When we started using a thread to catch the PERF_RECORD_BPF_EVENT meta
data events to then ask the kernel for further info (BTF, etc) for BPF
programs shortly after they get loaded, we forgot to use
unshare(CLONE_FS) as was done in:
868a832918 ("perf top: Support lookup of symbols in other mount namespaces.")
Do it so that we can enter the namespaces to read the build-ids at the
end of a 'perf record' session for the DSOs that had hits.
Before:
Starting a 'stress-ng --cpus 8' inside a container and then, outside the
container running:
# perf record -a --namespaces sleep 5
# perf buildid-list | grep stress-ng
#
We would end up with a 'perf.data' file that had no entry in its
build-id table for the /usr/bin/stress-ng binary inside the container
that got tons of PERF_RECORD_SAMPLEs.
After:
# perf buildid-list | grep stress-ng
f2ed02c68341183a124b9b0f6e2e6c493c465b29 /usr/bin/stress-ng
#
Then its just a matter of making sure that that binary debuginfo package
gets available in a place that 'perf report' will look at build-id keyed
ELF files, which, in my case, on a f30 notebook, was a matter of
installing the debuginfo file for the distro used in the container,
fedora 31:
# rpm -ivh http://fedora.c3sl.ufpr.br/linux/development/31/Everything/x86_64/debug/tree/Packages/s/stress-ng-debuginfo-0.07.29-10.fc31.x86_64.rpm
Then, because perf currently looks for those debuginfo files (richer ELF
symtab) inside that namespace (look at the setns calls):
openat(AT_FDCWD, "/proc/self/ns/mnt", O_RDONLY) = 137
openat(AT_FDCWD, "/proc/13169/ns/mnt", O_RDONLY) = 139
setns(139, CLONE_NEWNS) = 0
stat("/usr/bin/stress-ng", {st_mode=S_IFREG|0755, st_size=3065416, ...}) = 0
openat(AT_FDCWD, "/usr/bin/stress-ng", O_RDONLY) = 140
fcntl(140, F_GETFD) = 0
fstat(140, {st_mode=S_IFREG|0755, st_size=3065416, ...}) = 0
mmap(NULL, 3065416, PROT_READ, MAP_PRIVATE, 140, 0) = 0x7ff2fdc5b000
munmap(0x7ff2fdc5b000, 3065416) = 0
close(140) = 0
stat("stress-ng-0.07.29-10.fc31.x86_64.debug", 0x7fff45d71260) = -1 ENOENT (No such file or directory)
stat("/usr/bin/stress-ng-0.07.29-10.fc31.x86_64.debug", 0x7fff45d71260) = -1 ENOENT (No such file or directory)
stat("/usr/bin/.debug/stress-ng-0.07.29-10.fc31.x86_64.debug", 0x7fff45d71260) = -1 ENOENT (No such file or directory)
stat("/usr/lib/debug/usr/bin/stress-ng-0.07.29-10.fc31.x86_64.debug", 0x7fff45d71260) = -1 ENOENT (No such file or directory)
stat("/root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29", 0x7fff45d711e0) = -1 ENOENT (No such file or directory)
To only then go back to the "host" namespace to look just in the users's
~/.debug cache:
setns(137, CLONE_NEWNS) = 0
chdir("/root") = 0
close(137) = 0
close(139) = 0
stat("/root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf", 0x7fff45d732e0) = -1 ENOENT (No such file or directory)
It continues to fail to resolve symbols:
# perf report | grep stress-ng | head -5
9.50% stress-ng-cpu stress-ng [.] 0x0000000000021ac1
8.58% stress-ng-cpu stress-ng [.] 0x0000000000021ab4
8.51% stress-ng-cpu stress-ng [.] 0x0000000000021489
7.17% stress-ng-cpu stress-ng [.] 0x00000000000219b6
3.93% stress-ng-cpu stress-ng [.] 0x0000000000021478
#
To overcome that we use:
# perf buildid-cache -v --add /usr/lib/debug/usr/bin/stress-ng-0.07.29-10.fc31.x86_64.debug
Adding f2ed02c68341183a124b9b0f6e2e6c493c465b29 /usr/lib/debug/usr/bin/stress-ng-0.07.29-10.fc31.x86_64.debug: Ok
#
# ls -la /root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf
-rw-r--r--. 3 root root 2401184 Jul 27 07:03 /root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf
# file /root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf
/root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf: ELF 64-bit LSB shared object, x86-64, version 1 (SYSV), dynamically linked, interpreter \004, BuildID[sha1]=f2ed02c68341183a124b9b0f6e2e6c493c465b29, for GNU/Linux 3.2.0, with debug_info, not stripped, too many notes (256)
#
Now it finally works:
# perf report | grep stress-ng | head -5
23.59% stress-ng-cpu stress-ng [.] ackermann
23.33% stress-ng-cpu stress-ng [.] is_prime
17.36% stress-ng-cpu stress-ng [.] stress_cpu_sieve
6.08% stress-ng-cpu stress-ng [.] stress_cpu_correlate
3.55% stress-ng-cpu stress-ng [.] queens_try
#
I'll make sure that it looks for the build-id keyed files in both the
"host" namespace (the namespace the user running 'perf record' was a the
time of the recording) and in the container namespace, as it shouldn't
matter where a content based key lookup finds the ELF file to use in
resolving symbols, etc.
Reported-by: Karl Rister <krister@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Brendan Gregg <brendan.d.gregg@gmail.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Krister Johansen <kjlx@templeofstupid.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Song Liu <songliubraving@fb.com>
Cc: Stanislav Fomichev <sdf@google.com>
Cc: Thomas-Mich Richter <tmricht@linux.vnet.ibm.com>
Fixes: 657ee55319 ("perf evlist: Introduce side band thread")
Link: https://lkml.kernel.org/n/tip-g79k0jz41adiaeuqud742t2l@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The event group feature links relevant hist entries among events so that
they can be displayed together. During the link process, each hist
entry in non-leader events is connected to a hist entry in the leader
event. This is done in order of events specified in the command line so
it assumes that events are linked in the order.
But 'perf top' can break the assumption since it does the link process
multiple times. For example, a hist entry can be in the third event
only at first so it's linked after the leader. Some time later, second
event has a hist entry for it and it'll be linked after the entry of the
third event.
This makes the code compilicated to deal with such unordered entries.
This patch simply unlink all the entries after it's printed so that they
can assume the correct order after the repeated link process. Also it'd
be easy to deal with decaying old entries IMHO.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Reported-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lkml.kernel.org/r/20190827231555.121411-2-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently perf top only decays entries in a selected evsel. I don't
know whether it's intended (maybe due to performance reason?) but anyway
it might show incorrect output when event group is used since users will
see leader event is decayed but others are not.
This patch moves the decay code into perf_top__resort_hists() so that
stdio and TUI code shared the logic.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lkml.kernel.org/r/20190827231555.121411-1-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
If PEBS declares ability to output its data to Intel PT stream, use the
aux_output attribute bit to enable PEBS data output to PT. This requires
a PT event to be present and scheduled in the same context. Unlike the
DS area, the kernel does not extract PEBS records from the PT stream to
generate corresponding records in the perf stream, because that would
require real time in-kernel PT decoding, which is not feasible. The PMI,
however, can still be used.
The output setting is per-CPU, so all PEBS events must be either writing
to PT or to the DS area, therefore, in case of conflict, the conflicting
event will fail to schedule, allowing the rotation logic to alternate
between the PEBS->PT and PEBS->DS events.
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: kan.liang@linux.intel.com
Link: https://lkml.kernel.org/r/20190806084606.4021-3-alexander.shishkin@linux.intel.com
In some cases, ordinary (non-AUX) events can generate data for AUX events.
For example, PEBS events can come out as records in the Intel PT stream
instead of their usual DS records, if configured to do so.
One requirement for such events is to consistently schedule together, to
ensure that the data from the "AUX output" events isn't lost while their
corresponding AUX event is not scheduled. We use grouping to provide this
guarantee: an "AUX output" event can be added to a group where an AUX event
is a group leader, and provided that the former supports writing to the
latter.
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: kan.liang@linux.intel.com
Link: https://lkml.kernel.org/r/20190806084606.4021-2-alexander.shishkin@linux.intel.com
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
perf report:
Andi Kleen:
- Make --ns time sort key output column wide enough for nanoseconds.
perf script:
Gustavo A. R. Silva:
- Fix memory leaks in list_scripts()
perf tests:
James Clark:
- Fixes hang in zstd compression test by changing the source of random data.
perf trace:
Arnaldo Carvalho de Melo:
- augmented_raw_syscalls.c BPF helper improvements.
Benjamin Peterson:
- Fix off-by-one error in ioctl cmd->string table.
libperf:
Jiri Olsa:
- Move most PERF_RECORD_ structs to perf/event.h.
headers:
Arnaldo Carvalho de Melo:
- Move cacheline related routines to separate source files.
- Move record_opts and other record declarations to separate files.
- Explicitly add some more needed headers here and there.
Souptick Joarder:
- Remove some duplicate include directives.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Move the PERF_RECORD_KSYMBOL event definition into libperf's event.h
header include.
In order to keep libperf simple, we switch 'u64/u32/u16/u8' types used
events to their generic '__u*' versions.
Perf added 'u*' types mainly to ease up printing __u64 values
as stated in the linux/types.h comment:
/*
* We define u64 as uint64_t for every architecture
* so that we can print it with "%"PRIx64 without getting warnings.
*
* typedef __u64 u64;
* typedef __s64 s64;
*/
Add and use new PRI_lu64 and PRI_lx64 macros for that. Use extra '_' to
ease up the reading and differentiate them from standard PRI*64 macros.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20190825181752.722-11-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Move the PERF_RECORD_THROTTLE event definition into libperf's event.h
header include.
In order to keep libperf simple, we switch 'u64/u32/u16/u8' types used
events to their generic '__u*' versions.
Perf added 'u*' types mainly to ease up printing __u64 values as stated
in the linux/types.h comment:
/*
* We define u64 as uint64_t for every architecture
* so that we can print it with "%"PRIx64 without getting warnings.
*
* typedef __u64 u64;
* typedef __s64 s64;
*/
Add and use new PRI_lu64 and PRI_lx64 macros for that. Use extra '_' to
ease up the reading and differentiate them from standard PRI*64 macros.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20190825181752.722-10-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Move the PERF_RECORD_READ event definition to libperf's event.h header
include.
In order to keep libperf simple, we switch 'u64/u32/u16/u8' types used
events to their generic '__u*' versions.
Perf added 'u*' types mainly to ease up printing __u64 values
as stated in the linux/types.h comment:
/*
* We define u64 as uint64_t for every architecture
* so that we can print it with "%"PRIx64 without getting warnings.
*
* typedef __u64 u64;
* typedef __s64 s64;
*/
Add and use new PRI_lu64 and PRI_lx64 macros for that. Use extra '_' to
ease up the reading and differentiate them from standard PRI*64 macros.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20190825181752.722-9-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Move the PERF_RECORD_LOST_SAMPLES event definition into libperf's
event.h header include.
In order to keep libperf simple, we switch 'u64/u32/u16/u8' types used
events to their generic '__u*' versions.
Perf added 'u*' types mainly to ease up printing __u64 values
as stated in the linux/types.h comment:
/*
* We define u64 as uint64_t for every architecture
* so that we can print it with "%"PRIx64 without getting warnings.
*
* typedef __u64 u64;
* typedef __s64 s64;
*/
Add and use new PRI_lu64 and PRI_lx64 macros for that. Use extra '_' to
ease up the reading and differentiate them from standard PRI*64 macros.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20190825181752.722-8-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
