| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
In the earlier fix for the memory overrun of id arrays I managed to typo
the wrong event in the fix.
Of course we need to close the current event in the loop, not the
original failing event.
The same test case as in the original patch still passes.
Fixes: 7834fa948beb ("perf evlist: Fix access of freed id arrays")
Signed-off-by: Andi Kleen <[email protected]>
Cc: Jiri Olsa <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
The build of file libperf-jvmti.so succeeds but the resulting
object fails to load:
# ~/linux/tools/perf/perf record -k mono -- java \
-XX:+PreserveFramePointer \
-agentpath:/root/linux/tools/perf/libperf-jvmti.so \
hog 100000 123450
Error occurred during initialization of VM
Could not find agent library /root/linux/tools/perf/libperf-jvmti.so
in absolute path, with error:
/root/linux/tools/perf/libperf-jvmti.so: undefined symbol: _ctype
Add the missing _ctype symbol into the build script.
Fixes: 79743bc927f6 ("perf jvmti: Link against tools/lib/string.o to have weak strlcpy()")
Signed-off-by: Thomas Richter <[email protected]>
Cc: Heiko Carstens <[email protected]>
Cc: Vasily Gorbik <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Just like strace has:
# trace -s sleep 1
Summary of events:
sleep (32370), 80 events, 93.0%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
nanosleep 1 0 1000.402 1000.402 1000.402 1000.402 0.00%
mmap 8 0 0.023 0.002 0.003 0.004 8.49%
close 5 0 0.015 0.001 0.003 0.009 51.39%
mprotect 4 0 0.014 0.002 0.003 0.005 16.95%
openat 3 0 0.013 0.003 0.004 0.005 14.29%
munmap 1 0 0.010 0.010 0.010 0.010 0.00%
read 4 0 0.005 0.001 0.001 0.002 16.83%
brk 4 0 0.004 0.001 0.001 0.002 20.82%
access 1 1 0.004 0.004 0.004 0.004 0.00%
fstat 3 0 0.003 0.001 0.001 0.001 12.17%
lseek 3 0 0.003 0.001 0.001 0.001 11.45%
arch_prctl 2 1 0.002 0.001 0.001 0.001 2.30%
execve 1 0 0.000 0.000 0.000 0.000 0.00%
#
# perf trace -S sleep 1
? ... [continued]: execve()) = 0
0.028 brk(brk: NULL) = 0x559f5bd96000
0.033 arch_prctl(option: 0x3001, arg2: 0x7ffda8b715a0) = -1 EINVAL (Invalid argument)
0.046 access(filename: "/etc/ld.so.preload", mode: R) = -1 ENOENT (No such file or directory)
0.055 openat(dfd: CWD, filename: "/etc/ld.so.cache", flags: RDONLY|CLOEXEC) = 3
0.060 fstat(fd: 3, statbuf: 0x7ffda8b707a0) = 0
0.062 mmap(addr: NULL, len: 134346, prot: READ, flags: PRIVATE, fd: 3, off: 0) = 0x7f3aedfc4000
0.066 close(fd: 3) = 0
0.079 openat(dfd: CWD, filename: "/lib64/libc.so.6", flags: RDONLY|CLOEXEC) = 3
0.085 read(fd: 3, buf: 0x7ffda8b70948, count: 832) = 832
0.088 lseek(fd: 3, offset: 792, whence: SET) = 792
0.090 read(fd: 3, buf: 0x7ffda8b70810, count: 68) = 68
0.093 fstat(fd: 3, statbuf: 0x7ffda8b707f0) = 0
0.095 mmap(addr: NULL, len: 8192, prot: READ|WRITE, flags: PRIVATE|ANONYMOUS) = 0x7f3aedfc2000
0.101 lseek(fd: 3, offset: 792, whence: SET) = 792
0.103 read(fd: 3, buf: 0x7ffda8b70450, count: 68) = 68
0.105 lseek(fd: 3, offset: 864, whence: SET) = 864
0.107 read(fd: 3, buf: 0x7ffda8b70470, count: 32) = 32
0.110 mmap(addr: NULL, len: 1857472, prot: READ, flags: PRIVATE|DENYWRITE, fd: 3, off: 0) = 0x7f3aeddfc000
0.114 mprotect(start: 0x7f3aede1e000, len: 1679360, prot: NONE) = 0
0.121 mmap(addr: 0x7f3aede1e000, len: 1363968, prot: READ|EXEC, flags: PRIVATE|FIXED|DENYWRITE, fd: 3, off: 0x22000) = 0x7f3aede1e000
0.127 mmap(addr: 0x7f3aedf6b000, len: 311296, prot: READ, flags: PRIVATE|FIXED|DENYWRITE, fd: 3, off: 0x16f000) = 0x7f3aedf6b000
0.131 mmap(addr: 0x7f3aedfb8000, len: 24576, prot: READ|WRITE, flags: PRIVATE|FIXED|DENYWRITE, fd: 3, off: 0x1bb000) = 0x7f3aedfb8000
0.138 mmap(addr: 0x7f3aedfbe000, len: 14272, prot: READ|WRITE, flags: PRIVATE|FIXED|ANONYMOUS) = 0x7f3aedfbe000
0.147 close(fd: 3) = 0
0.158 arch_prctl(option: SET_FS, arg2: 0x7f3aedfc3580) = 0
0.210 mprotect(start: 0x7f3aedfb8000, len: 16384, prot: READ) = 0
0.230 mprotect(start: 0x559f5b27d000, len: 4096, prot: READ) = 0
0.236 mprotect(start: 0x7f3aee00f000, len: 4096, prot: READ) = 0
0.240 munmap(addr: 0x7f3aedfc4000, len: 134346) = 0
0.300 brk(brk: NULL) = 0x559f5bd96000
0.302 brk(brk: 0x559f5bdb7000) = 0x559f5bdb7000
0.305 brk(brk: NULL) = 0x559f5bdb7000
0.310 openat(dfd: CWD, filename: "/usr/lib/locale/locale-archive", flags: RDONLY|CLOEXEC) = 3
0.315 fstat(fd: 3, statbuf: 0x7f3aedfbdac0) = 0
0.318 mmap(addr: NULL, len: 217750512, prot: READ, flags: PRIVATE, fd: 3, off: 0) = 0x7f3ae0e52000
0.325 close(fd: 3) = 0
0.358 nanosleep(rqtp: 0x7ffda8b714b0, rmtp: NULL) = 0
1000.622 close(fd: 1) = 0
1000.641 close(fd: 2) = 0
1000.664 exit_group(error_code: 0) = ?
Summary of events:
sleep (722), 80 events, 93.0%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
nanosleep 1 0 1000.194 1000.194 1000.194 1000.194 0.00%
mmap 8 0 0.025 0.002 0.003 0.005 10.17%
close 5 0 0.018 0.001 0.004 0.010 50.18%
mprotect 4 0 0.016 0.003 0.004 0.006 16.81%
openat 3 0 0.011 0.003 0.004 0.004 6.57%
munmap 1 0 0.010 0.010 0.010 0.010 0.00%
brk 4 0 0.005 0.001 0.001 0.002 20.72%
read 4 0 0.005 0.001 0.001 0.002 16.71%
access 1 1 0.005 0.005 0.005 0.005 0.00%
fstat 3 0 0.004 0.001 0.001 0.002 14.82%
lseek 3 0 0.003 0.001 0.001 0.001 11.66%
arch_prctl 2 1 0.002 0.001 0.001 0.001 3.59%
execve 1 0 0.000 0.000 0.000 0.000 0.00%
#
Works for system wide, e.g. for 1ms:
# perf trace -s -a sleep 0.001
Summary of events:
sleep (768), 94 events, 37.9%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
nanosleep 1 0 1.133 1.133 1.133 1.133 0.00%
execve 7 6 0.351 0.003 0.050 0.316 88.53%
mmap 8 0 0.024 0.002 0.003 0.004 8.86%
mprotect 4 0 0.017 0.003 0.004 0.006 16.02%
openat 3 0 0.013 0.004 0.004 0.005 8.34%
munmap 1 0 0.010 0.010 0.010 0.010 0.00%
brk 4 0 0.007 0.001 0.002 0.002 10.99%
close 5 0 0.005 0.001 0.001 0.002 11.69%
read 5 0 0.005 0.000 0.001 0.002 30.53%
access 1 1 0.004 0.004 0.004 0.004 0.00%
fstat 3 0 0.004 0.001 0.001 0.002 10.74%
lseek 3 0 0.003 0.001 0.001 0.001 10.20%
arch_prctl 2 1 0.002 0.001 0.001 0.001 3.34%
Web Content (21258), 46 events, 18.5%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
recvmsg 12 12 0.015 0.001 0.001 0.002 8.50%
futex 2 0 0.008 0.003 0.004 0.005 27.08%
poll 6 0 0.006 0.000 0.001 0.002 22.14%
read 2 0 0.006 0.002 0.003 0.003 26.08%
write 1 0 0.002 0.002 0.002 0.002 0.00%
Web Content (4365), 36 events, 14.5%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
recvmsg 10 10 0.015 0.001 0.002 0.003 11.83%
poll 5 0 0.006 0.000 0.001 0.002 28.44%
futex 2 0 0.005 0.001 0.003 0.004 48.29%
read 1 0 0.003 0.003 0.003 0.003 0.00%
Timer (21275), 14 events, 5.6%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
futex 6 1 0.240 0.000 0.040 0.149 64.58%
write 1 0 0.008 0.008 0.008 0.008 0.00%
Timer (4383), 14 events, 5.6%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
futex 6 2 0.186 0.000 0.031 0.181 96.45%
write 1 0 0.010 0.010 0.010 0.010 0.00%
Web Content (20354), 28 events, 11.3%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
recvmsg 8 8 0.010 0.001 0.001 0.002 15.24%
poll 4 0 0.004 0.000 0.001 0.002 35.68%
futex 1 0 0.003 0.003 0.003 0.003 0.00%
read 1 0 0.003 0.003 0.003 0.003 0.00%
Timer (20371), 10 events, 4.0%
syscall calls errors total min avg max stddev
(msec) (msec) (msec) (msec) (%)
--------------- -------- ------ -------- --------- --------- --------- ------
futex 4 1 0.077 0.000 0.019 0.075 95.46%
write 1 0 0.005 0.005 0.005 0.005 0.00%
[root@quaco ~]#
Cc: Adrian Hunter <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: Brendan Gregg <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Luis Cláudio Gonçalves <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: https://lkml.kernel.org/n/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
'perf record' has supported --all-kernel / --all-user to configure all
used events to run in kernel space or run in user space. But 'perf stat'
doesn't support these options.
It would be useful to support these options in 'perf stat' too to keep
the same semantics available in both tools.
Signed-off-by: Jin Yao <[email protected]>
Acked-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
The build of file libperf-jvmti.so succeeds but the resulting
object fails to load:
# ~/linux/tools/perf/perf record -k mono -- java \
-XX:+PreserveFramePointer \
-agentpath:/root/linux/tools/perf/libperf-jvmti.so \
hog 100000 123450
Error occurred during initialization of VM
Could not find agent library /root/linux/tools/perf/libperf-jvmti.so
in absolute path, with error:
/root/linux/tools/perf/libperf-jvmti.so: undefined symbol: _ctype
Add the missing _ctype symbol into the build script.
Fixes: 79743bc927f6 ("perf jvmti: Link against tools/lib/string.o to have weak strlcpy()")
Signed-off-by: Thomas Richter <[email protected]>
Cc: Heiko Carstens <[email protected]>
Cc: Vasily Gorbik <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Remove redirection of objdump's stderr to /dev/null to help diagnose
failures.
Fix the '--no-show-raw' flag to be '--no-show-raw-insn' which binutils
is permissive and allows, but fails with LLVM objdump.
Signed-off-by: Ian Rogers <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Jin Yao <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Song Liu <[email protected]>
Cc: Stephane Eranian <[email protected]>
Cc: [email protected]
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Avoiding a pipe allows objdump command failures to surface. Move to the
caller of symbol__parse_objdump_line the call to strim that removes
leading and trailing tabs. Add a new expand_tabs function that if a tab
is present allocate a new line in which tabs are expanded. In
symbol__parse_objdump_line the line had no leading spaces, so simplify
the line_ip processing.
Signed-off-by: Ian Rogers <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Jin Yao <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Song Liu <[email protected]>
Cc: Stephane Eranian <[email protected]>
Cc: [email protected]
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Simplify the objdump command by not piping the output of objdump through
grep. Instead, drop lines that match the grep pattern during the reading
loop.
Signed-off-by: Ian Rogers <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Jin Yao <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Song Liu <[email protected]>
Cc: Stephane Eranian <[email protected]>
Cc: [email protected]
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Reduce duplicated logic by using the subcmd library. Ensure when errors
occur they are reported to the caller. Before this patch, if no lines
are read the error status is 0.
Signed-off-by: Ian Rogers <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Jin Yao <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Song Liu <[email protected]>
Cc: Stephane Eranian <[email protected]>
Cc: [email protected]
Link: http://lore.kernel.org/lkml/[email protected]
Link: http://lore.kernel.org/lkml/[email protected]
[ merged follow up fix for NULL termination as in the 2nd link above ]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Objdump output is parsed using getline which allocates memory for the
read. Getline will realloc if the memory is too small, but currently the
line is always freed after the call.
Simplify parse_objdump_line by performing the reading in symbol__disassemble.
Signed-off-by: Ian Rogers <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Jin Yao <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Song Liu <[email protected]>
Cc: Stephane Eranian <[email protected]>
Cc: [email protected]
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
We received a user report that call-graph DWARF mode was enabled in
'perf record' but 'perf report' didn't unwind the callstack correctly.
The reason was, libunwind was not compiled in.
We can use 'perf -vv' to check the compiled libraries but it would be
valuable to report a warning to user directly (especially valuable for
a perf newbie).
The warning is:
Warning:
Please install libunwind development packages during the perf build.
Both TUI and stdio are supported.
Signed-off-by: Jin Yao <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
When executing the task exit testing case, perf gets stuck in an endless
loop this case and doesn't return back on Arm64 Juno board.
After digging into this issue, since Juno board has Arm's big.LITTLE
CPUs, thus the PMUs are not compatible between the big CPUs and little
CPUs. This leads to a PMU event that cannot be enabled properly when
the traced task is migrated from one variant's CPU to another variant.
Finally, the test case runs into infinite loop for cannot read out any
event data after return from polling.
Eventually, we need to work out formal solution to allow PMU events can
be freely migrated from one CPU variant to another, but this is a
difficult task and a different topic. This patch tries to fix the Perf
test case to avoid infinite loop, when the testing detects 1000 times
retrying for reading empty events, it will directly bail out and return
failure. This allows the Perf tool can continue its other test cases.
Signed-off-by: Leo Yan <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
When fail to mmap events in task exit case, it misses to set 'err' to
-1; thus the testing will not report failure for it.
This patch sets 'err' to -1 when fails to mmap events, thus Perf tool
can report correct result.
Fixes: d723a55096b8 ("perf test: Add test case for checking number of EXIT events")
Signed-off-by: Leo Yan <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
In the earlier fix for the memory overrun of id arrays I managed to typo
the wrong event in the fix.
Of course we need to close the current event in the loop, not the
original failing event.
The same test case as in the original patch still passes.
Fixes: 7834fa948beb ("perf evlist: Fix access of freed id arrays")
Signed-off-by: Andi Kleen <[email protected]>
Cc: Jiri Olsa <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
My earlier patch to just enable --reltime with --time was a little too
optimistic. The --time parsing would accept absolute time, which is
very confusing to the user.
Support relative time in --time parsing too. This only works with recent
perf record that records the first sample time. Otherwise we error out.
Fixes: 3714437d3fcc ("perf script: Allow --time with --reltime")
Signed-off-by: Andi Kleen <[email protected]>
Cc: Jiri Olsa <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
By using "make TCMALLOC=1" you can enable perf to be build for usage
with libtcmalloc.so (gperftools).
Get heap profile (tools/perf directory):
$ <install gperftools>
$ make TCMALLOC=1 DEBUG=1
$ HEAPPROFILE=/tmp/heapprof ./perf ...
$ pprof ./perf /tmp/heapprof.000*
(pprof) top
Total: 2335.5 MB
1735.1 74.3% 74.3% 1735.1 74.3% memdup
402.0 17.2% 91.5% 402.0 17.2% zalloc
140.2 6.0% 97.5% 145.8 6.2% map__new
33.6 1.4% 98.9% 33.6 1.4% symbol__new
12.4 0.5% 99.5% 12.4 0.5% alloc_event
6.2 0.3% 99.7% 6.2 0.3% nsinfo__new
5.5 0.2% 100.0% 5.5 0.2% nsinfo__copy
0.3 0.0% 100.0% 0.3 0.0% dso__new
0.1 0.0% 100.0% 0.1 0.0% do_read_string
0.0 0.0% 100.0% 0.0 0.0% __GI__IO_file_doallocate
See callstack:
$ pprof --pdf ./perf /tmp/heapprof.00* > callstack.pdf
$ pprof --web ./perf /tmp/heapprof.00*
Committer testing:
Install gperftools, on fedora:
# dnf install gperftools-devel
Then build:
$ make TCMALLOC=1 DEBUG=1 -C tools/perf O=/tmp/build/perf install-bin
Verify that it linked against the right library:
$ ldd ~/bin/perf | grep tcma
libtcmalloc.so.4 => /lib64/libtcmalloc.so.4 (0x00007fb2953a7000)
$
Run 'perf trace' system wide for 1 minute:
# HEAPPROFILE=/tmp/heapprof perf trace -a sleep 1m
<SNIP>
59985.524 ( 0.006 ms): Web Content/20354 recvmsg(fd: 9<socket:[1762817]>, msg: 0x7ffee5fdafb0) = -1 EAGAIN (Resource temporarily unavailable)
59985.536 ( 0.005 ms): Web Content/20354 recvmsg(fd: 9<socket:[1762817]>, msg: 0x7ffee5fdafc0) = -1 EAGAIN (Resource temporarily unavailable)
59981.956 (10.143 ms): SCTP timer/21716 ... [continued]: select()) = 0 (Timeout)
59985.549 ( ): Web Content/20354 poll(ufds: 0x7f1df38af180, nfds: 3, timeout_msecs: 4294967295) ...
0.926 (59999.481 ms): sleep/29764 ... [continued]: nanosleep()) = 0
59992.133 ( ): SCTP timer/21716 select(tvp: 0x7ff5bf7fee80) ...
60000.477 ( 0.009 ms): sleep/29764 close(fd: 1) = 0
60000.493 ( 0.005 ms): sleep/29764 close(fd: 2) = 0
60000.514 ( ): sleep/29764 exit_group() = ?
Dumping heap profile to /tmp/heapprof.0001.heap (Exiting, 3 MB in use)
[root@quaco ~]#
Install pprof:
# dnf install pprof
And run it:
# pprof ~/bin/perf /tmp/heapprof.0001.heap
Using local file /root/bin/perf.
Using local file /tmp/heapprof.0001.heap.
Welcome to pprof! For help, type 'help'.
(pprof) top
Total: 4.0 MB
1.7 42.0% 42.0% 2.2 54.1% map__new
0.9 23.3% 65.3% 0.9 23.3% zalloc
0.5 11.4% 76.7% 0.5 11.4% dso__new
0.2 5.6% 82.3% 0.3 8.5% trace__sys_enter
0.2 4.9% 87.2% 0.2 4.9% __GI___strdup
0.2 3.8% 91.0% 0.2 3.8% new_term
0.1 2.2% 93.2% 0.4 10.1% __perf_pmu__new_alias
0.0 1.0% 94.3% 0.0 1.2% event_read_fields
0.0 0.8% 95.1% 0.0 0.8% nsinfo__new
0.0 0.7% 95.8% 0.1 3.2% trace__read_syscall_info
(pprof)
Signed-off-by: Jiri Olsa <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
This patch prints the stddev and hist for the cycles diff of program
block. It can help us to understand if the cycles is noisy or not.
This patch is inspired by Andi Kleen's patch:
https://lwn.net/Articles/600471/
We create new option '--cycles-hist'.
Example:
perf record -b ./div
perf record -b ./div
perf diff -c cycles
# Baseline [Program Block Range] Cycles Diff Shared Object Symbol
# ........ .......................................................... .... ................. ............................
#
46.72% [div.c:40 -> div.c:40] 0 div [.] main
46.72% [div.c:42 -> div.c:44] 0 div [.] main
46.72% [div.c:42 -> div.c:39] 0 div [.] main
20.54% [random_r.c:357 -> random_r.c:394] 1 libc-2.27.so [.] __random_r
20.54% [random_r.c:357 -> random_r.c:380] 0 libc-2.27.so [.] __random_r
20.54% [random_r.c:388 -> random_r.c:388] 0 libc-2.27.so [.] __random_r
20.54% [random_r.c:388 -> random_r.c:391] 0 libc-2.27.so [.] __random_r
17.04% [random.c:288 -> random.c:291] 0 libc-2.27.so [.] __random
17.04% [random.c:291 -> random.c:291] 0 libc-2.27.so [.] __random
17.04% [random.c:293 -> random.c:293] 0 libc-2.27.so [.] __random
17.04% [random.c:295 -> random.c:295] 0 libc-2.27.so [.] __random
17.04% [random.c:295 -> random.c:295] 0 libc-2.27.so [.] __random
17.04% [random.c:298 -> random.c:298] 0 libc-2.27.so [.] __random
8.40% [div.c:22 -> div.c:25] 0 div [.] compute_flag
8.40% [div.c:27 -> div.c:28] 0 div [.] compute_flag
5.14% [rand.c:26 -> rand.c:27] 0 libc-2.27.so [.] rand
5.14% [rand.c:28 -> rand.c:28] 0 libc-2.27.so [.] rand
2.15% [rand@plt+0 -> rand@plt+0] 0 div [.] rand@plt
0.00% [kernel.kallsyms] [k] __x86_indirect_thunk_rax
0.00% [do_mmap+714 -> do_mmap+732] -10 [kernel.kallsyms] [k] do_mmap
0.00% [do_mmap+737 -> do_mmap+765] 1 [kernel.kallsyms] [k] do_mmap
0.00% [do_mmap+262 -> do_mmap+299] 0 [kernel.kallsyms] [k] do_mmap
0.00% [__x86_indirect_thunk_r15+0 -> __x86_indirect_thunk_r15+0] 7 [kernel.kallsyms] [k] __x86_indirect_thunk_r15
0.00% [native_sched_clock+0 -> native_sched_clock+119] -1 [kernel.kallsyms] [k] native_sched_clock
0.00% [native_write_msr+0 -> native_write_msr+16] -13 [kernel.kallsyms] [k] native_write_msr
When we enable the option '--cycles-hist', the output is
perf diff -c cycles --cycles-hist
# Baseline [Program Block Range] Cycles Diff stddev/Hist Shared Object Symbol
# ........ .......................................................... .... ................. ................. ............................
#
46.72% [div.c:40 -> div.c:40] 0 ± 37.8% ▁█▁▁██▁█ div [.] main
46.72% [div.c:42 -> div.c:44] 0 ± 49.4% ▁▁▂█▂▂▂▂ div [.] main
46.72% [div.c:42 -> div.c:39] 0 ± 24.1% ▃█▂▄▁▃▂▁ div [.] main
20.54% [random_r.c:357 -> random_r.c:394] 1 ± 33.5% ▅▂▁█▃▁▂▁ libc-2.27.so [.] __random_r
20.54% [random_r.c:357 -> random_r.c:380] 0 ± 39.4% ▁▁█▁██▅▁ libc-2.27.so [.] __random_r
20.54% [random_r.c:388 -> random_r.c:388] 0 libc-2.27.so [.] __random_r
20.54% [random_r.c:388 -> random_r.c:391] 0 ± 41.2% ▁▃▁▂█▄▃▁ libc-2.27.so [.] __random_r
17.04% [random.c:288 -> random.c:291] 0 ± 48.8% ▁▁▁▁███▁ libc-2.27.so [.] __random
17.04% [random.c:291 -> random.c:291] 0 ±100.0% ▁█▁▁▁▁▁▁ libc-2.27.so [.] __random
17.04% [random.c:293 -> random.c:293] 0 ±100.0% ▁█▁▁▁▁▁▁ libc-2.27.so [.] __random
17.04% [random.c:295 -> random.c:295] 0 ±100.0% ▁█▁▁▁▁▁▁ libc-2.27.so [.] __random
17.04% [random.c:295 -> random.c:295] 0 libc-2.27.so [.] __random
17.04% [random.c:298 -> random.c:298] 0 ± 75.6% ▃█▁▁▁▁▁▁ libc-2.27.so [.] __random
8.40% [div.c:22 -> div.c:25] 0 ± 42.1% ▁▃▁▁███▁ div [.] compute_flag
8.40% [div.c:27 -> div.c:28] 0 ± 41.8% ██▁▁▄▁▁▄ div [.] compute_flag
5.14% [rand.c:26 -> rand.c:27] 0 ± 37.8% ▁▁▁████▁ libc-2.27.so [.] rand
5.14% [rand.c:28 -> rand.c:28] 0 libc-2.27.so [.] rand
2.15% [rand@plt+0 -> rand@plt+0] 0 div [.] rand@plt
0.00% [kernel.kallsyms] [k] __x86_indirect_thunk_rax
0.00% [do_mmap+714 -> do_mmap+732] -10 [kernel.kallsyms] [k] do_mmap
0.00% [do_mmap+737 -> do_mmap+765] 1 [kernel.kallsyms] [k] do_mmap
0.00% [do_mmap+262 -> do_mmap+299] 0 [kernel.kallsyms] [k] do_mmap
0.00% [__x86_indirect_thunk_r15+0 -> __x86_indirect_thunk_r15+0] 7 [kernel.kallsyms] [k] __x86_indirect_thunk_r15
0.00% [native_sched_clock+0 -> native_sched_clock+119] -1 ± 38.5% ▄█▁ [kernel.kallsyms] [k] native_sched_clock
0.00% [native_write_msr+0 -> native_write_msr+16] -13 ± 47.1% ▁█▇▃▁▁ [kernel.kallsyms] [k] native_write_msr
v8:
---
Rebase to perf/core branch
v7:
---
1. v6 got Jiri's ACK.
2. Rebase to latest perf/core branch.
v6:
---
1. Jiri provides better code for using data__hpp_register() in ui_init().
Use this code in v6.
v5:
---
1. Refine the use of data__hpp_register() in ui_init() according to
Jiri's suggestion.
v4:
---
1. Rename the new option from '--noisy' to '--cycles-hist'
2. Remove the option '-n'.
3. Only update the spark value and stats when '--cycles-hist' is enabled.
4. Remove the code of printing '..'.
v3:
---
1. Move the histogram to a separate column
2. Move the svals[] out of struct stats
v2:
---
Jiri got a compile error,
CC builtin-diff.o
builtin-diff.c: In function ‘compute_cycles_diff’:
builtin-diff.c:712:10: error: taking the absolute value of unsigned type ‘u64’ {aka ‘long unsigned int’} has no effect [-Werror=absolute-value]
712 | labs(pair->block_info->cycles_spark[i] -
| ^~~~
Because the result of u64 - u64 is still u64. Now we change the type of
cycles_spark[] to s64.
Signed-off-by: Jin Yao <[email protected]>
Acked-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Andi reported that 'make DEBUG=1' does not propagate to the libbperf
code. It's true also for the other flags. Changing the code to propagate
the global build flags to libperf compilation.
Reported-by: Andi Kleen <[email protected]>
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Introduce the perf_evlist__filter_pollfd function and export it in the
perf/evlist.h header, so that libperf users can check if the descriptor
is still alive.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add a static perf_evlist__purge() function to purge evsels from a evlist.
Add also perf_evlist__for_each_entry_safe() which is used by
perf_evlist__purge().
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add the perf_evlist__exit() function, so far it's not exported and added
only for internal use for perf and libperf.
USe it to release cpus/threads and pollfd array.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
It's needed in libperf only, so move it to the perf_evlist__mmap_ops()
function.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Currently when a new map is mmapped we set its refcnt to 2 in the
perf_evlist_mmap_ops::mmap callback.
Every mmap gets its refcnt set to 2 when it's first mmaped:
- 1 for the current user, which will be taken out by a call to
perf_evlist__munmap_filtered(), where we find out there's
no more data comming from kernel to this mmap.
- 1 for the drain code where in perf_mmap__consume() the mmap
is released if it is empty.
Move this common setup into libperf's generic code before the mmap
callback is called.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Switch to the libperf mmap interface by calling directly
perf_evlist__mmap_ops() and removing perf's evlist__mmap_per_*
functions.
By switching to libperf perf_evlist__mmap() we need to operate over
'struct perf_mmap' in evlist__add_pollfd, so make the related changes
there.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add the perf_evlist__mmap_cb_mmap() function to call perf specific
mmap__mmap() function during perf_evlist__mmap_ops() call.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add the perf_evlist__mmap_cb_get() function to return 'struct perf_mmap'
object during perf_evlist__mmap_ops() call.
The array of 'struct mmap' is allocated via evlist__alloc_mmap(), in
this callback we simply returns pointer to the base object.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add perf_evlist__mmap_cb_idx function to call auxtrace_mmap_params__set_idx()
on each new index during perf_evlist__mmap_ops call.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add the perf_evlist_mmap_ops::mmap callback to be called in
mmap_per_evsel() to actually mmap the map.
Add libperf's perf_evlist__mmap_cb_mmap() function as libperf's mmap
callback.
New mmaped map gets refcount set to 2 in mmap__mmap(), we follow that in
mmap callback. We will move this to common place after we switch to
perf_evlist__mmap().
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add the perf_evlist_mmap_ops::get callback to be called in
mmap_per_evsel() to get/allocate the 'struct perf_mmap' object.
Add the libperf's perf_evlist__mmap_cb_get() function as libperf's get
callback.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add the perf_evlist_mmap_ops::idx callback to be called in
mmap_per_cpu() and mmap_per_thread() with current cpu and thread
indexes.
It's used by current aux code, so perf will use this callback to set the
aux index.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
To be able to pass specific callbacks to evlist's mmap.
There will be a specific call to this function from perf's
evlist__mmap() and libperf's perf_evlist__mmap() functions in following
changes.
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Jiri Olsa <[email protected]>
|
|
Add libperf's version of perf_evlist__mmap()/munmap() functions and
exporting them in the perf/evlist.h header.
It's the backbone of what we have in perf code. The following changes
will add needed callbacks and then we'll finally switch the perf code to
use libperf's version.
Add mmap/mmap_ovw 'struct perf_mmap' object arrays to hold maps for
libperf's evlist.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__read_event() from tools/perf to libperf and export it in
the perf/mmap.h header.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__read_init() from tools/perf to libperf and export it in
the perf/mmap.h header.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__read_init() from tools/perf to libperf and export it in
perf/mmap.h header.
And add pr_debug2()/pr_debug3() macros support, because the code is
using them.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__consume() vrom tools/perf to libperf and export it in
the perf/mmap.h header.
Move also the needed helpers perf_mmap__write_tail(),
perf_mmap__read_head() and perf_mmap__empty().
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
We will move this code to libperf shortly, so we need to free it of
'struct auxtrace_mmap' usage, because it won't be available in libperf
(for now).
The perf_event_mmap_page::aux_size is set when the aux mmap is mapped,
so the check is equivalent.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__put() from tools/perf to libperf.
Once perf_mmap__put() is moved, we need a way to call application
related unmap code (AIO and aux related code for eprf), when the map
goes away.
Add the perf_mmap::unmap callback to do that.
The unmap path from perf is:
perf_mmap__put (libperf)
perf_mmap__munmap (libperf)
map->unmap_cb -> perf_mmap__unmap_cb (perf)
mmap__munmap (perf)
Committer notes:
Add missing linux/kernel.h to tools/perf/lib/mmap.c to get the BUG_ON
definition.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__unmap() from tools/perf to libperf, to internal header
internal/mmap.h. It will be used in the following patches. And rename
the existing perf's function to mmap__munmap().
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__get() from tools/perf to libperf in the internal header
internal/mmap.h.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__mmap() from tools/perf to libperf, it will be used in
the following patches. And rename the existing perf's function to
mmap__mmap().
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Move perf_mmap__mmap_len() from tools/perf wto libperf, it will be used
in the following patches. And rename the existing perf's function to
mmap__mmap_len().
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add libperf's version of mmap params 'struct perf_mmap_param' object
with the basics: 'prot' and 'mask'. Encapsulate it in the current
'struct mmap_params' object.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Add perf_mmap__init() function to initialize 'struct perf_mmap' objects.
Add it to a new mmap.c source file, that will carry all the mmap related
functions.
Signed-off-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Michael Petlan <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Being const + weak breaks with some compilers that constant-propagate
from the weak symbol. This behavior is outside of the specification, but
in LLVM is chosen to match GCC's behavior.
LLVM's implementation was set in this patch:
https://github.com/llvm/llvm-project/commit/f49573d1eedcf1e44893d5a062ac1b72c8419646
A const + weak symbol is set to be weak_odr:
https://llvm.org/docs/LangRef.html
ODR is one definition rule, and given there is one constant definition
constant-propagation is possible. It is possible to get this code to
miscompile with LLVM when applying link time optimization. As compilers
become more aggressive, this is likely to break in more instances.
Move the definition of sample_reg_masks to the conditional part of
perf_regs.h and guard usage with HAVE_PERF_REGS_SUPPORT. This avoids the
weak symbol.
Fix an issue when HAVE_PERF_REGS_SUPPORT isn't defined from patch v1.
In v3, add perf_regs.c for architectures that HAVE_PERF_REGS_SUPPORT but
don't declare sample_regs_masks.
Further notes:
Jiri asked:
"Is this just a precaution or you actualy saw some breakage?"
Ian answered:
"We saw a breakage with clang with thinlto enabled for linking. Our
compiler team had recently seen, and were surprised by, a similar issue
and were able to dig out the weak ODR issue."
Signed-off-by: Ian Rogers <[email protected]>
Reviewed-by: Nick Desaulniers <[email protected]>
Acked-by: Jiri Olsa <[email protected]>
Cc: Albert Ou <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Alexey Budankov <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: [email protected]
Cc: Guo Ren <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: [email protected]
Cc: Mao Han <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Palmer Dabbelt <[email protected]>
Cc: Paul Walmsley <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Stephane Eranian <[email protected]>
Link: http://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Continuing from the previous cset comment, now that filter expression
works:
# perf trace -e msr:* --filter="msr!=FS_BASE && msr != IA32_TSC_DEADLINE && msr != 0x830 && msr != 0x83f && msr !=IA32_SPEC_CTRL" --filter-pids 3750
0.000 Timer/5033 msr:write_msr(msr: SYSCALL_MASK, val: 292608)
0.009 Timer/5033 msr:write_msr(msr: LSTAR, val: -1398800368)
0.010 Timer/5033 msr:write_msr(msr: TSC_AUX, val: 4)
0.050 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
45.661 gnome-terminal/12595 msr:write_msr(msr: SYSCALL_MASK, val: 292608)
45.672 gnome-terminal/12595 msr:write_msr(msr: LSTAR, val: -1398800368)
45.675 gnome-terminal/12595 msr:write_msr(msr: TSC_AUX, val: 3)
54.852 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
130.508 Timer/4050 msr:write_msr(msr: SYSCALL_MASK, val: 292608)
130.527 Timer/4050 msr:write_msr(msr: LSTAR, val: -1398800368)
130.531 Timer/4050 msr:write_msr(msr: TSC_AUX, val: 3)
140.924 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
164.738 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
603.578 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
620.809 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
690.115 JS Watchdog/4259 msr:write_msr(msr: SYSCALL_MASK, val: 292608)
690.136 JS Watchdog/4259 msr:write_msr(msr: LSTAR, val: -1398800368)
690.141 JS Watchdog/4259 msr:write_msr(msr: TSC_AUX, val: 3)
690.186 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
759.016 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
^C[root@quaco ~]#
Or look at the first 3 write_msr events for that IA32_TSC_DEADLINE to learn why
it happens so often:
# perf trace --max-events=3 --max-stack=8 -e msr:* --filter="msr==IA32_TSC_DEADLINE" --filter-pids 3750
0.000 :0/0 msr:write_msr(msr: IA32_TSC_DEADLINE, val: 19296732550862)
do_trace_write_msr ([kernel.kallsyms])
do_trace_write_msr ([kernel.kallsyms])
lapic_next_deadline ([kernel.kallsyms])
clockevents_program_event ([kernel.kallsyms])
hrtimer_interrupt ([kernel.kallsyms])
smp_apic_timer_interrupt ([kernel.kallsyms])
apic_timer_interrupt ([kernel.kallsyms])
cpuidle_enter_state ([kernel.kallsyms])
32.646 :0/0 msr:write_msr(msr: IA32_TSC_DEADLINE, val: 19296800134158)
do_trace_write_msr ([kernel.kallsyms])
do_trace_write_msr ([kernel.kallsyms])
lapic_next_deadline ([kernel.kallsyms])
clockevents_program_event ([kernel.kallsyms])
hrtimer_start_range_ns ([kernel.kallsyms])
tick_nohz_restart_sched_tick ([kernel.kallsyms])
tick_nohz_idle_exit ([kernel.kallsyms])
do_idle ([kernel.kallsyms])
32.802 :0/0 msr:write_msr(msr: IA32_TSC_DEADLINE, val: 19297507436922)
do_trace_write_msr ([kernel.kallsyms])
do_trace_write_msr ([kernel.kallsyms])
lapic_next_deadline ([kernel.kallsyms])
clockevents_program_event ([kernel.kallsyms])
hrtimer_try_to_cancel ([kernel.kallsyms])
hrtimer_cancel ([kernel.kallsyms])
tick_nohz_restart_sched_tick ([kernel.kallsyms])
tick_nohz_idle_exit ([kernel.kallsyms])
#
And if some of the strings can't be found:
# trace -e msr:* --filter="msr!=SPECULATIVE_EXECUTION_PROBLEMS_SOLUTION && msr != IA32_TSC_DEADLINE && msr != 0x830 && msr != 0x83f && msr !=IA32_SPEC_CTRL" --filter-pids 3750
"SPECULATIVE_EXECUTION_PROBLEMS_SOLUTION" not found for "msr" in "msr:read_msr", can't set filter "(msr!=SPECULATIVE_EXECUTION_PROBLEMS_SOLUTION && msr != IA32_TSC_DEADLINE && msr != 0x830 && msr != 0x83f && msr !=IA32_SPEC_CTRL) && (common_pid != 28131 && common_pid != 3750)"
#
Next step is to automatically wire up the pre-existing strarrays, which there
are quite a few.
The strtoul() methods will be further enhanced to allow for looking at other
arguments in a syscall/tracepoint, just like going from integer to string
(scnprintf methods), so that those "val" lines for the msr tracepoints can be
properly formatted or even resolved into some string.
Cc: Adrian Hunter <[email protected]>
Cc: Brendan Gregg <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Luis Cláudio Gonçalves <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: https://lkml.kernel.org/n/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
So that one can try things like:
# perf trace -e msr:* --filter="msr!=FS_BASE && msr != IA32_TSC_DEADLINE && msr != 0x830 && msr != 0x83f && msr !=IA32_SPEC_CTRL" --filter-pids 3750
That, at this point in the patchset, without any strtoul in place for
tracepoint arguments, will result in:
No resolver (strtoul) for "msr" in "msr:read_msr", can't set filter "(msr!=FS_BASE && msr != IA32_TSC_DEADLINE && msr != 0x830 && msr != 0x83f && msr !=IA32_SPEC_CTRL) && (common_pid != 25407 && common_pid != 3750)"
#
See you in the next cset!
Cc: Adrian Hunter <[email protected]>
Cc: Brendan Gregg <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Luis Cláudio Gonçalves <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: https://lkml.kernel.org/n/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
And also for 'struct strarray', since its needed to implement
strarrays__strtoul(). This just traverses the entries and when finding a
match, returns (offset + index), i.e. the value associated with the
searched string.
E.g. "EFER" (MSR_EFER) returns:
# grep -w EFER -B2 /tmp/build/perf/trace/beauty/generated/x86_arch_MSRs_array.c
#define x86_64_specific_MSRs_offset 0xc0000080
static const char *x86_64_specific_MSRs[] = {
[0xc0000080 - x86_64_specific_MSRs_offset] = "EFER",
#
0xc0000080
This will be auto-attached to 'struct syscall_arg_fmt' entries
associated with strarrays as soon as we add a ->strarray and ->strarrays
to 'struct syscall_arg_fmt'.
Cc: Adrian Hunter <[email protected]>
Cc: Brendan Gregg <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Luis Cláudio Gonçalves <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: https://lkml.kernel.org/n/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
This will go from a string to a number, so that filter expressions can
be constructed with strings and then, before applying the tracepoint
filters (or eBPF, in the future) we can map those strings to numbers.
The first one will be for 'msr' tracepoint arguments, but real quickly
we will be able to reuse all strarrays for that.
Cc: Adrian Hunter <[email protected]>
Cc: Brendan Gregg <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Luis Cláudio Gonçalves <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: https://lkml.kernel.org/n/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|
|
Similar to what is in 'perf record', works just like there:
# perf trace -e msr:*
328.297 :0/0 msr:write_msr(msr: FS_BASE, val: 140240388381888)
328.302 :0/0 msr:write_msr(msr: FS_BASE, val: 140240388381888)
328.306 :0/0 msr:write_msr(msr: FS_BASE, val: 140240388381888)
328.317 :0/0 msr:write_msr(msr: FS_BASE, val: 140240388381888)
328.322 :0/0 msr:write_msr(msr: FS_BASE, val: 140240388381888)
328.327 :0/0 msr:write_msr(msr: FS_BASE, val: 140240388381888)
328.331 :0/0 msr:write_msr(msr: FS_BASE, val: 140240388381888)
328.336 :0/0 msr:write_msr(msr: FS_BASE, val: 140240388381888)
328.340 :0/0 ^Cmsr:write_msr(msr: FS_BASE, val: 140240388381888)
#
So, for a system wide trace session looking at the write_msr tracepoint
we see a flood of MSR_FS_BASE, we need to get the number for that:
# grep FS_BASE /tmp/build/perf/trace/beauty/generated/x86_arch_MSRs_array.c
[0xc0000100 - x86_64_specific_MSRs_offset] = "FS_BASE",
#
And then use it in a filter:
# perf trace -e msr:* --filter="msr!=0xc0000100"
<SNIP>
942.177 :0/0 msr:write_msr(msr: IA32_TSC_DEADLINE, val: 3056931068232)
942.199 :0/0 msr:write_msr(msr: IA32_TSC_DEADLINE, val: 3057135655252)
942.203 :0/0 msr:write_msr(msr: IA32_TSC_DEADLINE, val: 3056931068222)
942.231 :0/0 msr:write_msr(msr: IA32_TSC_DEADLINE, val: 3056998373022)
942.241 :0/0 msr:write_msr(msr: IA32_TSC_DEADLINE, val: 3056931068236)
<SNIP>
#
Ok, lets filter that too, too noisy:
# grep TSC_DEADLINE /tmp/build/perf/trace/beauty/generated/x86_arch_MSRs_array.c
[0x000006E0] = "IA32_TSC_DEADLINE",
#
# perf trace -e msr:* --filter="msr!=0xc0000100 && msr!=0x6e0" -a sleep 0.1
0.000 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
0.066 CPU 0/KVM/4895 msr:write_msr(msr: IA32_SPEC_CTRL, val: 6)
0.070 CPU 0/KVM/4895 msr:write_msr(msr: 0x830, val: 34359740667)
0.099 CPU 0/KVM/4895 msr:read_msr(msr: IA32_SYSENTER_ESP, val: -2199021993472)
0.100 CPU 0/KVM/4895 msr:read_msr(msr: IA32_APICBASE, val: 4276096000)
0.101 CPU 0/KVM/4895 msr:read_msr(msr: IA32_DEBUGCTLMSR)
0.109 :0/0 msr:write_msr(msr: IA32_SPEC_CTRL)
1.000 :0/0 msr:write_msr(msr: 0x830, val: 17179871485)
18.893 :0/0 msr:write_msr(msr: 0x83f, val: 246)
28.810 :0/0 msr:write_msr(msr: 0x830, val: 68719479037)
40.117 CPU 0/KVM/4895 msr:write_msr(msr: IA32_SPEC_CTRL, val: 6)
40.127 CPU 0/KVM/4895 msr:read_msr(msr: IA32_DEBUGCTLMSR)
40.139 CPU 0/KVM/4895 msr:write_msr(msr: LSTAR, val: -2130661312)
40.141 CPU 0/KVM/4895 msr:write_msr(msr: SYSCALL_MASK, val: 14080)
40.142 CPU 0/KVM/4895 msr:write_msr(msr: TSC_AUX)
40.144 CPU 0/KVM/4895 msr:write_msr(msr: KERNEL_GS_BASE)
40.147 CPU 0/KVM/4895 msr:write_msr(msr: IA32_SPEC_CTRL)
40.148 CPU 0/KVM/4895 msr:write_msr(msr: IA32_FLUSH_CMD, val: 1)
40.151 CPU 0/KVM/4895 msr:write_msr(msr: IA32_SPEC_CTRL, val: 6)
^C
#
One can combine that with filtering pids as well:
# perf trace -e msr:* --filter="msr!=0xc0000100 && msr!=0x6e0" --filter-pids 4895 -a sleep 0.09
0.000 :0/0 msr:write_msr(msr: 0x830, val: 4294969597)
0.291 gnome-terminal/2790 msr:write_msr(msr: SYSCALL_MASK, val: 292608)
0.294 gnome-terminal/2790 msr:write_msr(msr: LSTAR, val: -1935671280)
0.295 gnome-terminal/2790 msr:write_msr(msr: TSC_AUX, val: 6)
10.940 gnome-terminal/2790 msr:write_msr(msr: 0x830, val: 4294969597)
15.943 gnome-shell/2096 msr:write_msr(msr: 0x830, val: 4294969597)
16.975 :0/0 msr:write_msr(msr: 0x830, val: 4294969597)
19.560 :0/0 msr:write_msr(msr: 0x83f, val: 246)
25.162 :0/0 msr:read_msr(msr: IA32_TSC_ADJUST)
25.807 JS Watchdog/3635 msr:write_msr(msr: IA32_SPEC_CTRL, val: 6)
25.820 :0/0 msr:write_msr(msr: IA32_SPEC_CTRL)
25.941 gnome-terminal/2790 msr:write_msr(msr: 0x830, val: 4294969597)
26.941 gnome-terminal/2790 msr:write_msr(msr: 0x830, val: 4294969597)
29.942 gnome-terminal/2790 msr:write_msr(msr: 0x830, val: 4294969597)
45.313 :0/0 msr:write_msr(msr: 0x83f, val: 246)
56.945 gnome-terminal/2790 msr:write_msr(msr: 0x830, val: 4294969597)
60.946 gnome-terminal/2790 msr:write_msr(msr: 0x830, val: 4294969597)
74.096 JS Watchdog/8971 msr:write_msr(msr: IA32_SPEC_CTRL, val: 6)
74.130 :0/0 msr:write_msr(msr: IA32_SPEC_CTRL)
79.673 :0/0 msr:write_msr(msr: 0x83f, val: 246)
79.947 gnome-terminal/2790 msr:write_msr(msr: 0x830, val: 17179871485)
#
Or for just a pid, with callchains:
# grep SYSCALL_MAS /tmp/build/perf/trace/beauty/generated/x86_arch_MSRs_array.c
[0xc0000084 - x86_64_specific_MSRs_offset] = "SYSCALL_MASK",
# perf trace -e msr:* --filter="msr==0xc0000084" --pid 2790 --call-graph=dwarf
0.000 gnome-terminal/2790 msr:write_msr(msr: SYSCALL_MASK, val: 292608)
do_trace_write_msr ([kernel.kallsyms])
do_trace_write_msr ([kernel.kallsyms])
kvm_on_user_return ([kvm])
fire_user_return_notifiers ([kernel.kallsyms])
exit_to_usermode_loop ([kernel.kallsyms])
do_syscall_64 ([kernel.kallsyms])
entry_SYSCALL_64 ([kernel.kallsyms])
__GI___poll (inlined)
9299.073 gnome-terminal/2790 msr:write_msr(msr: SYSCALL_MASK, val: 292608)
do_trace_write_msr ([kernel.kallsyms])
do_trace_write_msr ([kernel.kallsyms])
kvm_on_user_return ([kvm])
fire_user_return_notifiers ([kernel.kallsyms])
exit_to_usermode_loop ([kernel.kallsyms])
do_syscall_64 ([kernel.kallsyms])
entry_SYSCALL_64 ([kernel.kallsyms])
__GI___poll (inlined)
9348.374 gnome-terminal/2790 msr:write_msr(msr: SYSCALL_MASK, val: 292608)
do_trace_write_msr ([kernel.kallsyms])
do_trace_write_msr ([kernel.kallsyms])
kvm_on_user_return ([kvm])
fire_user_return_notifiers ([kernel.kallsyms])
exit_to_usermode_loop ([kernel.kallsyms])
do_syscall_64 ([kernel.kallsyms])
entry_SYSCALL_64 ([kernel.kallsyms])
__GI___poll (inlined)
<SNIP>
#
Ok, just another form of KVM to emit MSRs :-)
Next step: elliminate those greps by getting the filter expression,
looking for arg names, then for the arrays associated with it to do a
reverse lookup.
Also allow those filters to be associated with strace-like syscall
names.
After that: augment the 'val' arg for 'msr:write_msr' based on the first
arg, 'msr'.
Then, do that with eBPF too, not just with tracepoint filters.
Cc: Adrian Hunter <[email protected]>
Cc: Brendan Gregg <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Luis Cláudio Gonçalves <[email protected]>
Cc: Marcelo Tosatti <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: https://lkml.kernel.org/n/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
|