25.14 us 25.14 us 2725038 EventManager_De
12 23.09 us 9.30 us 1.92 us 0 swapper
1 20.18 us 20.18 us 20.18 us 2725033 EventManager_De--- tools/perf/builtin-lock.c | 13 ++---- tools/perf/util/bpf_lock_contention.c | 40 ++++++++++++++++-- .../perf/util/bpf_skel/lock_contention.bpf.c | 41 +++++++++++++++++-- tools/perf/util/lock-contention.h | 1 + 4 files changed, 78 insertions(+), 17 deletions(-)
tools/perf/builtin-lock.c +3 -10
@@ -12,6 +12,7 @@#include "util/target.h" #include "util/callchain.h" #include "util/lock-contention.h" +#include "util/bpf_skel/lock_data.h" #include <subcmd/pager.h> #include <subcmd/parse-options.h>
@@ -61,11 +62,7 @@ static int max_stack_depth = CONTENTION_STACK_DEPTH;static int stack_skip =CONTENTION_STACK_SKIP ;static int print_nr_entries =INT_MAX /2 ; -static enum { -LOCK_AGGR_ADDR , -LOCK_AGGR_TASK , -LOCK_AGGR_CALLER , -} aggr_mode =LOCK_AGGR_ADDR ; +static enum lock_aggr_mode aggr_mode =LOCK_AGGR_ADDR ;static struct thread_stat *thread_stat_find (u32 tid) {
@@ -1619,6 +1616,7 @@ static int __cmd_contention(int argc, const char **argv) .map_nr_entries = bpf_map_entries, .max_stack = max_stack_depth, .stack_skip = stack_skip, + .aggr_mode = show_thread_stats ? LOCK_AGGR_TASK : LOCK_AGGR_CALLER, }; session =perf_session__new (use_bpf ?NULL : &data, &eops);
@@ -1691,11 +1689,6 @@ static int __cmd_contention(int argc, const char **argv)if (select_key (true)) goto out_delete; -if (show_thread_stats) - aggr_mode =LOCK_AGGR_TASK ; -else - aggr_mode =LOCK_AGGR_CALLER ; -if (use_bpf) {lock_contention_start ();if (argc)
tools/perf/util/bpf_lock_contention.c +37 -3
@@ -5,6 +5,7 @@#include "util/map.h" #include "util/symbol.h" #include "util/target.h" +#include "util/thread.h" #include "util/thread_map.h" #include "util/lock-contention.h" #include <linux/zalloc.h>
@@ -30,10 +31,17 @@ int lock_contention_prepare(struct lock_contention *con) }bpf_map__set_value_size (skel->maps.stacks, con->max_stack *sizeof (u64)); -bpf_map__set_max_entries (skel->maps.stacks, con->map_nr_entries);bpf_map__set_max_entries (skel->maps.lock_stat, con->map_nr_entries);bpf_map__set_max_entries (skel->maps.tstamp, con->map_nr_entries); +if (con->aggr_mode ==LOCK_AGGR_TASK ) { +bpf_map__set_max_entries (skel->maps.task_data, con->map_nr_entries); +bpf_map__set_max_entries (skel->maps.stacks,1 ); + }else { +bpf_map__set_max_entries (skel->maps.task_data,1 ); +bpf_map__set_max_entries (skel->maps.stacks, con->map_nr_entries); + } +if (target__has_cpu (target)) ncpus =perf_cpu_map__nr (evlist->core.user_requested_cpus);if (target__has_task (target))
@@ -82,7 +90,9 @@ int lock_contention_prepare(struct lock_contention *con)bpf_map_update_elem (fd, &pid, &val,BPF_ANY ); } +/* these don't work well if in the rodata section */ skel->bss->stack_skip = con->stack_skip; + skel->bss->aggr_mode = con->aggr_mode;lock_contention_bpf__attach (skel);return 0 ;
@@ -102,7 +112,7 @@ int lock_contention_stop(void)int lock_contention_read (struct lock_contention *con) { -int fd, stack, err =0 ; +int fd, stack, task_fd, err =0 ;struct contention_key *prev_key, key;struct contention_data data = {};struct lock_stat *st =NULL ;
@@ -112,6 +122,7 @@ int lock_contention_read(struct lock_contention *con) fd =bpf_map__fd (skel->maps.lock_stat); stack =bpf_map__fd (skel->maps.stacks); + task_fd =bpf_map__fd (skel->maps.task_data); con->lost = skel->bss->lost;
@@ -119,6 +130,13 @@ int lock_contention_read(struct lock_contention *con)if (stack_trace ==NULL )return -1 ; +if (con->aggr_mode ==LOCK_AGGR_TASK ) { +struct thread *idle =__machine__findnew_thread (machine, +/*pid=*/ 0 , +/*tid=*/ 0 ); +thread__set_comm (idle,"swapper" ,/*timestamp=*/ 0 ); + } + prev_key =NULL ;while (!bpf_map_get_next_key (fd, prev_key, &key)) {struct map *kmap;
@@ -143,6 +161,22 @@ int lock_contention_read(struct lock_contention *con) st->flags = data.flags; +if (con->aggr_mode ==LOCK_AGGR_TASK ) { +struct contention_task_data task; +struct thread *t; + + st->addr = key.stack_or_task_id; + +/* do not update idle comm which contains CPU number */ +if (st->addr) { +bpf_map_lookup_elem (task_fd, &key, &task); + t =__machine__findnew_thread (machine,/*pid=*/ -1 , + key.stack_or_task_id); +thread__set_comm (t, task.comm,/*timestamp=*/ 0 ); + } + goto next; + } +bpf_map_lookup_elem (stack, &key, stack_trace);/* skip lock internal functions */
@@ -175,7 +209,7 @@ int lock_contention_read(struct lock_contention *con)if (st->callstack ==NULL )break ; } - +next:hlist_add_head (&st->hash_entry, con->result); prev_key = &key;
tools/perf/util/bpf_skel/lock_contention.bpf.c +37 -4
72 lines not on this page · raw message
tools/perf/util/lock-contention.h +1 -0
@@ -117,6 +117,7 @@ struct lock_contention {int lost;int max_stack;int stack_skip; +int aggr_mode; };#ifdef HAVE_BPF_SKEL
Signature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Namhyung Kim
maintainerPATCH 3/4 perf lock contention: Add -l/--lock-addr option
In reply to Namhyung KimThe -l/--lock-addr option is to implement per-lock-instance contention stat using LOCK_AGGR_ADDR. It displays lock address and optionally symbol name if exists.
$ sudo ./perf lock con -abl sleep 1 contended total wait max wait avg wait address symbol
1 36.28 us 36.28 us 36.28 us ffff92615d6448b8
9 10.91 us 1.84 us 1.21 us ffffffffbaed50c0 rcu_state
1 10.49 us 10.49 us 10.49 us ffff9262ac4f0c80
8 4.68 us 1.67 us 585 ns ffffffffbae07a40 jiffies_lock
3 3.03 us 1.45 us 1.01 us ffff9262277861e0
1 924 ns 924 ns 924 ns ffff926095ba9d20
1 436 ns 436 ns 436 ns ffff9260bfda4f60--- tools/perf/Documentation/perf-lock.txt | 4 + tools/perf/builtin-lock.c | 84 +++++++++++++++---- tools/perf/util/bpf_lock_contention.c | 23 +++-- .../perf/util/bpf_skel/lock_contention.bpf.c | 17 +++- tools/perf/util/bpf_skel/lock_data.h | 2 +- 5 files changed, 102 insertions(+), 28 deletions(-)
tools/perf/Documentation/perf-lock.txt +4 -0
@@ -168,6 +168,10 @@ CONTENTION OPTIONS --entries=<value>:: Display this many entries. +-l:: +--lock-addr:: + Show lock contention stat by address + SEE ALSO --------
tools/perf/builtin-lock.c +67 -17
152 lines not on this page · raw message
tools/perf/util/bpf_lock_contention.c +17 -6
@@ -137,11 +137,15 @@ int lock_contention_read(struct lock_contention *con)thread__set_comm (idle,"swapper" ,/*timestamp=*/ 0 ); } +/* make sure it loads the kernel map */ +map__load (maps__first (machine->kmaps)); + prev_key =NULL ;while (!bpf_map_get_next_key (fd, prev_key, &key)) {struct map *kmap;struct symbol *sym;int idx =0 ; +s32 stack_id;/* to handle errors in the loop body */ err =-1 ;
@@ -160,24 +164,31 @@ int lock_contention_read(struct lock_contention *con) st->avg_wait_time = data.total_time / data.count; st->flags = data.flags; + st->addr = key.aggr_key;if (con->aggr_mode ==LOCK_AGGR_TASK ) {struct contention_task_data task;struct thread *t; - - st->addr = key.stack_or_task_id; +int pid = key.aggr_key;/* do not update idle comm which contains CPU number */ if (st->addr) { -bpf_map_lookup_elem (task_fd, &key, &task); - t =__machine__findnew_thread (machine,/*pid=*/ -1 , - key.stack_or_task_id); +bpf_map_lookup_elem (task_fd, &pid, &task); + t =__machine__findnew_thread (machine,/*pid=*/ -1 , pid);thread__set_comm (t, task.comm,/*timestamp=*/ 0 ); } goto next; } -bpf_map_lookup_elem (stack, &key, stack_trace); +if (con->aggr_mode ==LOCK_AGGR_ADDR ) { + sym =machine__find_kernel_symbol (machine, st->addr, &kmap); +if (sym) + st->name =strdup (sym->name); + goto next; + } + + stack_id = key.aggr_key; +bpf_map_lookup_elem (stack, &stack_id, stack_trace);/* skip lock internal functions */ while (machine__is_lock_function(machine, stack_trace[idx]) &&
tools/perf/util/bpf_skel/lock_contention.bpf.c +13 -4
@@ -168,11 +168,20 @@ int contention_end(u64 *ctx) duration =bpf_ktime_get_ns () - pelem->timestamp; -if (aggr_mode ==LOCK_AGGR_CALLER ) { - key.stack_or_task_id = pelem->stack_id; - }else { - key.stack_or_task_id = pid; +switch (aggr_mode) { +case LOCK_AGGR_CALLER : + key.aggr_key = pelem->stack_id; +break ; +case LOCK_AGGR_TASK : + key.aggr_key = pid;update_task_data (pid); +break ; +case LOCK_AGGR_ADDR : + key.aggr_key = pelem->lock; +break ; +default : +/* should not happen */ +return 0 ; } data =bpf_map_lookup_elem (&lock_stat, &key);
tools/perf/util/bpf_skel/lock_data.h +1 -1
@@ -4,7 +4,7 @@#define UTIL_BPF_SKEL_LOCK_DATA_H struct contention_key { -s32 stack_or_task_id; +u64 aggr_key;/* can be stack_id, pid or lock addr */ };#define TASK_COMM_LEN 16
Signature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Namhyung Kim
maintainerPATCH 4/4 perf test: Update perf lock contention test
In reply to Namhyung KimAdd test cases for the task and addr aggregation modes.
$ sudo ./perf test -v contention 86: kernel lock contention analysis test : --- start --- test child forked, pid 680006 Testing perf lock record and perf lock contention Testing perf lock contention --use-bpf Testing perf lock record and perf lock contention at the same time Testing perf lock contention --threads Testing perf lock contention --lock-addr test child finished with 0 ---- end ---- kernel lock contention analysis test: Ok
--- tools/perf/tests/shell/lock_contention.sh | 48 +++++++++++++++++++++++ 1 file changed, 48 insertions(+)
tools/perf/tests/shell/lock_contention.sh +48 -0
@@ -77,10 +77,58 @@ test_record_concurrent() fi } +test_aggr_task() +{ + echo "Testing perf lock contention --threads" + perf lock contention -i ${perfdata} -t -E 1 -q 2> ${result} + if [ $(cat "${result}" | wc -l) != "1" ]; then + echo "[Fail] Recorded result count is not 1:" $(cat "${result}" | wc -l) + err=1 + exit + fi + + if ! perf lock con -b true > /dev/null 2>&1 ; then + return + fi + + # the perf lock contention output goes to the stderr + perf lock con -a -b -t -E 1 -q -- perf bench sched messaging > /dev/null 2> ${result} + if [ $(cat "${result}" | wc -l) != "1" ]; then + echo "[Fail] BPF result count is not 1:" $(cat "${result}" | wc -l) + err=1 + exit + fi +} + +test_aggr_addr() +{ + echo "Testing perf lock contention --lock-addr" + perf lock contention -i ${perfdata} -l -E 1 -q 2> ${result} + if [ $(cat "${result}" | wc -l) != "1" ]; then + echo "[Fail] Recorded result count is not 1:" $(cat "${result}" | wc -l) + err=1 + exit + fi + + if ! perf lock con -b true > /dev/null 2>&1 ; then + return + fi + + # the perf lock contention output goes to the stderr + perf lock con -a -b -t -E 1 -q -- perf bench sched messaging > /dev/null 2> ${result} + if [ $(cat "${result}" | wc -l) != "1" ]; then + echo "[Fail] BPF result count is not 1:" $(cat "${result}" | wc -l) + err=1 + exit + fi +} + check test_record test_bpf test_record_concurrent +test_aggr_task +test_aggr_addr exit ${err}
Signature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Em Fri, Dec 09, 2022 at 11:07:23AM -0800, Namhyung Kim escreveu:
Quoted 3 lines
Hello,
This patchset adds two more aggregation modes for perf lock contention.
Thanks, applied.
- Arnaldo
Quoted 54 lines
The first one is the per-task mode with -t/--threads option. The option was there already but it remained broken with BPF for a while. Now it supports the output with and without BPF.
# perf lock contention -a -b -t -- sleep 1 contended total wait max wait avg wait pid comm11 11.85 us 2.23 us 1.08 us 0 swapper 2 11.13 us 10.22 us 5.56 us 749053 ThreadPoolForeg 1 8.15 us 8.15 us 8.15 us 321353 Chrome_ChildIOT 2 2.73 us 1.77 us 1.37 us 320761 Chrome_ChildIOT 1 1.40 us 1.40 us 1.40 us 320502 chrome 1 379 ns 379 ns 379 ns 321227 chromeThe other one is the per-lock-instance mode with -l/--lock-addr option. If the lock has a symbol, it will be displayed as well.
# perf lock contention -a -b -l -- sleep 1 contended total wait max wait avg wait address symbol3 4.79 us 2.33 us 1.60 us ffffffffbaed50c0 rcu_state 4 4.19 us 1.62 us 1.05 us ffffffffbae07a40 jiffies_lock 1 1.94 us 1.94 us 1.94 us ffff9262277861e0 1 387 ns 387 ns 387 ns ffff9260bfda4f60It's based on the current acme/tmp.perf/core branch. You can find the code in the 'perf/lock-con-aggr-v1' branch in
git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.gitThanks,
NamhyungNamhyung Kim (4):
perf lock contention: Addlock_data.hfor common data perf lock contention: Implement -t/--threads option for BPF perf lock contention: Add -l/--lock-addr option perf test: Update perf lock contention testtools/perf/Documentation/perf-lock.txt | 4 + tools/perf/builtin-lock.c | 95 ++++++++++++++----- tools/perf/tests/shell/lock_contention.sh | 48 ++++++++++ tools/perf/util/bpf_lock_contention.c | 72 ++++++++++---- .../perf/util/bpf_skel/lock_contention.bpf.c | 67 +++++++++---- tools/perf/util/bpf_skel/lock_data.h | 30 ++++++ tools/perf/util/lock-contention.h | 1 + 7 files changed, 255 insertions(+), 62 deletions(-) create mode 100644 tools/perf/util/bpf_skel/lock_data.hbase-commit:
b22802e295a80ec16e355d7208d2fbbd7bbc1b7aSignature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Signature · 1 line
- Arnaldo