- 05 Jun, 2023 7 commits
-
-
Ian Rogers authored
Currently the & and | operators are only used in metric thresholds like (from the tma_retiring metric): tma_retiring > 0.7 | tma_heavy_operations > 0.1 Thresholds are always computed when present, but a lack of events may mean the threshold can't be computed. This happens with the option --metric-no-threshold for say the metric tma_retiring on Tigerlake model CPUs. To fully compute the threshold tma_heavy_operations is needed and it needs the extra events of IDQ.MS_UOPS, UOPS_DECODED.DEC0, cpu/UOPS_DECODED.DEC0,cmask=1/ and IDQ.MITE_UOPS. So --metric-no-threshold is a useful option to reduce the number of events needed and potentially multiplexing of events. Rather than just fail threshold computations like this, we may know a result from just the left or right-hand side. So, for tma_retiring if its value is "> 0.7" we know it is over the threshold. This allows the metric to have the threshold coloring, when possible, without all the counters being programmed. Reviewed-by:
Kan Liang <kan.liang@linux.intel.com> Signed-off-by:
Ian Rogers <irogers@google.com> Tested-by:
Jiri Olsa <jolsa@kernel.org> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Ahmad Yasin <ahmad.yasin@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrii Nakryiko <andrii@kernel.org> Cc: Caleb Biggers <caleb.biggers@intel.com> Cc: Eduard Zingerman <eddyz87@gmail.com> Cc: Edward Baker <edward.baker@intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: James Clark <james.clark@arm.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Perry Taylor <perry.taylor@intel.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Samantha Alt <samantha.alt@intel.com> Cc: Stephane Eranian <eranian@google.com> Cc: Weilin Wang <weilin.wang@intel.com> Link: https://lore.kernel.org/r/20230519063719.1029596-1-irogers@google.comSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Tiezhu Yang authored
In order to print the numerical entries of the syscall table, there is no need to call the host compiler to build and then run a program, this can be done directly by the shell script. This is similar with commit 9854e7ad ("perf arm64: Simplify mksyscalltbl"). For now, the mksyscalltbl file of LoongArch is almost same with arm64. Reviewed-by:
Huacai Chen <chenhuacai@loongson.cn> Reviewed-by:
Leo Yan <leo.yan@linaro.org> Signed-off-by:
Tiezhu Yang <yangtiezhu@loongson.cn> Acked-by:
-
Tiezhu Yang authored
Like x86, powerpc, mips and s390, use max_nr which is a digital number to define SYSCALLTBL_ARM64_MAX_ID. Reviewed-by:
-
Tiezhu Yang authored
After commit 9854e7ad ("perf arm64: Simplify mksyscalltbl"), in the generated syscall table file syscalls.c, there exist some __NR3264_ prefixed syscall numbers such as [__NR3264_ftruncate], it looks like not so good, just do some small filter operations to handle __NR3264_ prefixed syscall number as a digital number. Without this patch: [__NR3264_ftruncate] = "ftruncate", With this patch: [46] = "ftruncate", Suggested-by:
Alexander Kapshuk <alexander.kapshuk@gmail.com> Reviewed-by:
-
Tiezhu Yang authored
After commit 9854e7ad ("perf arm64: Simplify mksyscalltbl") it has been removed the temporary C program and used shell to generate syscall table, so let us rename create_table_from_c() to create_sc_table() to avoid confusion. Suggested-by:
-
Tiezhu Yang authored
syscalltbl_*[] should never be changing, let us declare it as const. Suggested-by:
-
Arnaldo Carvalho de Melo authored
Without this we were not getting the thousands separator for big numbers. Noticed while developing 'perf bench uprobe', but the use of %' predates that, for instance 'perf bench syscall' uses it. Before: # perf bench uprobe all # Running uprobe/baseline benchmark... # Executed 1000 usleep(1000) calls Total time: 1054082243ns 1054082.243000 nsecs/op # After: # perf bench uprobe all # Running uprobe/baseline benchmark... # Executed 1,000 usleep(1000) calls Total time: 1,053,715,144ns 1,053,715.144000 nsecs/op # Fixes: c2a08203 ("perf bench: Add basic syscall benchmark") Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Andre Fredette <anfredet@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Dave Tucker <datucker@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Derek Barbosa <debarbos@redhat.com> Cc: Ian Rogers <irogers@google.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Tiezhu Yang <yangtiezhu@loongson.cn> Link: https://lore.kernel.org/lkml/ZH3lcepZ4tBYr1jv@kernel.orgSigned-off-by:
-
- 01 Jun, 2023 7 commits
-
-
Namhyung Kim authored
On large systems, it's common that PID/TID is bigger than 5-digit and it makes the output unaligned. Let's increase the width to 7. Before: $ perf script ... swapper 0 [006] 1540823.803935: 1369324 cycles:P: ffffffff9c755588 ktime_get+0x18 ([kernel.kallsyms]) gvfsd-dnssd 95114 [004] 1540823.804164: 1643871 cycles:P: ffffffff9cfdca5c __get_user_8+0x1c ([kernel.kallsyms]) perf-exec 1558582 [000] 1540823.804209: 1018714 cycles:P: ffffffff9c924ab9 __slab_free+0x9 ([kernel.kallsyms]) nmcli 1558589 [007] 1540823.804384: 1859212 cycles:P: 7f70537a8ad8 __strchrnul_evex+0x18 (/usr/lib/x86_64-linux-gnu/libc.so.6> sleep 1558582 [000] 1540823.804456: 987425 cycles:P: 7fd35bb27b30 _dl_init+0x0 (/usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2> dbus-daemon 3043 [003] 1540823.804575: 1564465 cycles:P: ffffffff9cb2bb70 llist_add_batch+0x0 ([kernel.kallsyms]) gdbus 1558592 [001] 1540823.804766: 1315219 cycles:P: ffffffff9c797b2e audit_filter_syscall+0x9e ([kernel.kallsyms]) NetworkManager 3452 [005] 1540823.805301: 1558782 cycles:P: 7fa957737748 g_bit_lock+0x58 (/usr/lib/x86_64-linux-gnu/libglib-2.0.so.0.7400.5> After: $ perf script ... swapper 0 [006] 1540823.803935: 1369324 cycles:P: ffffffff9c755588 ktime_get+0x18 ([kernel.kallsyms]) gvfsd-dnssd 95114 [004] 1540823.804164: 1643871 cycles:P: ffffffff9cfdca5c __get_user_8+0x1c ([kernel.kallsyms]) perf-exec 1558582 [000] 1540823.804209: 1018714 cycles:P: ffffffff9c924ab9 __slab_free+0x9 ([kernel.kallsyms]) nmcli 1558589 [007] 1540823.804384: 1859212 cycles:P: 7f70537a8ad8 __strchrnul_evex+0x18 (/usr/lib/x86_64-linux-gnu/libc.so.6> sleep 1558582 [000] 1540823.804456: 987425 cycles:P: 7fd35bb27b30 _dl_init+0x0 (/usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2> dbus-daemon 3043 [003] 1540823.804575: 1564465 cycles:P: ffffffff9cb2bb70 llist_add_batch+0x0 ([kernel.kallsyms]) gdbus 1558592 [001] 1540823.804766: 1315219 cycles:P: ffffffff9c797b2e audit_filter_syscall+0x9e ([kernel.kallsyms]) NetworkManager 3452 [005] 1540823.805301: 1558782 cycles:P: 7fa957737748 g_bit_lock+0x58 (/usr/lib/x86_64-linux-gnu/libglib-2.0.so.0.7400.5> Reviewer notes: Adrian added: "Might be worth noting that currently the biggest PID_MAX_LIMIT is 2^22 so pids don't get bigger than 7 digits presently" $ echo $((2 ** 22)) 4194304 $ echo -n $((2 ** 22)) | wc -c 7 $ Signed-off-by:Namhyung Kim <namhyung@kernel.org> Acked-by:
Adrian Hunter <adrian.hunter@intel.com> Tested-by:
-
Ian Rogers authored
Don't just check the raw PMU type, the only core PMU on homogeneous x86, check raw and all dynamically added PMUs. Extend the perf_pmu__warn_invalid_config to check all 4 config values. Rather than process the format list once per event, store the computed masks for each config value. Don't ignore the mask being zero, which is likely for config2 and config3, add config_masks_present so config values can be ignored only when no format information is present. Signed-off-by:
-
Ian Rogers authored
Avoid scanning format list for each event parsed. Signed-off-by:
-
Ian Rogers authored
With PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE events opening on multiple PMUs, the test expectations need updating to test for multiple events. TODOs are added to document existing hybrid perf bugs. Tested on hybrid alderlake and non-hybrid tigerlake. Signed-off-by:
-
Ian Rogers authored
Numeric events are either raw events or those with ABI defined numbers matched by the lexer. PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE events should wildcard match on hybrid systems. So "cycles" should match each PMU type with an extended type, not just PERF_TYPE_HARDWARE. Change wildcard matching to add the event even if wildcard PMU scanning fails, there will be no extended type but this best matches previous behavior. Only set the extended type when the event type supports it and when perf_pmus__supports_extended_type is true. This new function returns true if >1 core PMU and avoids potential errors on older kernels. Modify evsel__compute_group_pmu_name using a helper perf_pmu__is_software to determine when grouping should occur. Try to use PMUs, and evsel__find_pmu, as being more dependable than evsel->pmu_name. Set a parse events error if a hardware term's PMU lookup fails, to provide extra diagnostics. Fixes: 8bc75f69 ("perf parse-events: Support wildcards on raw events") Reported-by:
-
Ian Rogers authored
It is often useful to know not just the attribute and perf_event_open() details when opening an evsel, but also the evsel's name. Add this debug output for verbose 3 so that it won't interfere with the current verbose 2 output. Signed-off-by:
-
Ian Rogers authored
Flip the return value correcting a bug. Fixes: 6b9da260 ("perf pmu: Remove is_pmu_hybrid") Reported-by:
-
- 31 May, 2023 1 commit
-
-
Arnaldo Carvalho de Melo authored
perf tools fixes for v6.4: 2nd batch - Fix BPF CO-RE naming convention for checking the availability of fields on 'union perf_mem_data_src' on the running kernel. - Remove the use of llvm-strip on BPF skel object files, not needed, fixes a build breakage when the llvm package, that contains it in most distros, isn't installed. - Fix tools that use both evsel->{bpf_counter_list,bpf_filters}, removing them from a union. - Remove extra "--" from the 'perf ftrace latency' --use-nsec option, previously it was working only when using the '-n' alternative. - Don't stop building when both binutils-devel and a C++ compiler isn't available to compile the alternative C++ demangle support code, disable that feature instead. - Sync the linux/in.h and coresight-pmu.h header copies with the kernel sources. - Fix relative include path to cs-etm.h. Signed-off-by:
-
- 30 May, 2023 1 commit
-
-
Ian Rogers authored
Missed function rename from pmu_have_event to perf_pmus__have_event made the perf build fail on powerpc. Committer notes: The perf_pmus__have_event() is declared in util/pmus.h, so use it instead of by now needless util/pmu.h. Fixes: 1eaf496e ("perf pmu: Separate pmu and pmus") Reported-by:
Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by:
-
- 28 May, 2023 18 commits
-
-
Ian Rogers authored
Use a single stack allocated buffer and avoid 8,192 bytes in .bss. Signed-off-by:
-
Ian Rogers authored
Avoid two static paths that contributed 8,192 bytes to .bss are only used duing the perf parse pmu test. This change helps FORTIFY triggering 2 warnings like: ``` tests/pmu.c: In function ‘test__pmu’: tests/pmu.c:121:43: error: ‘%s’ directive output may be truncated writing up to 4095 bytes into a region of size 4090 [-Werror=format-truncation=] 121 | snprintf(buf, sizeof(buf), "rm -f %s/*\n", dir); ``` So make buf a little larger. Signed-off-by:
-
Ian Rogers authored
Move the cgroupfs_cache_entry 4128 byte array out of .bss. Signed-off-by:
-
Ian Rogers authored
Avoid 16,384 bytes in .bss by stack allocating two bitmaps. Signed-off-by:
-
Ian Rogers authored
Avoid 4 static arrays for paths, pass in a char[] buffer to use. Makes mkpath thread safe for the small number of users. Also removes 16,384 bytes from .bss. Signed-off-by:
-
Ian Rogers authored
Avoid 14,432 bytes in .bss by dynamically allocating params. Signed-off-by:
-
Ian Rogers authored
Allocate start time and state arrays when command starts rather than using 114,688 bytes in .bss. Signed-off-by:
-
Ian Rogers authored
lockhash_table is 32,768 bytes in .bss, make it a memory allocation so that the space is freed for non-lock perf commands. Signed-off-by:
-
Ian Rogers authored
Avoid a PATH_MAX array in __daemon (the .data section) by dynamically allocating the memory. Signed-off-by:
-
Ian Rogers authored
tracing_mnt was set but never written. tracing_events_path was set and read on errors paths, but its value is exactly tracing_path with a "/events" appended, so we can derive the value in the error paths. There appears to have been a missing "/" when tracing_events_path was initialized. Signed-off-by:
-
Ian Rogers authored
Change struct fs to have a pointer to a dynamically allocated array rather than an array. This reduces the size of fs__entries from 24,768 bytes to 240 bytes. Read paths into a stack allocated array and strdup. Fix off-by-1 fscanf %<num>s in fs__read_mounts caught by address sanitizer. Signed-off-by:
-
Ian Rogers authored
Allows the movement of 46,072 bytes from .data to .data.rel.ro. Signed-off-by:
-
Ian Rogers authored
Allows the movement of 33,128 bytes from .data to .data.rel.ro. Signed-off-by:
-
Ian Rogers authored
This allows the movement of 5,808 bytes from .data to .rodata. Signed-off-by:
-
Ian Rogers authored
This allows the movement of some sizeable data arrays (168,624 bytes) to .data.relro. Without PIE or the strings it could be moved to .rodata. Signed-off-by:
-
Ian Rogers authored
Avoid a large static array, dynamically allocate the nodes avoiding a hard coded limited as well. Signed-off-by:
-
Ian Rogers authored
Address/memory sanitizer was reporting issues in evsel__group_pmu_name because the for_each_group_evsel loop didn't terminate when the head was reached, the head would then be cast and accessed as an evsel leading to invalid memory accesses. Fix for_each_group_member and for_each_group_evsel to terminate at the list head. Note, evsel__group_pmu_name no longer iterates the group, but the problem is present regardless. Fixes: 717e263f ("perf report: Show group description when event group is enabled") Signed-off-by:
-
Ian Rogers authored
Previously the evsel__group_pmu_name would iterate the evsel's group, however, the list of evsels aren't yet sorted and so the loop may terminate prematurely. It is also not desirable to iterate the list of evsels during list_sort as the list may be broken. Precompute the group_pmu_name for the evsel before sorting, as part of the computation and only if necessary, iterate the whole list looking for group members so that being sorted isn't necessary. Move the group pmu name computation to parse-events.c given the closer dependency on the behavior of parse_events__sort_events_and_fix_groups. Fixes: 7abf0bcc ("perf evsel: Add function to compute group PMU name") Signed-off-by:
-
- 27 May, 2023 6 commits
-
-
Ian Rogers authored
Users have been removed or switched to using pmu->is_core with perf_pmus__num_core_pmus() > 1. Reviewed-by:
-
Ian Rogers authored
perf_pmus__has_hybrid was used to detect when there was >1 core PMU, this can be achieved with perf_pmus__num_core_pmus that doesn't depend upon is_pmu_hybrid and PMU name comparisons. When modifying the function calls take the opportunity to improve comments, enable/simplify tests that were previously failing for hybrid but now pass and to simplify generic code. Reviewed-by:
-
Ian Rogers authored
Add perf_pmus__num_core_pmus that will count core PMUs holding the result in a static. Reuse for perf_pmus__num_mem_pmus. Reviewed-by:
-
Ian Rogers authored
perf_pmus__find_by_type may be called for something like a raw event, in which case the PMU isn't guaranteed to have been looked up. Add a second check to make sure all PMUs are loaded. Reviewed-by:
-
Ian Rogers authored
perf_pmus__scan will process every directory in sysfs to see if it is a PMU, attempting to add it if not already in the pmus list. Add two booleans to record whether this scanning has been done for core or all PMUs. Skip scanning in the event that scanning has already occurred. Reviewed-by:
-
Ian Rogers authored
Scanning all PMUs is expensive as all PMUs sysfs entries are loaded, benchmarking shows more than 4x the cost: ``` $ perf bench internals pmu-scan -i 1000 Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 989.231 usec (+- 1.535 usec) Average PMU scanning took: 4309.425 usec (+- 74.322 usec) ``` Add new perf_pmus__scan_core routine that scans just core PMUs. Replace perf_pmus__scan calls with perf_pmus__scan_core when non-core PMUs are being ignored. Reviewed-by:
-
