Motivation behind this feature is to aggregate the data at the LLC level
for chiplet based processors which currently do not expose the chiplet
details in sysfs cpu topology information.
For the completeness of the feature, the series adds ability to
aggregate data at any cache level. Following is the example of the
output on a dual socket Zen3 processor with 2 x 64C/128T containing 8
chiplet per socket.
$ sudo perf stat --per-cache -a -e ls_dmnd_fills_from_sys.ext_cache_remote --\
taskset -c 0-15,64-79,128-143,192-207\
perf bench sched messaging -p -t -l 100000 -g 8
# Running 'sched/messaging' benchmark:
# 20 sender and receiver threads per group
# 8 groups == 320 threads run
Total time: 7.648 [sec]
Performance counter stats for 'system wide':
S0-D0-L3-ID0 16 17,145,912 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID8 16 14,977,628 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID16 16 262,539 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID24 16 3,140 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID32 16 27,403 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID40 16 17,026 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID48 16 7,292 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID56 16 2,464 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID64 16 22,489,306 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID72 16 21,455,257 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID80 16 11,619 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID88 16 30,978 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID96 16 37,628 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID104 16 13,594 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID112 16 10,164 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID120 16 11,259 ls_dmnd_fills_from_sys.ext_cache_remote
7.779171484 seconds time elapsed
The series also adds support for perf stat record and perf stat report
to aggregate data at various cache levels. Following is an example of
recording with aggregation at L2 level and reporting the same data with
aggregation at L3 level.
$ sudo perf stat record --per-cache=L2 -a -e ls_dmnd_fills_from_sys.ext_cache_remote --\
taskset -c 0-15,64-79,128-143,192-207\
perf bench sched messaging -p -t -l 100000 -g 8
# Running 'sched/messaging' benchmark:
# 20 sender and receiver threads per group
# 8 groups == 320 threads run
Total time: 7.318 [sec]
Performance counter stats for 'system wide':
S0-D0-L2-ID0 2 2,171,980 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L2-ID1 2 2,048,494 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L2-ID2 2 2,120,293 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L2-ID3 2 2,224,725 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L2-ID4 2 2,021,618 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L2-ID5 2 1,995,331 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L2-ID6 2 2,163,029 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L2-ID7 2 2,104,623 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L2-ID8 2 1,948,776 ls_dmnd_fills_from_sys.ext_cache_remote
...
S0-D0-L2-ID63 2 2,648 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID64 2 2,963,323 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID65 2 2,856,629 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID66 2 2,901,725 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID67 2 3,046,120 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID68 2 2,637,971 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID69 2 2,680,029 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID70 2 2,672,259 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID71 2 2,638,768 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID72 2 3,308,642 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID73 2 3,064,473 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID74 2 3,023,379 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID75 2 2,975,119 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID76 2 2,952,677 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID77 2 2,981,695 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID78 2 3,455,916 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID79 2 2,959,540 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L2-ID80 2 4,977 ls_dmnd_fills_from_sys.ext_cache_remote
...
S1-D1-L2-ID127 2 3,359 ls_dmnd_fills_from_sys.ext_cache_remote
7.451725897 seconds time elapsed
$ sudo perf stat report --per-cache=L3
Performance counter stats for '...':
S0-D0-L3-ID0 16 16,850,093 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID8 16 16,001,493 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID16 16 301,011 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID24 16 26,276 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID32 16 48,958 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID40 16 43,799 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID48 16 16,771 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID56 16 12,544 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID64 16 22,396,824 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID72 16 24,721,441 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID80 16 29,426 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID88 16 54,348 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID96 16 41,557 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID104 16 10,084 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID112 16 14,361 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID120 16 24,446 ls_dmnd_fills_from_sys.ext_cache_remote
7.451725897 seconds time elapsed
The aggregate at S0-D0-L3-ID0 is the sum of S0-D0-L2-ID0 to S0-D0-L3-ID7
as L3 containing CPU0 contains the L2 instance of CPU0 to CPU7.
[New in v3 - Handling IDs differently compared to v2]
Cache IDs are now derived from the shared_cpus_list file in the cache
topology. This allows for --per-cache aggregation of data on a kernel
which does not expose the cache instance ID in the sysfs. Running perf
stat will give the following output on the same system with cache
instance ID hidden:
$ ls /sys/devices/system/cpu/cpu0/cache/index0/
coherency_line_size level number_of_sets physical_line_partition
shared_cpu_list shared_cpu_map size type uevent
ways_of_associativity
$ sudo perf stat --per-cache -a -e ls_dmnd_fills_from_sys.ext_cache_remote --\
taskset -c 0-15,64-79,128-143,192-207\
perf bench sched messaging -p -t -l 100000 -g 8
# Running 'sched/messaging' benchmark:
# 20 sender and receiver threads per group
# 8 groups == 320 threads run
Total time: 6.949 [sec]
Performance counter stats for 'system wide':
S0-D0-L3-ID0 16 5,297,615 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID8 16 4,347,868 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID16 16 416,593 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID24 16 4,346 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID32 16 5,506 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID40 16 15,845 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID48 16 24,164 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID56 16 4,543 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID64 16 41,610,374 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID72 16 38,393,688 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID80 16 22,188 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID88 16 22,918 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID96 16 39,230 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID104 16 6,236 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID112 16 66,846 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID120 16 72,713 ls_dmnd_fills_from_sys.ext_cache_remote
7.098471410 seconds time elapsed
This series makes breaking change when saving the aggregation details as
the cache level needs to be saved along with the aggregation method.
This RFC assumes that caches at same level will be shared by same set of
threads. The implementation will run into an issue if, say L1i is thread
local, but L1d is shared by the SMT siblings on the core. I'm seeking
clarification from the community about the same and potential solutions
if processors where such a scenario exist.
This series cleanly applies on top perf-tool branch from Arnaldo's tree
(https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/log/?h=perf-tools)
at:
commit e8d018dd0257 ("Linux 6.3-rc3")
--
Changelog:
o v2->v3:
- Dropped patches 1 and 2 that saved and retrieved the cache instance
ID when saving the cache data.
- The above is unnecessary as the IDs are being derived from the first
online CPU in the cache domain for a given cache instance.
- Improvements to handling cases where a cache level is not present
but the level is allowed by MAX_CACHE_LVL.
- Updated details in cover letter.
o v1->v2
- Set cache instance ID to 0 if the file cannot be read.
- Fix cache level parsing function.
- Updated details in cover letter.
--
K Prateek Nayak (2):
perf: Extract building cache level for a CPU into separate function
perf: Add option for --per-cache aggregation
tools/lib/perf/include/perf/cpumap.h | 5 +
tools/lib/perf/include/perf/event.h | 3 +-
tools/perf/Documentation/perf-stat.txt | 16 ++
tools/perf/builtin-stat.c | 144 +++++++++++++++++-
.../tests/shell/lib/perf_json_output_lint.py | 4 +-
tools/perf/tests/shell/stat+csv_output.sh | 14 ++
tools/perf/tests/shell/stat+json_output.sh | 13 ++
tools/perf/util/cpumap.c | 118 ++++++++++++++
tools/perf/util/cpumap.h | 28 ++++
tools/perf/util/event.c | 7 +-
tools/perf/util/header.c | 62 +++++---
tools/perf/util/header.h | 4 +
tools/perf/util/stat-display.c | 17 +++
tools/perf/util/stat-shadow.c | 1 +
tools/perf/util/stat.h | 2 +
tools/perf/util/synthetic-events.c | 1 +
16 files changed, 409 insertions(+), 30 deletions(-)
--
2.34.1
build_caches() builds the complete cache topology of the system by
iterating over all CPU, building and comparing cache levels of each CPU,
keeping only the unique ones at the end.
Extract the unit that build the cache levels for a single CPU into a
separate function. Expose this function, and the MAX_CACHE_LVL value to
be used elsewhere in perf too.
Signed-off-by: K Prateek Nayak <[email protected]>
---
Changelog:
o v2->v3
- Patch 3 from previous series with a comment over the new
build_caches_for_cpu() method.
o v1->v2
- No Changes.
---
tools/perf/util/header.c | 62 +++++++++++++++++++++++++---------------
tools/perf/util/header.h | 4 +++
2 files changed, 43 insertions(+), 23 deletions(-)
diff --git a/tools/perf/util/header.c b/tools/perf/util/header.c
index 404d816ca124..5c5a9d23979b 100644
--- a/tools/perf/util/header.c
+++ b/tools/perf/util/header.c
@@ -1213,38 +1213,54 @@ static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c)
fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map);
}
-#define MAX_CACHE_LVL 4
-
-static int build_caches(struct cpu_cache_level caches[], u32 *cntp)
+/*
+ * Build caches levels for a particular CPU from the data in
+ * /sys/devices/system/cpu/cpu<cpu>/cache/
+ * The cache level data is stored in caches[] from index at
+ * *cntp.
+ */
+int build_caches_for_cpu(u32 cpu, struct cpu_cache_level caches[], u32 *cntp)
{
- u32 i, cnt = 0;
- u32 nr, cpu;
u16 level;
- nr = cpu__max_cpu().cpu;
+ for (level = 0; level < MAX_CACHE_LVL; level++) {
+ struct cpu_cache_level c;
+ int err;
+ u32 i;
- for (cpu = 0; cpu < nr; cpu++) {
- for (level = 0; level < MAX_CACHE_LVL; level++) {
- struct cpu_cache_level c;
- int err;
+ err = cpu_cache_level__read(&c, cpu, level);
+ if (err < 0)
+ return err;
- err = cpu_cache_level__read(&c, cpu, level);
- if (err < 0)
- return err;
+ if (err == 1)
+ break;
- if (err == 1)
+ for (i = 0; i < *cntp; i++) {
+ if (cpu_cache_level__cmp(&c, &caches[i]))
break;
+ }
- for (i = 0; i < cnt; i++) {
- if (cpu_cache_level__cmp(&c, &caches[i]))
- break;
- }
+ if (i == *cntp) {
+ caches[*cntp] = c;
+ *cntp = *cntp + 1;
+ } else
+ cpu_cache_level__free(&c);
+ }
- if (i == cnt)
- caches[cnt++] = c;
- else
- cpu_cache_level__free(&c);
- }
+ return 0;
+}
+
+static int build_caches(struct cpu_cache_level caches[], u32 *cntp)
+{
+ u32 nr, cpu, cnt = 0;
+
+ nr = cpu__max_cpu().cpu;
+
+ for (cpu = 0; cpu < nr; cpu++) {
+ int ret = build_caches_for_cpu(cpu, caches, &cnt);
+
+ if (ret)
+ return ret;
}
*cntp = cnt;
return 0;
diff --git a/tools/perf/util/header.h b/tools/perf/util/header.h
index e3861ae62172..94cf2ffb6e60 100644
--- a/tools/perf/util/header.h
+++ b/tools/perf/util/header.h
@@ -177,7 +177,11 @@ int do_write(struct feat_fd *fd, const void *buf, size_t size);
int write_padded(struct feat_fd *fd, const void *bf,
size_t count, size_t count_aligned);
+#define MAX_CACHE_LVL 4
+
int is_cpu_online(unsigned int cpu);
+int build_caches_for_cpu(u32 cpu, struct cpu_cache_level caches[], u32 *cntp);
+
/*
* arch specific callback
*/
--
2.34.1
Processors based on chiplet architecture, such as AMD EPYC and Hygon do
not expose the chiplet details in the sysfs CPU topology information.
However, this information can be derived from the per CPU cache level
information from the sysfs.
perf stat has already supported aggregation based on topology
information using core ID, socket ID, etc. It'll be useful to aggregate
based on the cache topology to detect problems like imbalance and
cache-to-cache sharing at various cache levels.
This patch adds support for "--per-cache" option for aggregation at a
particular cache level. Also update the docs and related test. The
output will be like:
$ sudo perf stat --per-cache=L3 -a -e ls_dmnd_fills_from_sys.ext_cache_remote -- sleep 5
Performance counter stats for 'system wide':
S0-D0-L3-ID0 16 12,644,599 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID8 16 13,847,598 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID16 16 223,592 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID24 16 7,775 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID32 16 31,302 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID40 16 17,722 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID48 16 8,272 ls_dmnd_fills_from_sys.ext_cache_remote
S0-D0-L3-ID56 16 5,765 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID64 16 18,227,173 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID72 16 20,926,878 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID80 16 13,705 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID88 16 24,062 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID96 16 27,891 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID104 16 13,480 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID112 16 12,656 ls_dmnd_fills_from_sys.ext_cache_remote
S1-D1-L3-ID120 16 21,881 ls_dmnd_fills_from_sys.ext_cache_remote
Also support perf stat record and perf stat report with the ability to
specify a different cache level to aggregate data at when running perf
stat report.
Suggested-by: Gautham R. Shenoy <[email protected]>
Signed-off-by: K Prateek Nayak <[email protected]>
---
Changelog:
o v2->v3
- Cache IDs are now derived from the shared_cpus_list making
aggregation possible on older kernels that do not expose the IDs.
- Updated the comments based on the new ID assignment strategy.
- Better handle the case when specifying a level is possible as it is
less than or equal to MAX_CACHE_LVL but it does not exist on the
machine. In such cased ID will be -1.
$ sudo perf stat --per-cache=L4 -a -e cycles -- sleep 5
Performance counter stats for 'system wide':
S0-D0-L4-ID-1 128 51,328,613 cycles
S1-D1-L4-ID-1 128 125,132,221 cycles
o v1->v2
- Fix cache level parsing logic. Previously, giving "--per-cache=2" was
considered valid but that was not the intention. Only parse strings
of form LX or lX where x is the cache level and can range from 1 to
MAX_CACHE_LVL.
---
tools/lib/perf/include/perf/cpumap.h | 5 +
tools/lib/perf/include/perf/event.h | 3 +-
tools/perf/Documentation/perf-stat.txt | 16 ++
tools/perf/builtin-stat.c | 144 +++++++++++++++++-
.../tests/shell/lib/perf_json_output_lint.py | 4 +-
tools/perf/tests/shell/stat+csv_output.sh | 14 ++
tools/perf/tests/shell/stat+json_output.sh | 13 ++
tools/perf/util/cpumap.c | 118 ++++++++++++++
tools/perf/util/cpumap.h | 28 ++++
tools/perf/util/event.c | 7 +-
tools/perf/util/stat-display.c | 17 +++
tools/perf/util/stat-shadow.c | 1 +
tools/perf/util/stat.h | 2 +
tools/perf/util/synthetic-events.c | 1 +
14 files changed, 366 insertions(+), 7 deletions(-)
diff --git a/tools/lib/perf/include/perf/cpumap.h b/tools/lib/perf/include/perf/cpumap.h
index 3f43f770cdac..8724dde79342 100644
--- a/tools/lib/perf/include/perf/cpumap.h
+++ b/tools/lib/perf/include/perf/cpumap.h
@@ -11,6 +11,11 @@ struct perf_cpu {
int cpu;
};
+struct perf_cache {
+ int cache_lvl;
+ int cache;
+};
+
struct perf_cpu_map;
LIBPERF_API struct perf_cpu_map *perf_cpu_map__dummy_new(void);
diff --git a/tools/lib/perf/include/perf/event.h b/tools/lib/perf/include/perf/event.h
index ad47d7b31046..f3ceb2f96593 100644
--- a/tools/lib/perf/include/perf/event.h
+++ b/tools/lib/perf/include/perf/event.h
@@ -378,7 +378,8 @@ enum {
PERF_STAT_CONFIG_TERM__AGGR_MODE = 0,
PERF_STAT_CONFIG_TERM__INTERVAL = 1,
PERF_STAT_CONFIG_TERM__SCALE = 2,
- PERF_STAT_CONFIG_TERM__MAX = 3,
+ PERF_STAT_CONFIG_TERM__AGGR_LEVEL = 3,
+ PERF_STAT_CONFIG_TERM__MAX = 4,
};
struct perf_record_stat_config_entry {
diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/perf-stat.txt
index 18abdc1dce05..5784131146a1 100644
--- a/tools/perf/Documentation/perf-stat.txt
+++ b/tools/perf/Documentation/perf-stat.txt
@@ -308,6 +308,14 @@ use --per-die in addition to -a. (system-wide). The output includes the
die number and the number of online processors on that die. This is
useful to gauge the amount of aggregation.
+--per-cache::
+Aggregate counts per cache instance for system-wide mode measurements. By
+default, the aggregation happens for the cache level at the highest index
+in the system. To specify a particular level, mention the cache level
+alongside the option in the format [Ll][1-9][0-9]*. For example:
+Using option "--per-cache=l3" or "--per-cache=L3" will aggregate the
+information at the boundary of the level 3 cache in the system.
+
--per-core::
Aggregate counts per physical processor for system-wide mode measurements. This
is a useful mode to detect imbalance between physical cores. To enable this mode,
@@ -379,6 +387,14 @@ Aggregate counts per processor socket for system-wide mode measurements.
--per-die::
Aggregate counts per processor die for system-wide mode measurements.
+--per-cache::
+Aggregate counts per cache instance for system-wide mode measurements. By
+default, the aggregation happens for the cache level at the highest index
+in the system. To specify a particular level, mention the cache level
+alongside the option in the format [Ll][1-9][0-9]*. For example: Using
+option "--per-cache=l3" or "--per-cache=L3" will aggregate the
+information at the boundary of the level 3 cache in the system.
+
--per-core::
Aggregate counts per physical processor for system-wide mode measurements.
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index fa7c40956d0f..744890898c13 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -203,6 +203,7 @@ struct perf_stat {
struct perf_cpu_map *cpus;
struct perf_thread_map *threads;
enum aggr_mode aggr_mode;
+ u32 aggr_level;
};
static struct perf_stat perf_stat;
@@ -210,8 +211,9 @@ static struct perf_stat perf_stat;
static volatile sig_atomic_t done = 0;
-static struct perf_stat_config stat_config = {
+struct perf_stat_config stat_config = {
.aggr_mode = AGGR_GLOBAL,
+ .aggr_level = MAX_CACHE_LVL + 1,
.scale = true,
.unit_width = 4, /* strlen("unit") */
.run_count = 1,
@@ -1160,6 +1162,55 @@ static int parse_hybrid_type(const struct option *opt,
return 0;
}
+static int parse_cache_level(const struct option *opt,
+ const char *str,
+ int unset __maybe_unused)
+{
+ int level;
+ u32 *aggr_mode = (u32 *)opt->value;
+ u32 *aggr_level = (u32 *)opt->data;
+
+ /*
+ * If no string is specified, aggregate based on the topology of
+ * Last Level Cache (LLC). Since the LLC level can change from
+ * architecture to architecture, set level greater than
+ * MAX_CACHE_LVL which will be interpreted as LLC.
+ */
+ if (str == NULL) {
+ level = MAX_CACHE_LVL + 1;
+ goto out;
+ }
+
+ /*
+ * The format to specify cache level is LX or lX where X is the
+ * cache level.
+ */
+ if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
+ pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
+ MAX_CACHE_LVL,
+ MAX_CACHE_LVL);
+ return -EINVAL;
+ }
+
+ level = atoi(&str[1]);
+ if (level < 1) {
+ pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
+ MAX_CACHE_LVL,
+ MAX_CACHE_LVL);
+ return -EINVAL;
+ }
+
+ if (level > MAX_CACHE_LVL) {
+ pr_err("perf only supports max cache level of %d.\n"
+ "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
+ return -EINVAL;
+ }
+out:
+ *aggr_mode = AGGR_CACHE;
+ *aggr_level = level;
+ return 0;
+}
+
static struct option stat_options[] = {
OPT_BOOLEAN('T', "transaction", &transaction_run,
"hardware transaction statistics"),
@@ -1237,6 +1288,9 @@ static struct option stat_options[] = {
"aggregate counts per processor socket", AGGR_SOCKET),
OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
"aggregate counts per processor die", AGGR_DIE),
+ OPT_CALLBACK_OPTARG(0, "per-cache", &stat_config.aggr_mode, &stat_config.aggr_level,
+ "cache level", "aggregate count at this cache level (Default: LLC)",
+ parse_cache_level),
OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
"aggregate counts per physical processor core", AGGR_CORE),
OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
@@ -1298,6 +1352,7 @@ static struct option stat_options[] = {
static const char *const aggr_mode__string[] = {
[AGGR_CORE] = "core",
+ [AGGR_CACHE] = "cache",
[AGGR_DIE] = "die",
[AGGR_GLOBAL] = "global",
[AGGR_NODE] = "node",
@@ -1319,6 +1374,12 @@ static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __m
return aggr_cpu_id__die(cpu, /*data=*/NULL);
}
+static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
+ struct perf_cpu cpu)
+{
+ return aggr_cpu_id__cache(cpu, /*data=*/NULL);
+}
+
static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
struct perf_cpu cpu)
{
@@ -1371,6 +1432,12 @@ static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *con
return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
}
+static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
+ struct perf_cpu cpu)
+{
+ return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
+}
+
static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
struct perf_cpu cpu)
{
@@ -1402,6 +1469,8 @@ static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
return aggr_cpu_id__socket;
case AGGR_DIE:
return aggr_cpu_id__die;
+ case AGGR_CACHE:
+ return aggr_cpu_id__cache;
case AGGR_CORE:
return aggr_cpu_id__core;
case AGGR_NODE:
@@ -1425,6 +1494,8 @@ static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
return perf_stat__get_socket_cached;
case AGGR_DIE:
return perf_stat__get_die_cached;
+ case AGGR_CACHE:
+ return perf_stat__get_cache_id_cached;
case AGGR_CORE:
return perf_stat__get_core_cached;
case AGGR_NODE:
@@ -1537,6 +1608,60 @@ static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, voi
return id;
}
+static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
+ u32 cache_level, struct aggr_cpu_id *id)
+{
+ int i;
+ int caches_cnt = env->caches_cnt;
+ struct cpu_cache_level *caches = env->caches;
+
+ id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0: cache_level;
+ id->cache = -1;
+
+ if (!caches_cnt)
+ return;
+
+ for (i = caches_cnt - 1; i > -1; --i) {
+ struct perf_cpu_map *cpu_map;
+ int map_contains_cpu;
+
+ /*
+ * If user has not specified a level, find the fist level with
+ * the cpu in the map. Since building the map is expensive, do
+ * this only if levels match.
+ */
+ if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
+ continue;
+
+ cpu_map = perf_cpu_map__new(caches[i].map);
+ map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
+ perf_cpu_map__put(cpu_map);
+
+ if (map_contains_cpu != -1) {
+ id->cache_lvl = caches[i].level;
+ id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
+ return;
+ }
+ }
+}
+
+static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
+ void *data)
+{
+ struct perf_env *env = data;
+ struct aggr_cpu_id id = aggr_cpu_id__empty();
+
+ if (cpu.cpu != -1) {
+ u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
+
+ id.socket = env->cpu[cpu.cpu].socket_id;
+ id.die = env->cpu[cpu.cpu].die_id;
+ perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
+ }
+
+ return id;
+}
+
static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
{
struct perf_env *env = data;
@@ -1605,6 +1730,12 @@ static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *confi
return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
}
+static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
+ struct perf_cpu cpu)
+{
+ return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
+}
+
static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
struct perf_cpu cpu)
{
@@ -1636,6 +1767,8 @@ static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
return perf_env__get_socket_aggr_by_cpu;
case AGGR_DIE:
return perf_env__get_die_aggr_by_cpu;
+ case AGGR_CACHE:
+ return perf_env__get_cache_aggr_by_cpu;
case AGGR_CORE:
return perf_env__get_core_aggr_by_cpu;
case AGGR_NODE:
@@ -1659,6 +1792,8 @@ static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
return perf_stat__get_socket_file;
case AGGR_DIE:
return perf_stat__get_die_file;
+ case AGGR_CACHE:
+ return perf_stat__get_cache_file;
case AGGR_CORE:
return perf_stat__get_core_file;
case AGGR_NODE:
@@ -2207,7 +2342,8 @@ static struct perf_stat perf_stat = {
.stat = perf_event__process_stat_event,
.stat_round = process_stat_round_event,
},
- .aggr_mode = AGGR_UNSET,
+ .aggr_mode = AGGR_UNSET,
+ .aggr_level = 0,
};
static int __cmd_report(int argc, const char **argv)
@@ -2219,6 +2355,10 @@ static int __cmd_report(int argc, const char **argv)
"aggregate counts per processor socket", AGGR_SOCKET),
OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
"aggregate counts per processor die", AGGR_DIE),
+ OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
+ "cache level",
+ "aggregate count at this cache level (Default: LLC)",
+ parse_cache_level),
OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
"aggregate counts per physical processor core", AGGR_CORE),
OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
diff --git a/tools/perf/tests/shell/lib/perf_json_output_lint.py b/tools/perf/tests/shell/lib/perf_json_output_lint.py
index 97598d14e532..62489766b93c 100644
--- a/tools/perf/tests/shell/lib/perf_json_output_lint.py
+++ b/tools/perf/tests/shell/lib/perf_json_output_lint.py
@@ -14,6 +14,7 @@ ap.add_argument('--system-wide', action='store_true')
ap.add_argument('--event', action='store_true')
ap.add_argument('--per-core', action='store_true')
ap.add_argument('--per-thread', action='store_true')
+ap.add_argument('--per-cache', action='store_true')
ap.add_argument('--per-die', action='store_true')
ap.add_argument('--per-node', action='store_true')
ap.add_argument('--per-socket', action='store_true')
@@ -46,6 +47,7 @@ def check_json_output(expected_items):
'counter-value': lambda x: is_counter_value(x),
'cgroup': lambda x: True,
'cpu': lambda x: isint(x),
+ 'cache': lambda x: True,
'die': lambda x: True,
'event': lambda x: True,
'event-runtime': lambda x: isfloat(x),
@@ -82,7 +84,7 @@ try:
expected_items = 7
elif args.interval or args.per_thread or args.system_wide_no_aggr:
expected_items = 8
- elif args.per_core or args.per_socket or args.per_node or args.per_die:
+ elif args.per_core or args.per_socket or args.per_node or args.per_die or args.per_cache_instance:
expected_items = 9
else:
# If no option is specified, don't check the number of items.
diff --git a/tools/perf/tests/shell/stat+csv_output.sh b/tools/perf/tests/shell/stat+csv_output.sh
index 324fc9e6edd7..6cdf2fd386d5 100755
--- a/tools/perf/tests/shell/stat+csv_output.sh
+++ b/tools/perf/tests/shell/stat+csv_output.sh
@@ -26,6 +26,7 @@ function commachecker()
;; "--per-socket") exp=8
;; "--per-node") exp=8
;; "--per-die") exp=8
+ ;; "--per-cache") exp=8
esac
while read line
@@ -123,6 +124,18 @@ check_per_thread()
echo "[Success]"
}
+check_per_cache_instance()
+{
+ echo -n "Checking CSV output: per cache instance "
+ if ParanoidAndNotRoot 0
+ then
+ echo "[Skip] paranoid and not root"
+ return
+ fi
+ perf stat -x$csv_sep --per-cache -a true 2>&1 | commachecker --per-cache
+ echo "[Success]"
+}
+
check_per_die()
{
echo -n "Checking CSV output: per die "
@@ -197,6 +210,7 @@ if [ $skip_test -ne 1 ]
then
check_system_wide_no_aggr
check_per_core
+ check_per_cache_instance
check_per_die
check_per_socket
else
diff --git a/tools/perf/tests/shell/stat+json_output.sh b/tools/perf/tests/shell/stat+json_output.sh
index 2c4212c641ed..d79a6e0d4042 100755
--- a/tools/perf/tests/shell/stat+json_output.sh
+++ b/tools/perf/tests/shell/stat+json_output.sh
@@ -100,6 +100,18 @@ check_per_thread()
echo "[Success]"
}
+check_per_cache_instance()
+{
+ echo -n "Checking json output: per cache_instance "
+ if ParanoidAndNotRoot 0
+ then
+ echo "[Skip] paranoia and not root"
+ return
+ fi
+ perf stat -j --per-cache -a true 2>&1 | $PYTHON $pythonchecker --per-cache
+ echo "[Success]"
+}
+
check_per_die()
{
echo -n "Checking json output: per die "
@@ -174,6 +186,7 @@ if [ $skip_test -ne 1 ]
then
check_system_wide_no_aggr
check_per_core
+ check_per_cache_instance
check_per_die
check_per_socket
else
diff --git a/tools/perf/util/cpumap.c b/tools/perf/util/cpumap.c
index 5e564974fba4..98dd308fcd21 100644
--- a/tools/perf/util/cpumap.c
+++ b/tools/perf/util/cpumap.c
@@ -3,6 +3,8 @@
#include "cpumap.h"
#include "debug.h"
#include "event.h"
+#include "header.h"
+#include "stat.h"
#include <assert.h>
#include <dirent.h>
#include <stdio.h>
@@ -227,6 +229,10 @@ static int aggr_cpu_id__cmp(const void *a_pointer, const void *b_pointer)
return a->socket - b->socket;
else if (a->die != b->die)
return a->die - b->die;
+ else if (a->cache_lvl != b->cache_lvl)
+ return a->cache_lvl - b->cache_lvl;
+ else if (a->cache != b->cache)
+ return a->cache - b->cache;
else if (a->core != b->core)
return a->core - b->core;
else
@@ -310,6 +316,112 @@ struct aggr_cpu_id aggr_cpu_id__die(struct perf_cpu cpu, void *data)
return id;
}
+extern struct perf_stat_config stat_config;
+
+int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
+{
+ int id;
+ struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
+
+ /*
+ * If the map contains no CPU, consider the current CPU to
+ * be the first online CPU in the cache domain else use the
+ * first online CPU of the cache domain as the ID.
+ */
+ if (perf_cpu_map__empty(cpu_map))
+ id = cpu.cpu;
+ else
+ id = perf_cpu_map__cpu(cpu_map, 0).cpu;
+
+ /* Free the perf_cpu_map used to find the cache ID */
+ perf_cpu_map__put(cpu_map);
+
+ return id;
+}
+int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
+{
+ int ret = 0;
+ struct cpu_cache_level caches[MAX_CACHE_LVL];
+ u32 cache_level = stat_config.aggr_level;
+ u32 i = 0, caches_cnt = 0;
+
+ cache->cache_lvl = (cache_level > MAX_CACHE_LVL)? 0: cache_level;
+ cache->cache = -1;
+
+ ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
+ if (ret) {
+ /*
+ * If caches_cnt is not 0, cpu_cache_level data
+ * was allocated when building the topology.
+ * Free the allocated data before returning.
+ */
+ if (caches_cnt)
+ goto free_caches;
+
+ return ret;
+ }
+
+ if (!caches_cnt)
+ return -1;
+
+ /*
+ * Save the data for the highest level if no
+ * level was specified by the user.
+ */
+ if (cache_level > MAX_CACHE_LVL) {
+ int max_level_index = 0;
+
+ for (i = 1; i < caches_cnt; ++i) {
+ if (caches[i].level > caches[max_level_index].level)
+ max_level_index = i;
+ }
+
+ cache->cache_lvl = caches[max_level_index].level;
+ cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
+
+ /* Reset i to 0 to free entire caches[] */
+ i = 0;
+ goto free_caches;
+ }
+
+ for (i = 0; i < caches_cnt; ++i) {
+ if (caches[i].level == cache_level) {
+ cache->cache_lvl = cache_level;
+ cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
+ }
+
+ cpu_cache_level__free(&caches[i]);
+ }
+
+free_caches:
+ /*
+ * Free all the allocated cpu_cache_level data.
+ */
+ while (i < caches_cnt)
+ cpu_cache_level__free(&caches[i++]);
+
+ return ret;
+}
+
+struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
+{
+ int ret;
+ struct aggr_cpu_id id;
+ struct perf_cache cache;
+
+ id = aggr_cpu_id__die(cpu, data);
+ if (aggr_cpu_id__is_empty(&id))
+ return id;
+
+ ret = cpu__get_cache_details(cpu, &cache);
+ if (ret)
+ return id;
+
+ id.cache_lvl = cache.cache_lvl;
+ id.cache = cache.cache;
+ return id;
+}
+
int cpu__get_core_id(struct perf_cpu cpu)
{
int value, ret = cpu__get_topology_int(cpu.cpu, "core_id", &value);
@@ -684,6 +796,8 @@ bool aggr_cpu_id__equal(const struct aggr_cpu_id *a, const struct aggr_cpu_id *b
a->node == b->node &&
a->socket == b->socket &&
a->die == b->die &&
+ a->cache_lvl == b->cache_lvl &&
+ a->cache == b->cache &&
a->core == b->core &&
a->cpu.cpu == b->cpu.cpu;
}
@@ -694,6 +808,8 @@ bool aggr_cpu_id__is_empty(const struct aggr_cpu_id *a)
a->node == -1 &&
a->socket == -1 &&
a->die == -1 &&
+ a->cache_lvl == -1 &&
+ a->cache == -1 &&
a->core == -1 &&
a->cpu.cpu == -1;
}
@@ -705,6 +821,8 @@ struct aggr_cpu_id aggr_cpu_id__empty(void)
.node = -1,
.socket = -1,
.die = -1,
+ .cache_lvl = -1,
+ .cache = -1,
.core = -1,
.cpu = (struct perf_cpu){ .cpu = -1 },
};
diff --git a/tools/perf/util/cpumap.h b/tools/perf/util/cpumap.h
index c2f5824a3a22..686ef904f57b 100644
--- a/tools/perf/util/cpumap.h
+++ b/tools/perf/util/cpumap.h
@@ -20,6 +20,13 @@ struct aggr_cpu_id {
int socket;
/** The die id as read from /sys/devices/system/cpu/cpuX/topology/die_id. */
int die;
+ /** The cache level as read from /sys/devices/system/cpu/cpuX/cache/indexY/level */
+ int cache_lvl;
+ /**
+ * The cache instance ID, which is the first CPU in the
+ * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
+ */
+ int cache;
/** The core id as read from /sys/devices/system/cpu/cpuX/topology/core_id. */
int core;
/** CPU aggregation, note there is one CPU for each SMT thread. */
@@ -76,6 +83,20 @@ int cpu__get_socket_id(struct perf_cpu cpu);
* /sys/devices/system/cpu/cpuX/topology/die_id for the given CPU.
*/
int cpu__get_die_id(struct perf_cpu cpu);
+/**
+ * Calculate the cache instance ID from the map in
+ * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
+ * Cache instance ID is the first CPU reported in the shared_cpu_list file.
+ */
+int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map);
+/**
+ * cpu__get_cache_id - Returns 0 if successful in populating the
+ * cache level and cache id. Cache level is read from
+ * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
+ * is the first CPU reported by
+ * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
+ */
+int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache);
/**
* cpu__get_core_id - Returns the core id as read from
* /sys/devices/system/cpu/cpuX/topology/core_id for the given CPU.
@@ -116,6 +137,13 @@ struct aggr_cpu_id aggr_cpu_id__socket(struct perf_cpu cpu, void *data);
* aggr_cpu_id_get_t.
*/
struct aggr_cpu_id aggr_cpu_id__die(struct perf_cpu cpu, void *data);
+/**
+ * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
+ * level, die and socket populated with the cache instache ID, cache level,
+ * die and socket for cpu. The function signature is compatible with
+ * aggr_cpu_id_get_t.
+ */
+struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data);
/**
* aggr_cpu_id__core - Create an aggr_cpu_id with the core, die and socket
* populated with the core, die and socket for cpu. The function signature is
diff --git a/tools/perf/util/event.c b/tools/perf/util/event.c
index 1fa14598b916..faf0df3c5b95 100644
--- a/tools/perf/util/event.c
+++ b/tools/perf/util/event.c
@@ -135,9 +135,10 @@ void perf_event__read_stat_config(struct perf_stat_config *config,
config->__val = event->data[i].val; \
break;
- CASE(AGGR_MODE, aggr_mode)
- CASE(SCALE, scale)
- CASE(INTERVAL, interval)
+ CASE(AGGR_MODE, aggr_mode)
+ CASE(SCALE, scale)
+ CASE(INTERVAL, interval)
+ CASE(AGGR_LEVEL, aggr_level)
#undef CASE
default:
pr_warning("unknown stat config term %" PRI_lu64 "\n",
diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c
index 1b5cb20efd23..82ec668bc3ba 100644
--- a/tools/perf/util/stat-display.c
+++ b/tools/perf/util/stat-display.c
@@ -36,6 +36,7 @@
static int aggr_header_lens[] = {
[AGGR_CORE] = 18,
+ [AGGR_CACHE] = 22,
[AGGR_DIE] = 12,
[AGGR_SOCKET] = 6,
[AGGR_NODE] = 6,
@@ -46,6 +47,7 @@ static int aggr_header_lens[] = {
static const char *aggr_header_csv[] = {
[AGGR_CORE] = "core,cpus,",
+ [AGGR_CACHE] = "cache,cpus,",
[AGGR_DIE] = "die,cpus,",
[AGGR_SOCKET] = "socket,cpus,",
[AGGR_NONE] = "cpu,",
@@ -56,6 +58,7 @@ static const char *aggr_header_csv[] = {
static const char *aggr_header_std[] = {
[AGGR_CORE] = "core",
+ [AGGR_CACHE] = "cache",
[AGGR_DIE] = "die",
[AGGR_SOCKET] = "socket",
[AGGR_NONE] = "cpu",
@@ -193,6 +196,10 @@ static void print_aggr_id_std(struct perf_stat_config *config,
case AGGR_CORE:
snprintf(buf, sizeof(buf), "S%d-D%d-C%d", id.socket, id.die, id.core);
break;
+ case AGGR_CACHE:
+ snprintf(buf, sizeof(buf), "S%d-D%d-L%d-ID%d",
+ id.socket, id.die, id.cache_lvl, id.cache);
+ break;
case AGGR_DIE:
snprintf(buf, sizeof(buf), "S%d-D%d", id.socket, id.die);
break;
@@ -239,6 +246,10 @@ static void print_aggr_id_csv(struct perf_stat_config *config,
fprintf(output, "S%d-D%d-C%d%s%d%s",
id.socket, id.die, id.core, sep, nr, sep);
break;
+ case AGGR_CACHE:
+ fprintf(config->output, "S%d-D%d-L%d-ID%d%s%d%s",
+ id.socket, id.die, id.cache_lvl, id.cache, sep, nr, sep);
+ break;
case AGGR_DIE:
fprintf(output, "S%d-D%d%s%d%s",
id.socket, id.die, sep, nr, sep);
@@ -284,6 +295,10 @@ static void print_aggr_id_json(struct perf_stat_config *config,
fprintf(output, "\"core\" : \"S%d-D%d-C%d\", \"aggregate-number\" : %d, ",
id.socket, id.die, id.core, nr);
break;
+ case AGGR_CACHE:
+ fprintf(output, "\"cache\" : \"S%d-D%d-L%d-ID%d\", \"aggregate-number\" : %d, ",
+ id.socket, id.die, id.cache_lvl, id.cache, nr);
+ break;
case AGGR_DIE:
fprintf(output, "\"die\" : \"S%d-D%d\", \"aggregate-number\" : %d, ",
id.socket, id.die, nr);
@@ -1126,6 +1141,7 @@ static void print_header_interval_std(struct perf_stat_config *config,
case AGGR_NODE:
case AGGR_SOCKET:
case AGGR_DIE:
+ case AGGR_CACHE:
case AGGR_CORE:
fprintf(output, "#%*s %-*s cpus",
INTERVAL_LEN - 1, "time",
@@ -1422,6 +1438,7 @@ void evlist__print_counters(struct evlist *evlist, struct perf_stat_config *conf
switch (config->aggr_mode) {
case AGGR_CORE:
+ case AGGR_CACHE:
case AGGR_DIE:
case AGGR_SOCKET:
case AGGR_NODE:
diff --git a/tools/perf/util/stat-shadow.c b/tools/perf/util/stat-shadow.c
index 806b32156459..f080905a3ece 100644
--- a/tools/perf/util/stat-shadow.c
+++ b/tools/perf/util/stat-shadow.c
@@ -20,6 +20,7 @@
* AGGR_GLOBAL: Use CPU 0
* AGGR_SOCKET: Use first CPU of socket
* AGGR_DIE: Use first CPU of die
+ * AGGR_CACHE: Use first CPU of cache level instance
* AGGR_CORE: Use first CPU of core
* AGGR_NONE: Use matching CPU
* AGGR_THREAD: Not supported?
diff --git a/tools/perf/util/stat.h b/tools/perf/util/stat.h
index bf1794ebc916..848b3b3f5819 100644
--- a/tools/perf/util/stat.h
+++ b/tools/perf/util/stat.h
@@ -74,6 +74,7 @@ enum aggr_mode {
AGGR_GLOBAL,
AGGR_SOCKET,
AGGR_DIE,
+ AGGR_CACHE,
AGGR_CORE,
AGGR_THREAD,
AGGR_UNSET,
@@ -139,6 +140,7 @@ typedef struct aggr_cpu_id (*aggr_get_id_t)(struct perf_stat_config *config, str
struct perf_stat_config {
enum aggr_mode aggr_mode;
+ u32 aggr_level;
bool scale;
bool no_inherit;
bool identifier;
diff --git a/tools/perf/util/synthetic-events.c b/tools/perf/util/synthetic-events.c
index 9ab9308ee80c..2fe648be1e7d 100644
--- a/tools/perf/util/synthetic-events.c
+++ b/tools/perf/util/synthetic-events.c
@@ -1373,6 +1373,7 @@ int perf_event__synthesize_stat_config(struct perf_tool *tool,
ADD(AGGR_MODE, config->aggr_mode)
ADD(INTERVAL, config->interval)
ADD(SCALE, config->scale)
+ ADD(AGGR_LEVEL, config->aggr_level)
WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
"stat config terms unbalanced\n");
--
2.34.1
On Wed, Apr 12, 2023 at 11:22 PM K Prateek Nayak <[email protected]> wrote:
>
> Processors based on chiplet architecture, such as AMD EPYC and Hygon do
> not expose the chiplet details in the sysfs CPU topology information.
> However, this information can be derived from the per CPU cache level
> information from the sysfs.
>
> perf stat has already supported aggregation based on topology
> information using core ID, socket ID, etc. It'll be useful to aggregate
> based on the cache topology to detect problems like imbalance and
> cache-to-cache sharing at various cache levels.
>
> This patch adds support for "--per-cache" option for aggregation at a
> particular cache level. Also update the docs and related test. The
> output will be like:
>
> $ sudo perf stat --per-cache=L3 -a -e ls_dmnd_fills_from_sys.ext_cache_remote -- sleep 5
>
> Performance counter stats for 'system wide':
>
> S0-D0-L3-ID0 16 12,644,599 ls_dmnd_fills_from_sys.ext_cache_remote
> S0-D0-L3-ID8 16 13,847,598 ls_dmnd_fills_from_sys.ext_cache_remote
> S0-D0-L3-ID16 16 223,592 ls_dmnd_fills_from_sys.ext_cache_remote
> S0-D0-L3-ID24 16 7,775 ls_dmnd_fills_from_sys.ext_cache_remote
> S0-D0-L3-ID32 16 31,302 ls_dmnd_fills_from_sys.ext_cache_remote
> S0-D0-L3-ID40 16 17,722 ls_dmnd_fills_from_sys.ext_cache_remote
> S0-D0-L3-ID48 16 8,272 ls_dmnd_fills_from_sys.ext_cache_remote
> S0-D0-L3-ID56 16 5,765 ls_dmnd_fills_from_sys.ext_cache_remote
> S1-D1-L3-ID64 16 18,227,173 ls_dmnd_fills_from_sys.ext_cache_remote
> S1-D1-L3-ID72 16 20,926,878 ls_dmnd_fills_from_sys.ext_cache_remote
> S1-D1-L3-ID80 16 13,705 ls_dmnd_fills_from_sys.ext_cache_remote
> S1-D1-L3-ID88 16 24,062 ls_dmnd_fills_from_sys.ext_cache_remote
> S1-D1-L3-ID96 16 27,891 ls_dmnd_fills_from_sys.ext_cache_remote
> S1-D1-L3-ID104 16 13,480 ls_dmnd_fills_from_sys.ext_cache_remote
> S1-D1-L3-ID112 16 12,656 ls_dmnd_fills_from_sys.ext_cache_remote
> S1-D1-L3-ID120 16 21,881 ls_dmnd_fills_from_sys.ext_cache_remote
>
> Also support perf stat record and perf stat report with the ability to
> specify a different cache level to aggregate data at when running perf
> stat report.
>
> Suggested-by: Gautham R. Shenoy <[email protected]>
> Signed-off-by: K Prateek Nayak <[email protected]>
> ---
> Changelog:
> o v2->v3
> - Cache IDs are now derived from the shared_cpus_list making
> aggregation possible on older kernels that do not expose the IDs.
> - Updated the comments based on the new ID assignment strategy.
> - Better handle the case when specifying a level is possible as it is
> less than or equal to MAX_CACHE_LVL but it does not exist on the
> machine. In such cased ID will be -1.
>
> $ sudo perf stat --per-cache=L4 -a -e cycles -- sleep 5
>
> Performance counter stats for 'system wide':
>
> S0-D0-L4-ID-1 128 51,328,613 cycles
> S1-D1-L4-ID-1 128 125,132,221 cycles
>
> o v1->v2
> - Fix cache level parsing logic. Previously, giving "--per-cache=2" was
> considered valid but that was not the intention. Only parse strings
> of form LX or lX where x is the cache level and can range from 1 to
> MAX_CACHE_LVL.
> ---
> tools/lib/perf/include/perf/cpumap.h | 5 +
> tools/lib/perf/include/perf/event.h | 3 +-
> tools/perf/Documentation/perf-stat.txt | 16 ++
> tools/perf/builtin-stat.c | 144 +++++++++++++++++-
> .../tests/shell/lib/perf_json_output_lint.py | 4 +-
> tools/perf/tests/shell/stat+csv_output.sh | 14 ++
> tools/perf/tests/shell/stat+json_output.sh | 13 ++
I think you can break this patch apart. You can add the feature, then
enable the CSV test, then the json test, etc. When adding the feature
you can also probably split the changes of "struct aggr_cpu_id" from
the display/enablement logic. Overall it looks good!
Thanks,
Ian
> tools/perf/util/cpumap.c | 118 ++++++++++++++
> tools/perf/util/cpumap.h | 28 ++++
> tools/perf/util/event.c | 7 +-
> tools/perf/util/stat-display.c | 17 +++
> tools/perf/util/stat-shadow.c | 1 +
> tools/perf/util/stat.h | 2 +
> tools/perf/util/synthetic-events.c | 1 +
> 14 files changed, 366 insertions(+), 7 deletions(-)
>
> diff --git a/tools/lib/perf/include/perf/cpumap.h b/tools/lib/perf/include/perf/cpumap.h
> index 3f43f770cdac..8724dde79342 100644
> --- a/tools/lib/perf/include/perf/cpumap.h
> +++ b/tools/lib/perf/include/perf/cpumap.h
> @@ -11,6 +11,11 @@ struct perf_cpu {
> int cpu;
> };
>
> +struct perf_cache {
> + int cache_lvl;
> + int cache;
> +};
> +
> struct perf_cpu_map;
>
> LIBPERF_API struct perf_cpu_map *perf_cpu_map__dummy_new(void);
> diff --git a/tools/lib/perf/include/perf/event.h b/tools/lib/perf/include/perf/event.h
> index ad47d7b31046..f3ceb2f96593 100644
> --- a/tools/lib/perf/include/perf/event.h
> +++ b/tools/lib/perf/include/perf/event.h
> @@ -378,7 +378,8 @@ enum {
> PERF_STAT_CONFIG_TERM__AGGR_MODE = 0,
> PERF_STAT_CONFIG_TERM__INTERVAL = 1,
> PERF_STAT_CONFIG_TERM__SCALE = 2,
> - PERF_STAT_CONFIG_TERM__MAX = 3,
> + PERF_STAT_CONFIG_TERM__AGGR_LEVEL = 3,
> + PERF_STAT_CONFIG_TERM__MAX = 4,
> };
>
> struct perf_record_stat_config_entry {
> diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/perf-stat.txt
> index 18abdc1dce05..5784131146a1 100644
> --- a/tools/perf/Documentation/perf-stat.txt
> +++ b/tools/perf/Documentation/perf-stat.txt
> @@ -308,6 +308,14 @@ use --per-die in addition to -a. (system-wide). The output includes the
> die number and the number of online processors on that die. This is
> useful to gauge the amount of aggregation.
>
> +--per-cache::
> +Aggregate counts per cache instance for system-wide mode measurements. By
> +default, the aggregation happens for the cache level at the highest index
> +in the system. To specify a particular level, mention the cache level
> +alongside the option in the format [Ll][1-9][0-9]*. For example:
> +Using option "--per-cache=l3" or "--per-cache=L3" will aggregate the
> +information at the boundary of the level 3 cache in the system.
> +
> --per-core::
> Aggregate counts per physical processor for system-wide mode measurements. This
> is a useful mode to detect imbalance between physical cores. To enable this mode,
> @@ -379,6 +387,14 @@ Aggregate counts per processor socket for system-wide mode measurements.
> --per-die::
> Aggregate counts per processor die for system-wide mode measurements.
>
> +--per-cache::
> +Aggregate counts per cache instance for system-wide mode measurements. By
> +default, the aggregation happens for the cache level at the highest index
> +in the system. To specify a particular level, mention the cache level
> +alongside the option in the format [Ll][1-9][0-9]*. For example: Using
> +option "--per-cache=l3" or "--per-cache=L3" will aggregate the
> +information at the boundary of the level 3 cache in the system.
> +
> --per-core::
> Aggregate counts per physical processor for system-wide mode measurements.
>
> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> index fa7c40956d0f..744890898c13 100644
> --- a/tools/perf/builtin-stat.c
> +++ b/tools/perf/builtin-stat.c
> @@ -203,6 +203,7 @@ struct perf_stat {
> struct perf_cpu_map *cpus;
> struct perf_thread_map *threads;
> enum aggr_mode aggr_mode;
> + u32 aggr_level;
> };
>
> static struct perf_stat perf_stat;
> @@ -210,8 +211,9 @@ static struct perf_stat perf_stat;
>
> static volatile sig_atomic_t done = 0;
>
> -static struct perf_stat_config stat_config = {
> +struct perf_stat_config stat_config = {
> .aggr_mode = AGGR_GLOBAL,
> + .aggr_level = MAX_CACHE_LVL + 1,
> .scale = true,
> .unit_width = 4, /* strlen("unit") */
> .run_count = 1,
> @@ -1160,6 +1162,55 @@ static int parse_hybrid_type(const struct option *opt,
> return 0;
> }
>
> +static int parse_cache_level(const struct option *opt,
> + const char *str,
> + int unset __maybe_unused)
> +{
> + int level;
> + u32 *aggr_mode = (u32 *)opt->value;
> + u32 *aggr_level = (u32 *)opt->data;
> +
> + /*
> + * If no string is specified, aggregate based on the topology of
> + * Last Level Cache (LLC). Since the LLC level can change from
> + * architecture to architecture, set level greater than
> + * MAX_CACHE_LVL which will be interpreted as LLC.
> + */
> + if (str == NULL) {
> + level = MAX_CACHE_LVL + 1;
> + goto out;
> + }
> +
> + /*
> + * The format to specify cache level is LX or lX where X is the
> + * cache level.
> + */
> + if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
> + pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
> + MAX_CACHE_LVL,
> + MAX_CACHE_LVL);
> + return -EINVAL;
> + }
> +
> + level = atoi(&str[1]);
> + if (level < 1) {
> + pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
> + MAX_CACHE_LVL,
> + MAX_CACHE_LVL);
> + return -EINVAL;
> + }
> +
> + if (level > MAX_CACHE_LVL) {
> + pr_err("perf only supports max cache level of %d.\n"
> + "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
> + return -EINVAL;
> + }
> +out:
> + *aggr_mode = AGGR_CACHE;
> + *aggr_level = level;
> + return 0;
> +}
> +
> static struct option stat_options[] = {
> OPT_BOOLEAN('T', "transaction", &transaction_run,
> "hardware transaction statistics"),
> @@ -1237,6 +1288,9 @@ static struct option stat_options[] = {
> "aggregate counts per processor socket", AGGR_SOCKET),
> OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
> "aggregate counts per processor die", AGGR_DIE),
> + OPT_CALLBACK_OPTARG(0, "per-cache", &stat_config.aggr_mode, &stat_config.aggr_level,
> + "cache level", "aggregate count at this cache level (Default: LLC)",
> + parse_cache_level),
> OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
> "aggregate counts per physical processor core", AGGR_CORE),
> OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
> @@ -1298,6 +1352,7 @@ static struct option stat_options[] = {
>
> static const char *const aggr_mode__string[] = {
> [AGGR_CORE] = "core",
> + [AGGR_CACHE] = "cache",
> [AGGR_DIE] = "die",
> [AGGR_GLOBAL] = "global",
> [AGGR_NODE] = "node",
> @@ -1319,6 +1374,12 @@ static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __m
> return aggr_cpu_id__die(cpu, /*data=*/NULL);
> }
>
> +static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
> + struct perf_cpu cpu)
> +{
> + return aggr_cpu_id__cache(cpu, /*data=*/NULL);
> +}
> +
> static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
> struct perf_cpu cpu)
> {
> @@ -1371,6 +1432,12 @@ static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *con
> return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
> }
>
> +static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
> + struct perf_cpu cpu)
> +{
> + return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
> +}
> +
> static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
> struct perf_cpu cpu)
> {
> @@ -1402,6 +1469,8 @@ static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
> return aggr_cpu_id__socket;
> case AGGR_DIE:
> return aggr_cpu_id__die;
> + case AGGR_CACHE:
> + return aggr_cpu_id__cache;
> case AGGR_CORE:
> return aggr_cpu_id__core;
> case AGGR_NODE:
> @@ -1425,6 +1494,8 @@ static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
> return perf_stat__get_socket_cached;
> case AGGR_DIE:
> return perf_stat__get_die_cached;
> + case AGGR_CACHE:
> + return perf_stat__get_cache_id_cached;
> case AGGR_CORE:
> return perf_stat__get_core_cached;
> case AGGR_NODE:
> @@ -1537,6 +1608,60 @@ static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, voi
> return id;
> }
>
> +static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
> + u32 cache_level, struct aggr_cpu_id *id)
> +{
> + int i;
> + int caches_cnt = env->caches_cnt;
> + struct cpu_cache_level *caches = env->caches;
> +
> + id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0: cache_level;
> + id->cache = -1;
> +
> + if (!caches_cnt)
> + return;
> +
> + for (i = caches_cnt - 1; i > -1; --i) {
> + struct perf_cpu_map *cpu_map;
> + int map_contains_cpu;
> +
> + /*
> + * If user has not specified a level, find the fist level with
> + * the cpu in the map. Since building the map is expensive, do
> + * this only if levels match.
> + */
> + if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
> + continue;
> +
> + cpu_map = perf_cpu_map__new(caches[i].map);
> + map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
> + perf_cpu_map__put(cpu_map);
> +
> + if (map_contains_cpu != -1) {
> + id->cache_lvl = caches[i].level;
> + id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
> + return;
> + }
> + }
> +}
> +
> +static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
> + void *data)
> +{
> + struct perf_env *env = data;
> + struct aggr_cpu_id id = aggr_cpu_id__empty();
> +
> + if (cpu.cpu != -1) {
> + u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
> +
> + id.socket = env->cpu[cpu.cpu].socket_id;
> + id.die = env->cpu[cpu.cpu].die_id;
> + perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
> + }
> +
> + return id;
> +}
> +
> static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
> {
> struct perf_env *env = data;
> @@ -1605,6 +1730,12 @@ static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *confi
> return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
> }
>
> +static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
> + struct perf_cpu cpu)
> +{
> + return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
> +}
> +
> static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
> struct perf_cpu cpu)
> {
> @@ -1636,6 +1767,8 @@ static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
> return perf_env__get_socket_aggr_by_cpu;
> case AGGR_DIE:
> return perf_env__get_die_aggr_by_cpu;
> + case AGGR_CACHE:
> + return perf_env__get_cache_aggr_by_cpu;
> case AGGR_CORE:
> return perf_env__get_core_aggr_by_cpu;
> case AGGR_NODE:
> @@ -1659,6 +1792,8 @@ static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
> return perf_stat__get_socket_file;
> case AGGR_DIE:
> return perf_stat__get_die_file;
> + case AGGR_CACHE:
> + return perf_stat__get_cache_file;
> case AGGR_CORE:
> return perf_stat__get_core_file;
> case AGGR_NODE:
> @@ -2207,7 +2342,8 @@ static struct perf_stat perf_stat = {
> .stat = perf_event__process_stat_event,
> .stat_round = process_stat_round_event,
> },
> - .aggr_mode = AGGR_UNSET,
> + .aggr_mode = AGGR_UNSET,
> + .aggr_level = 0,
> };
>
> static int __cmd_report(int argc, const char **argv)
> @@ -2219,6 +2355,10 @@ static int __cmd_report(int argc, const char **argv)
> "aggregate counts per processor socket", AGGR_SOCKET),
> OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
> "aggregate counts per processor die", AGGR_DIE),
> + OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
> + "cache level",
> + "aggregate count at this cache level (Default: LLC)",
> + parse_cache_level),
> OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
> "aggregate counts per physical processor core", AGGR_CORE),
> OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
> diff --git a/tools/perf/tests/shell/lib/perf_json_output_lint.py b/tools/perf/tests/shell/lib/perf_json_output_lint.py
> index 97598d14e532..62489766b93c 100644
> --- a/tools/perf/tests/shell/lib/perf_json_output_lint.py
> +++ b/tools/perf/tests/shell/lib/perf_json_output_lint.py
> @@ -14,6 +14,7 @@ ap.add_argument('--system-wide', action='store_true')
> ap.add_argument('--event', action='store_true')
> ap.add_argument('--per-core', action='store_true')
> ap.add_argument('--per-thread', action='store_true')
> +ap.add_argument('--per-cache', action='store_true')
> ap.add_argument('--per-die', action='store_true')
> ap.add_argument('--per-node', action='store_true')
> ap.add_argument('--per-socket', action='store_true')
> @@ -46,6 +47,7 @@ def check_json_output(expected_items):
> 'counter-value': lambda x: is_counter_value(x),
> 'cgroup': lambda x: True,
> 'cpu': lambda x: isint(x),
> + 'cache': lambda x: True,
> 'die': lambda x: True,
> 'event': lambda x: True,
> 'event-runtime': lambda x: isfloat(x),
> @@ -82,7 +84,7 @@ try:
> expected_items = 7
> elif args.interval or args.per_thread or args.system_wide_no_aggr:
> expected_items = 8
> - elif args.per_core or args.per_socket or args.per_node or args.per_die:
> + elif args.per_core or args.per_socket or args.per_node or args.per_die or args.per_cache_instance:
> expected_items = 9
> else:
> # If no option is specified, don't check the number of items.
> diff --git a/tools/perf/tests/shell/stat+csv_output.sh b/tools/perf/tests/shell/stat+csv_output.sh
> index 324fc9e6edd7..6cdf2fd386d5 100755
> --- a/tools/perf/tests/shell/stat+csv_output.sh
> +++ b/tools/perf/tests/shell/stat+csv_output.sh
> @@ -26,6 +26,7 @@ function commachecker()
> ;; "--per-socket") exp=8
> ;; "--per-node") exp=8
> ;; "--per-die") exp=8
> + ;; "--per-cache") exp=8
> esac
>
> while read line
> @@ -123,6 +124,18 @@ check_per_thread()
> echo "[Success]"
> }
>
> +check_per_cache_instance()
> +{
> + echo -n "Checking CSV output: per cache instance "
> + if ParanoidAndNotRoot 0
> + then
> + echo "[Skip] paranoid and not root"
> + return
> + fi
> + perf stat -x$csv_sep --per-cache -a true 2>&1 | commachecker --per-cache
> + echo "[Success]"
> +}
> +
> check_per_die()
> {
> echo -n "Checking CSV output: per die "
> @@ -197,6 +210,7 @@ if [ $skip_test -ne 1 ]
> then
> check_system_wide_no_aggr
> check_per_core
> + check_per_cache_instance
> check_per_die
> check_per_socket
> else
> diff --git a/tools/perf/tests/shell/stat+json_output.sh b/tools/perf/tests/shell/stat+json_output.sh
> index 2c4212c641ed..d79a6e0d4042 100755
> --- a/tools/perf/tests/shell/stat+json_output.sh
> +++ b/tools/perf/tests/shell/stat+json_output.sh
> @@ -100,6 +100,18 @@ check_per_thread()
> echo "[Success]"
> }
>
> +check_per_cache_instance()
> +{
> + echo -n "Checking json output: per cache_instance "
> + if ParanoidAndNotRoot 0
> + then
> + echo "[Skip] paranoia and not root"
> + return
> + fi
> + perf stat -j --per-cache -a true 2>&1 | $PYTHON $pythonchecker --per-cache
> + echo "[Success]"
> +}
> +
> check_per_die()
> {
> echo -n "Checking json output: per die "
> @@ -174,6 +186,7 @@ if [ $skip_test -ne 1 ]
> then
> check_system_wide_no_aggr
> check_per_core
> + check_per_cache_instance
> check_per_die
> check_per_socket
> else
> diff --git a/tools/perf/util/cpumap.c b/tools/perf/util/cpumap.c
> index 5e564974fba4..98dd308fcd21 100644
> --- a/tools/perf/util/cpumap.c
> +++ b/tools/perf/util/cpumap.c
> @@ -3,6 +3,8 @@
> #include "cpumap.h"
> #include "debug.h"
> #include "event.h"
> +#include "header.h"
> +#include "stat.h"
> #include <assert.h>
> #include <dirent.h>
> #include <stdio.h>
> @@ -227,6 +229,10 @@ static int aggr_cpu_id__cmp(const void *a_pointer, const void *b_pointer)
> return a->socket - b->socket;
> else if (a->die != b->die)
> return a->die - b->die;
> + else if (a->cache_lvl != b->cache_lvl)
> + return a->cache_lvl - b->cache_lvl;
> + else if (a->cache != b->cache)
> + return a->cache - b->cache;
> else if (a->core != b->core)
> return a->core - b->core;
> else
> @@ -310,6 +316,112 @@ struct aggr_cpu_id aggr_cpu_id__die(struct perf_cpu cpu, void *data)
> return id;
> }
>
> +extern struct perf_stat_config stat_config;
> +
> +int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
> +{
> + int id;
> + struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
> +
> + /*
> + * If the map contains no CPU, consider the current CPU to
> + * be the first online CPU in the cache domain else use the
> + * first online CPU of the cache domain as the ID.
> + */
> + if (perf_cpu_map__empty(cpu_map))
> + id = cpu.cpu;
> + else
> + id = perf_cpu_map__cpu(cpu_map, 0).cpu;
> +
> + /* Free the perf_cpu_map used to find the cache ID */
> + perf_cpu_map__put(cpu_map);
> +
> + return id;
> +}
> +int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
> +{
> + int ret = 0;
> + struct cpu_cache_level caches[MAX_CACHE_LVL];
> + u32 cache_level = stat_config.aggr_level;
> + u32 i = 0, caches_cnt = 0;
> +
> + cache->cache_lvl = (cache_level > MAX_CACHE_LVL)? 0: cache_level;
> + cache->cache = -1;
> +
> + ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
> + if (ret) {
> + /*
> + * If caches_cnt is not 0, cpu_cache_level data
> + * was allocated when building the topology.
> + * Free the allocated data before returning.
> + */
> + if (caches_cnt)
> + goto free_caches;
> +
> + return ret;
> + }
> +
> + if (!caches_cnt)
> + return -1;
> +
> + /*
> + * Save the data for the highest level if no
> + * level was specified by the user.
> + */
> + if (cache_level > MAX_CACHE_LVL) {
> + int max_level_index = 0;
> +
> + for (i = 1; i < caches_cnt; ++i) {
> + if (caches[i].level > caches[max_level_index].level)
> + max_level_index = i;
> + }
> +
> + cache->cache_lvl = caches[max_level_index].level;
> + cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
> +
> + /* Reset i to 0 to free entire caches[] */
> + i = 0;
> + goto free_caches;
> + }
> +
> + for (i = 0; i < caches_cnt; ++i) {
> + if (caches[i].level == cache_level) {
> + cache->cache_lvl = cache_level;
> + cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
> + }
> +
> + cpu_cache_level__free(&caches[i]);
> + }
> +
> +free_caches:
> + /*
> + * Free all the allocated cpu_cache_level data.
> + */
> + while (i < caches_cnt)
> + cpu_cache_level__free(&caches[i++]);
> +
> + return ret;
> +}
> +
> +struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
> +{
> + int ret;
> + struct aggr_cpu_id id;
> + struct perf_cache cache;
> +
> + id = aggr_cpu_id__die(cpu, data);
> + if (aggr_cpu_id__is_empty(&id))
> + return id;
> +
> + ret = cpu__get_cache_details(cpu, &cache);
> + if (ret)
> + return id;
> +
> + id.cache_lvl = cache.cache_lvl;
> + id.cache = cache.cache;
> + return id;
> +}
> +
> int cpu__get_core_id(struct perf_cpu cpu)
> {
> int value, ret = cpu__get_topology_int(cpu.cpu, "core_id", &value);
> @@ -684,6 +796,8 @@ bool aggr_cpu_id__equal(const struct aggr_cpu_id *a, const struct aggr_cpu_id *b
> a->node == b->node &&
> a->socket == b->socket &&
> a->die == b->die &&
> + a->cache_lvl == b->cache_lvl &&
> + a->cache == b->cache &&
> a->core == b->core &&
> a->cpu.cpu == b->cpu.cpu;
> }
> @@ -694,6 +808,8 @@ bool aggr_cpu_id__is_empty(const struct aggr_cpu_id *a)
> a->node == -1 &&
> a->socket == -1 &&
> a->die == -1 &&
> + a->cache_lvl == -1 &&
> + a->cache == -1 &&
> a->core == -1 &&
> a->cpu.cpu == -1;
> }
> @@ -705,6 +821,8 @@ struct aggr_cpu_id aggr_cpu_id__empty(void)
> .node = -1,
> .socket = -1,
> .die = -1,
> + .cache_lvl = -1,
> + .cache = -1,
> .core = -1,
> .cpu = (struct perf_cpu){ .cpu = -1 },
> };
> diff --git a/tools/perf/util/cpumap.h b/tools/perf/util/cpumap.h
> index c2f5824a3a22..686ef904f57b 100644
> --- a/tools/perf/util/cpumap.h
> +++ b/tools/perf/util/cpumap.h
> @@ -20,6 +20,13 @@ struct aggr_cpu_id {
> int socket;
> /** The die id as read from /sys/devices/system/cpu/cpuX/topology/die_id. */
> int die;
> + /** The cache level as read from /sys/devices/system/cpu/cpuX/cache/indexY/level */
> + int cache_lvl;
> + /**
> + * The cache instance ID, which is the first CPU in the
> + * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
> + */
> + int cache;
> /** The core id as read from /sys/devices/system/cpu/cpuX/topology/core_id. */
> int core;
> /** CPU aggregation, note there is one CPU for each SMT thread. */
> @@ -76,6 +83,20 @@ int cpu__get_socket_id(struct perf_cpu cpu);
> * /sys/devices/system/cpu/cpuX/topology/die_id for the given CPU.
> */
> int cpu__get_die_id(struct perf_cpu cpu);
> +/**
> + * Calculate the cache instance ID from the map in
> + * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
> + * Cache instance ID is the first CPU reported in the shared_cpu_list file.
> + */
> +int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map);
> +/**
> + * cpu__get_cache_id - Returns 0 if successful in populating the
> + * cache level and cache id. Cache level is read from
> + * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
> + * is the first CPU reported by
> + * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
> + */
> +int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache);
> /**
> * cpu__get_core_id - Returns the core id as read from
> * /sys/devices/system/cpu/cpuX/topology/core_id for the given CPU.
> @@ -116,6 +137,13 @@ struct aggr_cpu_id aggr_cpu_id__socket(struct perf_cpu cpu, void *data);
> * aggr_cpu_id_get_t.
> */
> struct aggr_cpu_id aggr_cpu_id__die(struct perf_cpu cpu, void *data);
> +/**
> + * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
> + * level, die and socket populated with the cache instache ID, cache level,
> + * die and socket for cpu. The function signature is compatible with
> + * aggr_cpu_id_get_t.
> + */
> +struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data);
> /**
> * aggr_cpu_id__core - Create an aggr_cpu_id with the core, die and socket
> * populated with the core, die and socket for cpu. The function signature is
> diff --git a/tools/perf/util/event.c b/tools/perf/util/event.c
> index 1fa14598b916..faf0df3c5b95 100644
> --- a/tools/perf/util/event.c
> +++ b/tools/perf/util/event.c
> @@ -135,9 +135,10 @@ void perf_event__read_stat_config(struct perf_stat_config *config,
> config->__val = event->data[i].val; \
> break;
>
> - CASE(AGGR_MODE, aggr_mode)
> - CASE(SCALE, scale)
> - CASE(INTERVAL, interval)
> + CASE(AGGR_MODE, aggr_mode)
> + CASE(SCALE, scale)
> + CASE(INTERVAL, interval)
> + CASE(AGGR_LEVEL, aggr_level)
> #undef CASE
> default:
> pr_warning("unknown stat config term %" PRI_lu64 "\n",
> diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c
> index 1b5cb20efd23..82ec668bc3ba 100644
> --- a/tools/perf/util/stat-display.c
> +++ b/tools/perf/util/stat-display.c
> @@ -36,6 +36,7 @@
>
> static int aggr_header_lens[] = {
> [AGGR_CORE] = 18,
> + [AGGR_CACHE] = 22,
> [AGGR_DIE] = 12,
> [AGGR_SOCKET] = 6,
> [AGGR_NODE] = 6,
> @@ -46,6 +47,7 @@ static int aggr_header_lens[] = {
>
> static const char *aggr_header_csv[] = {
> [AGGR_CORE] = "core,cpus,",
> + [AGGR_CACHE] = "cache,cpus,",
> [AGGR_DIE] = "die,cpus,",
> [AGGR_SOCKET] = "socket,cpus,",
> [AGGR_NONE] = "cpu,",
> @@ -56,6 +58,7 @@ static const char *aggr_header_csv[] = {
>
> static const char *aggr_header_std[] = {
> [AGGR_CORE] = "core",
> + [AGGR_CACHE] = "cache",
> [AGGR_DIE] = "die",
> [AGGR_SOCKET] = "socket",
> [AGGR_NONE] = "cpu",
> @@ -193,6 +196,10 @@ static void print_aggr_id_std(struct perf_stat_config *config,
> case AGGR_CORE:
> snprintf(buf, sizeof(buf), "S%d-D%d-C%d", id.socket, id.die, id.core);
> break;
> + case AGGR_CACHE:
> + snprintf(buf, sizeof(buf), "S%d-D%d-L%d-ID%d",
> + id.socket, id.die, id.cache_lvl, id.cache);
> + break;
> case AGGR_DIE:
> snprintf(buf, sizeof(buf), "S%d-D%d", id.socket, id.die);
> break;
> @@ -239,6 +246,10 @@ static void print_aggr_id_csv(struct perf_stat_config *config,
> fprintf(output, "S%d-D%d-C%d%s%d%s",
> id.socket, id.die, id.core, sep, nr, sep);
> break;
> + case AGGR_CACHE:
> + fprintf(config->output, "S%d-D%d-L%d-ID%d%s%d%s",
> + id.socket, id.die, id.cache_lvl, id.cache, sep, nr, sep);
> + break;
> case AGGR_DIE:
> fprintf(output, "S%d-D%d%s%d%s",
> id.socket, id.die, sep, nr, sep);
> @@ -284,6 +295,10 @@ static void print_aggr_id_json(struct perf_stat_config *config,
> fprintf(output, "\"core\" : \"S%d-D%d-C%d\", \"aggregate-number\" : %d, ",
> id.socket, id.die, id.core, nr);
> break;
> + case AGGR_CACHE:
> + fprintf(output, "\"cache\" : \"S%d-D%d-L%d-ID%d\", \"aggregate-number\" : %d, ",
> + id.socket, id.die, id.cache_lvl, id.cache, nr);
> + break;
> case AGGR_DIE:
> fprintf(output, "\"die\" : \"S%d-D%d\", \"aggregate-number\" : %d, ",
> id.socket, id.die, nr);
> @@ -1126,6 +1141,7 @@ static void print_header_interval_std(struct perf_stat_config *config,
> case AGGR_NODE:
> case AGGR_SOCKET:
> case AGGR_DIE:
> + case AGGR_CACHE:
> case AGGR_CORE:
> fprintf(output, "#%*s %-*s cpus",
> INTERVAL_LEN - 1, "time",
> @@ -1422,6 +1438,7 @@ void evlist__print_counters(struct evlist *evlist, struct perf_stat_config *conf
>
> switch (config->aggr_mode) {
> case AGGR_CORE:
> + case AGGR_CACHE:
> case AGGR_DIE:
> case AGGR_SOCKET:
> case AGGR_NODE:
> diff --git a/tools/perf/util/stat-shadow.c b/tools/perf/util/stat-shadow.c
> index 806b32156459..f080905a3ece 100644
> --- a/tools/perf/util/stat-shadow.c
> +++ b/tools/perf/util/stat-shadow.c
> @@ -20,6 +20,7 @@
> * AGGR_GLOBAL: Use CPU 0
> * AGGR_SOCKET: Use first CPU of socket
> * AGGR_DIE: Use first CPU of die
> + * AGGR_CACHE: Use first CPU of cache level instance
> * AGGR_CORE: Use first CPU of core
> * AGGR_NONE: Use matching CPU
> * AGGR_THREAD: Not supported?
> diff --git a/tools/perf/util/stat.h b/tools/perf/util/stat.h
> index bf1794ebc916..848b3b3f5819 100644
> --- a/tools/perf/util/stat.h
> +++ b/tools/perf/util/stat.h
> @@ -74,6 +74,7 @@ enum aggr_mode {
> AGGR_GLOBAL,
> AGGR_SOCKET,
> AGGR_DIE,
> + AGGR_CACHE,
> AGGR_CORE,
> AGGR_THREAD,
> AGGR_UNSET,
> @@ -139,6 +140,7 @@ typedef struct aggr_cpu_id (*aggr_get_id_t)(struct perf_stat_config *config, str
>
> struct perf_stat_config {
> enum aggr_mode aggr_mode;
> + u32 aggr_level;
> bool scale;
> bool no_inherit;
> bool identifier;
> diff --git a/tools/perf/util/synthetic-events.c b/tools/perf/util/synthetic-events.c
> index 9ab9308ee80c..2fe648be1e7d 100644
> --- a/tools/perf/util/synthetic-events.c
> +++ b/tools/perf/util/synthetic-events.c
> @@ -1373,6 +1373,7 @@ int perf_event__synthesize_stat_config(struct perf_tool *tool,
> ADD(AGGR_MODE, config->aggr_mode)
> ADD(INTERVAL, config->interval)
> ADD(SCALE, config->scale)
> + ADD(AGGR_LEVEL, config->aggr_level)
>
> WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
> "stat config terms unbalanced\n");
> --
> 2.34.1
>
Hello Ian,
Thank you for reviewing the series.
On 4/13/2023 10:18 PM, Ian Rogers wrote:
> On Wed, Apr 12, 2023 at 11:22 PM K Prateek Nayak <[email protected]> wrote:
>>
>> Processors based on chiplet architecture, such as AMD EPYC and Hygon do
>> not expose the chiplet details in the sysfs CPU topology information.
>> However, this information can be derived from the per CPU cache level
>> information from the sysfs.
>>
>> perf stat has already supported aggregation based on topology
>> information using core ID, socket ID, etc. It'll be useful to aggregate
>> based on the cache topology to detect problems like imbalance and
>> cache-to-cache sharing at various cache levels.
>>
>> This patch adds support for "--per-cache" option for aggregation at a
>> particular cache level. Also update the docs and related test. The
>> output will be like:
>>
>> $ sudo perf stat --per-cache=L3 -a -e ls_dmnd_fills_from_sys.ext_cache_remote -- sleep 5
>>
>> Performance counter stats for 'system wide':
>>
>> S0-D0-L3-ID0 16 12,644,599 ls_dmnd_fills_from_sys.ext_cache_remote
>> S0-D0-L3-ID8 16 13,847,598 ls_dmnd_fills_from_sys.ext_cache_remote
>> S0-D0-L3-ID16 16 223,592 ls_dmnd_fills_from_sys.ext_cache_remote
>> S0-D0-L3-ID24 16 7,775 ls_dmnd_fills_from_sys.ext_cache_remote
>> S0-D0-L3-ID32 16 31,302 ls_dmnd_fills_from_sys.ext_cache_remote
>> S0-D0-L3-ID40 16 17,722 ls_dmnd_fills_from_sys.ext_cache_remote
>> S0-D0-L3-ID48 16 8,272 ls_dmnd_fills_from_sys.ext_cache_remote
>> S0-D0-L3-ID56 16 5,765 ls_dmnd_fills_from_sys.ext_cache_remote
>> S1-D1-L3-ID64 16 18,227,173 ls_dmnd_fills_from_sys.ext_cache_remote
>> S1-D1-L3-ID72 16 20,926,878 ls_dmnd_fills_from_sys.ext_cache_remote
>> S1-D1-L3-ID80 16 13,705 ls_dmnd_fills_from_sys.ext_cache_remote
>> S1-D1-L3-ID88 16 24,062 ls_dmnd_fills_from_sys.ext_cache_remote
>> S1-D1-L3-ID96 16 27,891 ls_dmnd_fills_from_sys.ext_cache_remote
>> S1-D1-L3-ID104 16 13,480 ls_dmnd_fills_from_sys.ext_cache_remote
>> S1-D1-L3-ID112 16 12,656 ls_dmnd_fills_from_sys.ext_cache_remote
>> S1-D1-L3-ID120 16 21,881 ls_dmnd_fills_from_sys.ext_cache_remote
>>
>> Also support perf stat record and perf stat report with the ability to
>> specify a different cache level to aggregate data at when running perf
>> stat report.
>>
>> Suggested-by: Gautham R. Shenoy <[email protected]>
>> Signed-off-by: K Prateek Nayak <[email protected]>
>> ---
>> Changelog:
>> o v2->v3
>> - Cache IDs are now derived from the shared_cpus_list making
>> aggregation possible on older kernels that do not expose the IDs.
>> - Updated the comments based on the new ID assignment strategy.
>> - Better handle the case when specifying a level is possible as it is
>> less than or equal to MAX_CACHE_LVL but it does not exist on the
>> machine. In such cased ID will be -1.
>>
>> $ sudo perf stat --per-cache=L4 -a -e cycles -- sleep 5
>>
>> Performance counter stats for 'system wide':
>>
>> S0-D0-L4-ID-1 128 51,328,613 cycles
>> S1-D1-L4-ID-1 128 125,132,221 cycles
>>
>> o v1->v2
>> - Fix cache level parsing logic. Previously, giving "--per-cache=2" was
>> considered valid but that was not the intention. Only parse strings
>> of form LX or lX where x is the cache level and can range from 1 to
>> MAX_CACHE_LVL.
>> ---
>> tools/lib/perf/include/perf/cpumap.h | 5 +
>> tools/lib/perf/include/perf/event.h | 3 +-
>> tools/perf/Documentation/perf-stat.txt | 16 ++
>> tools/perf/builtin-stat.c | 144 +++++++++++++++++-
>> .../tests/shell/lib/perf_json_output_lint.py | 4 +-
>> tools/perf/tests/shell/stat+csv_output.sh | 14 ++
>> tools/perf/tests/shell/stat+json_output.sh | 13 ++
>
> I think you can break this patch apart. You can add the feature, then
> enable the CSV test, then the json test, etc. When adding the feature
> you can also probably split the changes of "struct aggr_cpu_id" from
> the display/enablement logic.
I agree this patch is huge. I'll break it down as per your suggestion in
the next version where I also plan to drop the RFC tag.
> Overall it looks good!
Thank you again for taking a look at the series.
>
> Thanks,
> Ian
>
>> [..snip..]
>>
--
Thanks and Regards,
Prateek