- 24 Nov, 2022 13 commits
-
-
Namhyung Kim authored
It generated a broken JSON output when aggregation mode or cgroup is used with --metric-only option. Also get rid of the header line and make the output single line for each entry. It needs to know whether the current metric is the first one or not. So add 'first' field in the outstate and mark it false after printing. Before: # perf stat -a -j --metric-only true {"unit" : "GHz"}{"unit" : "insn per cycle"}{"unit" : "branch-misses of all branches"} {{"metric-value" : "0.797"}{"metric-value" : "1.65"}{"metric-value" : "0.89"} ^ # perf stat -a -j --metric-only --per-socket true {"unit" : "GHz"}{"unit" : "insn per cycle"}{"unit" : "branch-misses of all branches"} {"socket" : "S0", "aggregate-number" : 8, {"metric-value" : "0.295"}{"metric-value" : "1.88"}{"metric-value" : "0.64"} ^ After: # perf stat -a -j --metric-only true {"GHz" : "0.990", "insn per cycle" : "2.06", "branch-misses of all branches" : "0.59"} # perf stat -a -j --metric-only --per-socket true {"socket" : "S0", "aggregate-number" : 8, "GHz" : "0.439", "insn per cycle" : "2.14", "branch-misses of all branches" : "0.51"} Signed-off-by:Namhyung Kim <namhyung@kernel.org> Acked-by:
Ian Rogers <irogers@google.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Athira Jajeev <atrajeev@linux.vnet.ibm.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: James Clark <james.clark@arm.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Kan Liang <kan.liang@linux.intel.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com> Link: https://lore.kernel.org/r/20221123180208.2068936-14-namhyung@kernel.orgSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Namhyung Kim authored
Now most of the print functions take a pointer to the struct outstate. We have one in the evlist__print_counters() and pass it through the child functions. Signed-off-by:
-
Namhyung Kim authored
It always passes a pointer to rt_stat as it's the only one. Let's not pass it and directly refer it in the printout(). Signed-off-by:
-
Namhyung Kim authored
The printout() takes a lot of arguments and sets an outstate with the value. Instead, we can fill the outstate first and then pass it to reduce the number of arguments. Signed-off-by:
-
Namhyung Kim authored
It passes prefix and cgroup pointers but the outstate already has them. Let's pass the outstate pointer instead. Signed-off-by:
-
Namhyung Kim authored
This is a preparation for the later cleanup. No functional changes intended. Signed-off-by:
-
Namhyung Kim authored
This is a minor cleanup and preparation for the later change. Signed-off-by:
-
Namhyung Kim authored
It already passes the stat_config argument, then it can find the value in the config. No need to pass it separately. Signed-off-by:
-
Namhyung Kim authored
It always passes a whitespace to the function, thus we can just add it to the function body. Furthermore, it's only used in the normal output mode. Well, actually CSV used it but it doesn't need to since we don't care about the indentation or alignment in the CSV output. Signed-off-by:
-
Namhyung Kim authored
It should not use sprintf() anymore. Let's pass the buffer size and use the safer scnprintf() instead. Signed-off-by:
-
Namhyung Kim authored
We don't care about the alignment in the CSV output as it's intended for machine processing. Let's get rid of it to make the output more compact. Before: # perf stat -a --summary -I 1 -x, true 0.001149309,219.20,msec,cpu-clock,219322251,100.00,219.200,CPUs utilized 0.001149309,144,,context-switches,219241902,100.00,656.935,/sec 0.001149309,38,,cpu-migrations,219173705,100.00,173.358,/sec 0.001149309,61,,page-faults,219093635,100.00,278.285,/sec 0.001149309,10679310,,cycles,218746228,100.00,0.049,GHz 0.001149309,6288296,,instructions,218589869,100.00,0.59,insn per cycle 0.001149309,1386904,,branches,218428851,100.00,6.327,M/sec 0.001149309,56863,,branch-misses,218219951,100.00,4.10,of all branches summary,219.20,msec,cpu-clock,219322251,100.00,20.025,CPUs utilized summary,144,,context-switches,219241902,100.00,656.935,/sec summary,38,,cpu-migrations,219173705,100.00,173.358,/sec summary,61,,page-faults,219093635,100.00,278.285,/sec summary,10679310,,cycles,218746228,100.00,0.049,GHz summary,6288296,,instructions,218589869,100.00,0.59,insn per cycle summary,1386904,,branches,218428851,100.00,6.327,M/sec summary,56863,,branch-misses,218219951,100.00,4.10,of all branches After: 0.001148449,224.75,msec,cpu-clock,224870589,100.00,224.747,CPUs utilized 0.001148449,176,,context-switches,224775564,100.00,783.103,/sec 0.001148449,38,,cpu-migrations,224707428,100.00,169.079,/sec 0.001148449,61,,page-faults,224629326,100.00,271.416,/sec 0.001148449,12172071,,cycles,224266368,100.00,0.054,GHz 0.001148449,6901907,,instructions,224108764,100.00,0.57,insn per cycle 0.001148449,1515655,,branches,223946693,100.00,6.744,M/sec 0.001148449,70027,,branch-misses,223735385,100.00,4.62,of all branches summary,224.75,msec,cpu-clock,224870589,100.00,21.066,CPUs utilized summary,176,,context-switches,224775564,100.00,783.103,/sec summary,38,,cpu-migrations,224707428,100.00,169.079,/sec summary,61,,page-faults,224629326,100.00,271.416,/sec summary,12172071,,cycles,224266368,100.00,0.054,GHz summary,6901907,,instructions,224108764,100.00,0.57,insn per cycle summary,1515655,,branches,223946693,100.00,6.744,M/sec summary,70027,,branch-misses,223735385,100.00,4.62,of all branches Signed-off-by: -
Namhyung Kim authored
It matches to the prefix (interval timestamp), so better to have them together. No functional change intended. Signed-off-by:
-
Namhyung Kim authored
It missed the 'else' keyword after checking json output mode. Fixes: 41cb8752 ("perf stat: Split print_cgroup() function") Signed-off-by:
-
- 23 Nov, 2022 21 commits
-
-
Namhyung Kim authored
It caused some troubles when a lock inside kmalloc is contended because task local storage would allocate memory using kmalloc. It'd create a recusion and even crash in my system. There could be a couple of workarounds but I think the simplest one is to use a pre-allocated hash map. We could fix the task local storage to use the safe BPF allocator, but it takes time so let's change this until it happens actually. Signed-off-by:
Martin KaFai Lau <martin.lau@kernel.org> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Blake Jones <blakejones@google.com> Cc: Chris Li <chriscli@google.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Song Liu <song@kernel.org> Cc: bpf@vger.kernel.org Link: https://lore.kernel.org/r/20221118190109.1512674-1-namhyung@kernel.orgSigned-off-by:
-
John Garry authored
Update my address. Signed-off-by:
John Garry <john.g.garry@oracle.com> Acked-by:
Will Deacon <will@kernel.org> Cc: Ian Rogers <irogers@google.com> Link: https://lore.kernel.org/r/20221121113018.1899426-1-john.g.garry@oracle.comSigned-off-by:
-
Michael Petlan authored
Using precise flag with br_inst_retired.near_call causes the test fail on KVM guests, even when the guests have PMU forwarding enabled and the event itself is supported. Remove the precise flag in order to make the test work on KVM guests. Signed-off-by:
Michael Petlan <mpetlan@redhat.com> Acked-by:
-
Namhyung Kim authored
With perf inject -b, it synthesizes build-id event for DSOs. But it missed to set the size and resulted in having trailing zeros. As perf record sets the size in write_build_id(), let's set the size here as well. Signed-off-by:
Adrian Hunter <adrian.hunter@intel.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: https://lore.kernel.org/r/20221119002750.1568027-1-namhyung@kernel.orgSigned-off-by:
-
Naveen N. Rao authored
On IBM Power9, perf watchpoint tests fail since no hardware breakpoints are available. Detect this by checking the error returned by perf_event_open() and skip the tests in that case. Reported-by:
Disha Goel <disgoel@linux.vnet.ibm.com> Signed-off-by:
Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com> Acked-by:
-
Leo Yan authored
augmented_raw_syscalls.c defines the bpf map 'syscalls' which is initialized by perf tool in user space to indicate which system calls are enabled for tracing, on the other flip eBPF program relies on the map to filter out the trace events which are not enabled. The map also includes a field 'string_args_len[6]' which presents the string length if the corresponding argument is a string type. Now the map 'syscalls' is not used, bpf program doesn't use it as filter anymore, this is replaced by using the function bpf_tail_call() and PROG_ARRAY syscalls map. And we don't need to explicitly set the string length anymore, bpf_probe_read_str() is smart to copy the string and return string length. Therefore, it's safe to remove the bpf map 'syscalls'. To consolidate the code, this patch removes the definition of map 'syscalls' from augmented_raw_syscalls.c and drops code for using the map in the perf trace. Note, since function trace__set_ev_qualifier_bpf_filter() is removed, calling trace__init_syscall_bpf_progs() from it is also removed. We don't need to worry it because trace__init_syscall_bpf_progs() is still invoked from trace__init_syscalls_bpf_prog_array_maps() for initialization the system call's bpf program callback. After: # perf trace -e examples/bpf/augmented_raw_syscalls.c,open* --max-events 10 perf stat --quiet sleep 0.001 openat(AT_FDCWD, "/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libm.so.6", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libelf.so.1", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libdw.so.1", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libunwind.so.8", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libunwind-aarch64.so.8", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libcrypto.so.3", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libslang.so.2", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libperl.so.5.34", O_RDONLY|O_CLOEXEC) = 3 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libc.so.6", O_RDONLY|O_CLOEXEC) = 3 # perf trace -e examples/bpf/augmented_raw_syscalls.c --max-events 10 perf stat --quiet sleep 0.001 ... [continued]: execve()) = 0 brk(NULL) = 0xaaaab1d28000 faccessat(-100, "/etc/ld.so.preload", 4) = -1 ENOENT (No such file or directory) openat(AT_FDCWD, "/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3 close(3</usr/lib/aarch64-linux-gnu/libcrypto.so.3>) = 0 openat(AT_FDCWD, "/lib/aarch64-linux-gnu/libm.so.6", O_RDONLY|O_CLOEXEC) = 3 read(3</usr/lib/aarch64-linux-gnu/libcrypto.so.3>, 0xfffff33f70d0, 832) = 832 munmap(0xffffb5519000, 28672) = 0 munmap(0xffffb55b7000, 32880) = 0 mprotect(0xffffb55a6000, 61440, PROT_NONE) = 0 Signed-off-by:
Leo Yan <leo.yan@linaro.org> Tested-by:
-
Leo Yan authored
The local variable 'syscall' is not used anymore, remove it. Signed-off-by:
-
Leo Yan authored
On Arm64 a case is perf tools fails to find the corresponding trace point folder for system calls listed in the table 'syscalltbl_arm64', e.g. the generated system call table contains "lookup_dcookie" but we cannot find out the matched trace point folder for it. We need to figure out if there have any issue for the generated system call table, on the other hand, we need to handle the case when trace point folder is missed under sysfs, this patch sets the flag syscall::nonexistent as true and returns the error from trace__read_syscall_info(). Another problem is for trace__syscall_info(), it returns two different values if a system call doesn't exist: at the first time calling trace__syscall_info() it returns NULL when the system call doesn't exist, later if call trace__syscall_info() again for the same missed system call, it returns pointer of syscall. trace__syscall_info() checks the condition 'syscalls.table[id].name == NULL', but the name will be assigned in the first invoking even the system call is not found. So checking system call's name in trace__syscall_info() is not the right thing to do, this patch simply checks flag syscall::nonexistent to make decision if a system call exists or not, finally trace__syscall_info() returns the consistent result (NULL) if a system call doesn't existed. Fixes: b8b1033f ("perf trace: Mark syscall ids that are not allocated to avoid unnecessary error messages") Signed-off-by:
-
Leo Yan authored
When a system call is not detected, the reason is either because the system call ID is out of scope or failure to find the corresponding path in the sysfs, trace__read_syscall_info() returns zero. Finally, without returning an error value it introduces confusion for the caller. This patch lets the function trace__read_syscall_info() to return -EEXIST when a system call doesn't exist. Fixes: b8b1033f ("perf trace: Mark syscall ids that are not allocated to avoid unnecessary error messages") Signed-off-by:
-
Leo Yan authored
This patch defines a macro RAW_SYSCALL_ARGS_NUM to replace the open coded number '6'. Signed-off-by:
-
Ian Rogers authored
Output events and metrics in a JSON format by overriding the print callbacks. Currently other command line options aren't supported and metrics are repeated once per metric group. Committer testing: $ perf list cache List of pre-defined events (to be used in -e or -M): L1-dcache-load-misses [Hardware cache event] L1-dcache-loads [Hardware cache event] L1-dcache-prefetches [Hardware cache event] L1-icache-load-misses [Hardware cache event] L1-icache-loads [Hardware cache event] branch-load-misses [Hardware cache event] branch-loads [Hardware cache event] dTLB-load-misses [Hardware cache event] dTLB-loads [Hardware cache event] iTLB-load-misses [Hardware cache event] iTLB-loads [Hardware cache event] $ perf list --json cache [ { "Unit": "cache", "EventName": "L1-dcache-load-misses", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "L1-dcache-loads", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "L1-dcache-prefetches", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "L1-icache-load-misses", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "L1-icache-loads", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "branch-load-misses", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "branch-loads", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "dTLB-load-misses", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "dTLB-loads", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "iTLB-load-misses", "EventType": "Hardware cache event" }, { "Unit": "cache", "EventName": "iTLB-loads", "EventType": "Hardware cache event" } ] $ Signed-off-by: -
Ian Rogers authored
Rather than controlling the list output with passed flags, add callbacks that are called when an event or metric are encountered. State is passed to the callback so that command line options can be respected, alternatively the callbacks can be changed. Fix a few bugs: - wordwrap to columns metric descriptions and expressions; - remove unnecessary whitespace after PMU event names; - the metric filter is a glob but matched using strstr which will always fail, switch to using a proper globmatch, - the detail flag gives details for extra kernel PMU events like branch-instructions. In metricgroup.c switch from struct mep being a rbtree of metricgroups containing a list of metrics, to the tree directly containing all the metrics. In general the alias for a name is passed to the print routine rather than being contained in the name with OR. Committer notes: Check the asprint() return to address this on fedora 36: util/print-events.c: In function ‘print_sdt_events’: util/print-events.c:183:33: error: ignoring return value of ‘asprintf’ declared with attribute ‘warn_unused_result’ [-Werror=unused-result] 183 | asprintf(&evt_name, "%s@%s(%.12s)", sdt_name->s, path, bid); | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ cc1: all warnings being treated as errors $ gcc --version | head -1 gcc (GCC) 12.2.1 20220819 (Red Hat 12.2.1-2) $ Fix ps.pmu_glob setting when dealing with *:* events, it was being left with a freed pointer that then at the end of cmd_list() would be double freed. Check if pmu_name is NULL in default_print_event() before calling strglobmatch(pmu_name, ...) to avoid a segfault. Signed-off-by: -
Ian Rogers authored
The tools/lib includes fixes break LIBTRACEVENT_DYNAMIC as the makefile erroneously had dependencies on building libtraceevent even when not linking with it. This change fixes the issues with LIBTRACEEVENT_DYNAMIC by making the built files optional. Signed-off-by:
-
Namhyung Kim authored
So that it can get rid of requirement of a compiler. $ sudo ./perf test -v 109 109: Test data symbol : --- start --- test child forked, pid 844526 Recording workload... [ perf record: Woken up 2 times to write data ] [ perf record: Captured and wrote 0.354 MB /tmp/__perf_test.perf.data.GFeZO (4847 samples) ] Cleaning up files... test child finished with 0 ---- end ---- Test data symbol: Ok Signed-off-by:
James Clark <james.clark@arm.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Athira Jajeev <atrajeev@linux.vnet.ibm.com> Cc: German Gomez <german.gomez@arm.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Leo Yan <leo.yan@linaro.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Zhengjun Xing <zhengjun.xing@linux.intel.com> Link: https://lore.kernel.org/r/20221116233854.1596378-13-namhyung@kernel.orgSigned-off-by:
-
Namhyung Kim authored
The datasym workload is to check if perf mem command gets the data addresses precisely. This is needed for data symbol test. $ perf test -w datasym I had to keep the buf1 in the data section, otherwise it could end up in the BSS and was mmaped as a separate //anon region, then it was not symbolized at all. It needs to be fixed separately. Committer notes: Add a -U _FORTIFY_SOURCE to the datasym CFLAGS, as the main perf flags set it and it requires building with optimization, and this new test has a -O0. Signed-off-by:
-
Namhyung Kim authored
So that it can get rid of requirement of a compiler. Also rename the symbols to match with the perf test workload. Signed-off-by:
German Gomez <german.gomez@arm.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Athira Jajeev <atrajeev@linux.vnet.ibm.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Leo Yan <leo.yan@linaro.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Zhengjun Xing <zhengjun.xing@linux.intel.com> Link: https://lore.kernel.org/r/20221116233854.1596378-11-namhyung@kernel.orgSigned-off-by:
-
Namhyung Kim authored
The brstack is to run different kinds of branches repeatedly. This is necessary for brstack test case to verify if it has correct branch info. $ perf test -w brstack I renamed the internal functions to have brstack_ prefix as it's too generic name. Add a -U_FORTIFY_SOURCE to the brstack CFLAGS, as the main perf flags set it and it requires building with optimization, and this new test has a -O0. Signed-off-by:
-
Namhyung Kim authored
So that it can get rid of requirement of a compiler. I've also removed killall as it'll kill perf process now and run the test workload for 10 sec instead. Signed-off-by:
-
Namhyung Kim authored
The sqrtloop creates a child process to run an infinite loop calling sqrt() with rand(). This is needed for ARM SPE fork test. $ perf test -w sqrtloop It can take an optional argument to specify how long it will run in seconds (default: 1). Committer notes: Explicitely ignored the sqrt() return to fix the build on systems where the compiler complains it isn't being used. And added a sqrtloop specific CFLAGS to disable optimizations to make this a bit more robust wrt dead code elimination. Doing that a -U_FORTIFY_SOURCE needs to be added, as -O0 is incompatible with it. Signed-off-by:
-
Namhyung Kim authored
So that it can get rid of requirement of a compiler. Reviewed-by:
-
Namhyung Kim authored
The leafloop workload is to run an infinite loop in the test_leaf function. This is needed for the ARM fp callgraph test to verify if it gets the correct callchains. $ perf test -w leafloop Committer notes: Add a: -U_FORTIFY_SOURCE to the leafloop CFLAGS as the main perf flags set it and it requires building with optimization, and this new test has a -O0. Signed-off-by:
-
- 20 Nov, 2022 6 commits
-
-
Namhyung Kim authored
So that it can get rid of requirements for a compiler. $ sudo ./perf test -v 92 92: perf record tests : --- start --- test child forked, pid 740204 Basic --per-thread mode test Basic --per-thread mode test [Success] Register capture test Register capture test [Success] Basic --system-wide mode test Basic --system-wide mode test [Success] Basic target workload test Basic target workload test [Success] test child finished with 0 ---- end ---- perf record tests: Ok Signed-off-by:
-
Namhyung Kim authored
The thloop is similar to noploop but runs in two threads. This is needed to verify perf record --per-thread to handle multi-threaded programs properly. $ perf test -w thloop It also takes an optional argument to specify runtime in seconds (default: 1). Signed-off-by:
-
Namhyung Kim authored
So that it can get rid of requirement of a compiler. Also define and use more local symbols to ease future changes. $ sudo ./perf test -v pipe 87: perf pipe recording and injection test : --- start --- test child forked, pid 748003 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.000 MB - ] 748014 748014 -1 |perf [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.000 MB - ] 99.83% perf perf [.] noploop [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.000 MB - ] 99.85% perf perf [.] noploop [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.160 MB /tmp/perf.data.2XYPdw (4007 samples) ] 99.83% perf perf [.] noploop test child finished with 0 ---- end ---- perf pipe recording and injection test: Ok Signed-off-by: -
Namhyung Kim authored
The -w/--workload option is to run a simple workload used by testing. This adds a basic framework to run the workloads and 'noploop' workload as an example. $ perf test -w noploop The noploop does a loop doing nothing (NOP) for a second by default. It can have an optional argument to specify the time in seconds. Signed-off-by:
-
Ian Rogers authored
Add libtraceevent to the name so that this install_headers build appears different to similar targets in different libraries. Add ; after kbuffer.h install target for consistency. Signed-off-by:
-
Ian Rogers authored
Add libsubcmd to the name so that this install_headers build appears different to similar targets in different libraries. Signed-off-by:
-
