- 25 Oct, 2021 7 commits
-
-
Alexey Bayduraev authored
Separate the reading code of a single event to a new reader__read_event() function. Suggested-by:
Jiri Olsa <jolsa@kernel.org> Reviewed-by:
Jiri Olsa <jolsa@redhat.com> Reviewed-by:
Riccardo Mancini <rickyman7@gmail.com> Signed-off-by:
Alexey Bayduraev <alexey.v.bayduraev@linux.intel.com> Tested-by:
Namhyung Kim <namhyung@gmail.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Antonov <alexander.antonov@linux.intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Alexei Budankov <abudankov@huawei.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: https://lore.kernel.org/r/ffe570d937138dd24f282978ce7ed9c46a06ff9b.1634113027.git.alexey.v.bayduraev@linux.intel.comSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Alexey Bayduraev authored
Move the unmapping code to reader__mmap(), so that the mmap code is located together. Move the head/file_offset computation to reader__mmap(), so all the offset computation is located together and in one place only. Suggested-by:
-
Alexey Bayduraev authored
Move the mapping code into a separate reader__mmap() function. Suggested-by:
-
Alexey Bayduraev authored
Separate init/release code into reader__init() and reader__release_decomp() functions. Remove a duplicate call to ui_progress__init_size(), the same call can be found in __perf_session__process_events(). For multiple traces ui_progress should be initialized by total size before reader__init() calls. Suggested-by:
-
Alexey Bayduraev authored
Introduce a decompressor data structure with pointers to decomp objects and to zstd object. We cannot just move session->zstd_data to decomp_data as session->zstd_data is not only used for decompression. Adding decompressor data object to reader object and introducing active_decomp into perf_session object to select current decompressor. Thus decompression could be executed separately for each data file. Reviewed-by:
-
Alexey Bayduraev authored
We need all the state info about reader in separate object to load data from multiple files, so we can keep multiple readers at the same time. Moving all items that need to be kept from reader__process_events to the reader object. Introducing mmap_cur to keep current mapping. Suggested-by:
-
Adrian Hunter authored
Originally, software only supported redirecting at most one PEBS event to Intel PT (PEBS-via-PT) because it was not able to differentiate one event from another. To overcome that, add support for the PERF_RECORD_AUX_OUTPUT_HW_ID side-band event. Committer notes: Cast the pointer arg to for_each_set_bit() to (unsigned long *), to fix the build on 32-bit systems. Reviewed-by:
Alexander Shishkin <alexander.shishkin@linux.intel.com> Reviewed-by:
Andi Kleen <ak@linux.intel.com> Signed-off-by:
Adrian Hunter <adrian.hunter@intel.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Kan Liang <kan.liang@linux.intel.com> Cc: Leo Yan <leo.yan@linaro.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: x86@kernel.org Link: http://lore.kernel.org/lkml/20210907163903.11820-4-adrian.hunter@intel.comSigned-off-by:
-
- 20 Oct, 2021 25 commits
-
-
Adrian Hunter authored
The PERF_RECORD_AUX_OUTPUT_HW_ID event provides a way to match AUX output data like Intel PT PEBS-via-PT back to the event that it came from, by providing a hardware ID that is present in the AUX output. Reviewed-by:
-
Andrew Kilroy authored
This is so they are categorised in the perf list output. The pmus all exist in the armv8-common-and-microarch.json and arm-recommended.json files, so this commit places them into each category's own file under tools/perf/pmu-events/arch/arm64/arm/neoverse-v1 Also add the Neoverse V1 to the arm64 mapfile Reviewed-by:
John Garry <john.garry@huawei.com> Signed-off-by:
Andrew Kilroy <andrew.kilroy@arm.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Leo Yan <leo.yan@linaro.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Poirier <mathieu.poirier@linaro.org> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Will Deacon <will@kernel.org> Cc: linux-arm-kernel@lists.infradead.org Link: https://lore.kernel.org/r/20211006081106.8649-3-andrew.kilroy@arm.comSigned-off-by:
-
Andrew Kilroy authored
Add new armv8 common events for use by Arm Neoverse V1 cores in a later commit. These are defined in the ArmV8 architecture reference manual available from https://developer.arm.com/documentation/ddi0487/gb/?lang=enReviewed-by:
-
Andrew Kilroy authored
There are some syntactical mistakes in the json files for the Cortex A76 N1 (Neoverse N1). This was obstructing parsing from an external tool. This patch fixes the erroneous placement of commas causing the problems. Reviewed-by:
-
Ian Rogers authored
By allowing modifiers on metrics we can, for example, gather the same metric for kernel and user mode. On a SkylakeX with TopDownL1 this gives: $ perf stat -M TopDownL1:u,TopDownL1:k -a sleep 2 Performance counter stats for 'system wide': 849,855,577 uops_issued.any:k # 0.06 Bad_Speculation:k # 0.51 Backend_Bound:k (16.71%) 1,995,257,996 cycles:k # 7981031984.00 SLOTS:k # 0.35 Frontend_Bound:k # 0.08 Retiring:k (16.71%) 2,791,940,753 idq_uops_not_delivered.core:k (16.71%) 641,961,928 uops_retired.retire_slots:k (16.71%) 72,239,337 int_misc.recovery_cycles:k (16.71%) 2,294,413,647 uops_issued.any:u # 0.04 Bad_Speculation:u # 0.39 Backend_Bound:u (16.78%) 1,333,248,940 cycles:u # 5332995760.00 SLOTS:u # 0.16 Frontend_Bound:u # 0.40 Retiring:u (16.78%) 858,517,081 idq_uops_not_delivered.core:u (16.78%) 2,153,789,582 uops_retired.retire_slots:u (16.78%) 19,373,627 int_misc.recovery_cycles:u (16.78%) 31,503,661 cpu_clk_unhalted.one_thread_active:k # 0.18 CoreIPC_SMT:k (16.73%) 315,454,104 inst_retired.any:k # 315454104.00 Instructions:k (16.73%) 42,533,729 cpu_clk_unhalted.ref_xclk:k (16.73%) 2,043,119,037 cpu_clk_unhalted.thread:k (16.73%) 28,843,803 cpu_clk_unhalted.one_thread_active:u # 1.55 CoreIPC_SMT:u (16.60%) 2,153,353,869 inst_retired.any:u # 2153353869.00 Instructions:u (16.60%) 28,844,743 cpu_clk_unhalted.ref_xclk:u (16.60%) 1,387,544,378 cpu_clk_unhalted.thread:u (16.60%) 308,031,603 inst_retired.any:k # 0.15 CoreIPC:k (33.19%) 2,036,774,753 cycles:k (33.19%) 1,994,344,281 inst_retired.any:u # 1.59 CoreIPC:u (33.18%) 1,251,538,227 cycles:u (33.18%) 2.000342948 seconds time elapsed Modifiers are naively copy and pasted on to events, this can yield errors like: $ perf stat -M Kernel_Utilization:k -a sleep 2 event syntax error: '..d.thread:k/kk,cpu_clk_unhalted.thread/metric-id=cpu_clk_unhalted.thread/k..' \___ Bad modifier Usage: perf stat [<options>] [<command>] -M, --metrics <metric/metric group list> monitor specified metrics or metric groups (separated by ,) When modifiers are present with constraints, from --metric-no-group or the NMI watchdog, they are no longer placed in the same set - which may miss deduplicating events. Signed-off-by:Ian Rogers <irogers@google.com> Acked-by:
-
Ian Rogers authored
Previously the broken modifier causes a usage message to printed but nothing else. After: $ perf stat -e 'cycles:kk' -a sleep 2 event syntax error: 'cycles:kk' \___ Bad modifier Run 'perf list' for a list of valid events Usage: perf stat [<options>] [<command>] -e, --event <event> event selector. use 'perf list' to list available events $ perf stat -e '{instructions,cycles}:kk' -a sleep 2 event syntax error: '..ns,cycles}:kk' \___ Bad modifier Run 'perf list' for a list of valid events Usage: perf stat [<options>] [<command>] -e, --event <event> event selector. use 'perf list' to list available events Signed-off-by: -
Ian Rogers authored
There's no clear reason for the inconsistency that stems from the initial commit. Signed-off-by:
-
Ian Rogers authored
Previously find_evsel_group was trying to share events while mark-sweeping to eliminate unused events, this was complicated and had issues around uncore events and grouped sharing. This was further complicated by the event string being created while metrics and metric groups were being added, with the string affecting the evlist order. This change moves deduplication before event parsing. Ungrouped events are placed in a single combined set. Groups are checked to see if an earlier (larger) group can support their events. As the deduplication and sharing detection is done on metric IDs before parsing, wildcard expansion problems with uncore events are avoided. Overall the code is simpler while working better. An example of failing to deduplicate can be seen with a list of metrics like the following, where in the after case multiplexing has been avoided: Before: $ perf stat -M Bad_Speculation,Backend_Bound,Frontend_Bound,Retiring -a sleep 2 Performance counter stats for 'system wide': 959,620,872 uops_issued.any # 0.06 Bad_Speculation (50.03%) 2,163,072,261 cycles # 0.09 Retiring (50.03%) 735,827,436 uops_retired.retire_slots (50.03%) 74,676,484 int_misc.recovery_cycles (50.03%) 987,062,794 uops_issued.any # 0.50 Backend_Bound (49.97%) 2,203,734,187 cycles # 0.35 Frontend_Bound (49.97%) 3,085,016,091 idq_uops_not_delivered.core (49.97%) 758,599,232 uops_retired.retire_slots (49.97%) 75,807,526 int_misc.recovery_cycles (49.97%) 2.002103760 seconds time elapsed After: $ sudo perf stat -M Bad_Speculation,Backend_Bound,Frontend_Bound,Retiring -a sleep 2 Performance counter stats for 'system wide': 769,694,676 uops_issued.any # 0.08 Bad_Speculation # 0.41 Backend_Bound 1,087,548,633 cycles # 0.38 Frontend_Bound # 0.14 Retiring 1,642,085,777 idq_uops_not_delivered.core 603,112,590 uops_retired.retire_slots 43,787,854 int_misc.recovery_cycles 2.003844383 seconds time elapsed Signed-off-by: -
Ian Rogers authored
Add a helper that returns true if all the IDs in needles are present in haystack. Later this will be used in sharing events between metrics. Signed-off-by:
-
Ian Rogers authored
For a metric like IPC a group of events like {instructions,cycles}:W would be formed. If the events names were changed in parsing then the metric expression parser would fail to find them. This change makes the event encoding be something like: {instructions/metric-id=instructions/, cycles/metric-id=cycles/} and then uses the evsel's stable metric-id value to locate the events. This fixes the case that an event is restricted to user because of the paranoia setting: $ echo 2 > /proc/sys/kernel/perf_event_paranoid $ perf stat -M IPC /bin/true Performance counter stats for '/bin/true': 150,298 inst_retired.any:u # 0.77 IPC 187,095 cpu_clk_unhalted.thread:u 0.002042731 seconds time elapsed 0.000000000 seconds user 0.002377000 seconds sys Adding the metric-id as a qualifier has a complication in that qualifiers will become embedded in qualifiers. For example, msr/tsc/ could become msr/tsc,metric-id=msr/tsc// which will fail parse-events. To solve this problem the metric is encoded and decoded for the metric-id with !<num> standing in for an encoded value. Previously ! wasn't parsed. With this msr/tsc/ becomes msr/tsc,metric-id=msr!3tsc!3/ The metric expression parser is changed so that @ isn't changed to /, instead this is done when the ID is encoded for parse events. metricgroup__add_metric_non_group() and metricgroup__add_metric_weak_group() need to inject the metric-id qualifier, so to avoid repetition they are merged into a single metricgroup__build_event_string with error codes more rigorously checked. stat-shadow's prepare_metric() uses the metric-id to match the metricgroup code. As "metric-id=..." is added to all events, it is adding during testing with the fake PMU. This complicates pmu_str_check code as PE_PMU_EVENT_FAKE won't match as part of a configuration. The testing fake PMU case is fixed so that if a known qualifier with an ! is parsed then it isn't reported as a fake PMU. This is sufficient to pass all testing but it and the original mechanism are somewhat brittle. Signed-off-by: -
Ian Rogers authored
An event like inst_retired.any on an Intel skylake is found in the pmu-events code created from the pipeline event JSON. The event is an alias for cpu/event=0xc0,period=2000003/ and parse-events recognizes the event with the token PE_KERNEL_PMU_EVENT. The parser doesn't currently allow extra configuration on such events, except for modifiers, so: $ perf stat -e inst_retired.any// /bin/true event syntax error: 'inst_retired.any//' \___ parser error Run 'perf list' for a list of valid events Usage: perf stat [<options>] [<command>] -e, --event <event> event selector. use 'perf list' to list available events This patch adds configuration to these events which can be useful for a number of parameters like name and call-graph: $ sudo perf record -e inst_retired.any/call-graph=lbr/ -a sleep 1 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 1.856 MB perf.data (44 samples) ] It is necessary for the metric code so that we may add metric-id values to these events before they are parsed. Signed-off-by: -
Ian Rogers authored
Add a new "metric-id" term to events so that metric parsing can set an ID that can be reliably looked up. Metric parsing currently will turn a metric like "instructions/cycles" into a parse events string of "{instructions,cycles}:W". However, parse-events may change "instructions" into "instructions:u" if perf_event_paranoid=2. When this happens expr__resolve_id currently fails as stat-shadow adds the ID "instructions:u" to match with the counter value and the metric tries to look up the ID just "instructions". A later patch will use the new term. An example of the current problem: $ echo -1 > /proc/sys/kernel/perf_event_paranoid $ perf stat -M IPC /bin/true Performance counter stats for '/bin/true': 1,217,161 inst_retired.any # 0.97 IPC 1,250,389 cpu_clk_unhalted.thread 0.002064773 seconds time elapsed 0.002378000 seconds user 0.000000000 seconds sys $ echo 2 > /proc/sys/kernel/perf_event_paranoid $ perf stat -M IPC /bin/true Performance counter stats for '/bin/true': 150,298 inst_retired.any:u # nan IPC 187,095 cpu_clk_unhalted.thread:u 0.002042731 seconds time elapsed 0.000000000 seconds user 0.002377000 seconds sys Note: nan IPC is printed as an effect of "perf metric: Use NAN for missing event IDs." but earlier versions of perf just fail with a parse error and display no value. Signed-off-by: -
Ian Rogers authored
The evsel name is strdup-ed before assignment and so can be const. A later change will add another similar string. Using const makes it clearer that these are not out arguments. Signed-off-by:
-
Ian Rogers authored
Don't build a list and then turn to an array, just directly build the array. The size of the array is known due to the search for a duplicate. Signed-off-by:
-
Ian Rogers authored
Add documentation as part of code tidying. Signed-off-by:
-
Ian Rogers authored
Document the data structures maintained by metricgroup.c and used by stat-shadow.c for metric output. Signed-off-by:
-
Ian Rogers authored
Modify resolution. Rather than resolving a list of metrics, resolve a metric immediately after it is added. This simplifies knowing the root of the metric's tree so that IDs may be associated with it. A bug in the current implementation is that all the IDs were placed on the first metric in a metric group. Rather than maintain data on IDs' parents to detect cycles, maintain a list of visited metrics and detect cycles if the same metric is visited twice. Only place the root metric onto the list of metrics. Signed-off-by:
-
Ian Rogers authored
If a metric references other metrics then the same other metrics may be referenced more than once, but the events and metric ref are only needed once. An example of this is in tests/parse-metric.c where DCache_L2_Hits references the metric DCache_L2_All_Hits twice, once directly and once through DCache_L2_All. Signed-off-by:
-
Ian Rogers authored
Metrics are complex enough that a new/free reduces the risk of memory leaks. Move static functions used in new. Reviewed-by:
-
Ian Rogers authored
Documentation to make current functionality clearer. Rename a variable called 'metric' to 'metric_name' as it can be ambiguous as to whether a string is the name of a metric or the expression. Signed-off-by:
-
Ian Rogers authored
The runtime value is needed when recursively parsing metrics, currently a value of 1 is passed which is incorrect. Rather than add more arguments to the bison parser, add runtime to the context. Fix call sites not to pass a value. The runtime value is defaulted to 0, which is arbitrary. In some places this replaces a value of 1, which was also arbitrary. This shouldn't affect anything other than PPC. The use of 0 or 1 shouldn't matter as a proper runtime value would be needed in a case that it did matter. Signed-off-by:
-
Ian Rogers authored
Make lookup nature of data structures clearer through their type. Reduce scope of architecture specific pmu_event tables by making them static. Suggested-by:
-
Ian Rogers authored
Make lookup nature of data structures clearer through their type. Reviewed-by:
-
Ian Rogers authored
The pmu_events_map is generated at compile time and used for lookup. For testing purposes we need to swap the map being used. Having the pmu_events_map be non-const is misleading as it may be an out argument. Make it const and update uses so they work on const too. Reviewed-by:
-
Ian Rogers authored
Add list_sort.[ch] from the main kernel tree. The linux/bug.h #include is removed due to conflicting definitions. Add check-headers and modify perf build accordingly. MANIFEST and python-ext-sources fixes suggested by Arnaldo. Suggested-by:
Arnaldo Carvalho de Melo <acme@kernel.org> Signed-off-by:
-
- 08 Oct, 2021 8 commits
-
-
Ian Rogers authored
Since: https://lore.kernel.org/lkml/20200708183919.4141023-1-irogers@google.com/ The output option works for 'perf kmem', however, it must appear after 'record'. This is different to 'stat' where '-i' for the input must appear before. Try to capture this complication in the man page. Signed-off-by:
-
James Clark authored
This is to ensure that the PMU event files can always be parsed by other tools. Testing ======= * There are no errors when parsing files for all architectures: # pmu-events/jevents nds32 pmu-events/arch/ test # pmu-events/jevents s390 pmu-events/arch/ test # pmu-events/jevents powerpc pmu-events/arch/ test # pmu-events/jevents arm64 pmu-events/arch/ test # pmu-events/jevents test pmu-events/arch/ test # pmu-events/jevents x86 pmu-events/arch/ test * Trailing and leading commas now cause a parse error * Double commas now cause a parse error * Compilation and parsing works with strict mode disabled and enabled * A diff of the output files shows no changes Reviewed-by:James Clark <james.clark@arm.com> Acked-by:
-
James Clark authored
Return an error when a trailing comma is found or a new item is encountered before a comma or an opening brace. This ensures that the perf JSON files conform more closely to the spec at https://www.json.orgReviewed-by:
-
James Clark authored
Remove trailing commas. A later commit will make the parser more strict and these will not be valid anymore. Reviewed-by:
-
Guo Zhengkui authored
There is a "#include <sys/file.h>" in line 10, so remove a duplicate one in line 1124. Signed-off-by:
Guo Zhengkui <guozhengkui@vivo.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lore.kernel.org/lkml/20211006062235.6364-1-guozhengkui@vivo.comSigned-off-by:
-
Riccardo Mancini authored
This patch introduces a new inline function to convert a timeval to usec. This function will be used also in the next patch. Signed-off-by:
-
Riccardo Mancini authored
This patch adds a new function in util/mmap.c to duplicate a mmap_cpu_mask. This new function will be used in patches in the workqueue patchkit. Signed-off-by:
-
Riccardo Mancini authored
Since 7074674e ("perf cpumap: Maintain cpumaps ordered and without dups") perf_cpu_map elements are sorted in ascending order. This patch improves perf_cpu_map__idx() by using a binary search. Signed-off-by:
-
