- 17 Feb, 2016 2 commits
-
-
Borislav Petkov authored
Start moving the Intel bits. Signed-off-by:
Borislav Petkov <bp@suse.de> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Link: http://lkml.kernel.org/r/1455098123-11740-2-git-send-email-bp@alien8.deSigned-off-by:
Ingo Molnar <mingo@kernel.org>
-
Ingo Molnar authored
Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo: User visible changes: - Make 'perf record' collect CPU cache info in the perf.data file header: $ perf record usleep 1 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.017 MB perf.data (7 samples) ] $ perf report --header-only -I | tail -10 | head -8 # CPU cache info: # L1 Data 32K [0-1] # L1 Instruction 32K [0-1] # L1 Data 32K [2-3] # L1 Instruction 32K [2-3] # L2 Unified 256K [0-1] # L2 Unified 256K [2-3] # L3 Unified 4096K [0-3] $ Will be used in 'perf c2c' and eventually in 'perf diff' to allow, for instance running the same workload in multiple machines and then when using 'diff' show the hardware difference. (Jiri Olsa) - 'perf stat' now shows shadow metrics (insn per cycle, etc) in interval mode too. E.g: # perf stat -I 1000 -e instructions,cycles sleep 1 # time counts unit events 1.000215928 519,620 instructions # 0.69 insn per cycle 1.000215928 752,003 cycles <SNIP> Infrastructure changes: - libapi now can also use pr_{warning,info,debug}() and that can be set by tools using it (Jiri Olsa) - libapi adopts filename__read_str() from perf, adds sysfs__read_str() (Jiri Olsa) - Add check for java alternatives cmd in jvmti Makefile, so that it manages to automatically find the right path for the JDK devel files in Ubuntu like systems in addition to Fedora like ones (Stephane Eranian) Signed-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com> Signed-off-by:
-
- 16 Feb, 2016 11 commits
-
-
Andi Kleen authored
Move the running/noise printing into printout to avoid duplicated code in the callers. v2: Merged with other patches. Remove unnecessary hunk. Readd hunk that ended in earlier patch. v3: Fix noise/running output in CSV mode v4: Merge with later patch that also moves not supported printing. Signed-off-by:Andi Kleen <ak@linux.intel.com> Acked-by:
Jiri Olsa <jolsa@kernel.org> Cc: Stephane Eranian <eranian@google.com> Link: http://lkml.kernel.org/r/1454173616-17710-4-git-send-email-andi@firstfloor.orgSigned-off-by:
-
Andi Kleen authored
Now that we can modify the metrics printout functions easily, it's straight forward to support metric printing for interval mode. All that is needed is to print the time stamp on every new line. Pass the prefix into the context and print it out. v2: Move wrong hunk to here. Committer note: Before: [root@jouet ~]# perf stat -I 1000 -e instructions,cycles sleep 1 # time counts unit events 1.000168216 538,913 instructions 1.000168216 748,765 cycles 1.000660048 153,741 instructions 1.000660048 214,066 cycles After: # perf stat -I 1000 -e instructions,cycles sleep 1 # time counts unit events 1.000215928 519,620 instructions # 0.69 insn per cycle 1.000215928 752,003 cycles 1.000946033 148,502 instructions # 0.33 insn per cycle 1.000946033 160,104 cycles Signed-off-by:Jiri Olsa <jolsa@redhat.com> Tested-by:
-
Andi Kleen authored
Abstract the printing of shadow metrics. Instead of every metric calling fprintf directly and taking care of indentation, use two call backs: one to print metrics and another to start a new line. This will allow adding metrics to CSV mode and also using them for other purposes. The computation of padding is now done in the central callback, instead of every metric doing it manually. This makes it easier to add new metrics. v2: Refactor functions, printout now does more. Move shadow printing. Improve fallback callbacks. Don't use void * callback data. v3: Remove unnecessary hunk. Add typedef for new_line v4: Remove unnecessary hunk. Don't print metrics for CSV/interval mode yet. Move printout change to separate patch. v5: Fix bisect bugs. Avoid bogus frontend cycles printing. Fix indentation in different aggregation modes. v6: Delay newline handling Signed-off-by:
-
Jiri Olsa authored
Storing CPU cache details under perf data. It's stored as new HEADER_CACHE feature and it's displayed under header info with -I option: $ perf report --header-only -I ... # CPU cache info: # L1 Data 32K [0-1] # L1 Instruction 32K [0-1] # L1 Data 32K [2-3] # L1 Instruction 32K [2-3] # L2 Unified 256K [0-1] # L2 Unified 256K [2-3] # L3 Unified 4096K [0-3] ... All distinct caches are stored/displayed. Signed-off-by:
-
Jiri Olsa authored
Setting libapi debug output functions to use perf functions. Signed-off-by:
-
Arnaldo Carvalho de Melo authored
Adhering to the naming convention used when va_args is in a printf like function, e.g. stdio.h. Cc: Jiri Olsa <jolsa@kernel.org> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: David Ahern <dsahern@gmail.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Wang Nan <wangnan0@huawei.com> Link: http://lkml.kernel.org/n/tip-b5l3wt77ct28dcnriguxtvn6@git.kernel.orgSigned-off-by:
-
Jiri Olsa authored
Adding sysfs__read_str function to ease up reading string files from sysfs. New interface is: int sysfs__read_str(const char *entry, char **buf, size_t *sizep); Signed-off-by:
-
Jiri Olsa authored
We already moved similar functions in here, also it'll be useful for sysfs__read_str addition in following patch. Signed-off-by:
-
Jiri Olsa authored
Adding support for warning/info/debug output within libapi code. Adding following macros: pr_warning(fmt, ...) pr_info(fmt, ...) pr_debug(fmt, ...) Also adding libapi_set_print function to set above functions. This will be used in perf to set standard debug handlers for libapi. Adding 2 header files: debug.h - to be used outside libapi, contains libapi_set_print interface debug-internal.h - to be used within libapi, contains pr_warning/pr_info/pr_debug definitions Signed-off-by: -
Stephane Eranian authored
This patch modifies the jvmti makefile to check if the /usr/sbin/java-update-alternatives utility is present. If so, then use it, if not then use the altenatives command. This helps handle the difference between Ubuntu and Fedora Linux distributions. Signed-off-by:
Stephane Eranian <eranian@google.com> Tested-by:
-
Ingo Molnar authored
Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo: User visible changes: - Do not print trailing spaces in the hists browser (top, report) to avoid line wrapping issues when long C++ demangled functions are sampled (Arnaldo Carvalho de Melo) - Allow 'perf config' to show --system or --user settings (Taeung Song) - Add better warning about the need to install the audit-lib-python package when using perf python scripts (Taeung Song) - Fix symbol resolution when kernel modules files are only in the build id cache (~/.debug) (Wang Nan) Build fixes: - Fix 'perf test' build on older systems where 'signal' is reserved (Arnaldo Carvalho de Melo) Infrastructure changes: - Free the terms list_head in parse_events__free_terms(), also unlink the entries when deleting them (Wang Nan) - Fix releasing event_class in 'perf data' fixing integration with libbabeltrace (Wang Nan) - Add EXTRA_LDFLAGS option to Makefile (Zubair Lutfullah Kakakhel) Signed-off-by:
-
- 15 Feb, 2016 1 commit
-
-
Arnaldo Carvalho de Melo authored
fixing the following problems, for instance, on RHEL6.7: CC /tmp/build/perf/tests/bp_signal.o cc1: warnings being treated as errors tests/bp_signal.c: In function ‘__event’: tests/bp_signal.c:106: error: declaration of ‘signal’ shadows a global declaration /usr/include/signal.h:101: error: shadowed declaration is here tests/bp_signal.c: In function ‘bp_event’: tests/bp_signal.c:144: error: declaration of ‘signal’ shadows a global declaration /usr/include/signal.h:101: error: shadowed declaration is here tests/bp_signal.c: In function ‘wp_event’: tests/bp_signal.c:149: error: declaration of ‘signal’ shadows a global declaration /usr/include/signal.h:101: error: shadowed declaration is here mv: cannot stat `/tmp/build/perf/tests/.bp_signal.o.tmp': No such file or directory make[3]: *** [/tmp/build/perf/tests/bp_signal.o] Error 1 make[2]: *** [tests] Error 2 make[1]: *** [/tmp/build/perf/perf-in.o] Error 2 make[1]: *** Waiting for unfinished jobs.... Reported-by:Vinson Lee <vlee@freedesktop.org> Cc: Alexei Starovoitov <ast@kernel.org> Cc: Brendan Gregg <brendan.d.gregg@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: He Kuang <hekuang@huawei.com> Cc: Li Zefan <lizefan@huawei.com> Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Wang Nan <wangnan0@huawei.com> Cc: Will Deacon <will.deacon@arm.com> Cc: pi3orama@163.com Fixes: 8fd34e1c ("perf test: Improve bp_signal") Link: http://lkml.kernel.org/n/tip-wlpx6tik1b0jirlkw64bv400@git.kernel.orgSigned-off-by:
-
- 12 Feb, 2016 12 commits
-
-
Wang Nan authored
A new patch in libbabeltrace [1] reveals a object leak problem in 'perf data' CTF support: perf code never releases the event_class which is allocated in add_event() and stored in evsel's private field. If libbabeltrace has the above patch applied, leaking event_class prevents the writer from being destroyed and flushing metadata. For example: $ perf record ls perf.data [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.012 MB perf.data (12 samples) ] $ perf data convert --to-ctf ./out.ctf [ perf data convert: Converted 'perf.data' into CTF data './out.ctf' ] [ perf data convert: Converted and wrote 0.000 MB (12 samples) ] $ cat ./out.ctf/metadata $ ls -l ./out.ctf/metadata -rw-r----- 1 w00229757 mm 0 Jan 27 10:49 ./out.ctf/metadata The correct result should be: ... $ cat ./out.ctf/metadata /* CTF 1.8 */ trace { [SNIP] $ ls -l ./out.ctf/metadata -rw-r----- 1 w00229757 mm 2446 Jan 27 10:52 ./out.ctf/metadata The full story is: Patch [1] of babeltrace redesigns its reference counting scheme. In that patch: * writer <- trace (bt_ctf_writer_create) * trace <- stream_class (bt_ctf_trace_add_stream_class) * stream_class <- event_class (bt_ctf_stream_class_add_event_class) ('<-' means 'is a parent of') Holding of event_class causes reference count of corresponding 'writer' to increase through parent chain. Perf expects that 'writer' is released (so metadata is flushed) through bt_ctf_writer_put() in ctf_writer__cleanup(). However, since it never releases event_class, the reference of 'writer' won't be dropped, so bt_ctf_writer_put() won't lead to the release of writer. Before this CTF patch, !(writer <- trace). Even with event_class leaking, the writer ends up being released. [1] https://github.com/efficios/babeltrace/commit/e6a8e8e4744633807083a077ff9f101eb97d9801Signed-off-by:Wang Nan <wangnan0@huawei.com> Acked-by:
-
Arnaldo Carvalho de Melo authored
To follow convention used in other tools/perf/ areas. Also remove the need to check if it is NULL before calling the destructor, again, to follow convention that goes back to free(). Cc: Alexei Starovoitov <ast@kernel.org> Cc: He Kuang <hekuang@huawei.com> cc: Jiri Olsa <jolsa@kernel.org> Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Wang Nan <wangnan0@huawei.com> Cc: Zefan Li <lizefan@huawei.com> Cc: pi3orama@163.com Link: http://lkml.kernel.org/n/tip-w6owu7rb8a46gvunlinxaqwx@git.kernel.orgSigned-off-by:
-
Wang Nan authored
Fixing a leak, since code calling parse_events__free_terms() expect it to free the list_head too. Signed-off-by:
-
Arnaldo Carvalho de Melo authored
In these two cases, a 'perf test' entry and in the PMU code the list_head is on the stack, so we can't use perf_event__free_terms() (soon to be renamed to perf_event_terms__delete()), because it will free the list_head as well. Acked-by:
-
Arnaldo Carvalho de Melo authored
Purges 'struct parse_event_term' entries from a list_head. Some users need this because they don't allocate space for the list head, it maybe on the stack or embedded into some other struct. Next patch will convert users that need just purging and then the perf_events__free_terms() routine will free the list head as well, finally being renamed to perf_events_terms__delete(). Acked-by:
-
Wang Nan authored
We were just freeing them, better unlink and init its nodes to catch bugs faster if we keep dangling references to them. Signed-off-by:
-
Arnaldo Carvalho de Melo authored
We were doing column alignment in the format function for each cell, returning a string padded with spaces so that when the next column is printed the cursor is at its column alignment. This ends up needlessly printing trailing spaces, do it at the format iterator, that is where we know if it is needed, i.e. if there is more columns to be printed. This eliminates the need for triming lines when doing a dump using 'P' in the TUI browser and also produces far saner results with things like piping 'perf report' to 'less'. Right now only the formatters for sym->name and the 'locked' column (perf mem report), that are the ones that end up at the end of lines in the default 'perf report', 'perf top' and 'perf mem report' tools, the others will be done in a subsequent patch. In the end the 'width' parameter for the formatters now mean, in 'printf' terms, the 'precision', where before it was the field 'width'. Reported-by:
Dave Jones <davej@codemonkey.org.uk> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Link: http://lkml.kernel.org/n/tip-s7iwl2gj23w92l6tibnrcqzr@git.kernel.orgSigned-off-by:
-
Arnaldo Carvalho de Melo authored
Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Wang Nan <wangnan0@huawei.com> Link: http://lkml.kernel.org/n/tip-4j67nvlfwbnkg85b969ewnkr@git.kernel.orgSigned-off-by:
-
Taeung Song authored
To print syscall names, the audit-libs-python package is required.. If not installed, it prints this error string: # perf script syscall-counts Install the audit-libs-python package to get syscall names. But the package name is different in Ubuntu, mention that in the error message, similar to a error message of util/trace-event-scripting.c: # perf script syscall-counts Install the audit-libs-python package to get syscall names. For example: # apt-get install python-audit (Ubuntu) # yum install audit-libs-python (Fedora) etc. Signed-off-by:Taeung Song <treeze.taeung@gmail.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Link: http://lkml.kernel.org/r/1455018790-13425-1-git-send-email-treeze.taeung@gmail.comSigned-off-by:
-
Zubair Lutfullah Kakakhel authored
To compile for little-endian systems, you need to pass -EL to CC and LD. EXTRA_CFLAGS works to pass -EL to CC. Add EXTRA_LDFLAGS to pass -EL to LD. Signed-off-by:
Zubair Lutfullah Kakakhel <Zubair.Kakakhel@imgtec.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1455024818-15842-1-git-send-email-Zubair.Kakakhel@imgtec.comSigned-off-by:
-
Wang Nan authored
Before this patch, if a sample is triggered inside a module not in /lib/modules/`uname -r`/, even if the module is in buildid-cache, 'perf report' will still be unable to find the correct symbol. For example: # rm -rf ~/.debug/ # perf buildid-cache -a ./mymodule.ko # perf probe -m ./mymodule.ko -a get_mymodule_val Added new event: probe:get_mymodule_val (on get_mymodule_val in mymodule) You can now use it in all perf tools, such as: perf record -e probe:get_mymodule_val -aR sleep 1 # perf record -e probe:get_mymodule_val cat /proc/mymodule mymodule:3 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.011 MB perf.data (1 samples) ] # perf report --stdio [SNIP] # # Overhead Command Shared Object Symbol # ........ ....... ................ ...................... # 100.00% cat [mymodule] [k] 0x0000000000000001 # perf report -vvvv --stdio dso__load_sym: adjusting symbol: st_value: 0 sh_addr: 0 sh_offset: 0x70 symbol__new: get_mymodule_val 0x70-0x8a [SNIP] This is caused by dso__load() -> dso__load_sym(). In dso__load(), kmod is true only when its file is found in some well know directories. All files loaded from buildid-cache are treated as user programs. Following dso__load_sym() set map->pgoff incorrectly. This patch gives kernel modules in buildid-cache a chance to adjust value of kmod. After dso__load() get the type of symbols, if it is buildid, check the last 3 chars of original filename against '.ko', and adjust the value of kmod if the file is a kernel module. Signed-off-by: -
Taeung Song authored
The '--system' option means $(sysconfdir)/perfconfig and '--user' means $HOME/.perfconfig. If none is used, both system and user config file are read. E.g.: # perf config [<file-option>] [options] With an specific config file: # perf config --user | --system or both user and system config file: # perf config Signed-off-by:
-
- 09 Feb, 2016 8 commits
-
-
Martin KaFai Lau authored
When doing ebpf+kprobe on some hot TCP functions (e.g. tcp_rcv_established), the kprobe_dispatcher() function shows up in 'perf report'. In kprobe_dispatcher(), there is a lot of cache bouncing on 'tk->nhit++'. 'tk->nhit' and 'tk->tp.flags' also share the same cacheline. perf report (cycles:pp): 8.30% ipv4_dst_check 4.74% copy_user_enhanced_fast_string 3.93% dst_release 2.80% tcp_v4_rcv 2.31% queued_spin_lock_slowpath 2.30% _raw_spin_lock 1.88% mlx4_en_process_rx_cq 1.84% eth_get_headlen 1.81% ip_rcv_finish ~~~~ 1.71% kprobe_dispatcher ~~~~ 1.55% mlx4_en_xmit 1.09% __probe_kernel_read perf report after patch: 9.15% ipv4_dst_check 5.00% copy_user_enhanced_fast_string 4.12% dst_release 2.96% tcp_v4_rcv 2.50% _raw_spin_lock 2.39% queued_spin_lock_slowpath 2.11% eth_get_headlen 2.03% mlx4_en_process_rx_cq 1.69% mlx4_en_xmit 1.19% ip_rcv_finish 1.12% __probe_kernel_read 1.02% ehci_hcd_cleanup Signed-off-by:
Martin KaFai Lau <kafai@fb.com> Acked-by:
Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Acked-by:
Steven Rostedt <rostedt@goodmis.org> Cc: Alexei Starovoitov <alexei.starovoitov@gmail.com> Cc: Josef Bacik <jbacik@fb.com> Cc: Kernel Team <kernel-team@fb.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454531308-2441898-1-git-send-email-kafai@fb.comSigned-off-by:
-
Ingo Molnar authored
Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo: User visible fixes: - Handle spaces in file names obtained from /proc/pid/maps (Marcin Ślusarz) New features: - Improved support for Java, using the JVMTI agent library to do jitdumps that then will be inserted in synthesized PERF_RECORD_MMAP2 events via 'perf inject' pointed to synthesized ELF files stored in ~/.debug and keyed with build-ids, to allow symbol resolution and even annotation with source line info, see the changeset comments to see how to use it (Stephane Eranian) Documentation changes: - Document mmore variables in the 'perf config' man page (Taeung Song) Infrastructure changes: - Improve a bit the 'make -C tools/perf build-test' output (Arnaldo Carvalho de Melo) - Do 'build-test' in parallel, using 'make -j' (Arnaldo Carvalho de Melo) - Fix handling of 'clean' in multi-target make invokations for parallell builds (Jiri Olsa) Signed-off-by:
-
Borislav Petkov authored
Signed-off-by:
-
Borislav Petkov authored
Signed-off-by:
-
Borislav Petkov authored
Signed-off-by:
-
Borislav Petkov authored
We distribute those in vendor subdirs, starting with .../events/amd/. Signed-off-by:
-
Borislav Petkov authored
Also, keep the churn at minimum by adjusting the include "perf_event.h" when each file gets moved. Signed-off-by:
-
Ingo Molnar authored
Signed-off-by:
-
- 05 Feb, 2016 6 commits
-
-
Stephane Eranian authored
This patch adds source line information support to perf for jitted code. The source line info must be emitted by the runtime, such as JVMTI. Perf injects extract the source line info from the jitdump file and adds the corresponding .debug_lines section in the ELF image generated for each jitted function. The source line enables matching any address in the profile with a source file and line number. The improvement is visible in perf annotate with the source code displayed alongside the assembly code. The dwarf code leverages the support from OProfile which is also released under GPLv2. Copyright 2007 OProfile authors. Signed-off-by:
-
Stephane Eranian authored
This is a standalone JVMTI library to help profile Java jitted code with perf record/perf report. The library is not installed or compiled automatically by perf Makefile. It is not used directly by perf. It is arch agnostic and has been tested on X86 and ARM. It needs to be used with a Java runtime, such as OpenJDK, as follows: $ java -agentpath:libjvmti.so ....... See the "Committer Notes" below on how to build it. When used this way, java will generate a jitdump binary file in $HOME/.debug/java/jit/java-jit-* This binary dump file contains information to help symbolize and annotate jitted code. The jitdump information must be injected into the perf.data file using: $ perf inject --jit -i perf.data -o perf.data.jitted This injects the MMAP records to cover the jitted code and also generates one ELF image for each jitted function. The ELF images are created in the same subdir as the jitdump file. The MMAP records point there too. Then, to visualize the function or asm profile, simply use the regular perf commands: $ perf report -i perf.data.jitted or $ perf annotate -i perf.data.jitted JVMTI agent code adapted from the OProfile's opagent code. This version of the JVMTI agent is using the CLOCK_MONOTONIC as the time source to timestamp jit samples. To correlate with perf_events samples, it needs to run on kernel 4.0.0-rc5+ or later with the following commit from Peter Zijlstra: 34f43927 ("perf: Add per event clockid support") With this patch recording jitted code is done as follows: $ perf record -k mono -- java -agentpath:libjvmti.so ....... -------------------------------------------------------------------------- Committer Notes: Extended testing instructions: $ cd tools/perf/jvmti/ $ dnf install java-devel $ make Then, create some simple java stuff to record some samples: $ cat hello.java public class hello { public static void main(String[] args) { System.out.println("Hello, World"); } } $ javac hello.java $ java hello Hello, World $ And then record it using this jvmti thing: $ perf record -k mono java -agentpath:/home/acme/git/linux/tools/perf/jvmti/libjvmti.so hello java: jvmti: jitdump in /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jit-1908.dump Hello, World [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.030 MB perf.data (268 samples) ] $ Now lets insert the PERF_RECORD_MMAP2 records to point jitted mmaps to files created by the agent: $ perf inject --jit -i perf.data -o perf.data.jitted And finally see that it did its job: $ perf report -D -i perf.data.jitted | grep PERF_RECORD_MMAP2 | tail -5 79197149129422 0xfe10 [0xa0]: PERF_RECORD_MMAP2 1908/1923: [0x7f172428bd60(0x80) @ 0x40 fd:02 1840554 1]: --xs /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-283.so 79197149235701 0xfeb0 [0xa0]: PERF_RECORD_MMAP2 1908/1923: [0x7f172428ba60(0x180) @ 0x40 fd:02 1840555 1]: --xs /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-284.so 79197149250558 0xff50 [0xa0]: PERF_RECORD_MMAP2 1908/1923: [0x7f172428b860(0x180) @ 0x40 fd:02 1840556 1]: --xs /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-285.so 79197149714746 0xfff0 [0xa0]: PERF_RECORD_MMAP2 1908/1923: [0x7f172428b660(0x180) @ 0x40 fd:02 1840557 1]: --xs /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-286.so 79197149806558 0x10090 [0xa0]: PERF_RECORD_MMAP2 1908/1923: [0x7f172428b460(0x180) @ 0x40 fd:02 1840558 1]: --xs /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-287.so $ So: $ perf report -D -i perf.data | grep PERF_RECORD_MMAP2 | wc -l Failed to open /tmp/perf-1908.map, continuing without symbols 21 $ perf report -D -i perf.data.jitted | grep PERF_RECORD_MMAP2 | wc -l 307 $ echo $((307 - 21)) 286 $ 286 extra PERF_RECORD_MMAP2 records. All for thise tiny, with just one function, ELF files: $ file /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-9.so /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-9.so: ELF 64-bit LSB shared object, x86-64, version 1 (SYSV), corrupted program header size, BuildID[sha1]=ae54a2ebc3ecf0ba547bfc8cabdea1519df5203f, not stripped $ readelf -sw /home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-9.so Symbol table '.symtab' contains 2 entries: Num: Value Size Type Bind Vis Ndx Name 0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND 1: 0000000000000040 9 FUNC LOCAL DEFAULT 1 atomic_cmpxchg_long $ Inserted into the build-id cache: $ ls -la ~/.debug/.build-id/ae/54a2ebc3ecf0ba547bfc8cabdea1519df5203f lrwxrwxrwx. 1 acme acme 111 Feb 5 11:30 /home/acme/.debug/.build-id/ae/54a2ebc3ecf0ba547bfc8cabdea1519df5203f -> ../../home/acme/.debug/jit/java-jit-20160205.XXWIEDls/jitted-1908-9.so/ae54a2ebc3ecf0ba547bfc8cabdea1519df5203f Note: check why 'file' reports that 'corrupted program header size'. With a stupid java hog to do some profiling: $ cat hog.java public class hog { private static double do_something_else(int i) { double total = 0; while (i > 0) { total += Math.log(i--); } return total; } private static double do_something(int i) { double total = 0; while (i > 0) { total += Math.sqrt(i--) + do_something_else(i / 100); } return total; } public static void main(String[] args) { System.out.println(String.format("%s=%f & %f", args[0], do_something(Integer.parseInt(args[0])), do_something_else(Integer.parseInt(args[1])))); } } $ javac hog.java $ perf record -F 10000 -g -k mono java -agentpath:/home/acme/git/linux/tools/perf/jvmti/libjvmti.so hog 100000 2345000 java: jvmti: jitdump in /home/acme/.debug/jit/java-jit-20160205.XX4sqd14/jit-8670.dump 100000=291561592.669602 & 32050989.778714 [ perf record: Woken up 6 times to write data ] [ perf record: Captured and wrote 1.536 MB perf.data (12538 samples) ] $ perf inject --jit -i perf.data -o perf.data.jitted Looking at the 'perf report' TUI, at one expanded callchain leading to the jitted code: $ perf report --no-children -i perf.data.jitted Samples: 12K of event 'cycles:pp', Event count (approx.): 3829569932 Overhead Comm Shared Object Symbol - 93.38% java jitted-8670-291.so [.] class hog.do_something_else(int) class hog.do_something_else(int) - Interpreter - 75.86% call_stub JavaCalls::call_helper jni_invoke_static jni_CallStaticVoidMethod JavaMain start_thread - 17.52% JavaCalls::call_helper jni_invoke_static jni_CallStaticVoidMethod JavaMain start_thread Signed-off-by:
-
Stephane Eranian authored
This patch adds a --jit/-j option to perf inject. This options injects MMAP records into the perf.data file to cover the jitted code mmaps. It also emits ELF images for each function in the jidump file. Those images are created where the jitdump file is. The MMAP records point to that location as well. Typical flow: $ perf record -k mono -- java -agentpath:libpjvmti.so java_class $ perf inject --jit -i perf.data -o perf.data.jitted $ perf report -i perf.data.jitted Note that jitdump.h support is not limited to Java, it works with any jitted environment modified to emit the jitdump file format, include those where code can be jitted multiple times and moved around. The jitdump.h format is adapted from the Oprofile project. The genelf.c (ELF binary generation) depends on MD5 hash encoding for the buildid. To enable this, libssl-dev must be installed. If not, then genelf.c defaults to using urandom to generate the buildid, which is not ideal. The Makefile auto-detects the presence on libssl-dev. This version mmaps the jitdump file to create a marker MMAP record in the perf.data file. The marker is used to detect jitdump and cause perf inject to inject the jitted mmaps and generate ELF images for jitted functions. In V8, the following fixes and changes were made among other things: - the jidump header format include a new flags field to be used to carry information about the configuration of the runtime agent. Contributed by: Adrian Hunter <adrian.hunter@intel.com> - Fix mmap pgoff: MMAP event pgoff must be the offset within the ELF file at which the code resides. Contributed by: Adrian Hunter <adrian.hunter@intel.com> - Fix ELF virtual addresses: perf tools expect the ELF virtual addresses of dynamic objects to match the file offset. Contributed by: Adrian Hunter <adrian.hunter@intel.com> - JIT MMAP injection does not obey finished_round semantics. JIT MMAP injection injects all MMAP events in one go, so it does not obey finished_round semantics, so drop the finished_round events from the output perf.data file. Contributed by: Adrian Hunter <adrian.hunter@intel.com> Signed-off-by: -
Arnaldo Carvalho de Melo authored
To make sure the mmap records are ordered correctly and so that the correct especially due to jitted code mmaps. We cannot generate the buildid hit list and inject the jit mmaps (will come right after this patch) in at the same time for now. Signed-off-by:
-
Stephane Eranian authored
Will be used to generate build-ids in the jitdump code. Signed-off-by:
-
Stephane Eranian authored
Add Java function descriptor demangling support. Something bfd cannot do. Use the JAVA_DEMANGLE_NORET flag to avoid decoding the return type of functions. Signed-off-by:
-
