- 03 Apr, 2020 15 commits
-
-
Stephane Eranian authored
This patch extends the perf script --symbols option to filter on hexadecimal addresses in addition to symbol names. This makes it easier to handle cases where symbols are aliased. With this patch, it is possible to mix and match symbols and hexadecimal addresses using the --symbols option. $ perf script --symbols=noploop,0x4007a0 Signed-off-by:
Stephane Eranian <eranian@google.com> Reviewed-by:
Ian Rogers <irogers@google.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/20200325220802.15039-1-irogers@google.comSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Arnaldo Carvalho de Melo authored
In d10ec006 ("perf hists browser: Allow passing an initial hotkey") the hist_entry__title() call was cut'n'pasted to a function where the 'title' variable is a pointer, not an array, so the sizeof(title) continues syntactically valid but ends up reducing the real size of the buffer where to format the first line in the screen to 8 bytes, which makes the formatting at the title at each refresh to produce just the string "Samples ", duh, fix it by passing the size of the buffer. Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Jin Yao <yao.jin@linux.intel.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Fixes: d10ec006 ("perf hists browser: Allow passing an initial hotkey") Link: http://lore.kernel.org/lkml/20200330154314.GB4576@kernel.orgSigned-off-by:
-
Jin Yao authored
It would be nice if we can use a hotkey in perf top browser to select a event for sorting. For example: perf top --group -e cycles,instructions,cache-misses Samples Overhead Shared Object Symbol 40.03% 45.71% 0.03% div [.] main 20.46% 14.67% 0.21% libc-2.27.so [.] __random_r 20.01% 19.54% 0.02% libc-2.27.so [.] __random 9.68% 10.68% 0.00% div [.] compute_flag 4.32% 4.70% 0.00% libc-2.27.so [.] rand 3.84% 3.43% 0.00% div [.] rand@plt 0.05% 0.05% 2.33% libc-2.27.so [.] __strcmp_sse2_unaligned 0.04% 0.08% 2.43% perf [.] perf_hpp__is_dynamic_en 0.04% 0.02% 6.64% perf [.] rb_next 0.04% 0.01% 3.87% perf [.] dso__find_symbol 0.04% 0.04% 1.77% perf [.] sort__dso_cmp When user press hotkey '2' (event index, starting from 0), it indicates to sort output by the third event in group (cache-misses). Samples Overhead Shared Object Symbol 4.07% 1.28% 6.68% perf [.] rb_next 3.57% 3.98% 4.11% perf [.] __hists__insert_output 3.67% 11.24% 3.60% perf [.] perf_hpp__is_dynamic_e 3.67% 3.20% 3.20% perf [.] hpp__sort_overhead 0.81% 0.06% 3.01% perf [.] dso__find_symbol 1.62% 5.47% 2.51% perf [.] hists__match 2.70% 1.86% 2.47% libc-2.27.so [.] _int_malloc 0.19% 0.00% 2.29% [kernel] [k] copy_page 0.41% 0.32% 1.98% perf [.] hists__decay_entries 1.84% 3.67% 1.68% perf [.] sort__dso_cmp 0.16% 0.00% 1.63% [kernel] [k] clear_page_erms Now the output is sorted by cache-misses. v2: --- Zero the history if hotkey is pressed. Signed-off-by:Jin Yao <yao.jin@linux.intel.com> Suggested-by:
-
Jin Yao authored
'perf report' supports the option --group-sort-idx, which sorts the output by the event at the index n in event group. For example: perf record -e cycles,instructions,cache-misses perf report --group --group-sort-idx 2 --stdio The perf-report output is sorted by cache-misses. This patch supports --group-sort-idx in perf-top. For example: perf top --group -e cycles,instructions,cache-misses --group-sort-idx 2 The perf-top output is sorted by cache-misses. Signed-off-by:
-
Kemeng Shi authored
During execution of command 'perf report' in my arm64 virtual machine, this error message is showed: failed to process sample __symbol__inc_addr_samples(860): ENOMEM! sym->name=__this_module, start=0x1477100, addr=0x147dbd8, end=0x80002000, func: 0 The error is caused with path: cmd_report __cmd_report perf_session__process_events __perf_session__process_events ordered_events__flush __ordered_events__flush oe->deliver (ordered_events__deliver_event) perf_session__deliver_event machines__deliver_event perf_evlist__deliver_sample tool->sample (process_sample_event) hist_entry_iter__add iter->add_entry_cb(hist_iter__report_callback) hist_entry__inc_addr_samples symbol__inc_addr_samples __symbol__inc_addr_samples h = annotated_source__histogram(src, evidx) (NULL) annotated_source__histogram failed is caused with path: ... hist_entry__inc_addr_samples symbol__inc_addr_samples symbol__hists annotated_source__alloc_histograms src->histograms = calloc(nr_hists, sizeof_sym_hist) (failed) Calloc failed as the symbol__size(sym) is too huge. As show in error message: start=0x1477100, end=0x80002000, size of symbol is about 2G. This is the same problem as 'perf annotate: Fix s390 gap between kernel end and module start (b9c0a649)'. Perf gets symbol information from /proc/kallsyms in __dso__load_kallsyms. A part of symbol in /proc/kallsyms from my virtual machine is as follows: #cat /proc/kallsyms | sort ... ffff000001475080 d rpfilter_mt_reg [ip6t_rpfilter] ffff000001475100 d $d [ip6t_rpfilter] ffff000001475100 d __this_module [ip6t_rpfilter] ffff000080080000 t _head ffff000080080000 T _text ffff000080080040 t pe_header ... Take line 'ffff000001475100 d __this_module [ip6t_rpfilter]' as example. The start and end of symbol are both set to ffff000001475100 in dso__load_all_kallsyms. Then symbols__fixup_end will set the end of symbol to next big address to ffff000001475100 in /proc/kallsyms, ffff000080080000 in this example. Then sizeof of symbol will be about 2G and cause the problem. The start of module in my machine is ffff000000a62000 t $x [dm_mod] The start of kernel in my machine is ffff000080080000 t _head There is a big gap between end of module and begin of kernel if a samll amount of memory is used by module. And the last symbol in module will have a large address range as caotaining the big gap. Give that the module and kernel text segment sequence may change in the future, fix this by limiting range of last symbol in module and kernel to 4K in arch arm64. Signed-off-by:Kemeng Shi <shikemeng@huawei.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Hewenliang <hewenliang4@huawei.com> Cc: Hu Shiyuan <hushiyuan@huawei.com> Cc: Ian Rogers <irogers@google.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Richter <tmricht@linux.ibm.com> Link: http://lore.kernel.org/lkml/33fd24c4-0d5a-9d93-9b62-dffa97c992ca@huawei.com [ refreshed the patch on current codebase, added string.h include as strchr() is used ] Signed-off-by:
-
Arnaldo Carvalho de Melo authored
When one does: $ make -C tools/perf build-test The makefile in tools/perf/tests/ will, just like the main one, detect how many cores are in the system and use it with -j. Sometimes we may need to override that, for instance, when using icecream or distcc to use multiple machines in the build process, then we need to, as with the main makefile, use: $ make JOBS=N -C tools/perf build-test Fix the tests makefile to honour that. Acked-by:
Jiri Olsa <jolsa@kernel.org> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Namhyung Kim <namhyung@kernel.org> Link: http://lore.kernel.org/lkml/20200330130301.GA31702@kernel.orgSigned-off-by:
-
Namhyung Kim authored
The --show-cgroup-events option is to print CGROUP events in the output like others. Committer testing: [root@seventh ~]# perf record --all-cgroups --namespaces /wb/cgtest [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.039 MB perf.data (487 samples) ] [root@seventh ~]# perf script --show-cgroup-events | grep PERF_RECORD_CGROUP -B2 -A2 swapper 0 0.000000: PERF_RECORD_CGROUP cgroup: 1 / perf 12145 11200.440730: 1 cycles: ffffffffb900d58b __intel_pmu_enable_all.constprop.0+0x3b (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) perf 12145 11200.440733: 1 cycles: ffffffffb900d58b __intel_pmu_enable_all.constprop.0+0x3b (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) -- cgtest 12145 11200.440739: 193472 cycles: ffffffffb90f6fbc commit_creds+0x1fc (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) cgtest 12145 11200.440790: 2691608 cycles: 7fa2cb43019b _dl_sysdep_start+0x7cb (/usr/lib64/ld-2.29.so) cgtest 12145 11200.440962: PERF_RECORD_CGROUP cgroup: 83 /sub cgtest 12147 11200.441054: 1 cycles: ffffffffb900d58b __intel_pmu_enable_all.constprop.0+0x3b (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) cgtest 12147 11200.441057: 1 cycles: ffffffffb900d58b __intel_pmu_enable_all.constprop.0+0x3b (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) -- cgtest 12148 11200.441103: 10227 cycles: ffffffffb9a0153d end_repeat_nmi+0x48 (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) cgtest 12148 11200.441106: 273295 cycles: ffffffffb99ecbc7 copy_page+0x7 (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) cgtest 12147 11200.441133: PERF_RECORD_CGROUP cgroup: 88 /sub/cgrp1 cgtest 12147 11200.441143: 2788845 cycles: ffffffffb94676c2 security_genfs_sid+0x102 (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) cgtest 12148 11200.441162: PERF_RECORD_CGROUP cgroup: 93 /sub/cgrp2 cgtest 12148 11200.441182: 2669546 cycles: 401020 _init+0x20 (/wb/cgtest) cgtest 12149 11200.441247: 1 cycles: ffffffffb900d58b __intel_pmu_enable_all.constprop.0+0x3b (/lib/modules/5.6.0-rc6-00008-gfe2413eefd7f/build/vmlinux) [root@seventh ~]# Signed-off-by:Namhyung Kim <namhyung@kernel.org> Tested-by:
-
Namhyung Kim authored
The --all-cgroups option is to enable cgroup profiling support. It tells kernel to record CGROUP events in the ring buffer so that 'perf top' can identify task/cgroup association later. Committer testing: Use: # perf top --all-cgroups -s cgroup_id,cgroup,pid Signed-off-by:
-
Namhyung Kim authored
The --all-cgroups option is to enable cgroup profiling support. It tells kernel to record CGROUP events in the ring buffer so that perf report can identify task/cgroup association later. [root@seventh ~]# perf record --all-cgroups --namespaces /wb/cgtest [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.042 MB perf.data (558 samples) ] [root@seventh ~]# perf report --stdio -s cgroup_id,cgroup,pid # To display the perf.data header info, please use --header/--header-only options. # # # Total Lost Samples: 0 # # Samples: 558 of event 'cycles' # Event count (approx.): 458017341 # # Overhead cgroup id (dev/inode) Cgroup Pid:Command # ........ ..................... .......... ............... # 33.15% 4/0xeffffffb /sub 9615:looper0 32.83% 4/0xf00002f5 /sub/cgrp2 9620:looper2 32.79% 4/0xf00002f4 /sub/cgrp1 9619:looper1 0.35% 4/0xf00002f5 /sub/cgrp2 9618:cgtest 0.34% 4/0xf00002f4 /sub/cgrp1 9617:cgtest 0.32% 4/0xeffffffb / 9615:looper0 0.11% 4/0xeffffffb /sub 9617:cgtest 0.10% 4/0xeffffffb /sub 9618:cgtest # # (Tip: Sample related events with: perf record -e '{cycles,instructions}:S') # [root@seventh ~]# Signed-off-by: -
Namhyung Kim authored
Synthesize cgroup events by iterating cgroup filesystem directories. The cgroup event only saves the portion of cgroup path after the mount point and the cgroup id (which actually is a file handle). Signed-off-by:
-
Namhyung Kim authored
The cgroup sort key is to show cgroup membership of each task. Currently it shows full path in the cgroupfs (not relative to the root of cgroup namespace) since it'd be more intuitive IMHO. Otherwise root cgroup in different namespaces will all show same name - "/". The cgroup sort key should come before cgroup_id otherwise sort_dimension__add() will match it to cgroup_id as it only matches with the given substring. For example it will look like following. Note that record patch adding --all-cgroups patch will come later. $ perf record -a --namespace --all-cgroups cgtest [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.208 MB perf.data (4090 samples) ] $ perf report -s cgroup_id,cgroup,pid ... # Overhead cgroup id (dev/inode) Cgroup Pid:Command # ........ ..................... .......... ............... # 93.96% 0/0x0 / 0:swapper 1.25% 3/0xeffffffb / 278:looper0 0.86% 3/0xf000015f /sub/cgrp1 280:cgtest 0.37% 3/0xf0000160 /sub/cgrp2 281:cgtest 0.34% 3/0xf0000163 /sub/cgrp3 282:cgtest 0.22% 3/0xeffffffb /sub 278:looper0 0.20% 3/0xeffffffb / 280:cgtest 0.15% 3/0xf0000163 /sub/cgrp3 285:looper3 Signed-off-by: -
Namhyung Kim authored
Each cgroup is kept in the perf_env's cgroup_tree sorted by the cgroup id. Hist entries have cgroup id can compare it directly and later it can be used to find a group name using this tree. Signed-off-by:
-
Namhyung Kim authored
Implement basic functionality to support cgroup tracking. Each cgroup can be identified by inode number which can be read from userspace too. The actual cgroup processing will come in the later patch. Reported-by:
kernel test robot <rong.a.chen@intel.com> Signed-off-by:
-
Namhyung Kim authored
The file handle (FHANDLE) support is configurable so some systems might not have it. So add a config feature item to check it on build time so that we don't add the cgroup tracking feature based on that. Committer notes: Had to make the test use the same construct as its later use in synthetic-events.c, in the next patch in this series. i.e. make it be: struct { struct file_handle fh; uint64_t cgroup_id; } handle; To cope with: CC /tmp/build/perf/util/cloexec.o util/synthetic-events.c:428:22: error: field 'fh' with CC /tmp/build/perf/util/call-path.o variable sized type 'struct file_handle' not at the end of a struct or class is a GNU extension [-Werror,-Wgnu-variable-sized-type-not-at-end] struct file_handle fh; ^ 1 error generated. Deal with this at some point, i.e. investigate if the right thing is to remove that -Wgnu-variable-sized-type-not-at-end from our CFLAGS, for now do the test the same way as it is used looks more sensible. Reported-by:Arnaldo Carvalho de Melo <acme@kernel.org> Signed-off-by:
-
Arnaldo Carvalho de Melo authored
We'll need it for the cgroup patches, and its better to have it in a separate patch in case we need to later revert the cgroup patches. I.e. without this we have: [root@five ~]# perf test -v python 19: 'import perf' in python : --- start --- test child forked, pid 148447 Traceback (most recent call last): File "<stdin>", line 1, in <module> ImportError: /tmp/build/perf/python/perf.cpython-37m-x86_64-linux-gnu.so: undefined symbol: down_write test child finished with -1 ---- end ---- 'import perf' in python: FAILED! [root@five ~]# Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.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: http://lore.kernel.org/lkml/20200403123606.GC23243@kernel.orgSigned-off-by:
-
- 02 Apr, 2020 1 commit
-
-
Namhyung Kim authored
To get the changes in: 6546b19f ("perf/core: Add PERF_SAMPLE_CGROUP feature") 96aaab68 ("perf/core: Add PERF_RECORD_CGROUP event") This silences this perf tools build warning: Warning: Kernel ABI header at 'tools/include/uapi/linux/perf_event.h' differs from latest version at 'include/uapi/linux/perf_event.h' diff -u tools/include/uapi/linux/perf_event.h include/uapi/linux/perf_event.h This update is a prerequisite to adding support for the HW index of raw branch records. Signed-off-by:
-
- 27 Mar, 2020 5 commits
-
-
Namhyung Kim authored
The PERF_SAMPLE_CGROUP bit is to save (perf_event) cgroup information in the sample. It will add a 64-bit id to identify current cgroup and it's the file handle in the cgroup file system. Userspace should use this information with PERF_RECORD_CGROUP event to match which cgroup it belongs. I put it before PERF_SAMPLE_AUX for simplicity since it just needs a 64-bit word. But if we want bigger samples, I can work on that direction too. Committer testing: $ pahole perf_sample_data | grep -w cgroup -B5 -A5 /* --- cacheline 4 boundary (256 bytes) was 56 bytes ago --- */ struct perf_regs regs_intr; /* 312 16 */ /* --- cacheline 5 boundary (320 bytes) was 8 bytes ago --- */ u64 stack_user_size; /* 328 8 */ u64 phys_addr; /* 336 8 */ u64 cgroup; /* 344 8 */ /* size: 384, cachelines: 6, members: 22 */ /* padding: 32 */ }; $ Signed-off-by:
Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by:
Tejun Heo <tj@kernel.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Zefan Li <lizefan@huawei.com> Link: http://lore.kernel.org/lkml/20200325124536.2800725-3-namhyung@kernel.orgSigned-off-by:
-
Namhyung Kim authored
To support cgroup tracking, add CGROUP event to save a link between cgroup path and id number. This is needed since cgroups can go away when userspace tries to read the cgroup info (from the id) later. The attr.cgroup bit was also added to enable cgroup tracking from userspace. This event will be generated when a new cgroup becomes active. Userspace might need to synthesize those events for existing cgroups. Committer testing: From the resulting kernel, using /sys/kernel/btf/vmlinux: $ pahole perf_event_attr | grep -w cgroup -B5 -A1 __u64 write_backward:1; /* 40:27 8 */ __u64 namespaces:1; /* 40:28 8 */ __u64 ksymbol:1; /* 40:29 8 */ __u64 bpf_event:1; /* 40:30 8 */ __u64 aux_output:1; /* 40:31 8 */ __u64 cgroup:1; /* 40:32 8 */ __u64 __reserved_1:31; /* 40:33 8 */ $ Reported-by:
kbuild test robot <lkp@intel.com> Signed-off-by:
-
Hagen Paul Pfeifer authored
For some kind of analysis a deltatime output is more human friendly and reduce the cognitive load for further analysis. The following output demonstrate the new option "deltatime": calculate the time difference in relation to the previous event. $ perf script --deltatime test 2525 [001] 0.000000: sdt_libev:ev_add: (5635e72a5ebd) test 2525 [001] 0.000091: sdt_libev:epoll_wait_enter: (5635e72a76a9) test 2525 [001] 1.000051: sdt_libev:epoll_wait_return: (5635e72a772e) arg1=1 test 2525 [001] 0.000685: sdt_libev:ev_add: (5635e72a5ebd) test 2525 [001] 0.000048: sdt_libev:epoll_wait_enter: (5635e72a76a9) test 2525 [001] 1.000104: sdt_libev:epoll_wait_return: (5635e72a772e) arg1=1 test 2525 [001] 0.003895: sdt_libev:epoll_wait_enter: (5635e72a76a9) test 2525 [001] 0.996034: sdt_libev:epoll_wait_return: (5635e72a772e) arg1=1 test 2525 [001] 0.000058: sdt_libev:epoll_wait_enter: (5635e72a76a9) test 2525 [001] 1.000004: sdt_libev:epoll_wait_return: (5635e72a772e) arg1=1 test 2525 [001] 0.000064: sdt_libev:epoll_wait_enter: (5635e72a76a9) test 2525 [001] 0.999934: sdt_libev:epoll_wait_return: (5635e72a772e) arg1=1 test 2525 [001] 0.000056: sdt_libev:epoll_wait_enter: (5635e72a76a9) test 2525 [001] 0.999930: sdt_libev:epoll_wait_return: (5635e72a772e) arg1=1 Committer testing: So go from default output to --reltime and then this new --deltatime, to contrast the various timestamp presentation modes for a random perf.data file I had laying around: [root@five ~]# perf script --reltime | head perf 442394 [000] 0.000000: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 0.000002: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 0.000004: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 0.000006: 128 cycles: ffffffff972415a1 perf_event_update_userpage+0x1 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 0.000009: 2597 cycles: ffffffff97463785 cap_task_setscheduler+0x5 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000036: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000038: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000040: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000041: 224 cycles: ffffffff9700a53a perf_ibs_handle_irq+0x1da (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000044: 4439 cycles: ffffffff97120d85 put_prev_entity+0x45 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) [root@five ~]# perf script --deltatime | head perf 442394 [000] 0.000000: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 0.000002: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 0.000001: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 0.000001: 128 cycles: ffffffff972415a1 perf_event_update_userpage+0x1 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 0.000002: 2597 cycles: ffffffff97463785 cap_task_setscheduler+0x5 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000027: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000002: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000001: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000001: 224 cycles: ffffffff9700a53a perf_ibs_handle_irq+0x1da (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 0.000002: 4439 cycles: ffffffff97120d85 put_prev_entity+0x45 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) [root@five ~]# perf script | head perf 442394 [000] 7600.157861: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 7600.157864: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 7600.157866: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 7600.157867: 128 cycles: ffffffff972415a1 perf_event_update_userpage+0x1 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [000] 7600.157870: 2597 cycles: ffffffff97463785 cap_task_setscheduler+0x5 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 7600.157897: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 7600.157900: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 7600.157901: 16 cycles: ffffffff9706e544 native_write_msr+0x4 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 7600.157903: 224 cycles: ffffffff9700a53a perf_ibs_handle_irq+0x1da (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) perf 442394 [001] 7600.157906: 4439 cycles: ffffffff97120d85 put_prev_entity+0x45 (/usr/lib/debug/lib/modules/5.5.10-200.fc31.x86_64/vmlinux) [root@five ~]# Andi suggested we better implement it as a new field, i.e. -F deltatime, like: [root@five ~]# perf script -F deltatime Invalid field requested. Usage: perf script [<options>] or: perf script [<options>] record <script> [<record-options>] <command> or: perf script [<options>] report <script> [script-args] or: perf script [<options>] <script> [<record-options>] <command> or: perf script [<options>] <top-script> [script-args] -F, --fields <str> comma separated output fields prepend with 'type:'. +field to add and -field to remove.Valid types: hw,sw,trace,raw,synth. Fields: comm,tid,pid,time,cpu,event,trace,ip,sym,dso,addr,symoff,srcline,period,iregs,uregs,brstack,brstacksym,flags,bpf-output,brstackinsn,brstackoff,callindent,insn,insnlen,synth,phys_addr,metric,misc,ipc [root@five ~]# I.e. we have -F for maximum flexibility: [root@five ~]# perf script -F comm,pid,cpu,time | head perf 442394 [000] 7600.157861: perf 442394 [000] 7600.157864: perf 442394 [000] 7600.157866: perf 442394 [000] 7600.157867: perf 442394 [000] 7600.157870: perf 442394 [001] 7600.157897: perf 442394 [001] 7600.157900: perf 442394 [001] 7600.157901: perf 442394 [001] 7600.157903: perf 442394 [001] 7600.157906: [root@five ~]# But since we already have --reltime, having --deltatime, documented one after the other is sensible. Signed-off-by:Hagen Paul Pfeifer <hagen@jauu.net> Tested-by:
-
Adrian Hunter authored
Add to the "x86 instruction decoder - new instructions" test the following instructions: incsspd incsspq rdsspd rdsspq saveprevssp rstorssp wrssd wrssq wrussd wrussq setssbsy clrssbsy endbr32 endbr64 And the "notrack" prefix for indirect calls and jumps. For information about the instructions, refer Intel Control-flow Enforcement Technology Specification May 2019 (334525-003). Committer testing: $ perf test instr 67: x86 instruction decoder - new instructions : Ok $ Then use verbose mode and check one of those new instructions: $ perf test -v instr |& grep saveprevssp Decoded ok: f3 0f 01 ea saveprevssp Decoded ok: f3 0f 01 ea saveprevssp $ Signed-off-by:
Adrian Hunter <adrian.hunter@intel.com> Acked-by:
Masami Hiramatsu <mhiramat@kernel.org> Tested-by:
Yu-cheng Yu <yu-cheng.yu@intel.com> Signed-off-by:
-
Yu-cheng Yu authored
Add the following CET instructions to the opcode map: INCSSP: Increment Shadow Stack pointer (SSP). RDSSP: Read SSP into a GPR. SAVEPREVSSP: Use "previous ssp" token at top of current Shadow Stack (SHSTK) to create a "restore token" on the previous (outgoing) SHSTK. RSTORSSP: Restore from a "restore token" to SSP. WRSS: Write to kernel-mode SHSTK (kernel-mode instruction). WRUSS: Write to user-mode SHSTK (kernel-mode instruction). SETSSBSY: Verify the "supervisor token" pointed by MSR_IA32_PL0_SSP, set the token busy, and set then Shadow Stack pointer(SSP) to the value of MSR_IA32_PL0_SSP. CLRSSBSY: Verify the "supervisor token" and clear its busy bit. ENDBR64/ENDBR32: Mark a valid 64/32 bit control transfer endpoint. Detailed information of CET instructions can be found in Intel Software Developer's Manual. Signed-off-by:Tony Luck <tony.luck@intel.com> Acked-by:
-
- 26 Mar, 2020 2 commits
-
-
He Zhe authored
The $(CC) passed to arch_errno_names.sh may include a series of parameters along with gcc itself. To avoid overwriting the following parameters of arch_errno_names.sh and break the build like below, we just pick up the first word of the $(CC). find: unknown predicate `-m64/arch' x86_64-wrs-linux-gcc: warning: '-x c' after last input file has no effect x86_64-wrs-linux-gcc: error: unrecognized command line option '-m64/include/uapi/asm-generic/errno.h' x86_64-wrs-linux-gcc: fatal error: no input files Signed-off-by:
He Zhe <zhe.he@windriver.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/1581618066-187262-2-git-send-email-zhe.he@windriver.comSigned-off-by:
-
Ian Rogers authored
Terms may have a NULL config in which case a strcmp will SEGV. This can be reproduced with: perf stat -e '*/event=?,nr/' sleep 1 Add a NULL check to avoid this. This was caught by LLVM's libfuzzer. Signed-off-by:
-
- 25 Mar, 2020 1 commit
-
-
Tony Jones authored
The method of unwinding for kernel space is defined by the kernel config, not by the value of --call-graph. Improve the documentation to reflect this. Signed-off-by:
Tony Jones <tonyj@suse.de> Link: http://lore.kernel.org/lkml/20200325164053.10177-1-tonyj@suse.deSigned-off-by:
-
- 24 Mar, 2020 16 commits
-
-
Ravi Bangoria authored
Perf gets dso details from two different sources. 1st, from builid headers in perf.data and 2nd from MMAP2 samples. Dso from buildid header does not have dso_id detail. And dso from MMAP2 samples does not have buildid information. If detail of the same dso is present at both the places, filename is common. Previously, __dsos__findnew_link_by_longname_id() used to compare only long or short names, but Commit 0e3149f8 ("perf dso: Move dso_id from 'struct map' to 'struct dso'") also added a dso_id comparison. Because of that, now perf is creating two different dso objects of the same file, one from buildid header (with dso_id but without buildid) and second from MMAP2 sample (with buildid but without dso_id). This is causing issues with archive, buildid-list etc subcommands. Fix this by comparing dso_id only when it's present. And incase dso is present in 'dsos' list without dso_id, inject dso_id detail as well. Before: $ sudo ./perf buildid-list -H 0000000000000000000000000000000000000000 /usr/bin/ls 0000000000000000000000000000000000000000 /usr/lib64/ld-2.30.so 0000000000000000000000000000000000000000 /usr/lib64/libc-2.30.so $ ./perf archive perf archive: no build-ids found After: $ ./perf buildid-list -H b6b1291d0cead046ed0fa5734037fa87a579adee /usr/bin/ls 641f0c90cfa15779352f12c0ec3c7a2b2b6f41e8 /usr/lib64/ld-2.30.so 675ace3ca07a0b863df01f461a7b0984c65c8b37 /usr/lib64/libc-2.30.so $ ./perf archive Now please run: $ tar xvf perf.data.tar.bz2 -C ~/.debug wherever you need to run 'perf report' on. Committer notes: Renamed is_empty_dso_id() to dso_id__empty() and inject_dso_id() to dso__inject_id() to keep namespacing consistent. Fixes: 0e3149f8 ("perf dso: Move dso_id from 'struct map' to 'struct dso'") Reported-by:
Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com> Signed-off-by:
Ravi Bangoria <ravi.bangoria@linux.ibm.com> Tested-by:
-
Christophe JAILLET authored
'snprintf' returns the number of characters which would be generated for the given input. If the returned value is *greater than* or equal to the buffer size, it means that the output has been truncated. Fix the overflow test accordingly. Fixes: 7780c25b ("perf tools: Allow ability to map cpus to nodes easily") Fixes: 92a7e127 ("perf cpumap: Add cpu__max_present_cpu()") Signed-off-by:
Christophe JAILLET <christophe.jaillet@wanadoo.fr> Suggested-by:
David Laight <David.Laight@ACULAB.COM> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Don Zickus <dzickus@redhat.com> Cc: He Zhe <zhe.he@windriver.com> Cc: Jan Stancek <jstancek@redhat.com> Cc: Jiri Olsa <jolsa@redhat.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: kernel-janitors@vger.kernel.org Link: http://lore.kernel.org/lkml/20200324070319.10901-1-christophe.jaillet@wanadoo.frSigned-off-by:
-
John Garry authored
Add creating event aliases to the pmu-events test. So currently we verify that the generated pmu-events.c is as expected for some test events. Now test that we generate aliases as expected for those events during normal operation. For that, we cycle through each HW PMU in the system, and use the test events to create aliases, and verify those against known, expected values. For core PMUs, we should create an alias for every event in test_cpu_events[]. However, for uncore PMUs, they need to be matched by the pmu_event.pmu member, so use test_uncore_events[]; so check the match beforehand with pmu_uncore_alias_match(). A sample run is as follows for my x86 machine: john@linux-3c19:~/linux> tools/perf/perf test -vv 10 10: PMU events : --- start --- ... testing PMU uncore_arb aliases: no events to match testing PMU cstate_pkg aliases: no events to match skipping testing PMU breakpoint testing aliases PMU uncore_cbox_1: matched event unc_cbo_xsnp_response.miss_eviction testing PMU uncore_cbox_1 aliases: pass testing PMU power aliases: no events to match testing aliases PMU cpu: matched event bp_l1_btb_correct testing aliases PMU cpu: matched event bp_l2_btb_correct testing aliases PMU cpu: matched event segment_reg_loads.any testing aliases PMU cpu: matched event dispatch_blocked.any testing aliases PMU cpu: matched event eist_trans testing PMU cpu aliases: pass testing PMU intel_pt aliases: no events to match skipping testing PMU software skipping testing PMU intel_bts testing PMU uncore_imc aliases: no events to match testing aliases PMU uncore_cbox_0: matched event unc_cbo_xsnp_response.miss_eviction testing PMU uncore_cbox_0 aliases: pass testing PMU cstate_core aliases: no events to match skipping testing PMU tracepoint testing PMU msr aliases: no events to match test child finished with 0 Signed-off-by:
John Garry <john.garry@huawei.com> Acked-by:
Jiri Olsa <jolsa@redhat.com> Tested-by:
-
John Garry authored
The perf pmu-events test will want to use pmu_uncore_alias_match(), so make it public. Signed-off-by:
-
John Garry authored
Add a function to decide whether a PMU is a core PMU. Signed-off-by:
-
John Garry authored
The initial test will verify that the test tables in generated pmu-events.c match against known, expected values. For known events added in pmu-events/arch/test, we need to add an entry in test_cpu_aliases_events[] or test_uncore_events[]. A sample run is as follows for x86: john@linux-3c19:~/linux> tools/perf/perf test -vv 10 10: PMU event aliases : --- start --- test child forked, pid 5316 testing event table bp_l1_btb_correct: pass testing event table bp_l2_btb_correct: pass testing event table segment_reg_loads.any: pass testing event table dispatch_blocked.any: pass testing event table eist_trans: pass testing event table uncore_hisi_ddrc.flux_wcmd: pass testing event table unc_cbo_xsnp_response.miss_eviction: pass test child finished with 0 ---- end ---- PMU event aliases: Ok Signed-off-by:
-
John Garry authored
Create pmu_add_cpu_aliases_map() from pmu_add_cpu_aliases(), so the caller can pass the map; the pmu-events test would use this since there would be no CPUID matching to a mapfile there. Signed-off-by:
-
John Garry authored
With the goal of supporting pmu-events test case, introduce support for a test events folder. These test events can be used for testing generation of pmu-event tables and alias creation for any arch. When running the pmu-events test case, these test events will be used as the platform-agnostic events, so aliases can be created per-PMU and validated against known expected values. To support the test events, add a "testcpu" entry in pmu_events_map[]. The pmu-events test will be able to lookup the events map for "testcpu", to verify the generated tables against expected values. The resultant generated pmu-events.c will now look like the following: struct pmu_event pme_ampere_emag[] = { { .name = "ldrex_spec", .event = "event=0x6c", .desc = "Exclusive operation spe...", .topic = "intrinsic", .long_desc = "Exclusive operation ...", }, ... }; struct pmu_event pme_test_cpu[] = { { .name = "uncore_hisi_ddrc.flux_wcmd", .event = "event=0x2", .desc = "DDRC write commands. Unit: hisi_sccl,ddrc ", .topic = "uncore", .long_desc = "DDRC write commands", .pmu = "hisi_sccl,ddrc", }, { .name = "unc_cbo_xsnp_response.miss_eviction", .event = "umask=0x81,event=0x22", .desc = "Unit: uncore_cbox A cross-core snoop resulted ...", .topic = "uncore", .long_desc = "A cross-core snoop resulted from L3 ...", .pmu = "uncore_cbox", }, { .name = "eist_trans", .event = "umask=0x0,period=200000,event=0x3a", .desc = "Number of Enhanced Intel SpeedStep(R) ...", .topic = "other", }, { .name = 0, }, }; struct pmu_events_map pmu_events_map[] = { ... { .cpuid = "0x00000000500f0000", .version = "v1", .type = "core", .table = pme_ampere_emag }, ... { .cpuid = "testcpu", .version = "v1", .type = "core", .table = pme_test_cpu, }, { .cpuid = 0, .version = 0, .type = 0, .table = 0, }, }; Signed-off-by: -
John Garry authored
Add some test PMU events. The events are randomly chosen from x86 and arm64 JSONs. The events include CPU and uncore events. Signed-off-by:
-
Jiri Olsa authored
Removing the extra 'SUBDIR' line from clean and doc build output. Because it's annoying.. ;-) Before: $ make clean ... SUBDIR Documentation CLEAN Documentation After: $ make clean ... CLEAN Documentation Before: $ make doc BUILD: Doing 'make -j8' parallel build SUBDIR Documentation ASCIIDOC perf-stat.html ... After: $ make doc BUILD: Doing 'make -j8' parallel build ASCIIDOC perf-stat.html ... Signed-off-by:
-
Arnaldo Carvalho de Melo authored
To get the changes in: 26776253 ("seg6: fix SRv6 L2 tunnels to use IANA-assigned protocol number") That ends up automatically adding the new IPPROTO_ETHERNET to the socket args beautifiers: $ tools/perf/trace/beauty/socket_ipproto.sh > before Apply this patch: $ tools/perf/trace/beauty/socket_ipproto.sh > after $ diff -u before after --- before 2020-03-19 11:48:36.876673819 -0300 +++ after 2020-03-19 11:49:00.148541377 -0300 @@ -6,6 +6,7 @@ [132] = "SCTP", [136] = "UDPLITE", [137] = "MPLS", + [143] = "ETHERNET", [17] = "UDP", [1] = "ICMP", [22] = "IDP", $ Addresses this tools/perf build warning: Warning: Kernel ABI header at 'tools/include/uapi/linux/in.h' differs from latest version at 'include/uapi/linux/in.h' diff -u tools/include/uapi/linux/in.h include/uapi/linux/in.h Cc: David S. Miller <davem@davemloft.net> Cc: Paolo Lungaroni <paolo.lungaroni@cnit.it> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Signed-off-by:
-
Vijay Thakkar authored
This patch updates the PMCs for AMD Zen1 core based processors (Family 17h; Models 0 through 2F) to be in accordance with PMCs as documented in the latest versions of the AMD Processor Programming Reference [1], [2] and [3]. Note that some events, such as FPU pipe assignment are missing in [1], and therefore [3] is included for full coverage of events. PMCs added: fpu_pipe_assignment.dual{0|1|2|3} fpu_pipe_assignment.total{0|1|2|3} ls_mab_alloc.dc_prefetcher ls_mab_alloc.stores ls_mab_alloc.loads bp_dyn_ind_pred bp_de_redirect PMC removed: ex_ret_cond_misp Cumulative counts, fpu_pipe_assignment.total and fpu_pipe_assignment.dual, existed in v1, but did expose port-level counters. ex_ret_cond_misp has been removed as it has been removed from the latest versions of the PPR, and when tested, always seems to sample zero as tested on a Ryzen 3400G system. [1]: Processor Programming Reference (PPR) for AMD Family 17h Models 01h,08h, Revision B2 Processors, 54945 Rev 3.03 - Jun 14, 2019. [2]: Processor Programming Reference (PPR) for AMD Family 17h Model 18h, Revision B1 Processors, 55570-B1 Rev 3.14 - Sep 26, 2019. [3]: OSRR for AMD Family 17h processors, Models 00h-2Fh, 56255 Rev 3.03 - July, 2018 All of the PPRs can be found at: https://bugzilla.kernel.org/show_bug.cgi?id=206537Signed-off-by:Vijay Thakkar <vijaythakkar@me.com> Acked-by:
Kim Phillips <kim.phillips@amd.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Jon Grimm <jon.grimm@amd.com> Cc: Martin Liška <mliska@suse.cz> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: vijay thakkar <vijaythakkar@me.com> Link: http://lore.kernel.org/lkml/20200318190002.307290-4-vijaythakkar@me.comSigned-off-by:
-
Vijay Thakkar authored
This patch adds PMU events for AMD Zen2 core based processors, namely, Matisse (model 71h), Castle Peak (model 31h) and Rome (model 2xh), as documented in the AMD Processor Programming Reference for Matisse [1]. The model number regex has been set to detect all the models under family 17 that do not match those of Zen1, as the range is larger for zen2. Zen2 adds some additional counters that are not present in Zen1 and events for them have been added in this patch. Some counters have also been removed for Zen2 thatwere previously present in Zen1 and have been confirmed to always sample zero on zen2. These added/removed counters have been omitted for brevity but can be found here: https://gist.github.com/thakkarV/5b12ca5fd7488eb2c42e451e40bdd5f3 Note that PPR for Zen2 [1] does not include some counters that were documented in the PPR for Zen1 based processors [2]. After having tested these counters, some of them that still work for zen2 systems have been preserved in the events for zen2. The counters that are omitted in [1] but are still measurable and non-zero on zen2 (tested on a Ryzen 3900X system) are the following: PMC 0x000 fpu_pipe_assignment.{total|total0|total1|total2|total3} PMC 0x004 fp_num_mov_elim_scal_op.* PMC 0x046 ls_tablewalker.* PMC 0x062 l2_latency.l2_cycles_waiting_on_fills PMC 0x063 l2_wcb_req.* PMC 0x06D l2_fill_pending.l2_fill_busy PMC 0x080 ic_fw32 PMC 0x081 ic_fw32_miss PMC 0x086 bp_snp_re_sync PMC 0x087 ic_fetch_stall.* PMC 0x08C ic_cache_inval.* PMC 0x099 bp_tlb_rel PMC 0x0C7 ex_ret_brn_resync PMC 0x28A ic_oc_mode_switch.* L3PMC 0x001 l3_request_g1.* L3PMC 0x006 l3_comb_clstr_state.* [1]: Processor Programming Reference (PPR) for AMD Family 17h Model 71h, Revision B0 Processors, 56176 Rev 3.06 - Jul 17, 2019 [2]: Processor Programming Reference (PPR) for AMD Family 17h Models 01h,08h, Revision B2 Processors, 54945 Rev 3.03 - Jun 14, 2019 All of the PPRs can be found at: https://bugzilla.kernel.org/show_bug.cgi?id=206537 Here are the results of running "fpu_pipe_assignment.total" events on my Ryzen 3900X family 17h model 71h system: Before this patch: $> perf list *fpu_pipe_assignment* List of pre-defined events (to be used in -e): After: $> perf list *fpu_pipe_assignment* floating point: fpu_pipe_assignment.total [Total number of fp uOps] fpu_pipe_assignment.total0 [Total number uOps assigned to pipe 0] fpu_pipe_assignment.total1 [Total number uOps assigned to pipe 1] fpu_pipe_assignment.total2 [Total number uOps assigned to pipe 2] fpu_pipe_assignment.total3 [Total number uOps assigned to pipe 3] Metric Groups: $> perf stat -e fpu_pipe_assignment.total sleep 1 Performance counter stats for 'sleep 1': 25,883 fpu_pipe_assignment.total 1.004145868 seconds time elapsed 0.001805000 seconds user 0.000000000 seconds sys Usage tests while running Linpackin the background: $> perf stat -I1000 -e fpu_pipe_assignment.total 1.000266796 79,313,191,516 fpu_pipe_assignment.total 2.000809630 68,091,474,430 fpu_pipe_assignment.total 3.001028115 52,925,023,174 fpu_pipe_assignment.total $> perf record -e fpu_pipe_assignment.total,fpu_pipe_assignment.total0 -a sleep 1 [ perf record: Woken up 9 times to write data ] [ perf record: Captured and wrote 4.031 MB perf.data (64764 samples) ] $> perf report --stdio --no-header | head -30 98.33% xhpl xhpl [.] dgemm_kernel 0.28% xhpl xhpl [.] dtrsm_kernel_LT 0.10% xhpl [kernel.kallsyms] [k] entry_SYSCALL_64 0.08% xhpl xhpl [.] idamax_k 0.07% baloo_file_extr liblmdb.so [.] mdb_mid2l_insert 0.06% xhpl xhpl [.] dgemm_itcopy 0.06% xhpl xhpl [.] dgemm_oncopy 0.06% xhpl [kernel.kallsyms] [k] __schedule 0.06% xhpl [kernel.kallsyms] [k] syscall_trace_enter 0.06% xhpl [kernel.kallsyms] [k] native_sched_clock 0.06% xhpl [kernel.kallsyms] [k] pick_next_task_fair 0.05% xhpl xhpl [.] blas_thread_server.llvm.15009391670273914865 0.04% xhpl [kernel.kallsyms] [k] do_syscall_64 0.04% xhpl [kernel.kallsyms] [k] yield_task_fair 0.04% xhpl libpthread-2.31.so [.] __pthread_mutex_unlock_usercnt 0.03% xhpl [kernel.kallsyms] [k] cpuacct_charge 0.03% xhpl [kernel.kallsyms] [k] syscall_return_via_sysret 0.03% xhpl libc-2.31.so [.] __sched_yield 0.03% xhpl [kernel.kallsyms] [k] __calc_delta $> perf annotate --stdio2 dgemm_kernel | egrep '^ {0,2}[0-9]+' -B2 -A2 sub $0x60,%rsp mov %rbx,(%rsp) 0.00 mov %rbp,0x8(%rsp) mov %r12,0x10(%rsp) 0.00 mov %r13,0x18(%rsp) mov %r14,0x20(%rsp) mov %r15,0x28(%rsp) -- mov %rdi,%r13 mov %rsi,0x28(%rsp) 0.00 mov %rdx,%r12 vmovsd %xmm0,0x30(%rsp) shl $0x3,%r10 mov 0x28(%rsp),%rax 0.00 xor %rdx,%rdx mov $0x18,%rdi div %rdi -- nop a0: mov %r12,%rax 0.00 shl $0x3,%rax mov %r8,%rdi lea (%r8,%rax,8),%r15 -- mov %r12,%rax nop 0.00 c0: vmovups (%rdi),%ymm1 0.09 vmovups 0x20(%rdi),%ymm2 0.02 vmovups (%r15),%ymm3 0.10 vmovups %ymm1,(%rsi) 0.07 vmovups %ymm2,0x20(%rsi) 0.07 vmovups %ymm3,0x40(%rsi) 0.06 add $0x40,%rdi add $0x40,%r15 add $0x60,%rsi 0.00 dec %rax ↑ jne c0 mov %r9,%r15 -- nop 110: lea 0x80(%rsp),%rsi 0.01 add $0x60,%rsi 0.03 mov %r12,%rax 0.00 sar $0x3,%rax cmp $0x2,%rax ↓ jl d26 prefetcht0 0x200(%rdi) 0.01 vmovups -0x60(%rsi),%ymm1 0.02 prefetcht0 0xa0(%rsi) 0.00 vbroadcastsd -0x80(%rdi),%ymm0 0.00 prefetcht0 0xe0(%rsi) 0.03 vmovups -0x40(%rsi),%ymm2 0.00 prefetcht0 0x120(%rsi) vmovups -0x20(%rsi),%ymm3 vmulpd %ymm0,%ymm1,%ymm4 0.01 prefetcht0 0x160(%rsi) vmulpd %ymm0,%ymm2,%ymm8 0.01 vmulpd %ymm0,%ymm3,%ymm12 0.02 prefetcht0 0x1a0(%rsi) 0.01 vbroadcastsd -0x78(%rdi),%ymm0 vmulpd %ymm0,%ymm1,%ymm5 0.01 vmulpd %ymm0,%ymm2,%ymm9 vmulpd %ymm0,%ymm3,%ymm13 0.01 vbroadcastsd -0x70(%rdi),%ymm0 vmulpd %ymm0,%ymm1,%ymm6 0.00 vmulpd %ymm0,%ymm2,%ymm10 0.00 add $0x60,%rsi ... snip ... nop 65e0: vmovddup -0x60(%rsi),%xmm2 0.00 vmovups -0x80(%rdi),%xmm0 vmovups -0x70(%rdi),%xmm1 0.00 vmovddup -0x58(%rsi),%xmm3 vfmadd231pd %xmm0,%xmm2,%xmm4 0.00 vfmadd231pd %xmm1,%xmm2,%xmm5 0.00 vfmadd231pd %xmm0,%xmm3,%xmm6 0.00 vfmadd231pd %xmm1,%xmm3,%xmm7 0.00 add $0x10,%rsi add $0x20,%rdi 0.00 dec %rax ↑ jne 65e0 nop nop 6620: vmovddup 0x30(%rsp),%xmm0 0.00 vmulpd %xmm0,%xmm4,%xmm4 0.00 vmulpd %xmm0,%xmm5,%xmm5 vmulpd %xmm0,%xmm6,%xmm6 vmulpd %xmm0,%xmm7,%xmm7 vaddpd (%r15),%xmm4,%xmm4 vaddpd 0x10(%r15),%xmm5,%xmm5 0.00 vaddpd (%r15,%r10,1),%xmm6,%xmm6 0.00 vaddpd 0x10(%r15,%r10,1),%xmm7,%xmm7 0.00 vmovups %xmm4,(%r15) vmovups %xmm5,0x10(%r15) 0.00 vmovups %xmm6,(%r15,%r10,1) vmovups %xmm7,0x10(%r15,%r10,1) add $0x20,%r15 -- lea (%r8,%rax,8),%r8 69d8: mov 0x20(%rsp),%r14 0.00 test $0x1,%r14 ↓ je 6d84 mov %r9,%r15 -- vbroadcastsd -0x28(%rsi),%ymm3 vfmadd231pd (%rdi),%ymm0,%ymm4 0.00 vfmadd231pd 0x20(%rdi),%ymm1,%ymm5 vfmadd231pd 0x40(%rdi),%ymm2,%ymm6 vfmadd231pd 0x60(%rdi),%ymm3,%ymm7 -- vmulpd %ymm0,%ymm4,%ymm4 vaddpd (%r15),%ymm4,%ymm4 0.00 vmovups %ymm4,(%r15) add $0x20,%r15 dec %r11 -- mov %rbx,%rsp mov (%rsp),%rbx 0.01 mov 0x8(%rsp),%rbp mov 0x10(%rsp),%r12 mov 0x18(%rsp),%r13 Signed-off-by:
-
Vijay Thakkar authored
This patch changes the previous blanket detection of AMD Family 17h processors to be more specific to Zen1 core based products only by replacing model detection regex pattern [[:xdigit:]]+ with ([12][0-9A-F]|[0-9A-F]), restricting to models 0 though 2f only. This change is required to allow for the addition of separate PMU events for Zen2 core based models in the following patches as those belong to family 17h but have different PMCs. Current PMU events directory has also been renamed to "amdzen1" from "amdfam17h" to reflect this specificity. Note that although this change does not break PMU counters for existing zen1 based systems, it does disable the current set of counters for zen2 based systems. Counters for zen2 have been added in the following patches in this patchset. Signed-off-by:
-
Kajol Jain authored
perf metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events Commit f01642e4 ("perf metricgroup: Support multiple events for metricgroup") introduced support for multiple events in a metric group. But with the current upstream, metric events names are not printed properly incase we try to run multiple metric groups with overlapping event. With current upstream version, incase of overlapping metric events issue is, we always start our comparision logic from start. So, the events which already matched with some metric group also take part in comparision logic. Because of that when we have overlapping events, we end up matching current metric group event with already matched one. For example, in skylake machine we have metric event CoreIPC and Instructions. Both of them need 'inst_retired.any' event value. As events in Instructions is subset of events in CoreIPC, they endup in pointing to same 'inst_retired.any' value. In skylake platform: command:# ./perf stat -M CoreIPC,Instructions -C 0 sleep 1 Performance counter stats for 'CPU(s) 0': 1,254,992,790 inst_retired.any # 1254992790.0 Instructions # 1.3 CoreIPC 977,172,805 cycles 1,254,992,756 inst_retired.any 1.000802596 seconds time elapsed command:# sudo ./perf stat -M UPI,IPC sleep 1 Performance counter stats for 'sleep 1': 948,650 uops_retired.retire_slots 866,182 inst_retired.any # 0.7 IPC 866,182 inst_retired.any 1,175,671 cpu_clk_unhalted.thread Patch fixes the issue by adding a new bool pointer 'evlist_used' to keep track of events which already matched with some group by setting it true. So, we skip all used events in list when we start comparision logic. Patch also make some changes in comparision logic, incase we get a match miss, we discard the whole match and start again with first event id in metric event. With this patch: In skylake platform: command:# ./perf stat -M CoreIPC,Instructions -C 0 sleep 1 Performance counter stats for 'CPU(s) 0': 3,348,415 inst_retired.any # 0.3 CoreIPC 11,779,026 cycles 3,348,381 inst_retired.any # 3348381.0 Instructions 1.001649056 seconds time elapsed command:# ./perf stat -M UPI,IPC sleep 1 Performance counter stats for 'sleep 1': 1,023,148 uops_retired.retire_slots # 1.1 UPI 924,976 inst_retired.any 924,976 inst_retired.any # 0.6 IPC 1,489,414 cpu_clk_unhalted.thread 1.003064672 seconds time elapsed Signed-off-by:
Kajol Jain <kjain@linux.ibm.com> Acked-by:
-
Jin Yao authored
There is a slight misalignment in -A -I output. For example: # perf stat -e cpu/event=cpu-cycles/ -a -A -I 1000 # time CPU counts unit events 1.000440863 CPU0 1,068,388 cpu/event=cpu-cycles/ 1.000440863 CPU1 875,954 cpu/event=cpu-cycles/ 1.000440863 CPU2 3,072,538 cpu/event=cpu-cycles/ 1.000440863 CPU3 4,026,870 cpu/event=cpu-cycles/ 1.000440863 CPU4 5,919,630 cpu/event=cpu-cycles/ 1.000440863 CPU5 2,714,260 cpu/event=cpu-cycles/ 1.000440863 CPU6 2,219,240 cpu/event=cpu-cycles/ 1.000440863 CPU7 1,299,232 cpu/event=cpu-cycles/ The value of counts is not aligned with the column "counts" and the event name is not aligned with the column "events". With this patch, the output is, # perf stat -e cpu/event=cpu-cycles/ -a -A -I 1000 # time CPU counts unit events 1.000423009 CPU0 997,421 cpu/event=cpu-cycles/ 1.000423009 CPU1 1,422,042 cpu/event=cpu-cycles/ 1.000423009 CPU2 484,651 cpu/event=cpu-cycles/ 1.000423009 CPU3 525,791 cpu/event=cpu-cycles/ 1.000423009 CPU4 1,370,100 cpu/event=cpu-cycles/ 1.000423009 CPU5 442,072 cpu/event=cpu-cycles/ 1.000423009 CPU6 205,643 cpu/event=cpu-cycles/ 1.000423009 CPU7 1,302,250 cpu/event=cpu-cycles/ Now output is aligned. Signed-off-by:
-
