- 31 Oct, 2022 4 commits
-
-
Arnaldo Carvalho de Melo authored
Leave just some forward declarations for pointers, move the includes to where they are really needed. Signed-off-by:Arnaldo Carvalho de Melo <acme@redhat.com>
-
Arnaldo Carvalho de Melo authored
Some places were including event.h just to get 'struct perf_sample', move it to a separate place so that we speed up a bit the build. Signed-off-by: -
Arnaldo Carvalho de Melo authored
map_symbol.h is needed because we have structs that contains 'struct addr_map_symbol', so add it, remove the others. Signed-off-by: -
Arnaldo Carvalho de Melo authored
In the bpf-prologue.h header we are just using pointers, so no need to include headers for that, just provide forward declarations for those types. Signed-off-by:
-
- 27 Oct, 2022 36 commits
-
-
Arnaldo Carvalho de Melo authored
Signed-off-by: -
James Clark authored
Use the global quiet variable everywhere so that all tools hide warnings in quiet mode and update the documentation to reflect this. 'perf probe' claimed that errors are not printed in quiet mode but I don't see this so remove it from the docs. Signed-off-by:
James Clark <james.clark@arm.com> Cc: Adrian Hunter <adrian.hunter@intel.com> 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: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: https://lore.kernel.org/r/20221018094137.783081-3-james.clark@arm.comSigned-off-by:
-
James Clark authored
Especially when CONFIG_LOCKDEP and other debug configs are enabled, Perf can print the following warning when running the "kernel lock contention analysis" test: Warning: Processed 1378918 events and lost 4 chunks! Check IO/CPU overload! Warning: Processed 4593325 samples and lost 70.00%! The test already supplies -q to run in quiet mode, so extend quiet mode to perf_stdio__warning() and also ui__warning() for consistency. This fixes the following failure due to the extra lines counted: perf test "lock cont" -vvv 82: kernel lock contention analysis test : --- start --- test child forked, pid 3125 Testing perf lock record and perf lock contention [Fail] Recorded result count is not 1: 9 test child finished with -1 ---- end ---- kernel lock contention analysis test: FAILED! Fixes: ec685de2 ("perf test: Add kernel lock contention test") Signed-off-by:
-
Namhyung Kim authored
It now has 4 sub tests and at least one of them should run. But once the TEST_SKIP (= 2) return value is set, it won't be overwritten unless there's a failure. I think we should return success when one or more tests are skipped but the remaining subtests are passed. So update the test code not to set the err variable when it skips the test. Reviewed-by:
Adrian Hunter <adrian.hunter@intel.com> Signed-off-by:
Namhyung Kim <namhyung@kernel.org> Acked-by:
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/20221020172643.3458767-9-namhyung@kernel.orgSigned-off-by:
-
Namhyung Kim authored
The --threads option changed the 'perf record' behavior significantly, so it'd be nice if we test it separately. Add --threads options with different argument in each test supported and check the result. Also update the cleanup routine because threads recording produces data in a directory. Reviewed-by:
-
Namhyung Kim authored
Add a subtest which profiles the given workload on the command line. As it's a minimal requirement, the test should run ok so it doesn't skip the test even if it failed to run the 'perf record' command. Reviewed-by:
-
Namhyung Kim authored
Add system wide recording test with the same pattern. It'd skip the test when it fails to run 'perf record'. For system-wide mode, it needs to avoid build-id collection and synthesis because the test only cares about the test program and kernel would generate the necessary events as the process starts. Reviewed-by:
-
Namhyung Kim authored
Just running the target program is not enough to test multi-thread target because it'd be racy perf vs target startup. I used the initial delay but it cannot guarantee for perf to see the thread. Instead, use wait_for_threads helper from shell/lib/waiting.sh to make sure it starts the sibling thread first. Then perf record can use -p option to profile the target process. Reviewed-by:
-
Namhyung Kim authored
If the system has cc it could build a test program with two threads and then use it for more detailed testing. Also it accepts an option to run a thread forever to ensure multi-thread runs. If cc is not found, it falls back to use the default value 'true'. Reviewed-by:
-
Namhyung Kim authored
Basically there are 3 issues: 1. quote shell expansion 2. do not use egrep 3. use upper case letters for signal names Reviewed-by:
-
Namhyung Kim authored
I think it's to support non-root user tests. But perf record can handle the case and fall back to a software event (cpu-clock). Practically this would affect when it's run on a VM, but it seems no reason to prevent running the test in the guest. Reviewed-by:
-
Adrian Hunter authored
Intel PT timestamps are not provided for every branch, let alone every instruction, so there can be many samples with the same timestamp. With per-cpu contexts, decoding is done for each CPU in turn, which can make it difficult to see what is happening on different CPUs at the same time. Currently the interleaving from perf script --itrace=i0ns is quite coarse grained. There are often long stretches executing on one CPU and nothing on another. Some people are interested in seeing what happened on multiple CPUs before a crash to debug races etc. To improve perf script interleaving for parallel execution, the intel-pt-events.py script has been enhanced to enable interleaving the output with the same timestamp from different CPUs. It is understood that interleaving is not perfect or causal. Add parameter --interleave [<n>] to interleave sample output for the same timestamp so that no more than n samples for a CPU are displayed in a row. 'n' defaults to 4. Note this only affects the order of output, and only when the timestamp is the same. Example: $ perf script intel-pt-events.py --insn-trace --interleave 3 ... bash 2267/2267 [004] 9323.692625625 563caa3c86f0 jz 0x563caa3c89c7 run_pending_traps+0x30 (/usr/bin/bash) IPC: 1.52 (38/25) bash 2267/2267 [004] 9323.692625625 563caa3c89c7 movq 0x118(%rsp), %rax run_pending_traps+0x307 (/usr/bin/bash) bash 2267/2267 [004] 9323.692625625 563caa3c89cf subq %fs:0x28, %rax run_pending_traps+0x30f (/usr/bin/bash) bash 2270/2270 [007] 9323.692625625 55dc58cabf02 jz 0x55dc58cabf48 unquoted_glob_pattern_p+0x102 (/usr/bin/bash) IPC: 1.56 (25/16) bash 2270/2270 [007] 9323.692625625 55dc58cabf04 cmp $0x5d, %al unquoted_glob_pattern_p+0x104 (/usr/bin/bash) bash 2270/2270 [007] 9323.692625625 55dc58cabf06 jnz 0x55dc58cabf10 unquoted_glob_pattern_p+0x106 (/usr/bin/bash) bash 2264/2264 [001] 9323.692625625 7fd556a4376c jbe 0x7fd556a43ac8 round_and_return+0x3fc (/usr/lib/x86_64-linux-gnu/libc.so.6) IPC: 4.30 (43/10) bash 2264/2264 [001] 9323.692625625 7fd556a43772 and $0x8, %edx round_and_return+0x402 (/usr/lib/x86_64-linux-gnu/libc.so.6) bash 2264/2264 [001] 9323.692625625 7fd556a43775 jnz 0x7fd556a43ac8 round_and_return+0x405 (/usr/lib/x86_64-linux-gnu/libc.so.6) bash 2267/2267 [004] 9323.692625625 563caa3c89d8 jnz 0x563caa3c8b11 run_pending_traps+0x318 (/usr/bin/bash) bash 2267/2267 [004] 9323.692625625 563caa3c89de add $0x128, %rsp run_pending_traps+0x31e (/usr/bin/bash) bash 2267/2267 [004] 9323.692625625 563caa3c89e5 popq %rbx run_pending_traps+0x325 (/usr/bin/bash) ... Signed-off-by:
-
Arnaldo Carvalho de Melo authored
Since commit c8978997 ("perf tools: Prevent out-of-bounds access to registers") the util/event.h header doesn't use anything from util/perf_regs.h, so drop it to untangle the header dependency tree a bit, speeding up compilation. Signed-off-by:
-
Arnaldo Carvalho de Melo authored
It was getting it via event.h, that doesn't need that include anymore and will drop it. Signed-off-by: -
Arnaldo Carvalho de Melo authored
It was getting indirectly, out of luck, add it. Signed-off-by: -
Arnaldo Carvalho de Melo authored
To reduce compile time and header dependency chains just add forward declarations for pointer types and include linux/types.h for u64. Signed-off-by: -
Yang Jihong authored
Add missing newline on pr_warning() call in home_perfconfig(). Before: # perf record File /home/yangjihong/.perfconfig not owned by current user or root, ignoring it.Couldn't synthesize bpf events. After: # perf record File /home/yangjihong/.perfconfig not owned by current user or root, ignoring it. Couldn't synthesize bpf events. Signed-off-by:
Yang Jihong <yangjihong1@huawei.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: https://lore.kernel.org/r/20221022092735.114967-4-yangjihong1@huawei.comSigned-off-by:
-
Yang Jihong authored
perf daemon supports start, signal, stop and ping subcommands, complete it Before: # perf daemon -h Usage: perf daemon start [<options>] or: perf daemon [<options>] -v, --verbose be more verbose -x, --field-separator[=<field separator>] print counts with custom separator --base <directory> base directory --config <config file> config file path After: # perf daemon -h Usage: perf daemon {start|signal|stop|ping} [<options>] or: perf daemon [<options>] -v, --verbose be more verbose -x, --field-separator[=<field separator>] print counts with custom separator --base <directory> base directory --config <config file> config file path Signed-off-by: -
Namhyung Kim authored
The aggr field in the struct perf_counts is to keep the aggregated value in the AGGR_GLOBAL for the old code. But it's not used anymore. Signed-off-by:
-
Namhyung Kim authored
The recent change in the perf stat broke the percore event display. Note that the aggr counts are already processed so that the every sibling thread in the same core will get the per-core counter values. Check percore evsels and skip the sibling threads in the display. Signed-off-by:
-
Namhyung Kim authored
Now aggr counts are ready for use. Convert the display routines to use the aggr counts and update the shadow stat with them. It doesn't need to aggregate counts or collect aliases anymore during the display. Get rid of now unused struct perf_aggr_thread_value. Note that there's a difference in the display order among the aggr mode. For per-core/die/socket/node aggregation, it shows relevant events in the same unit together, whereas global/thread/no aggregation it shows the same events for different units together. So it still uses separate codes to display them due to the ordering. One more thing to note is that it breaks per-core event display for now. The next patch will fix it to have identical output as of now. Signed-off-by:
-
Namhyung Kim authored
This function updates the shadow stats using the aggregated counts uniformly since it uses the aggr_counts for the every aggr mode. It'd have duplicate shadow stats for each items for now since the display routines will update them once again. But that'd be fine as it shows the average values and it'd be gone eventually. Signed-off-by:
-
Namhyung Kim authored
The perf_stat_process_percore() is to aggregate counts for an event per-core even if the aggr_mode is AGGR_NONE. This is enabled when user requested it on the command line. To handle that, it keeps the per-cpu counts at first. And then it aggregates the counts that have the same core id in the aggr->counts and updates the values for each cpu back. Later, per-core events will skip one of the CPUs unless percore-show-thread option is given. In that case, it can simply print all cpu stats with the updated (per-core) values. Signed-off-by:
-
Namhyung Kim authored
The perf_stat_merge_counters() is to aggregate the same events in different PMUs like in case of uncore or hybrid. The same logic is in the stat-display routines but I think it should be handled when it processes the event counters. As it works on the aggr_counters, it doesn't change the output yet. Signed-off-by:
-
Namhyung Kim authored
It'd do more processing with aggregation. Let's split the function so that it can be shared with by process_stat_round_event() too. Signed-off-by:
-
Namhyung Kim authored
The evsel->stats->aggr->count should be reset for interval processing since we want to use the values directly for display. Signed-off-by:
-
Namhyung Kim authored
In the process_stat_config_event() it sets the aggr_mode that means the earlier evlist__alloc_stats() cannot allocate the aggr counts due to the missing aggr_mode. Do it after setting the aggr_map using evlist__alloc_aggr_stats(). Signed-off-by:
-
Namhyung Kim authored
Per-thread aggregation doesn't use the CPU numbers but the logic should be the same. Initialize cpu_aggr_map separately for AGGR_THREAD and use thread map idx to aggregate counter values. Signed-off-by:
-
Namhyung Kim authored
It's possible to have 0 enabled/running time for some per-task or per-cgroup events since it's not scheduled on any CPU. Treating the whole event as failed would not work in this case. Thinking again, the code only existed when any CPU-level aggregation is enabled (like per-socket, per-core, ...). To make it clearer, factor out the condition check into the new evsel__count_has_error() function and add some comments. Signed-off-by:
-
Namhyung Kim authored
Add a logic to aggregate counter values to the new evsel->stats->aggr. This is not used yet so shadow stats are not updated. But later patch will convert the existing code to use it. With that, we don't need to handle AGGR_GLOBAL specially anymore. It can use the same logic with counts, prev_counts and aggr_counts. Signed-off-by:
-
Namhyung Kim authored
The perf_stat_config.aggr_map should have a correct size of the aggregation map. Use it to allocate aggr_counts. Also AGGR_NONE with per-core events can be tricky because it doesn't aggreate basically but it needs to do so for per-core events only. So only per-core evsels will have stats->aggr data. Note that other caller of evlist__alloc_stat() might not have stat_config or aggr_map. Signed-off-by:
-
Namhyung Kim authored
The perf_stat_aggr struct is to keep aggregated counter values and the states according to the aggregation mode. The number of entries is depends on the mode and this is a preparation for the later use. Signed-off-by:
-
Namhyung Kim authored
In case of no aggregation, it needs to keep the original (cpu) ordering in the aggr_map so that it can be in sync with the cpu map. This will make the code easier to handle AGGR_NONE similar to others. Signed-off-by:
-
Namhyung Kim authored
Likewise, add an aggr_id for cpu for none aggregation mode. This is not used actually yet but later code will use to unify the aggregation code. No functional change intended. Signed-off-by:
-
Namhyung Kim authored
To make the code simpler, I'd like to use the same aggregation code for the global mode. We can simply add an id function to return cpu 0 and use print_aggr(). No functional change intended. Signed-off-by:
-
Namhyung Kim authored
In the stat-display code, it needs to check if the current evsel is hybrid but it uses perf_pmu__has_hybrid() which can return true for non-hybrid event too. I think it's better to use evsel__is_hybrid(). Also remove a NULL check for the 'config' parameter in the hybrid_merge() since it's called after config->no_merge check. Signed-off-by:
-
