- 24 Aug, 2023 16 commits
-
-
Ian Rogers authored
Don't load sysfs aliases for a PMU when the PMU is first created, defer until an alias needs to be found. For the pmu-scan benchmark, average core PMU scanning is reduced by 30.8%, and average PMU scanning by 12.6%. Signed-off-by:
Ian Rogers <irogers@google.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Gaosheng Cui <cuigaosheng1@huawei.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: James Clark <james.clark@arm.com> Cc: Jing Zhang <renyu.zj@linux.alibaba.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: John Garry <john.g.garry@oracle.com> Cc: Kajol Jain <kjain@linux.ibm.com> Cc: Kan Liang <kan.liang@linux.intel.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ravi Bangoria <ravi.bangoria@amd.com> Cc: Rob Herring <robh@kernel.org> Link: https://lore.kernel.org/r/20230824041330.266337-17-irogers@google.comSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Ian Rogers authored
Event info is only needed when an event is parsed or when merging data from an JSON and sysfs event. Be lazy in its loading to reduce file accesses. Signed-off-by:
-
Ian Rogers authored
Scan sysfs PMU's type early so that format and aliases aren't attempted to be loaded if the PMU name is invalid. This is the case for event_pmu tokens in parse-events.y where a wildcard name is first assumed to be a PMU name. Signed-off-by:
-
Ian Rogers authored
Rather than scanning all JSON events and adding them when a PMU is created, add the alias when the JSON event is needed. Average core PMU scanning run time reduced by 60.2%. Average PMU scanning run time reduced by 15%. Page faults with no events reduced by 74 page faults, 4% of total. Signed-off-by:
-
Ian Rogers authored
Cache the JSON events table so that finding it isn't done per event/alias. Change the events table find so that when the PMU is given, if the PMU has no JSON events return null. Update usage to always use the PMU variable. Signed-off-by:
-
Ian Rogers authored
Rather than load all sysfs events then parsing all JSON events and merging with ones that already exist. When a sysfs event is loaded, look for a corresponding JSON event and merge immediately. To simplify the logic, early exit the perf_pmu__new_alias function if an alias is attempted to be added twice - as merging has already been explicitly handled. Fix the copying of terms to a merged alias and some ENOMEM paths. Signed-off-by:
-
Ian Rogers authored
The aliases list is part of the PMU. Rather than pass the aliases list, pass the full PMU simplifying some callbacks. Signed-off-by:
-
Ian Rogers authored
Rather than read a sysfs events file into a 256 byte char buffer, pass the FILE* directly to the lex/yacc parser. This avoids there being a maximum events file size. While changing the API, constify some arguments to remove unnecessary casts. Allocating the read buffer decreases the performance of pmu-scan by around 3%. Signed-off-by:
-
Ian Rogers authored
jevents stores events sorted by name. Add a find function that will binary search event names avoiding the need to linearly search through events. Add a test in tests/pmu-events.c. If the PMU or event aren't found -1000 is returned. If the event is found but no callback function given, 0 is returned. This allows the find function also act as a test for existence. Signed-off-by:
-
Ian Rogers authored
Pass the PMU to pmu_events_table__for_each_event so that entries that don't match don't need to be processed by callback. If a NULL PMU is passed then all PMUs are processed. 'perf bench internals pmu-scan's "Average PMU scanning" performance is reduced by about 5% on an Intel tigerlake. Signed-off-by:
-
Ian Rogers authored
Rather than scanning all PMUs for a counter name, scan the PMU associated with the evsel of the sample. This is done to remove a dependence on pmu-events.h. Signed-off-by:
-
Ian Rogers authored
Double setting information for an event would produce an error message associated with the PMU rather than the term that was double setting. Improve the error message to be on the term. Before: $ perf stat -e 'cpu/inst_retired.any,inst_retired.any/' true event syntax error: 'cpu/inst_retired.any,inst_retired.any/' \___ Bad event or PMU Unabled to find PMU or event on a PMU of 'cpu' Run 'perf list' for a list of valid events $ After: $ perf stat -e 'cpu/inst_retired.any,inst_retired.any/' true event syntax error: '..etired.any,inst_retired.any/' \___ Bad event or PMU Unabled to find PMU or event on a PMU of 'cpu' Initial error: event syntax error: '..etired.any,inst_retired.any/' \___ Attempt to set event's scale twice Run 'perf list' for a list of valid events Signed-off-by: -
Ian Rogers authored
Prior to this change a cpuid would map to a list of events where the PMU would be encoded alongside the event information. This change breaks apart each group of events so that there is a group per PMU. A new table is added with the PMU's name and the list of events, the original table now holding an array of these per PMU tables. These changes are to make it easier to get per PMU information about events, rather than the current approach of scanning all events. The perf binary size with BPF skeletons on x86 is reduced by about 1%. The unidentified PMU is now always expanded to "cpu". Signed-off-by:
-
Ian Rogers authored
Add extra underscore before "for" of pmu_events_table_for_each_event and pmu_metrics_table_for_each_metric. Signed-off-by:
-
Ian Rogers authored
In order to be able to lazily compute aliases/events for a PMU, move the struct perf_pmu_alias into pmu.c. Add perf_pmu__find_event and perf_pmu__for_each_event that take a callback that is called for the found event or for each event. The layout of struct pmu and the event/alias list is unchanged but the API is altered so that aliases are no longer directly accessed, allowing for later changes. Signed-off-by:
-
Ian Rogers authored
The sysfs format files are loaded eagerly in a PMU. Add a flag so that we create the format but only load the contents when necessary. Reduce the size of the value in struct perf_pmu_format and avoid holes so there is no additional space requirement. For "perf stat -e cycles true" this reduces the number of openat calls from 648 to 573 (about 12%). The benchmark pmu scan speed is improved by roughly 5%. Before: $ perf bench internals pmu-scan Computing performance of sysfs PMU event scan for 100 times Average core PMU scanning took: 1061.100 usec (+- 9.965 usec) Average PMU scanning took: 4725.300 usec (+- 260.599 usec) After: $ perf bench internals pmu-scan Computing performance of sysfs PMU event scan for 100 times Average core PMU scanning took: 989.170 usec (+- 6.873 usec) Average PMU scanning took: 4520.960 usec (+- 251.272 usec) Committer testing: On a AMD Ryzen 5950x: Before: $ perf bench internals pmu-scan -i1000 # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 563.466 usec (+- 1.008 usec) Average PMU scanning took: 1619.174 usec (+- 23.627 usec) $ perf stat -r5 perf bench internals pmu-scan -i1000 # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 583.401 usec (+- 2.098 usec) Average PMU scanning took: 1677.352 usec (+- 24.636 usec) # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 553.254 usec (+- 0.825 usec) Average PMU scanning took: 1635.655 usec (+- 24.312 usec) # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 557.733 usec (+- 0.980 usec) Average PMU scanning took: 1600.659 usec (+- 23.344 usec) # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 554.906 usec (+- 0.774 usec) Average PMU scanning took: 1595.338 usec (+- 23.288 usec) # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 551.798 usec (+- 0.967 usec) Average PMU scanning took: 1623.213 usec (+- 23.998 usec) Performance counter stats for 'perf bench internals pmu-scan -i1000' (5 runs): 3276.82 msec task-clock:u # 0.990 CPUs utilized ( +- 0.82% ) 0 context-switches:u # 0.000 /sec 0 cpu-migrations:u # 0.000 /sec 1008 page-faults:u # 307.615 /sec ( +- 0.04% ) 12049614778 cycles:u # 3.677 GHz ( +- 0.07% ) (83.34%) 117507478 stalled-cycles-frontend:u # 0.98% frontend cycles idle ( +- 0.33% ) (83.32%) 27106761 stalled-cycles-backend:u # 0.22% backend cycles idle ( +- 9.55% ) (83.36%) 33294953848 instructions:u # 2.76 insn per cycle # 0.00 stalled cycles per insn ( +- 0.03% ) (83.31%) 6849825049 branches:u # 2.090 G/sec ( +- 0.03% ) (83.37%) 71533903 branch-misses:u # 1.04% of all branches ( +- 0.20% ) (83.30%) 3.3088 +- 0.0302 seconds time elapsed ( +- 0.91% ) $ After: $ perf stat -r5 perf bench internals pmu-scan -i1000 # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 550.702 usec (+- 0.958 usec) Average PMU scanning took: 1566.577 usec (+- 22.747 usec) # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 548.315 usec (+- 0.555 usec) Average PMU scanning took: 1565.499 usec (+- 22.760 usec) # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 548.073 usec (+- 0.555 usec) Average PMU scanning took: 1586.097 usec (+- 23.299 usec) # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 561.184 usec (+- 2.709 usec) Average PMU scanning took: 1567.153 usec (+- 22.548 usec) # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 546.987 usec (+- 0.553 usec) Average PMU scanning took: 1562.814 usec (+- 22.729 usec) Performance counter stats for 'perf bench internals pmu-scan -i1000' (5 runs): 3170.86 msec task-clock:u # 0.992 CPUs utilized ( +- 0.22% ) 0 context-switches:u # 0.000 /sec 0 cpu-migrations:u # 0.000 /sec 1010 page-faults:u # 318.526 /sec ( +- 0.04% ) 11890047674 cycles:u # 3.750 GHz ( +- 0.14% ) (83.27%) 119090499 stalled-cycles-frontend:u # 1.00% frontend cycles idle ( +- 0.46% ) (83.40%) 32502449 stalled-cycles-backend:u # 0.27% backend cycles idle ( +- 8.32% ) (83.30%) 33119141261 instructions:u # 2.79 insn per cycle # 0.00 stalled cycles per insn ( +- 0.01% ) (83.37%) 6812816561 branches:u # 2.149 G/sec ( +- 0.01% ) (83.29%) 70157855 branch-misses:u # 1.03% of all branches ( +- 0.28% ) (83.38%) 3.19710 +- 0.00826 seconds time elapsed ( +- 0.26% ) $ Signed-off-by:
-
- 23 Aug, 2023 10 commits
-
-
Guilherme Amadio authored
This also puts an unconditional -Werror under control of WERROR. The clang includes added during the build can lead to a warning that may be turned into an error. In addition, hardened clang produces a warning about lack of support for -fstack-protector* options for the BPF target: clang -g -O2 -target bpf -Wall -Werror -Ilinux/tools/perf/util/bpf_skel/.tmp/.. \ -I -idirafter /usr/lib/llvm/16/bin/../../../../lib/clang/16/include -idirafter /usr/local/include \ -idirafter /usr/include -Ilinux/tools/include/uapi -c util/bpf_skel/bperf_follower.bpf.c \ -o linux/tools/perf/util/bpf_skel/.tmp/bperf_follower.bpf.o && llvm-strip -g linux/tools/perf/util/bpf_skel/.tmp/bperf_follower.bpf.o clang-16: error: /usr/lib/llvm/16/bin/../../../../lib/clang/16/include: 'linker' input unused [-Werror,-Wunused-command-line-argument] clang-16: error: ignoring '-fstack-protector-strong' option as it is not currently supported for target 'bpf' [-Werror,-Woption-ignored] make[1]: *** [Makefile.perf:1082: linux/tools/perf/util/bpf_skel/.tmp/bpf_prog_profiler.bpf.o] Error 1 Signed-off-by:Guilherme Amadio <amadio@gentoo.org> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Ian Rogers <irogers@google.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Link: https://lore.kernel.org/lkml/ZOZQ2LDA+3Wg8x2T@kernel.orgSigned-off-by:
-
Ian Rogers authored
Pass the pmu so the aliases and format list can be better abstracted and later lazily loaded. Signed-off-by:
-
Ian Rogers authored
Pass the PMU so the format list can be better abstracted and later lazily loaded. Signed-off-by:
-
Ian Rogers authored
Pass the pmu so the format list can be better abstracted and later lazily loaded. Signed-off-by:
-
Ian Rogers authored
Abstract the format list better, hiding it in the PMU, by changing perf_pmu__config_terms() the PMU rather than the format list in the PMU. Change the PMU test to pass a dummy PMU for this purpose. Changing the test allows perf_pmu__del_formats() to become static. Signed-off-by:
-
Ian Rogers authored
Move declaration from header file to pmu.y and make static. Signed-off-by:
-
Ian Rogers authored
Avoid having the function in the C and header file, as it is only used locally by pmu.y. Signed-off-by:
-
Ian Rogers authored
Rather than read a base path and append into a 2nd path, read the base path directly into output buffer and append to that. Signed-off-by:
-
Ian Rogers authored
Done to reduce dependencies on pmu-events.h. Signed-off-by:
-
Kajol Jain authored
Based on commit 7d54a4ac ("perf test: Skip watchpoint tests if no watchpoints available"), hardware breakpoints are not available for power9 platform and because of that 'perf bench breakpoint' run fails on power9 platform. Add code to check for the return value of perf_event_open() in the breakpoint run and skip the 'perf bench breakpoint' run, if hardware breakpoints are not available. Result on power9 system before patch changes: [command]# perf bench breakpoint thread perf_event_open: No such device Result on power9 system after patch changes: [command]# ./perf bench breakpoint thread Skipping perf bench breakpoint thread: No hardware support Reported-by:
Disha Goel <disgoel@linux.vnet.ibm.com> Signed-off-by:
Kajol Jain <kjain@linux.ibm.com> Acked-by:
Naveen N Rao <naveen@kernel.org> Cc: Athira Rajeev <atrajeev@linux.vnet.ibm.com> Cc: Disha Goel <disgoel@linux.ibm.com> Cc: Ian Rogers <irogers@google.com> Cc: Madhavan Srinivasan <maddy@linux.ibm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com> Cc: linuxppc-dev@lists.ozlabs.org Link: https://lore.kernel.org/r/20230823075103.190565-1-kjain@linux.ibm.comSigned-off-by:
-
- 22 Aug, 2023 1 commit
-
-
Arnaldo Carvalho de Melo authored
I noticed some error with: # perf list ex_ret_brn lzma: fopen failed on /usr/lib/modules/5.15.14-100.fc34.x86_64/kernel/net/bluetooth/bnep/bnep.ko.xz: 'No such file or directory' lzma: fopen failed on /usr/lib/modules/5.16.16-200.fc35.x86_64/kernel/drivers/gpu/drm/drm_kms_helper.ko.xz: 'No such file or directory' lzma: fopen failed on /usr/lib/modules/5.18.16-200.fc36.x86_64/kernel/arch/x86/crypto/crct10dif-pclmul.ko.xz: 'No such file or directory' lzma: fopen failed on /usr/lib/modules/5.16.16-200.fc35.x86_64/kernel/drivers/i2c/busses/i2c-piix4.ko.xz: 'No such file or directory' <BIG SNIP> Then using 'perf probe' + 'perf trace' to debug 'perf list', it seems its some inconsistency in the ~/.debug/ cache where broken build id symlinks that ends up making it try to uncompress some kernel modules using the lzma routines: 395.309 perf/3594447 probe_perf:lzma_decompress_to_file(__probe_ip: 6118448, input_string: "/usr/lib/modules/5.18.17-200.fc36.x86_64/kernel/drivers/nvme/host/nvme.ko.xz") lzma_decompress_to_file (/var/home/acme/bin/perf) filename__decompress (/var/home/acme/bin/perf) filename__read_build_id (/var/home/acme/bin/perf) filename__sprintf_build_id (inlined) build_id_cache__valid_id (inlined) build_id_cache__list_all (/var/home/acme/bin/perf) print_sdt_events (/var/home/acme/bin/perf) cmd_list (/var/home/acme/bin/perf) run_builtin (/var/home/acme/bin/perf) handle_internal_command (inlined) run_argv (inlined) main (/var/home/acme/bin/perf) __libc_start_call_main (/usr/lib64/libc.so.6) __libc_start_main@@GLIBC_2.34 (/usr/lib64/libc.so.6) _start (/var/home/acme/bin/perf) But callers of filename__decompress() already check its return and use pr_debug(), so be consistent and make functions it calls also use pr_debug(). Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Ian Rogers <irogers@google.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Link: https://lore.kernel.org/lkml/ZOUD0+GkuCVkYF7n@kernel.orgSigned-off-by:
-
- 21 Aug, 2023 8 commits
-
-
Kaige Ye authored
It is undefined behavior to pass NULL as snprintf()'s fmt argument. Here is an example to trigger the problem: $ perf stat --metric-only -x, -e instructions -- sleep 1 insn per cycle, Segmentation fault (core dumped) With this patch: $ perf stat --metric-only -x, -e instructions -- sleep 1 insn per cycle, , Reviewed-by:
Kaige Ye <ye@kaige.org> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: https://lore.kernel.org/r/01CA7674B690CA24+20230804020907.144562-2-ye@kaige.orgSigned-off-by:
-
Arnaldo Carvalho de Melo authored
perf bpf augmented_raw_syscalls: Add an assert to make sure sizeof(augmented_arg->value) is a power of two. Similar to what was done in the previous cset for sizeof(saddr), we need to make sure sizeof(augmented_arg->value) is a power of two to do bounds checking using &=: augmented_len &= sizeof(augmented_arg->value) - 1; Suggested-by:
-
Arnaldo Carvalho de Melo authored
We're using the BPF verifier suggestion: 22: (85) call bpf_probe_read#4 R2 min value is negative, either use unsigned or 'var &= const' That works only when const is a (power of two - 1) so add an assert to make sure that that is the case. Suggested-by: -
Ilkka Koskinen authored
Some of the events included in the ampereone/core-imp-def are not supported on AmpereOne, remove them. Signed-off-by:
Ilkka Koskinen <ilkka@os.amperecomputing.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Dave Kleikamp <dave.kleikamp@oracle.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: James Clark <james.clark@arm.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: John Garry <john.g.garry@oracle.com> Cc: Leo Yan <leo.yan@linaro.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mike Leach <mike.leach@linaro.org> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will@kernel.org> Cc: linux-arm-kernel@lists.infradead.org Link: https://lore.kernel.org/r/20230803211331.140553-5-ilkka@os.amperecomputing.comSigned-off-by:
-
Ilkka Koskinen authored
This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events. Signed-off-by:
-
Ilkka Koskinen authored
Per errata AC03_CPU_29, STALL_SLOT_FRONTEND, STALL_FRONTEND, and STALL events are not counting as expected. The follow up metrics patch will include correct way to calculate the impacted events. Reviewed-by:
John Garry <john.g.garry@oracle.com> Signed-off-by:
-
Ilkka Koskinen authored
amperene/cache.json file tried to include L1D_CACHE_LMISS while it doesn't exist in common-and-microarch.json. While this bug doesn't seem to cause issue in newer kernels with jevents.py script, it prevents building older perf tools with the backported patch. Fixes: a9650b7f ("perf vendor events arm64: Add AmpereOne core PMU events") Reported-by:
Dave Kleikamp <dave.kleikamp@oracle.com> Reviewed-by:
-
John Garry authored
Recently Ilkka reported that the JSONs for the AmpereOne arm64-based platform included a dud event which referenced a non-existent arch std event [0]. Previously in the times of jevents.c, we would raise an exception for this. This is still invalid, even though the current code just ignores such an event. Re-introduce code to raise an exception for when no definition exists to help catch as many invalid JSONs as possible. [0] https://lore.kernel.org/linux-perf-users/9e851e2a-26c7-ba78-cb20-be4337b2916a@oracle.com/Reviewed-by:
-
- 18 Aug, 2023 2 commits
-
-
Arnaldo Carvalho de Melo authored
perf trace: Use heuristic when deciding if a syscall tracepoint "const char *" field is really a string 'perf trace' tries to find BPF progs associated with a syscall that have a signature that is similar to syscalls without one to try and reuse, so, for instance, the 'open' signature can be reused with many other syscalls that have as its first arg a string. It uses the tracefs events format file for finding a signature that can be reused, but then comes the "write" syscall with its second argument as a "const char *": # cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_write/format name: sys_enter_write ID: 746 format: field:unsigned short common_type; offset:0; size:2; signed:0; field:unsigned char common_flags; offset:2; size:1; signed:0; field:unsigned char common_preempt_count; offset:3; size:1; signed:0; field:int common_pid; offset:4; size:4; signed:1; field:int __syscall_nr; offset:8; size:4; signed:1; field:unsigned int fd; offset:16; size:8; signed:0; field:const char * buf; offset:24; size:8; signed:0; field:size_t count; offset:32; size:8; signed:0; print fmt: "fd: 0x%08lx, buf: 0x%08lx, count: 0x%08lx", ((unsigned long)(REC->fd)), ((unsigned long)(REC->buf)), ((unsigned long)(REC->count)) # Which isn't a string (the man page for glibc has buf as "void *"), so we have to use the name of the argument as an heuristic, to consider a string just args that are "const char *" and that have in its name the "path", "file", etc substrings. With that now it reuses: [root@quaco ~]# perf trace -v --max-events=1 |& grep Reus Reusing "open" BPF sys_enter augmenter for "stat" Reusing "open" BPF sys_enter augmenter for "lstat" Reusing "open" BPF sys_enter augmenter for "access" Reusing "connect" BPF sys_enter augmenter for "accept" Reusing "sendto" BPF sys_enter augmenter for "recvfrom" Reusing "connect" BPF sys_enter augmenter for "bind" Reusing "connect" BPF sys_enter augmenter for "getsockname" Reusing "connect" BPF sys_enter augmenter for "getpeername" Reusing "open" BPF sys_enter augmenter for "execve" Reusing "open" BPF sys_enter augmenter for "truncate" Reusing "open" BPF sys_enter augmenter for "chdir" Reusing "open" BPF sys_enter augmenter for "mkdir" Reusing "open" BPF sys_enter augmenter for "rmdir" Reusing "open" BPF sys_enter augmenter for "creat" Reusing "open" BPF sys_enter augmenter for "link" Reusing "open" BPF sys_enter augmenter for "unlink" Reusing "open" BPF sys_enter augmenter for "symlink" Reusing "open" BPF sys_enter augmenter for "readlink" Reusing "open" BPF sys_enter augmenter for "chmod" Reusing "open" BPF sys_enter augmenter for "chown" Reusing "open" BPF sys_enter augmenter for "lchown" Reusing "open" BPF sys_enter augmenter for "mknod" Reusing "open" BPF sys_enter augmenter for "statfs" Reusing "open" BPF sys_enter augmenter for "pivot_root" Reusing "open" BPF sys_enter augmenter for "chroot" Reusing "open" BPF sys_enter augmenter for "acct" Reusing "open" BPF sys_enter augmenter for "swapon" Reusing "open" BPF sys_enter augmenter for "swapoff" Reusing "open" BPF sys_enter augmenter for "delete_module" Reusing "open" BPF sys_enter augmenter for "setxattr" Reusing "open" BPF sys_enter augmenter for "lsetxattr" Reusing "openat" BPF sys_enter augmenter for "fsetxattr" Reusing "open" BPF sys_enter augmenter for "getxattr" Reusing "open" BPF sys_enter augmenter for "lgetxattr" Reusing "openat" BPF sys_enter augmenter for "fgetxattr" Reusing "open" BPF sys_enter augmenter for "listxattr" Reusing "open" BPF sys_enter augmenter for "llistxattr" Reusing "open" BPF sys_enter augmenter for "removexattr" Reusing "open" BPF sys_enter augmenter for "lremovexattr" Reusing "fsetxattr" BPF sys_enter augmenter for "fremovexattr" Reusing "open" BPF sys_enter augmenter for "mq_open" Reusing "open" BPF sys_enter augmenter for "mq_unlink" Reusing "fsetxattr" BPF sys_enter augmenter for "add_key" Reusing "fremovexattr" BPF sys_enter augmenter for "request_key" Reusing "fremovexattr" BPF sys_enter augmenter for "inotify_add_watch" Reusing "fremovexattr" BPF sys_enter augmenter for "mkdirat" Reusing "fremovexattr" BPF sys_enter augmenter for "mknodat" Reusing "fremovexattr" BPF sys_enter augmenter for "fchownat" Reusing "fremovexattr" BPF sys_enter augmenter for "futimesat" Reusing "fremovexattr" BPF sys_enter augmenter for "newfstatat" Reusing "fremovexattr" BPF sys_enter augmenter for "unlinkat" Reusing "fremovexattr" BPF sys_enter augmenter for "linkat" Reusing "open" BPF sys_enter augmenter for "symlinkat" Reusing "fremovexattr" BPF sys_enter augmenter for "readlinkat" Reusing "fremovexattr" BPF sys_enter augmenter for "fchmodat" Reusing "fremovexattr" BPF sys_enter augmenter for "faccessat" Reusing "fremovexattr" BPF sys_enter augmenter for "utimensat" Reusing "connect" BPF sys_enter augmenter for "accept4" Reusing "fremovexattr" BPF sys_enter augmenter for "name_to_handle_at" Reusing "fremovexattr" BPF sys_enter augmenter for "renameat2" Reusing "open" BPF sys_enter augmenter for "memfd_create" Reusing "fremovexattr" BPF sys_enter augmenter for "execveat" Reusing "fremovexattr" BPF sys_enter augmenter for "statx" [root@quaco ~]# Reviewed-by:
-
Arnaldo Carvalho de Melo authored
It is possible to use 'perf trace' with tracepoints and in that case we can't initialize/use the augmented_raw_syscalls BPF skel. For instance, this usecase: # perf trace -e sched:*exec --max-events=5 ? ( ): NetworkManager/1183 ... [continued]: poll()) = 1 0.043 ( 0.007 ms): NetworkManager/1183 epoll_wait(epfd: 17<anon_inode:[eventpoll]>, events: 0x55555f90e920, maxevents: 6) = 0 0.060 ( 0.007 ms): NetworkManager/1183 write(fd: 3<anon_inode:[eventfd]>, buf: 0x7ffc5a27cd30, count: 8) = 8 0.073 ( 0.005 ms): NetworkManager/1183 epoll_wait(epfd: 24<anon_inode:[eventpoll]>, events: 0x7ffc5a27cd20, maxevents: 2) = 1 0.082 ( 0.010 ms): NetworkManager/1183 recvmmsg(fd: 26<socket:[30298]>, mmsg: 0x7ffc5a27caa0, vlen: 8) = 1 # Where we want to trace just some sched tracepoints ending in 'exec' ends up tracing all syscalls. Fix it by checking existing trace->trace_syscalls boolean to see if we need the augmenter. A followup patch will move those sections of code used only with the augmenter to separate functions, to get it cleaner and remove the goto, done just for reviewing purposes. With this patch in place the previous behaviour is restored: no syscalls when we have other events and no syscall names: [root@quaco ~]# perf probe do_filp_open "filename=pathname->name:string" Added new event: probe:do_filp_open (on do_filp_open with filename=pathname->name:string) You can now use it in all perf tools, such as: perf record -e probe:do_filp_open -aR sleep 1 [root@quaco ~]# perf trace --max-events=10 -e probe:do_filp_open sleep 1 0.000 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/etc/ld.so.cache") 0.056 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/lib64/libc.so.6") 0.481 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/lib/locale/locale-archive") 0.501 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/share/locale/locale.alias") 0.572 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/lib/locale/en_US.UTF-8/LC_IDENTIFICATION") 0.581 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/lib/locale/en_US.utf8/LC_IDENTIFICATION") 0.616 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/lib64/gconv/gconv-modules.cache") 0.656 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/lib/locale/en_US.UTF-8/LC_MEASUREMENT") 0.664 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/lib/locale/en_US.utf8/LC_MEASUREMENT") 0.696 sleep/455122 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/lib/locale/en_US.UTF-8/LC_TELEPHONE") [root@quaco ~]# As well as mixing syscalls with tracepoints, getting the syscall tracepoints used augmented using the BPF skel: [root@quaco ~]# perf trace --max-events=10 -e open*,probe:do_filp_open sleep 1 0.000 ( ): sleep/455124 openat(dfd: CWD, filename: "/etc/ld.so.cache", flags: RDONLY|CLOEXEC) ... 0.005 ( ): sleep/455124 probe:do_filp_open(__probe_ip: -1186560412, filename: "/etc/ld.so.cache") 0.000 ( 0.011 ms): sleep/455124 ... [continued]: openat()) = 3 0.031 ( ): sleep/455124 openat(dfd: CWD, filename: "/lib64/libc.so.6", flags: RDONLY|CLOEXEC) ... 0.033 ( ): sleep/455124 probe:do_filp_open(__probe_ip: -1186560412, filename: "/lib64/libc.so.6") 0.031 ( 0.006 ms): sleep/455124 ... [continued]: openat()) = 3 0.258 ( ): sleep/455124 openat(dfd: CWD, filename: "/usr/lib/locale/locale-archive", flags: RDONLY|CLOEXEC) ... 0.261 ( ): sleep/455124 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/lib/locale/locale-archive") 0.258 ( 0.006 ms): sleep/455124 ... [continued]: openat()) = -1 ENOENT (No such file or directory) 0.272 ( ): sleep/455124 openat(dfd: CWD, filename: "/usr/share/locale/locale.alias", flags: RDONLY|CLOEXEC) ... 0.273 ( ): sleep/455124 probe:do_filp_open(__probe_ip: -1186560412, filename: "/usr/share/locale/locale.alias") A final note: the probe:do_filp_open uses a kprobe (probably optimized as its in the start of a function) that uses the kprobe_tracer mechanism in the kernel to collect the pathname->name string and stash it into the tracepoint created by 'perf probe' for that: [root@quaco ~]# cat /sys/kernel/debug/tracing/kprobe_events p:probe/do_filp_open _text+4621920 filename=+0(+0(%si)):string [root@quaco ~]# While the syscalls:sys_enter_openat tracepoint gets its string from a BPF program attached to raw_syscalls:sys_enter that tail calls into another BPF program that knows the types for the openat syscall args and thus can bpf_probe_read it right after the normal sys_enter/sys_enter_openat tracepoint payload that comes prefixed with whatever perf_event_open asked for (CPU, timestamp, etc): [root@quaco ~]# bpftool prog | grep -E "sys_enter |sys_enter_opena" -A3 3176: tracepoint name sys_enter tag 0bc3fc9d11754ba1 gpl loaded_at 2023-08-17T12:32:20-0300 uid 0 xlated 272B jited 257B memlock 4096B map_ids 2462,2466,2463 btf_id 2976 -- 3180: tracepoint name sys_enter_opena tag 19dd077f00ec2f58 gpl loaded_at 2023-08-17T12:32:20-0300 uid 0 xlated 328B jited 206B memlock 4096B map_ids 2466,2465 btf_id 2976 [root@quaco ~]# Fixes: 5e6da6be ("perf trace: Migrate BPF augmentation to use a skeleton") Reviewed-by:
-
- 17 Aug, 2023 3 commits
-
-
Arnaldo Carvalho de Melo authored
While debugging a segfault on 'perf lock contention' without an available perf.data file I noticed that it was basically calling: perf_session__delete(ERR_PTR(-1)) Resulting in: (gdb) run lock contention Starting program: /root/bin/perf lock contention [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib64/libthread_db.so.1". failed to open perf.data: No such file or directory (try 'perf record' first) Initializing perf session failed Program received signal SIGSEGV, Segmentation fault. 0x00000000005e7515 in auxtrace__free (session=0xffffffffffffffff) at util/auxtrace.c:2858 2858 if (!session->auxtrace) (gdb) p session $1 = (struct perf_session *) 0xffffffffffffffff (gdb) bt #0 0x00000000005e7515 in auxtrace__free (session=0xffffffffffffffff) at util/auxtrace.c:2858 #1 0x000000000057bb4d in perf_session__delete (session=0xffffffffffffffff) at util/session.c:300 #2 0x000000000047c421 in __cmd_contention (argc=0, argv=0x7fffffffe200) at builtin-lock.c:2161 #3 0x000000000047dc95 in cmd_lock (argc=0, argv=0x7fffffffe200) at builtin-lock.c:2604 #4 0x0000000000501466 in run_builtin (p=0xe597a8 <commands+552>, argc=2, argv=0x7fffffffe200) at perf.c:322 #5 0x00000000005016d5 in handle_internal_command (argc=2, argv=0x7fffffffe200) at perf.c:375 #6 0x0000000000501824 in run_argv (argcp=0x7fffffffe02c, argv=0x7fffffffe020) at perf.c:419 #7 0x0000000000501b11 in main (argc=2, argv=0x7fffffffe200) at perf.c:535 (gdb) So just set it to NULL after using PTR_ERR(session) to decode the error as perf_session__delete(NULL) is supported. Fixes: eef4fee5 ("perf lock: Dynamically allocate lockhash_table") Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: K Prateek Nayak <kprateek.nayak@amd.com> Cc: Kan Liang <kan.liang@linux.intel.com> Cc: Leo Yan <leo.yan@linaro.org> Cc: Mamatha Inamdar <mamatha4@linux.vnet.ibm.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ravi Bangoria <ravi.bangoria@amd.com> Cc: Ross Zwisler <zwisler@chromium.org> Cc: Sean Christopherson <seanjc@google.com> Cc: Steven Rostedt (VMware) <rostedt@goodmis.org> Cc: Tiezhu Yang <yangtiezhu@loongson.cn> Cc: Yang Jihong <yangjihong1@huawei.com> Link: https://lore.kernel.org/lkml/ZN4R1AYfsD2J8lRs@kernel.orgSigned-off-by:
-
Arnaldo Carvalho de Melo authored
While debugging a segfault on 'perf lock contention' without an available perf.data file I noticed that it was basically calling: perf_session__delete(ERR_PTR(-1)) Resulting in: (gdb) run lock contention Starting program: /root/bin/perf lock contention [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib64/libthread_db.so.1". failed to open perf.data: No such file or directory (try 'perf record' first) Initializing perf session failed Program received signal SIGSEGV, Segmentation fault. 0x00000000005e7515 in auxtrace__free (session=0xffffffffffffffff) at util/auxtrace.c:2858 2858 if (!session->auxtrace) (gdb) p session $1 = (struct perf_session *) 0xffffffffffffffff (gdb) bt #0 0x00000000005e7515 in auxtrace__free (session=0xffffffffffffffff) at util/auxtrace.c:2858 #1 0x000000000057bb4d in perf_session__delete (session=0xffffffffffffffff) at util/session.c:300 #2 0x000000000047c421 in __cmd_contention (argc=0, argv=0x7fffffffe200) at builtin-lock.c:2161 #3 0x000000000047dc95 in cmd_lock (argc=0, argv=0x7fffffffe200) at builtin-lock.c:2604 #4 0x0000000000501466 in run_builtin (p=0xe597a8 <commands+552>, argc=2, argv=0x7fffffffe200) at perf.c:322 #5 0x00000000005016d5 in handle_internal_command (argc=2, argv=0x7fffffffe200) at perf.c:375 #6 0x0000000000501824 in run_argv (argcp=0x7fffffffe02c, argv=0x7fffffffe020) at perf.c:419 #7 0x0000000000501b11 in main (argc=2, argv=0x7fffffffe200) at perf.c:535 (gdb) So just set it to NULL after using PTR_ERR(session) to decode the error as perf_session__delete(NULL) is supported. The same problem was found in 'perf top' after an audit of all perf_session__new() failure handling. Fixes: 6ef81c55 ("perf session: Return error code for perf_session__new() function on failure") Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Alexey Budankov <alexey.budankov@linux.intel.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Jeremie Galarneau <jeremie.galarneau@efficios.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Kate Stewart <kstewart@linuxfoundation.org> Cc: Mamatha Inamdar <mamatha4@linux.vnet.ibm.com> Cc: Mukesh Ojha <mojha@codeaurora.org> Cc: Nageswara R Sastry <rnsastry@linux.vnet.ibm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com> Cc: Shawn Landden <shawn@git.icu> Cc: Song Liu <songliubraving@fb.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tzvetomir Stoyanov <tstoyanov@vmware.com> Link: https://lore.kernel.org/lkml/ZN4Q2rxxsL08A8rd@kernel.orgSigned-off-by:
-
James Clark authored
Apart from some slight naming and grouping differences, the new metrics are functionally the same as the existing ones. Any missing metrics were manually appended to the end of the auto generated file. For the events, the new data includes descriptions that may have product specific details and new groupings that will be consistent with other products. After generating the metrics from the telemetry repo [1], the following manual steps were performed: * Change the topdown expressions to compare on CPUID and use #slots so that the same data can be shared between N2 and V2. Apart from these modifications, the expressions now match more closely with the Arm telemetry data which will hopefully make future updates easier. * Append some metrics from the old N2/V2 data that aren't present in the telemetry data. These will possibly be added to the telemetry-solution repo at a later time: l3d_cache_mpki, l3d_cache_miss_rate, branch_pki, ipc_rate, spec_ipc, retired_rate, wasted_rate, branch_immed_spec_rate, branch_return_spec_rate, branch_indirect_spec_rate [1]: https://gitlab.arm.com/telemetry-solution/telemetry-solution/-/blob/main/data/pmu/cpu/neoverse/neoverse-n2.jsonSigned-off-by:James Clark <james.clark@arm.com> Reviewed-by:
-
