linux/tools/perf, branch v4.19

perf tools: Stop fallbacking to kallsyms for vdso symbols lookup

2018-10-17T18:56:15Z

David reports that: Perf has this hack where it uses the kernel symbol map as a backup when a symbol can't be found in the user's symbol table(s). This causes problems because the tests driving this code path use machine__kernel_ip(), and that is completely meaningless on Sparc. On sparc64 the kernel and user live in physically separate virtual address spaces, rather than a shared one. And the kernel lives at a virtual address that overlaps common userspace addresses. So this test passes almost all the time when a user symbol lookup fails. The consequence of this is that, if the unfound user virtual address in the sample doesn't match up to a kernel symbol either, we trigger things like this code in builtin-top.c: if (al.sym == NULL && al.map != NULL) { const char *msg = "Kernel samples will not be resolved.\n"; /* * As we do lazy loading of symtabs we only will know if the * specified vmlinux file is invalid when we actually have a * hit in kernel space and then try to load it. So if we get * here and there are _no_ symbols in the DSO backing the * kernel map, bail out. * * We may never get here, for instance, if we use -K/ * --hide-kernel-symbols, even if the user specifies an * invalid --vmlinux ;-) */ if (!machine->kptr_restrict_warned && !top->vmlinux_warned && __map__is_kernel(al.map) && map__has_symbols(al.map)) { if (symbol_conf.vmlinux_name) { char serr[256]; dso__strerror_load(al.map->dso, serr, sizeof(serr)); ui__warning("The %s file can't be used: %s\n%s", symbol_conf.vmlinux_name, serr, msg); } else { ui__warning("A vmlinux file was not found.\n%s", msg); } if (use_browser <= 0) sleep(5); top->vmlinux_warned = true; } } When I fire up a compilation on sparc, this triggers immediately. I'm trying to figure out what the "backup to kernel map" code is accomplishing. I see some language in the current code and in the changes that have happened in this area talking about vdso. Does that really happen? The vdso is mapped into userspace virtual addresses, not kernel ones. More history. This didn't cause problems on sparc some time ago, because the kernel IP check used to be "ip < 0" :-) Sparc kernel addresses are not negative. But now with machine__kernel_ip(), which works using the symbol table determined kernel address range, it does trigger. What it all boils down to is that on architectures like sparc, machine__kernel_ip() should always return false in this scenerio, and therefore this kind of logic: if (cpumode == PERF_RECORD_MISC_USER && machine && mg != &machine->kmaps && machine__kernel_ip(machine, al->addr)) { is basically invalid. PERF_RECORD_MISC_USER implies no kernel address can possibly match for the sample/event in question (no matter how hard you try!) :-) So, I thought something had changed and in the past we would somehow find that address in the kallsyms, but I couldn't find anything to back that up, the patch introducing this is over a decade old, lots of things changed, so I was just thinking I was missing something. I tried a gtod busy loop to generate vdso activity and added a 'perf probe' at that branch, on x86_64 to see if it ever gets hit: Made thread__find_map() noinline, as 'perf probe' in lines of inline functions seems to not be working, only at function start. (Masami?) # perf probe -x ~/bin/perf -L thread__find_map:57 57 if (cpumode == PERF_RECORD_MISC_USER && machine && 58 mg != &machine->kmaps && 59 machine__kernel_ip(machine, al->addr)) { 60 mg = &machine->kmaps; 61 load_map = true; 62 goto try_again; } } else { /* * Kernel maps might be changed when loading * symbols so loading * must be done prior to using kernel maps. */ 69 if (load_map) 70 map__load(al->map); 71 al->addr = al->map->map_ip(al->map, al->addr); # perf probe -x ~/bin/perf thread__find_map:60 Added new event: probe_perf:thread__find_map (on thread__find_map:60 in /home/acme/bin/perf) You can now use it in all perf tools, such as: perf record -e probe_perf:thread__find_map -aR sleep 1 # Then used this to see if, system wide, those probe points were being hit: # perf trace -e *perf:thread*/max-stack=8/ ^C[root@jouet ~]# No hits when running 'perf top' and: # cat gtod.c #include int main(void) { struct timeval tv; while (1) gettimeofday(&tv, 0); return 0; } [root@jouet c]# ./gtod ^C Pressed 'P' in 'perf top' and the [vdso] samples are there: 62.84% [vdso] [.] __vdso_gettimeofday 8.13% gtod [.] main 7.51% [vdso] [.] 0x0000000000000914 5.78% [vdso] [.] 0x0000000000000917 5.43% gtod [.] _init 2.71% [vdso] [.] 0x000000000000092d 0.35% [kernel] [k] native_io_delay 0.33% libc-2.26.so [.] __memmove_avx_unaligned_erms 0.20% [vdso] [.] 0x000000000000091d 0.17% [i2c_i801] [k] i801_access 0.06% firefox [.] free 0.06% libglib-2.0.so.0.5400.3 [.] g_source_iter_next 0.05% [vdso] [.] 0x0000000000000919 0.05% libpthread-2.26.so [.] __pthread_mutex_lock 0.05% libpixman-1.so.0.34.0 [.] 0x000000000006d3a7 0.04% [kernel] [k] entry_SYSCALL_64_trampoline 0.04% libxul.so [.] style::dom_apis::query_selector_slow 0.04% [kernel] [k] module_get_kallsym 0.04% firefox [.] malloc 0.04% [vdso] [.] 0x0000000000000910 I added a 'perf probe' to thread__find_map:69, and that surely got tons of hits, i.e. for every map found, just to make sure the 'perf probe' command was really working. In the process I noticed a bug, we're only have records for '[vdso]' for pre-existing commands, i.e. ones that are running when we start 'perf top', when we will generate the PERF_RECORD_MMAP by looking at /perf/PID/maps. I.e. like this, for preexisting processes with a vdso map, again, tracing for all the system, only pre-existing processes get a [vdso] map (when having one): [root@jouet ~]# perf probe -x ~/bin/perf __machine__addnew_vdso Added new event: probe_perf:__machine__addnew_vdso (on __machine__addnew_vdso in /home/acme/bin/perf) You can now use it in all perf tools, such as: perf record -e probe_perf:__machine__addnew_vdso -aR sleep 1 [root@jouet ~]# perf trace -e probe_perf:__machine__addnew_vdso/max-stack=8/ 0.000 probe_perf:__machine__addnew_vdso:(568eb3) __machine__addnew_vdso (/home/acme/bin/perf) map__new (/home/acme/bin/perf) machine__process_mmap2_event (/home/acme/bin/perf) machine__process_event (/home/acme/bin/perf) perf_event__process (/home/acme/bin/perf) perf_tool__process_synth_event (/home/acme/bin/perf) perf_event__synthesize_mmap_events (/home/acme/bin/perf) __event__synthesize_thread (/home/acme/bin/perf) The kernel is generating a PERF_RECORD_MMAP for vDSOs, but somehow 'perf top' is not getting those records while 'perf record' is: # perf record ~acme/c/gtod ^C[ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.076 MB perf.data (1499 samples) ] # perf report -D | grep PERF_RECORD_MMAP2 71293612401913 0x11b48 [0x70]: PERF_RECORD_MMAP2 25484/25484: [0x400000(0x1000) @ 0 fd:02 1137 541179306]: r-xp /home/acme/c/gtod 71293612419012 0x11be0 [0x70]: PERF_RECORD_MMAP2 25484/25484: [0x7fa4a2783000(0x227000) @ 0 fd:00 3146370 854107250]: r-xp /usr/lib64/ld-2.26.so 71293612432110 0x11c50 [0x60]: PERF_RECORD_MMAP2 25484/25484: [0x7ffcdb53a000(0x2000) @ 0 00:00 0 0]: r-xp [vdso] 71293612509944 0x11cb0 [0x70]: PERF_RECORD_MMAP2 25484/25484: [0x7fa4a23cd000(0x3b6000) @ 0 fd:00 3149723 262067164]: r-xp /usr/lib64/libc-2.26.so # # perf script | grep vdso | head gtod 25484 71293.612768: 2485554 cycles:ppp: 7ffcdb53a914 [unknown] ([vdso]) gtod 25484 71293.613576: 2149343 cycles:ppp: 7ffcdb53a917 [unknown] ([vdso]) gtod 25484 71293.614274: 1814652 cycles:ppp: 7ffcdb53aca8 __vdso_gettimeofday+0x98 ([vdso]) gtod 25484 71293.614862: 1669070 cycles:ppp: 7ffcdb53acc5 __vdso_gettimeofday+0xb5 ([vdso]) gtod 25484 71293.615404: 1451589 cycles:ppp: 7ffcdb53acc5 __vdso_gettimeofday+0xb5 ([vdso]) gtod 25484 71293.615999: 1269941 cycles:ppp: 7ffcdb53ace6 __vdso_gettimeofday+0xd6 ([vdso]) gtod 25484 71293.616405: 1177946 cycles:ppp: 7ffcdb53a914 [unknown] ([vdso]) gtod 25484 71293.616775: 1121290 cycles:ppp: 7ffcdb53ac47 __vdso_gettimeofday+0x37 ([vdso]) gtod 25484 71293.617150: 1037721 cycles:ppp: 7ffcdb53ace6 __vdso_gettimeofday+0xd6 ([vdso]) gtod 25484 71293.617478: 994526 cycles:ppp: 7ffcdb53ace6 __vdso_gettimeofday+0xd6 ([vdso]) # The patch is the obvious one and with it we also continue to resolve vdso symbols for pre-existing processes in 'perf top' and for all processes in 'perf record' + 'perf report/script'. Suggested-by: David Miller Acked-by: David Miller Cc: Adrian Hunter Cc: David Ahern Cc: Jiri Olsa Cc: Namhyung Kim Cc: Wang Nan Link: https://lkml.kernel.org/n/tip-cs7skq9pp0kjypiju6o7trse@git.kernel.org Signed-off-by: Arnaldo Carvalho de Melo

perf tools: Pass build flags to traceevent build

2018-10-16T17:57:59Z

So the extra user build flags are propagated to libtraceevent. Signed-off-by: Jiri Olsa Cc: "Herton R. Krzesinski" Cc: Steven Rostedt (VMware) Cc: Tzvetomir Stoyanov (VMware) Cc: Yordan Karadzhov (VMware) Link: http://lkml.kernel.org/r/20181016150614.21260-3-jolsa@kernel.org Signed-off-by: Arnaldo Carvalho de Melo

perf report: Don't crash on invalid inline debug information

2018-10-16T17:52:21Z

When the function name for an inline frame is invalid, we must not try to demangle this symbol, otherwise we crash with: #0 0x0000555555895c01 in bfd_demangle () #1 0x0000555555823262 in demangle_sym (dso=0x555555d92b90, elf_name=0x0, kmodule=0) at util/symbol-elf.c:215 #2 dso__demangle_sym (dso=dso@entry=0x555555d92b90, kmodule=, kmodule@entry=0, elf_name=elf_name@entry=0x0) at util/symbol-elf.c:400 #3 0x00005555557fef4b in new_inline_sym (funcname=0x0, base_sym=0x555555d92b90, dso=0x555555d92b90) at util/srcline.c:89 #4 inline_list__append_dso_a2l (dso=dso@entry=0x555555c7bb00, node=node@entry=0x555555e31810, sym=sym@entry=0x555555d92b90) at util/srcline.c:264 #5 0x00005555557ff27f in addr2line (dso_name=dso_name@entry=0x555555d92430 "/home/milian/.debug/.build-id/f7/186d14bb94f3c6161c010926da66033d24fce5/elf", addr=addr@entry=2888, file=file@entry=0x0, line=line@entry=0x0, dso=dso@entry=0x555555c7bb00, unwind_inlines=unwind_inlines@entry=true, node=0x555555e31810, sym=0x555555d92b90) at util/srcline.c:313 #6 0x00005555557ffe7c in addr2inlines (sym=0x555555d92b90, dso=0x555555c7bb00, addr=2888, dso_name=0x555555d92430 "/home/milian/.debug/.build-id/f7/186d14bb94f3c6161c010926da66033d24fce5/elf") at util/srcline.c:358 So instead handle the case where we get invalid function names for inlined frames and use a fallback '??' function name instead. While this crash was originally reported by Hadrien for rust code, I can now also reproduce it with trivial C++ code. Indeed, it seems like libbfd fails to interpret the debug information for the inline frame symbol name: $ addr2line -e /home/milian/.debug/.build-id/f7/186d14bb94f3c6161c010926da66033d24fce5/elf -if b48 main /usr/include/c++/8.2.1/complex:610 ?? /usr/include/c++/8.2.1/complex:618 ?? /usr/include/c++/8.2.1/complex:675 ?? /usr/include/c++/8.2.1/complex:685 main /home/milian/projects/kdab/rnd/hotspot/tests/test-clients/cpp-inlining/main.cpp:39 I've reported this bug upstream and also attached a patch there which should fix this issue: https://sourceware.org/bugzilla/show_bug.cgi?id=23715 Reported-by: Hadrien Grasland Signed-off-by: Milian Wolff Cc: Jin Yao Cc: Jiri Olsa Cc: Namhyung Kim Fixes: a64489c56c30 ("perf report: Find the inline stack for a given address") [ The above 'Fixes:' cset is where originally the problem was introduced, i.e. using a2l->funcname without checking if it is NULL, but this current patch fixes the current codebase, i.e. multiple csets were applied after a64489c56c30 before the problem was reported by Hadrien ] Link: http://lkml.kernel.org/r/20180926135207.30263-3-milian.wolff@kdab.com Signed-off-by: Arnaldo Carvalho de Melo

perf cpu_map: Align cpu map synthesized events properly.

2018-10-16T15:30:03Z

The size of the resulting cpu map can be smaller than a multiple of sizeof(u64), resulting in SIGBUS on cpus like Sparc as the next event will not be aligned properly. Signed-off-by: David S. Miller Cc: Jiri Olsa Cc: Kan Liang Fixes: 6c872901af07 ("perf cpu_map: Add cpu_map event synthesize function") Link: http://lkml.kernel.org/r/20181011.224655.716771175766946817.davem@davemloft.net Signed-off-by: Arnaldo Carvalho de Melo

perf tools: Fix use of alternatives to find JDIR

2018-10-16T15:06:47Z

When a build is run from something like a cron job, the user's $PATH is rather minimal, of note, not including /usr/sbin in my own case. Because of that, an automated rpm package build ultimately fails to find libperf-jvmti.so, because somewhere within the build, this happens... /bin/sh: alternatives: command not found /bin/sh: alternatives: command not found Makefile.config:849: No openjdk development package found, please install JDK package, e.g. openjdk-8-jdk, java-1.8.0-openjdk-devel ...and while the build continues, libperf-jvmti.so isn't built, and things fall down when rpm tries to find all the %files specified. Exact same system builds everything just fine when the job is launched from a login shell instead of a cron job, since alternatives is in $PATH, so openjdk is actually found. The test required to get into this section of code actually specifies the full path, as does a block just above it, so let's do that here too. Signed-off-by: Jarod Wilson Acked-by: Jiri Olsa Cc: Alexander Shishkin Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Stephane Eranian Cc: William Cohen Fixes: d4dfdf00d43e ("perf jvmti: Plug compilation into perf build") Link: http://lkml.kernel.org/r/20180906221812.11167-1-jarod@redhat.com Signed-off-by: Arnaldo Carvalho de Melo

perf evsel: Store ids for events with their own cpus perf_event__synthesize_event_update_cpus

2018-10-16T11:18:52Z

John reported crash when recording on an event under PMU with cpumask defined: root@localhost:~# ./perf_debug_ record -e armv8_pmuv3_0/br_mis_pred/ sleep 1 perf: Segmentation fault Obtained 9 stack frames. ./perf_debug_() [0x4c5ef8] [0xffff82ba267c] ./perf_debug_() [0x4bc5a8] ./perf_debug_() [0x419550] ./perf_debug_() [0x41a928] ./perf_debug_() [0x472f58] ./perf_debug_() [0x473210] ./perf_debug_() [0x4070f4] /lib/aarch64-linux-gnu/libc.so.6(__libc_start_main+0xe0) [0xffff8294c8a0] Segmentation fault (core dumped) We synthesize an update event that needs to touch the evsel id array, which is not defined at that time. Fixing this by forcing the id allocation for events with their own cpus. Reported-by: John Garry Signed-off-by: Jiri Olsa Tested-by: John Garry Cc: Alexander Shishkin Cc: Andi Kleen Cc: Mark Rutland Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Will Deacon Cc: linux-arm-kernel@lists.infradead.org Cc: linuxarm@huawei.com Fixes: bfd8f72c2778 ("perf record: Synthesize unit/scale/... in event update") Link: http://lkml.kernel.org/r/20181003212052.GA32371@krava Signed-off-by: Arnaldo Carvalho de Melo

perf vendor events intel: Fix wrong filter_band* values for uncore events

2018-10-11T14:13:23Z

Michael reported that he could not stat following event: $ perf stat -e unc_p_freq_ge_1200mhz_cycles -a -- ls event syntax error: '..e_1200mhz_cycles' \___ value too big for format, maximum is 255 Run 'perf list' for a list of valid events The event is unwrapped into: uncore_pcu/event=0xb,filter_band0=1200/ where filter_band0 format says it's one byte only: # cat uncore_pcu/format/filter_band0 config1:0-7 while JSON files specifies bigger number: "Filter": "filter_band0=1200", all the filter_band* formats show 1 byte width: # cat uncore_pcu/format/filter_band1 config1:8-15 # cat uncore_pcu/format/filter_band2 config1:16-23 # cat uncore_pcu/format/filter_band3 config1:24-31 The reason of the issue is that filter_band* values are supposed to be in 100Mhz units.. it's stated in the JSON help for the events, like: filter_band3=XXX, with XXX in 100Mhz units This patch divides the filter_band* values by 100, plus there's couple of changes that actually change the number completely, like: - "Filter": "edge=1,filter_band2=4000", + "Filter": "edge=1,filter_band2=30", Reported-by: Michael Petlan Signed-off-by: Jiri Olsa Acked-by: Andi Kleen Cc: Alexander Shishkin Cc: Kan Liang Cc: Namhyung Kim Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/20181010080339.GB15790@krava Signed-off-by: Arnaldo Carvalho de Melo

Revert "perf tools: Fix PMU term format max value calculation"

2018-10-09T13:48:55Z

This reverts commit ac0e2cd555373ae6f8f3a3ad3fbbf5b6d1e7aaaa. Michael reported an issue with oversized terms values assignment and I noticed there was actually a misunderstanding of the max value check in the past. The above commit's changelog says: If bit 21 is set, there is parsing issues as below. $ perf stat -a -e uncore_qpi_0/event=0x200002,umask=0x8/ event syntax error: '..pi_0/event=0x200002,umask=0x8/' \___ value too big for format, maximum is 511 But there's no issue there, because the event value is distributed along the value defined by the format. Even if the format defines separated bit, the value is treated as a continual number, which should follow the format definition. In above case it's 9-bit value with last bit separated: $ cat uncore_qpi_0/format/event config:0-7,21 Hence the value 0x200002 is correctly reported as format violation, because it exceeds 9 bits. It should have been 0x102 instead, which sets the 9th bit - the bit 21 of the format. $ perf stat -vv -a -e uncore_qpi_0/event=0x102,umask=0x8/ Using CPUID GenuineIntel-6-2D ... ------------------------------------------------------------ perf_event_attr: type 10 size 112 config 0x200802 sample_type IDENTIFIER ... Reported-by: Michael Petlan Signed-off-by: Jiri Olsa Cc: Alexander Shishkin Cc: Andi Kleen Cc: Kan Liang Cc: Namhyung Kim Cc: Peter Zijlstra Fixes: ac0e2cd55537 ("perf tools: Fix PMU term format max value calculation") Link: http://lkml.kernel.org/r/20181003072046.29276-1-jolsa@kernel.org Signed-off-by: Arnaldo Carvalho de Melo

perf record: Use unmapped IP for inline callchain cursors

2018-10-05T14:18:09Z

Only use the mapped IP to find inline frames, but keep using the unmapped IP for the callchain cursor. This ensures we properly show the unmapped IP when displaying a frame we received via the dso__parse_addr_inlines API for a module which does not contain sufficient debug symbols to show the srcline. This is another follow-up to commit 19610184693c ("perf script: Show virtual addresses instead of offsets"). Signed-off-by: Milian Wolff Acked-by: Jiri Olsa Tested-by: Ravi Bangoria Tested-by: Arnaldo Carvalho de Melo Cc: Jin Yao Cc: Namhyung Kim Cc: Sandipan Das Fixes: 19610184693c ("perf script: Show virtual addresses instead of offsets") Link: http://lkml.kernel.org/r/20180926135207.30263-2-milian.wolff@kdab.com Link: http://lkml.kernel.org/r/20181002073949.3297-1-milian.wolff@kdab.com [ Squashed a fix from Milian for a problem reported by Ravi, fixed up space damage ] Signed-off-by: Arnaldo Carvalho de Melo

perf python: Use -Wno-redundant-decls to build with PYTHON=python3

2018-10-05T14:11:55Z

When building in ClearLinux using 'make PYTHON=python3' with gcc 8.2.1 it fails with: GEN /tmp/build/perf/python/perf.so In file included from /usr/include/python3.7m/Python.h:126, from /git/linux/tools/perf/util/python.c:2: /usr/include/python3.7m/import.h:58:24: error: redundant redeclaration of ‘_PyImport_AddModuleObject’ [-Werror=redundant-decls] PyAPI_FUNC(PyObject *) _PyImport_AddModuleObject(PyObject *, PyObject *); ^~~~~~~~~~~~~~~~~~~~~~~~~ /usr/include/python3.7m/import.h:47:24: note: previous declaration of ‘_PyImport_AddModuleObject’ was here PyAPI_FUNC(PyObject *) _PyImport_AddModuleObject(PyObject *name, ^~~~~~~~~~~~~~~~~~~~~~~~~ cc1: all warnings being treated as errors error: command 'gcc' failed with exit status 1 And indeed there is a redundant declaration in that Python.h file, one with parameter names and the other without, so just add -Wno-error=redundant-decls to the python setup instructions. Now perf builds with gcc in ClearLinux with the following Dockerfile: # docker.io/acmel/linux-perf-tools-build-clearlinux:latest FROM docker.io/clearlinux:latest MAINTAINER Arnaldo Carvalho de Melo RUN swupd update && \ swupd bundle-add sysadmin-basic-dev RUN mkdir -m 777 -p /git /tmp/build/perf /tmp/build/objtool /tmp/build/linux && \ groupadd -r perfbuilder && \ useradd -m -r -g perfbuilder perfbuilder && \ chown -R perfbuilder.perfbuilder /tmp/build/ /git/ USER perfbuilder COPY rx_and_build.sh / ENV EXTRA_MAKE_ARGS=PYTHON=python3 ENTRYPOINT ["/rx_and_build.sh"] Now to figure out why the build fails with clang, that is present in the above container as detected by the rx_and_build.sh script: clang version 6.0.1 (tags/RELEASE_601/final) Target: x86_64-unknown-linux-gnu Thread model: posix InstalledDir: /usr/sbin make: Entering directory '/git/linux/tools/perf' BUILD: Doing 'make -j4' parallel build HOSTCC /tmp/build/perf/fixdep.o HOSTLD /tmp/build/perf/fixdep-in.o LINK /tmp/build/perf/fixdep Auto-detecting system features: ... dwarf: [ OFF ] ... dwarf_getlocations: [ OFF ] ... glibc: [ OFF ] ... gtk2: [ OFF ] ... libaudit: [ OFF ] ... libbfd: [ OFF ] ... libelf: [ OFF ] ... libnuma: [ OFF ] ... numa_num_possible_cpus: [ OFF ] ... libperl: [ OFF ] ... libpython: [ OFF ] ... libslang: [ OFF ] ... libcrypto: [ OFF ] ... libunwind: [ OFF ] ... libdw-dwarf-unwind: [ OFF ] ... zlib: [ OFF ] ... lzma: [ OFF ] ... get_cpuid: [ OFF ] ... bpf: [ OFF ] Makefile.config:331: *** No gnu/libc-version.h found, please install glibc-dev[el]. Stop. make[1]: *** [Makefile.perf:206: sub-make] Error 2 make: *** [Makefile:70: all] Error 2 make: Leaving directory '/git/linux/tools/perf' Cc: Adrian Hunter Cc: David Ahern Cc: Jiri Olsa Cc: Namhyung Kim Cc: Thiago Macieira Cc: Wang Nan Link: https://lkml.kernel.org/n/tip-c3khb9ac86s00qxzjrueomme@git.kernel.org Signed-off-by: Arnaldo Carvalho de Melo