BMC/Intel-BMC/linux.git/tools/perf/builtin-record.c, branch dev-5.2

perf parse-regs: Split parse_regs

2019-05-15T19:36:49+00:00

The available registers for --int-regs and --user-regs may be different, e.g. XMM registers. Split parse_regs into two dedicated functions for --int-regs and --user-regs respectively. Modify the warning message. "--user-regs=?" should be applied to show the available registers for --user-regs. Signed-off-by: Kan Liang Tested-by: Ravi Bangoria Cc: Andi Kleen Cc: Jiri Olsa Link: http://lkml.kernel.org/r/1557865174-56264-1-git-send-email-kan.liang@linux.intel.com [ Changed docs as suggested by Ravi and agreed by Kan ] Signed-off-by: Arnaldo Carvalho de Melo

perf record: Implement -z,--compression_level[=] option

2019-05-15T19:36:49+00:00

Implemented -z,--compression_level[=] option that enables compression of mmaped kernel data buffers content in runtime during perf record mode collection. Default option value is 1 (fastest compression). Compression overhead has been measured for serial and AIO streaming when profiling matrix multiplication workload: ------------------------------------------------------------- | SERIAL | AIO-1 | ----------------------------------------------------------------| |-z | OVH(x) | ratio(x) size(MiB) | OVH(x) | ratio(x) size(MiB) | |---------------------------------------------------------------| | 0 | 1,00 | 1,000 179,424 | 1,00 | 1,000 187,527 | | 1 | 1,04 | 8,427 181,148 | 1,01 | 8,474 188,562 | | 2 | 1,07 | 8,055 186,953 | 1,03 | 7,912 191,773 | | 3 | 1,04 | 8,283 181,908 | 1,03 | 8,220 191,078 | | 5 | 1,09 | 8,101 187,705 | 1,05 | 7,780 190,065 | | 8 | 1,05 | 9,217 179,191 | 1,12 | 6,111 193,024 | ----------------------------------------------------------------- OVH = (Execution time with -z N) / (Execution time with -z 0) ratio - compression ratio size - number of bytes that was compressed size ~= trace size x ratio Committer notes: Testing it I noticed that it failed to disable build id processing when compression is enabled, and as we'd have to uncompress everything to look for the PERF_RECORD_{MMAP,SAMPLE,etc} to figure out which build ids to read from DSOs, we better disable build id processing when compression is enabled, logging with pr_debug() when doing so: Original patch: # perf record -z2 ^C[ perf record: Woken up 1 times to write data ] 0x1746e0 [0x76]: failed to process type: 81 [Invalid argument] [ perf record: Captured and wrote 1.568 MB perf.data, compressed (original 0.452 MB, ratio is 3.995) ] # After auto-disabling build id processing when compression is enabled: $ perf record -z2 sleep 1 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.001 MB perf.data, compressed (original 0.001 MB, ratio is 2.292) ] $ perf record -v -z2 sleep 1 Compression enabled, disabling build id collection at the end of the session. [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.001 MB perf.data, compressed (original 0.001 MB, ratio is 2.305) ] $ Also, with parts of the patch originally after this one moved to just before this one we get: $ perf record -z2 sleep 1 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.001 MB perf.data, compressed (original 0.001 MB, ratio is 2.371) ] $ perf report -D | grep COMPRESS 0 0x1b8 [0x155]: PERF_RECORD_COMPRESSED: unhandled! 0 0x30d [0x80]: PERF_RECORD_COMPRESSED: unhandled! COMPRESSED events: 2 COMPRESSED events: 0 $ I.e. when faced with PERF_RECORD_COMPRESSED that we still have no code to process, we just show it as not being handled, skip them and continue, while before we had: $ perf report -D | grep COMPRESS 0x1b8 [0x169]: failed to process type: 81 [Invalid argument] Error: failed to process sample 0 0x1b8 [0x169]: PERF_RECORD_COMPRESSED $ Signed-off-by: Alexey Budankov Reviewed-by: Jiri Olsa Tested-by: Arnaldo Carvalho de Melo Cc: Alexander Shishkin Cc: Andi Kleen Cc: Namhyung Kim Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/9ff06518-ae63-a908-e44d-5d9e56dd66d9@linux.intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf record: Implement compression for AIO trace streaming

2019-05-15T19:36:49+00:00

Compression is implemented using the functions from zstd.c. As the memory to operate on the compression uses mmap->aio.data[] buffers. If Zstd streaming compression API fails for some reason the data to be compressed are just copied into the memory buffers using plain memcpy(). Compressed trace frame consists of an array of PERF_RECORD_COMPRESSED records. Each element of the array is not longer that PERF_SAMPLE_MAX_SIZE and consists of perf_event_header followed by the compressed chunk that is decompressed on the loading stage. perf_mmap__aio_push() is replaced by perf_mmap__push() which is now used in the both serial and AIO streaming cases. perf_mmap__push() is extended with positive return values to signify absence of data ready for processing. Signed-off-by: Alexey Budankov Reviewed-by: Jiri Olsa Cc: Alexander Shishkin Cc: Andi Kleen Cc: Namhyung Kim Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/77db2b2c-5d03-dbb0-aeac-c4dd92129ab9@linux.intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf record: Implement compression for serial trace streaming

2019-05-15T19:36:49+00:00

Compression is implemented using the functions from zstd.c. As the memory to operate on the compression uses mmap->data buffer. If Zstd streaming compression API fails for some reason the data to be compressed are just copied into the memory buffers using plain memcpy(). Compressed trace frame consists of an array of PERF_RECORD_COMPRESSED records. Each element of the array is not longer that PERF_SAMPLE_MAX_SIZE and consists of perf_event_header followed by the compressed chunk that is decompressed on the loading stage. Comitter notes: Undo some unnecessary line breaks, remove some unnecessary () around zstd_data to then just get its address, and fix conflicts with BPF_PROG_INFO/BPF_BTF patchkits. Signed-off-by: Alexey Budankov Reviewed-by: Jiri Olsa Cc: Alexander Shishkin Cc: Andi Kleen Cc: Namhyung Kim Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/744df43f-3932-2594-ddef-1e99a3cad03a@linux.intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf mmap: Implement dedicated memory buffer for data compression

2019-05-15T19:36:49+00:00

Implemented mmap data buffer that is used as the memory to operate on when compressing data in case of serial trace streaming. Signed-off-by: Alexey Budankov Reviewed-by: Jiri Olsa Cc: Alexander Shishkin Cc: Andi Kleen Cc: Namhyung Kim Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/49b31321-0f70-392b-9a4f-649d3affe090@linux.intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf record: Implement COMPRESSED event record and its attributes

2019-05-15T19:36:49+00:00

Implemented PERF_RECORD_COMPRESSED event, related data types, header feature and functions to write, read and print feature attributes from the trace header section. comp_mmap_len preserves the size of mmaped kernel buffer that was used during collection. comp_mmap_len size is used on loading stage as the size of decomp buffer for decompression of COMPRESSED events content. Committer notes: Fixed up conflict with BPF_PROG_INFO and BTF_BTF header features. Signed-off-by: Alexey Budankov Reviewed-by: Jiri Olsa Cc: Alexander Shishkin Cc: Andi Kleen Cc: Namhyung Kim Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/ebbaf031-8dda-3864-ebc6-7922d43ee515@linux.intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf session: Define 'bytes_transferred' and 'bytes_compressed' metrics

2019-05-15T19:36:49+00:00

Define 'bytes_transferred' and 'bytes_compressed' metrics to calculate ratio in the end of the data collection: compression ratio = bytes_transferred / bytes_compressed The 'bytes_transferred' metric accumulates the amount of bytes that was extracted from the mmaped kernel buffers for compression, while 'bytes_compressed' accumulates the amount of bytes that was received after applying compression. Signed-off-by: Alexey Budankov Reviewed-by: Jiri Olsa Cc: Alexander Shishkin Cc: Andi Kleen Cc: Namhyung Kim Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/1d4bf499-cb03-26dc-6fc6-f14fec7622ce@linux.intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf record: Fix suggestion to get list of registers usable with --user-regs and --intr-regs

2019-05-15T19:36:46+00:00

$ perf record -h -I Usage: perf record [] [] or: perf record [] -- [] -I, --intr-regs[=] sample selected machine registers on interrupt, use -I ? to list register names $ m $ perf record -I ? Workload failed: No such file or directory $ After: $ perf record -h -I Usage: perf record [] [] or: perf record [] -- [] -I, --intr-regs[=] sample selected machine registers on interrupt, use '-I?' to list register names $ $ perf record -I? available registers: AX BX CX DX SI DI BP SP IP FLAGS CS SS R8 R9 R10 R11 R12 R13 R14 R15 Usage: perf record [] [] or: perf record [] -- [] -I, --intr-regs[=] sample selected machine registers on interrupt, use '-I?' to list register names $ Cc: Adrian Hunter Cc: Alexander Shishkin Cc: Andi Kleen Cc: Jiri Olsa Cc: Kan Liang Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Stephane Eranian Fixes: bcc84ec65ad1 ("perf record: Add ability to name registers to record") Link: https://lkml.kernel.org/n/tip-r0xhfhy5radmkhhcbcfs5izf@git.kernel.org Signed-off-by: Arnaldo Carvalho de Melo

perf record: Implement --mmap-flush= option

2019-04-01T18:18:10+00:00

Implement a --mmap-flush option that specifies minimal number of bytes that is extracted from mmaped kernel buffer to store into a trace. The default option value is 1 byte what means every time trace writing thread finds some new data in the mmaped buffer the data is extracted, possibly compressed and written to a trace. $ tools/perf/perf record --mmap-flush 1024 -e cycles -- matrix.gcc $ tools/perf/perf record --aio --mmap-flush 1K -e cycles -- matrix.gcc The option is independent from -z setting, doesn't vary with compression level and can serve two purposes. The first purpose is to increase the compression ratio of a trace data. Larger data chunks are compressed more effectively so the implemented option allows specifying data chunk size to compress. Also at some cases executing more write syscalls with smaller data size can take longer than executing less write syscalls with bigger data size due to syscall overhead so extracting bigger data chunks specified by the option value could additionally decrease runtime overhead. The second purpose is to avoid self monitoring live-lock issue in system wide (-a) profiling mode. Profiling in system wide mode with compression (-a -z) can additionally induce data into the kernel buffers along with the data from monitored processes. If performance data rate and volume from the monitored processes is high then trace streaming and compression activity in the tool is also high. High tool process activity can lead to subtle live-lock effect when compression of single new byte from some of mmaped kernel buffer leads to generation of the next single byte at some mmaped buffer. So perf tool process ends up in endless self monitoring. Implemented synch parameter is the mean to force data move independently from the specified flush threshold value. Despite the provided flush value the tool needs capability to unconditionally drain memory buffers, at least in the end of the collection. Committer testing: Running with the default value, i.e. as soon as there is something to read go on consuming, we first write the synthesized events, small chunks of about 128 bytes: # perf trace -m 2048 --call-graph dwarf -e write -- perf record 101.142 ( 0.004 ms): perf/25821 write(fd: 3, buf: 0x210db60, count: 120) = 120 __libc_write (/usr/lib64/libpthread-2.28.so) ion (/home/acme/bin/perf) record__write (inlined) process_synthesized_event (/home/acme/bin/perf) perf_tool__process_synth_event (inlined) perf_event__synthesize_mmap_events (/home/acme/bin/perf) Then we move to reading the mmap buffers consuming the events put there by the kernel perf infrastructure: 107.561 ( 0.005 ms): perf/25821 write(fd: 3, buf: 0x7f1befc02000, count: 336) = 336 __libc_write (/usr/lib64/libpthread-2.28.so) ion (/home/acme/bin/perf) record__write (inlined) record__pushfn (/home/acme/bin/perf) perf_mmap__push (/home/acme/bin/perf) record__mmap_read_evlist (inlined) record__mmap_read_all (inlined) __cmd_record (inlined) cmd_record (/home/acme/bin/perf) 12919.953 ( 0.136 ms): perf/25821 write(fd: 3, buf: 0x7f1befc83150, count: 184984) = 184984 12920.094 ( 0.155 ms): perf/25821 write(fd: 3, buf: 0x7f1befc02150, count: 261816) = 261816 12920.253 ( 0.093 ms): perf/25821 write(fd: 3, buf: 0x7f1befb81120, count: 170832) = 170832 If we limit it to write only when more than 16MB are available for reading, it throttles that to a quarter of the --mmap-pages set for 'perf record', which by default get to 528384 bytes, found out using 'record -v': mmap flush: 132096 mmap size 528384B With that in place all the writes coming from record__mmap_read_evlist(), i.e. from the mmap buffers setup by the kernel perf infrastructure were at least 132096 bytes long. Trying with a bigger mmap size: perf trace -e write perf record -v -m 2048 --mmap-flush 16M 74982.928 ( 2.471 ms): perf/26500 write(fd: 3, buf: 0x7ff94a6cc000, count: 3580888) = 3580888 74985.406 ( 2.353 ms): perf/26500 write(fd: 3, buf: 0x7ff949ecb000, count: 3453256) = 3453256 74987.764 ( 2.629 ms): perf/26500 write(fd: 3, buf: 0x7ff9496ca000, count: 3859232) = 3859232 74990.399 ( 2.341 ms): perf/26500 write(fd: 3, buf: 0x7ff948ec9000, count: 3769032) = 3769032 74992.744 ( 2.064 ms): perf/26500 write(fd: 3, buf: 0x7ff9486c8000, count: 3310520) = 3310520 74994.814 ( 2.619 ms): perf/26500 write(fd: 3, buf: 0x7ff947ec7000, count: 4194688) = 4194688 74997.439 ( 2.787 ms): perf/26500 write(fd: 3, buf: 0x7ff9476c6000, count: 4029760) = 4029760 Was again limited to a quarter of the mmap size: mmap flush: 2098176 mmap size 8392704B A warning about that would be good to have but can be added later, something like: "max flush is a quarter of the mmap size, if wanting to bump the mmap flush further, bump the mmap size as well using -m/--mmap-pages" Also rename the 'sync' parameters to 'synch' to keep tools/perf building with older glibcs: cc1: warnings being treated as errors builtin-record.c: In function 'record__mmap_read_evlist': builtin-record.c:775: warning: declaration of 'sync' shadows a global declaration /usr/include/unistd.h:933: warning: shadowed declaration is here builtin-record.c: In function 'record__mmap_read_all': builtin-record.c:856: warning: declaration of 'sync' shadows a global declaration /usr/include/unistd.h:933: warning: shadowed declaration is here Signed-off-by: Alexey Budankov Reviewed-by: Jiri Olsa Tested-by: Arnaldo Carvalho de Melo Cc: Alexander Shishkin Cc: Andi Kleen Cc: Namhyung Kim Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/f6600d72-ecfa-2eb7-7e51-f6954547d500@linux.intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf tools: Save bpf_prog_info and BTF of new BPF programs

2019-03-21T14:27:04+00:00

To fully annotate BPF programs with source code mapping, 4 different information are needed: 1) PERF_RECORD_KSYMBOL 2) PERF_RECORD_BPF_EVENT 3) bpf_prog_info 4) btf This patch handles 3) and 4) for BPF programs loaded after 'perf record|top'. For timely process of these information, a dedicated event is added to the side band evlist. When PERF_RECORD_BPF_EVENT is received via the side band event, the polling thread gathers 3) and 4) vis sys_bpf and store them in perf_env. This information is saved to perf.data at the end of 'perf record'. Committer testing: The 'wakeup_watermark' member in 'struct perf_event_attr' is inside a unnamed union, so can't be used in a struct designated initialization with older gccs, get it out of that, isolating as 'attr.wakeup_watermark = 1;' to work with all gcc versions. We also need to add '--no-bpf-event' to the 'perf record' perf_event_attr tests in 'perf test', as the way that that test goes is to intercept the events being setup and looking if they match the fields described in the control files, since now it finds first the side band event used to catch the PERF_RECORD_BPF_EVENT, they all fail. With these issues fixed: Same scenario as for testing BPF programs loaded before 'perf record' or 'perf top' starts, only start the BPF programs after 'perf record|top', so that its information get collected by the sideband threads, the rest works as for the programs loaded before start monitoring. Add missing 'inline' to the bpf_event__add_sb_event() when HAVE_LIBBPF_SUPPORT is not defined, fixing the build in systems without binutils devel files installed. Signed-off-by: Song Liu Reviewed-by: Jiri Olsa Tested-by: Arnaldo Carvalho de Melo Cc: Alexei Starovoitov Cc: Daniel Borkmann Cc: Jiri Olsa Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Stanislav Fomichev Link: http://lkml.kernel.org/r/20190312053051.2690567-16-songliubraving@fb.com Signed-off-by: Arnaldo Carvalho de Melo