- 07 Apr, 2023 12 commits
-
-
Ian Rogers authored
Improve documentation around perf_pmu_alias pmu_name and on functions. Reduce the scope of pmu_uncore_alias_match to just file. Rename perf_pmu__valid_suffix to the more revealing perf_pmu__match_ignoring_suffix. Add a short-cut to perf_pmu__match_ignoring_suffix for PMU names that don't also have a socket value, and can therefore avoid a memory allocation. Signed-off-by:
Ian Rogers <irogers@google.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: James Clark <james.clark@arm.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> Cc: Ravi Bangoria <ravi.bangoria@amd.com> Cc: Rob Herring <robh@kernel.org> Cc: Sean Christopherson <seanjc@google.com> Cc: Suzuki Poulouse <suzuki.poulose@arm.com> Link: https://lore.kernel.org/r/20230406235256.2768773-1-irogers@google.comSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Ian Rogers authored
struct pmu_event has const char*s, only unit needs to be non-const for the sake of passing as an out argument to strtod(). Reduce the const casts from 4 down to 1. Signed-off-by:
-
Namhyung Kim authored
It doesn't delete data in the task_data and lock_stat maps. The data is kept there until it's consumed by userspace at the end. But it calls bpf_map_update_elem() again and again, and the data will be discarded if the map is full. This is not good. Worse, in the bpf_map_update_elem(), it keeps trying to get a new node even if the map was full. I guess it makes sense if it deletes some node like in the tstamp map (that's why I didn't make the change there). In a pre-allocated hash map, that means it'd iterate all CPU to check the freelist. And it has a bad performance impact on large machines. I've checked it on my 64 CPU machine with this. $ perf bench sched messaging -g 1000 # Running 'sched/messaging' benchmark: # 20 sender and receiver processes per group # 1000 groups == 40000 processes run Total time: 2.825 [sec] And I used the task mode, so that it can guarantee the map is full. The default map entry size is 16K and this workload has 40K tasks. Before: $ sudo ./perf lock con -abt -E3 -- perf bench sched messaging -g 1000 # Running 'sched/messaging' benchmark: # 20 sender and receiver processes per group # 1000 groups == 40000 processes run Total time: 11.299 [sec] contended total wait max wait avg wait pid comm 19284 3.51 s 3.70 ms 181.91 us 1305863 sched-messaging 243 84.09 ms 466.67 us 346.04 us 1336608 sched-messaging 177 66.35 ms 12.08 ms 374.88 us 1220416 node For some reason, it didn't report the data failures. But you can see the total time in the workload is increased a lot (2.8 -> 11.3). If it fails early when the map is full, it goes back to normal. After: $ sudo ./perf lock con -abt -E3 -- perf bench sched messaging -g 1000 # Running 'sched/messaging' benchmark: # 20 sender and receiver processes per group # 1000 groups == 40000 processes run Total time: 3.044 [sec] contended total wait max wait avg wait pid comm 18743 591.92 ms 442.96 us 31.58 us 1431454 sched-messaging 51 210.64 ms 207.45 ms 4.13 ms 1468724 sched-messaging 81 68.61 ms 65.79 ms 847.07 us 1463183 sched-messaging === output for debug === bad: 1164137, total: 2253341 bad rate: 51.66 % histogram of failure reasons task: 0 stack: 0 time: 0 data: 1164137 Signed-off-by:Namhyung Kim <namhyung@kernel.org> Acked-by:
-
Namhyung Kim authored
It needs callstacks for two reasons: * for stack aggregation mode, the map key is the stack id and it can also show the full stack traces when -v is used * for other aggregation modes, the stack filter can be used to limit lock contentions from known call paths The -v option is meaningful (in terms of stack trace) only for stack aggregation mode, so it should not set the save_callstack for other mode like with -t or -l options. I've noticed this with the following command line: $ sudo ./perf lock con -ablv -E 3 -M 16 -- ./perf bench sched messaging ... contended total wait max wait avg wait address symbol 88 4.59 ms 108.07 us 52.13 us ffff935757f46ec0 (spinlock) 33 905.22 us 73.67 us 27.43 us ffff935757f41700 (spinlock) 28 703.69 us 79.28 us 25.13 us ffff938a3d9b0c80 rq_lock (spinlock) === output for debug === bad: 12272, total: 12421 bad rate: 98.80 % histogram of failure reasons task: 8285 stack: 3987 <---------- here time: 0 data: 0 It should not have any failure on stacks since it doesn't use it. No functional change intended. Signed-off-by: -
Namhyung Kim authored
When -E option is used, it only prints the given number of entries but the event stat at the end should have the numbers for entire entries. Likewise, -S option will hide entries that don't have the named function in the callstack. Also update event stat for them. Signed-off-by:
-
Namhyung Kim authored
It's possible to fail to update the data when the lock_stat map is full. We should check that case and show the number at the end. $ sudo ./perf lock con -ablv -E3 -- ./perf bench sched messaging ... contended total wait max wait avg wait address symbol 6157 208.48 ms 69.29 us 33.86 us ffff934c001c1f00 (spinlock) 4030 72.04 ms 61.84 us 17.88 us ffff934c000415c0 (spinlock) 3201 50.30 ms 47.73 us 15.71 us ffff934c2eead850 (spinlock) === output for debug === bad: 0, total: 13388 bad rate: 0.00 % histogram of failure reasons task: 0 stack: 0 time: 0 data: 0 <----- added Signed-off-by: -
Namhyung Kim authored
The BPF hash map will align the map size to a power of 2. So 10k would be 16k anyway. Let's have the actual size to avoid confusions. Signed-off-by:
-
Namhyung Kim authored
Users often want to change the map size, let's add a short option (-M) for that. Signed-off-by:
-
Namhyung Kim authored
The get_type_flag() should check both str and name fields in the lock_type_table so that it can find the appropriate flag without retrying with ':R' or ':W' suffix from the caller. Also fix a typo in the rt-mutex. Signed-off-by:
-
Liam Howlett authored
Rename the fallthrough attribute to better align with the kernel version. Copy the definition from include/linux/compiler_attributes.h including the #else clause. Adding the #else clause allows the tools compiler.h header to drop the check for a definition entirely and keeps both definitions together. Change any __fallthrough statements to fallthrough anywhere it was used within perf. This allows other tools to use the same key word as the kernel. Committer notes: Did some missing conversions to: builtin-list.c Also included gtk.h before the 'fallthrough' definition in: tools/perf/ui/gtk/hists.c tools/perf/ui/gtk/helpline.c tools/perf/ui/gtk/browser.c As it is the arg name for a macro in glib.h: /var/home/acme/git/perf-tools-next/tools/include/linux/compiler-gcc.h:16:55: error: missing binary operator before token "(" 16 | # define fallthrough __attribute__((__fallthrough__)) | ^ /usr/include/glib-2.0/glib/gmacros.h:637:28: note: in expansion of macro ‘fallthrough’ 637 | #if g_macro__has_attribute(fallthrough) Reviewed-by:Miguel Ojeda <ojeda@kernel.org> Signed-off-by:
Liam Howlett <Liam.Howlett@oracle.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Miguel Ojeda <ojeda@kernel.org> Cc: Nathan Chancellor <nathan@kernel.org> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk> Cc: Tom Rix <trix@redhat.com> Cc: linux-sparse@vger.kernel.org <linux-sparse@vger.kernel.org> Cc: llvm@lists.linux.dev <llvm@lists.linux.dev> Link: https://lore.kernel.org/r/20221125154947.2163498-1-Liam.Howlett@oracle.comSigned-off-by:
-
Ian Rogers authored
Ensure fd is closed on error paths. Signed-off-by:
Jiri Olsa <jolsa@kernel.org> Acked-by:
-
Ian Rogers authored
By default bison uses global state for compatibility with yacc. Make the parser reentrant so that it may be used in asynchronous and multithreaded situations. Signed-off-by:
-
- 04 Apr, 2023 28 commits
-
-
Ian Rogers authored
Later changes will add reference count checking for struct map, start and end are frequently accessed variables. Add an accessor so that the reference count check is only necessary in one place. Signed-off-by:
-
Ian Rogers authored
Later changes will add reference count checking for struct map, with dso being the most frequently accessed variable. Add an accessor so that the reference count check is only necessary in one place. Additional changes: - add a dso variable to avoid repeated map__dso calls. - in builtin-mem.c dump_raw_samples, code only partially tested for dso == NULL. Make the possibility of NULL consistent. - in thread.c thread__memcpy fix use of spaces and use tabs. Committer notes: Did missing conversions on these files: tools/perf/arch/powerpc/util/skip-callchain-idx.c tools/perf/arch/powerpc/util/sym-handling.c tools/perf/ui/browsers/hists.c tools/perf/ui/gtk/annotate.c tools/perf/util/cs-etm.c tools/perf/util/thread.c tools/perf/util/unwind-libunwind-local.c tools/perf/util/unwind-libunwind.c Signed-off-by:
-
Ian Rogers authored
Introduce functions to access struct maps. These functions reduce the number of places reference counting is necessary. While tidying APIs do some small const-ification, in particlar to unwind_libunwind_ops. Committer notes: Fixed up tools/perf/util/unwind-libunwind.c: - return ops->get_entries(cb, arg, thread, data, max_stack); + return ops->get_entries(cb, arg, thread, data, max_stack, best_effort); Signed-off-by:
-
Ian Rogers authored
struct map is reference counted, having it also be a node in an red-black tree complicates the reference counting. Switch to having a map_rb_node which is a red-block tree node but points at the reference counted struct map. This reference is responsible for a single reference count. Committer notes: Fixed up tools/perf/util/unwind-libunwind-local.c to use map_rb_node as well. Signed-off-by:
-
Ian Rogers authored
Using a perf map as a list node is only done in symbol. Move the list_node struct into symbol as a single pointer to the map. This makes reference count behavior more obvious and easy to check. Committer notes: Some changes to reduce the number of lines touched by keeping, for instance, the 'new_map' variable and setting it to new_node->map, so that we keep more of the project history in place and keep as much as possible the value of the 'git blame' tool. Also use map__zput() when putting a struct members, so that when we free the container struct we can get use-after-free errors as NULL pointer derefs sometimes. Signed-off-by:
-
Ian Rogers authored
As reported by leak sanitizer. Signed-off-by:
-
Ian Rogers authored
The cxa demangle support isn't dependent on libelf and so we no longer need to disable demangling if libelf isn't present. Signed-off-by:
-
Ian Rogers authored
Ignore SIGPIPEs when addr2line is configured. Signed-off-by:
-
Ian Rogers authored
The sentinel value differs for llvm-addr2line. Configure this once and then detect when reading records. Signed-off-by:
-
Ian Rogers authored
Don't wrap stdin and stdout of subprocess with streams, use the api/io library for buffering. Signed-off-by:
-
Ian Rogers authored
Reads a line to allocated memory up to a newline following the getline API. Committer notes: It also adds this new function to the 'api io' 'perf test' entry: $ perf test "api io" 64: Test api io : Ok $ Signed-off-by:
-
Namhyung Kim authored
Intel-PT calls perf_pmu__scan_file() a lot, let's use relative address when it accesses multiple files at one place. Signed-off-by:
Adrian Hunter <adrian.hunter@intel.com> Acked-by:
-
Namhyung Kim authored
These two helpers will also use openat() to reduce the overhead with relative pathnames. Convert other functions in pmu_lookup() to use the new helpers. Committer testing: Before: ⬢[acme@toolbox perf-tools-next]$ perf bench internals pmu-scan # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 100 times Average PMU scanning took: 2729.040 usec (+- 7.117 usec) ⬢[acme@toolbox perf-tools-next]$ After: ⬢[acme@toolbox perf-tools-next]$ perf bench internals pmu-scan # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 100 times Average PMU scanning took: 2419.870 usec (+- 9.057 usec) ⬢[acme@toolbox perf-tools-next]$ Signed-off-by: -
Namhyung Kim authored
Likewise, x86 needs to traverse the PMU list to build alias. Let's use the new helpers to use relative paths. Signed-off-by:
-
Namhyung Kim authored
Likewise, it needs to traverse the pmu/caps directory, let's use openat() with the dirfd instead of open() using the absolute path. Signed-off-by:
-
Namhyung Kim authored
The PMU information is in the kernel sysfs so it needs to scan the directory to get the whole information like event aliases, formats and so on. During the traversal, it opens a lot of files and directories like below: dir = opendir("/sys/bus/event_source/devices"); while (dentry = readdir(dir)) { char buf[PATH_MAX]; snprintf(buf, sizeof(buf), "%s/%s", "/sys/bus/event_source/devices", dentry->d_name); fd = open(buf, O_RDONLY); ... } But this is not good since it needs to copy the string to build the absolute pathname, and it makes redundant pathname walk (from the /sys) unnecessarily. We can use openat(2) to open the file in the given directory. While it's not a problem ususally, it can be a problem when the kernel has contentions on the sysfs. Add a couple of new helper to return the file descriptor of PMU directory so that it can use it with relative paths. * perf_pmu__event_source_devices_fd() - returns a fd for the PMU root ("/sys/bus/event_source/devices") * perf_pmu__pathname_fd() - returns a fd for "<pmu>/<file>" under the PMU root Now the above code can be converted something like below: dirfd = perf_pmu__event_source_devices_fd(); dir = fdopendir(dirfd); while (dentry = readdir(dir)) { fd = openat(dirfd, dentry->d_name, O_RDONLY); ... } Signed-off-by: -
Namhyung Kim authored
The pmu-scan benchmark will repeatedly scan the sysfs to get the available PMU information. $ ./perf bench internals pmu-scan # Running 'internals/pmu-scan' benchmark: Computing performance of sysfs PMU event scan for 100 times Average PMU scanning took: 6850.990 usec (+- 48.445 usec) Signed-off-by: -
Namhyung Kim authored
It seems there's no function to delete the perf pmu struct. Add the perf_pmu__destroy() to do the job. While at it, add some more helper functions to delete pmu aliases and caps. Signed-off-by:
-
Namhyung Kim authored
In the parse_events_multi_pmu_add() it passes the 'config' variable twice to parse_events_term__num() - one for config and another for loc_term. I'm not sure about the second one as it's converted to YYLTYPE variable. Asan reports it like below: In function ‘parse_events_term__num’, inlined from ‘parse_events_multi_pmu_add’ at util/parse-events.c:1602:6: util/parse-events.c:2653:64: error: array subscript ‘YYLTYPE[0]’ is partly outside array bounds of ‘char[8]’ [-Werror=array-bounds] 2653 | .err_term = loc_term ? loc_term->first_column : 0, | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~ util/parse-events.c: In function ‘parse_events_multi_pmu_add’: util/parse-events.c:1587:15: note: object ‘config’ of size 8 1587 | char *config; | ^~~~~~ cc1: all warnings being treated as errors Signed-off-by: -
Namhyung Kim authored
Committer notes: Added missing #include <unistd.h> for the close() prototype to fix this on Alma Linux 8: 1 21.54 almalinux:8 : FAIL gcc version 8.5.0 20210514 (Red Hat 8.5.0-16) (GCC) util/print-events.c: In function 'print_tracepoint_events': util/print-events.c:103:4: error: implicit declaration of function 'close'; did you mean 'clone'? [-Werror=implicit-function-declaration] close(evt_fd); ^~~~~ clone Also use the newly added scandirat feature test to check if that function is available, providing a HAVE_SCANDIRAT_SUPPORT conditional warning to the user if it isn't available. Signed-off-by: -
Arnaldo Carvalho de Melo authored
We use it just when listing tracepoint events, and for root, so just emit a warning about it to get users to ask the library maintainers to implement it, as suggested in this systemd ticket: https://github.com/systemd/casync/issues/129 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/ZCwv4z5Dh%2FdHUMG6@kernel.org/Signed-off-by:
-
Adrian Hunter authored
After a standalone CBR (not associated with TSC), update the cycles reference timestamp and reset the cycle count, so that CYC timestamps are calculated relative to that point with the new frequency. Fixes: cc336186 ("perf tools: Add Intel PT support for decoding CYC packets") Signed-off-by:
-
Adrian Hunter authored
kallsyms is not completely in address order. In find_entire_kern_cb(), calculate the kernel end from the maximum address not the last symbol. Example: Before: $ sudo cat /proc/kallsyms | grep ' [twTw] ' | tail -1 ffffffffc00b8bd0 t bpf_prog_6deef7357e7b4530 [bpf] $ sudo cat /proc/kallsyms | grep ' [twTw] ' | sort | tail -1 ffffffffc15e0cc0 t iwl_mvm_exit [iwlmvm] $ perf.d093603a05aa record -v --kcore -e intel_pt// --filter 'filter *' -- uname |& grep filter Address filter: filter 0xffffffff93200000/0x2ceba000 After: $ perf.8fb0f7a01f8e record -v --kcore -e intel_pt// --filter 'filter *' -- uname |& grep filter Address filter: filter 0xffffffff93200000/0x2e3e2000 Fixes: 1b36c03e ("perf record: Add support for using symbols in address filters") Signed-off-by: -
Rob Herring authored
Arm SPEv1.3 adds new load/store operation subclasses for Memory Tagging Extension (MTE) and memory operations (MOPS). The memory operations are memcpy and memset. Add support for decoding these new subclasses in the raw decoding. Reviewed-by: Leo Yan <leo.yan@linaro.org Signed-off-by:
Rob Herring <robh@kernel.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Leo Yan <leo.yan@linaro.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/20230327162057.4057188-1-robh@kernel.orgSigned-off-by:
-
Mike Leach authored
When using dynamically assigned CoreSight trace IDs the drivers can output the ID / CPU association as a PERF_RECORD_AUX_OUTPUT_HW_ID packet. Update cs-etm decoder to handle this packet by setting the CPU/Trace ID mapping. Reviewed-by:
James Clark <james.clark@arm.com> Signed-off-by:
Mike Leach <mike.leach@linaro.org> Acked-by:
Suzuki Poulouse <suzuki.poulose@arm.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Darren Hart <darren@os.amperecomputing.com> Cc: Ganapatrao Kulkarni <gankulkarni@os.amperecomputing.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Leo Yan <leo.yan@linaro.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will@kernel.org> Cc: coresight@lists.linaro.org Cc: linux-arm-kernel@lists.infradead.org Link: https://lore.kernel.org/r/20230331055645.26918-2-mike.leach@linaro.orgSigned-off-by:
-
Mike Leach authored
Trace IDs are now dynamically allocated. Previously used the static association algorithm that is no longer used. The 'cpu * 2 + seed' was outdated and broken for systems with high core counts (>46). as it did not scale and was broken for larger core counts. Trace ID will now be sent in PERF_RECORD_AUX_OUTPUT_HW_ID record. Legacy ID algorithm renamed and retained for limited backward compatibility use. Reviewed-by:
-
Mike Leach authored
The information to associate Trace ID and CPU will be changing. Drivers will start outputting this as a hardware ID packet in the data file which if present will be used in preference to the AUXINFO values. To prepare for this we provide a helper functions to do the individual ID mapping, and one to extract the IDs from the completed metadata blocks. Reviewed-by:
-
Namhyung Kim authored
It can fail to collect lock stat from BPF for various reasons. For example, I've got a report that sometimes time calculation seems wrong in case of contended spinlocks. I suspect the time delta went negative for some reason. Count them separately and show in the output like below: $ sudo perf lock contention -abE5 sleep 10 contended total wait max wait avg wait type caller 13 785.61 us 79.36 us 60.43 us spinlock remove_wait_queue+0x14 10 469.02 us 87.51 us 46.90 us spinlock prepare_to_wait+0x27 9 289.09 us 69.08 us 32.12 us spinlock finish_wait+0x36 114 251.05 us 8.56 us 2.20 us spinlock try_to_wake_up+0x1f5 132 188.63 us 5.01 us 1.43 us spinlock __wake_up_common_lock+0x62 === output for debug === bad: 1, total: 279 bad rate: 0.36 % histogram of failure reasons task: 1 stack: 0 time: 0 Signed-off-by:
-
