The -i and -v options can be used in subcommands so enable cascading the
sched_options. This fixes the following inconvenience in 'perf sched':
$ perf sched -i perf.data.sched map
... (it works well) ...
$ perf sched map -i perf.data.sched
Error: unknown switch `i'
Usage: perf sched map [<options>]
--color-cpus <cpus>
highlight given CPUs in map
--color-pids <pids>
highlight given pids in map
--compact map output in compact mode
--cpus <cpus> display given CPUs in map
With this patch, the second command line works with the perf.data.sched
data file.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/20161024030003.28534-2-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The perf report/top on TUI supports horizontal scrolling using LEFT and
RIGHT keys.
But it calculate the number of columns incorrectly when hierarchy mode
is enabled so that keep pressing RIGHT key can make the output
disappeared.
In the hierarchy mode, all sort keys are collapsed into a single column,
so it needs to be applied when calculating column numbers.
Reported-and-Tested-by: Markus Trippelsdorf <markus@trippelsdorf.de>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20161024162110.17918-1-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Sebastian noted that overhead for worker thread ops (throughput)
accounting was producing 'perf' to appear in the profiles, consuming a
non-trivial (i.e. 13%) amount of CPU.
This is due to cacheline bouncing due to the increment of w->ops.
We can easily fix this by just working on a local copy and updating the
actual worker once done running, and ready to show the program summary.
There is no danger of the worker being concurrent, so we can trust that
no stale value is being seen by another thread.
This also gets rid of the unnecessary cache alignment hack; its not
worth it.
Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: http://lkml.kernel.org/r/1477342613-9938-2-git-send-email-dave@stgolabs.net
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Make the 'perf list' glob matching for vendor events case insensitive.
This allows to use the upper case vendor events with perf list too.
Now the following works:
% perf list LONGEST_LAT
...
cache:
longest_lat_cache.miss
[Core-originated cacheable demand requests missed LLC]
longest_lat_cache.reference
[Core-originated cacheable demand requests that refer to LLC]
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Suggested-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/1476899402-31460-1-git-send-email-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Instead of the one when another syscall takes place while another is being
processed (in another CPU, but we show it serialized, so need to "interrupt"
the other), and also when finally showing the sys_enter + sys_exit + duration,
where we were showing the sample->time for the sys_exit, duh.
Before:
# perf trace sleep 1
<SNIP>
0.373 ( 0.001 ms): close(fd: 3 ) = 0
1000.626 (1000.211 ms): nanosleep(rqtp: 0x7ffd6ddddfb0) = 0
1000.653 ( 0.003 ms): close(fd: 1 ) = 0
1000.657 ( 0.002 ms): close(fd: 2 ) = 0
1000.667 ( 0.000 ms): exit_group( )
#
After:
# perf trace sleep 1
<SNIP>
0.336 ( 0.001 ms): close(fd: 3 ) = 0
0.373 (1000.086 ms): nanosleep(rqtp: 0x7ffe303e9550) = 0
1000.481 ( 0.002 ms): close(fd: 1 ) = 0
1000.485 ( 0.001 ms): close(fd: 2 ) = 0
1000.494 ( 0.000 ms): exit_group( )
[root@jouet linux]#
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.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-ecbzgmu2ni6glc6zkw8p1zmx@git.kernel.org
Fixes: 752fde44fd ("perf trace: Support interrupted syscalls")
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
It popped up in perf testing that the worker consumes some amount of
CPU. It boils down to the increment of `ops` which causes cache line
bouncing between the individual threads.
This patch aligns the struct by 256 bytes to ensure that not a cache
line is shared among CPUs. 128 byte is the x86 worst case and grep says
that L1_CACHE_SHIFT is set to 8 on s390.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Davidlohr Bueso <dbueso@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20161016190803.3392-1-bigeasy@linutronix.de
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In the perf wiki todo-list[1], there is an entry regarding initial-delay
and 'perf trace'; the following small patch tries to fulfill this point.
It has been generated against the branch tip/perf/core.
It has only been implemented in the "trace__run" case.
Ex.:
$ sudo strace -- ./perf trace --delay 5 sleep 1 2>&1
...
fcntl(7, F_SETFL, O_RDONLY|O_NONBLOCK) = 0
ioctl(7, PERF_EVENT_IOC_ID, 0x7ffc8fd35718) = 0
ioctl(11, PERF_EVENT_IOC_SET_OUTPUT, 0x7) = 0
fcntl(11, F_SETFL, O_RDONLY|O_NONBLOCK) = 0
ioctl(11, PERF_EVENT_IOC_ID, 0x7ffc8fd35718) = 0
write(6, "\0", 1) = 1
close(6) = 0
nanosleep({0, 5000000}, NULL) = 0 # DELAY OF 5 MS BEFORE ENABLING THE EVENTS
ioctl(3, PERF_EVENT_IOC_ENABLE, 0) = 0
ioctl(4, PERF_EVENT_IOC_ENABLE, 0) = 0
ioctl(5, PERF_EVENT_IOC_ENABLE, 0) = 0
ioctl(7, PERF_EVENT_IOC_ENABLE, 0) = 0
...
[1]: https://perf.wiki.kernel.org/index.php/Todo
Signed-off-by: Alexis Berlemont <alexis.berlemont@gmail.com>
Suggested-and-Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20161010054328.4028-2-alexis.berlemont@gmail.com
[ Add entry to the manpage, cut'n'pasted from stat's and record's ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Here is a small patch which tries to fulfill a point in the perf todo
list:
* Make pressing 'V' multiple times to go on cycling thru various
verbosity levels in 'perf top', so that info that is present in
'perf top -v' can be obtained without having to restart the tool
(acme).
After a small grep in the code, the max verbosity level seems 3; so,
we cycle at 4; I did not dare define a MAX_VERBOSE_LEVEL constant.
Signed-off-by: Alexis Berlemont <alexis.berlemont@gmail.com>
Suggested-and-Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20161012214823.14324-2-alexis.berlemont@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When the jit_buf_desc contains unwinding information, it is emitted as
eh_frame unwinding sections in the DSOs generated by perf inject.
The unwinding information is required to unwind of JITed code which do
not maintain the frame pointer register during function calls. It can
be emitted by V8 / Chromium when the --perf_prof_unwinding_info is
passed to V8.
The eh_frame and eh_frame_hdr sections are emitted immediately after the
.text.
The .eh_frame is aligned at a 8-byte boundary, and .eh_frame_hdr at a
4-byte one. Since size of the .eh_frame is required to be a multiple of
the word size, which means there will never be additional padding
between it and the .eh_frame_hdr on machines where the word size is 4 or
8 bytes.
However, additional padding might be inserted between .text and
.eh_frame to reach the correct alignment, which will always be 8 bytes,
also on 32bit machines. The reasoning behind this choice is that 4 extra
bytes of padding worst case are not a large cost for the advantage of
removing word-size dependent offset calculations when emitting the
jitdump.
Signed-off-by: Stefano Sanfilippo <ssanfilippo@chromium.org>
Signed-off-by: Ross McIlroy <rmcilroy@chromium.org>
Reviewed-by: Stephane Eranian <eranian@google.com>
Cc: Anton Blanchard <anton@ozlabs.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1476356383-30100-8-git-send-email-eranian@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This record is intended to provide unwinding information in the
eh_frame format. This is required to unwind JITed code which
does not maintain the frame pointer register during function calls.
The eh_frame unwinding information can be emitted by V8 / Chromium
when the --perf_prof_unwinding_info is passed.
A record of type jr_code_unwinding_info comes before the jr_code_load
it referred to and contains both the .eh_frame and .eh_frame_hdr.
The fields in the header have the following meaning:
* unwinding_size: size of the eh_frame and eh_frame_hdr, necessary
for distinguishing the content from the padding.
* eh_frame_hdr_size: as the name says.
* mapped_size: size of the payload that was in memory at runtime.
typically unwinding_size if the .eh_frame_hdr and .eh_frame were
mapped, or 0 if they weren't. It should always be the former case,
since the .eh_frame is guaranteed to be mapped in memory. However,
certain JITs might want to inject an .eh_frame_hdr with an empty LUT
to trigger fp-based unwinding fallback in libunwind. The only part
of the .eh_frame_hdr that libunwind reads from remote memory is the
LUT, and since there is none, mapping the unwinding info in memory
is not necessary, and 0 in this field signifies that it wasn't.
This practical hack allows to save bytes in code memory for those
JIT compilers that might or might not maintain a valid frame pointer.
The payload that follows is assumed to contain first the .eh_frame and
then the .eh_header_hdr, with no padding between the two.
Signed-off-by: Stefano Sanfilippo <ssanfilippo@chromium.org>
Signed-off-by: Ross McIlroy <rmcilroy@chromium.org>
Reviewed-by: Stephane Eranian <eranian@google.com>
Cc: Anton Blanchard <anton@ozlabs.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1476356383-30100-7-git-send-email-eranian@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This patch modifies the build dependencies on the jitdump support in
perf. As it stands jitdump was wrongfully made dependent 100% on using
DWARF. However, the dwarf dependency, only exist if generating the
source line table in genelf_debug.c. The rest of the support does not
need DWARF.
This patch removes the dependency on DWARF for the entire jitdump
support. It keeps it only for the genelf_debug.c support.
Signed-off-by: Maciej Debski <maciejd@google.com>
Reviewed-by: Stephane Eranian <eranian@google.com>
Cc: Anton Blanchard <anton@ozlabs.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1476356383-30100-3-git-send-email-eranian@google.com
Fixes: e12b202f8f ("perf jitdump: Build only on supported archs")
[ Make it build only if NO_LIBELF isn't defined, as jitdump.o will only be built in that case ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When the loop body isn't executed at all, then the 'ret' local variable,
that is uninitialized will be used as the return value.
This triggers this error on Alpine Linux:
CC /tmp/build/perf/util/demangle-java.o
CC /tmp/build/perf/util/demangle-rust.o
CC /tmp/build/perf/util/jitdump.o
CC /tmp/build/perf/util/genelf.o
util/jitdump.c: In function 'jit_process':
util/jitdump.c:622:3: error: 'ret' may be used uninitialized in this function [-Werror=maybe-uninitialized]
fprintf(stderr, "injected: %s (%d)\n", path, ret);
^
util/jitdump.c:584:6: note: 'ret' was declared here
int ret;
^
FLEX /tmp/build/perf/util/parse-events-flex.c
/ $ gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/libexec/gcc/x86_64-alpine-linux-musl/5.3.0/lto-wrapper
Target: x86_64-alpine-linux-musl
Configured with: /home/buildozer/aports/main/gcc/src/gcc-5.3.0/configure --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info
+--build=x86_64-alpine-linux-musl --host=x86_64-alpine-linux-musl --target=x86_64-alpine-linux-musl --with-pkgversion='Alpine 5.3.0' --enable-checking=release
+--disable-fixed-point --disable-libstdcxx-pch --disable-multilib --disable-nls --disable-werror --disable-symvers --enable-__cxa_atexit --enable-esp
+--enable-cloog-backend --enable-languages=c,c++,objc,java,fortran,ada --disable-libssp --disable-libmudflap --disable-libsanitizer --enable-shared
+--enable-threads --enable-tls --with-system-zlib
Thread model: posix
gcc version 5.3.0 (Alpine 5.3.0)
But this so far got under the radar, not causing any build problem, till the
"perf jit: enable jitdump support without dwarf" gets applied, when the above
problem takes place, some combination of inlining or whatever, the problem
is real, so fix it by initializing the variable to zero.
Cc: Anton Blanchard <anton@ozlabs.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Maciej Debski <maciejd@google.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: https://lkml.kernel.org/r/20161013200437.GA12815@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
It can be useful to specify branch type state per event, for example if
we want to collect both software trace points and last branch PMU events
in a single collection. Currently this doesn't work because the software
trace point errors out with -b.
There was already a branch-type parameter to configure branch sample
types per event in the parser, but it was stubbed out. This patch
implements the necessary plumbing to actually enable it.
Now:
$ perf record -e sched:sched_switch,cpu/cpu-cycles,branch_type=any/ ...
works.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/1476306127-19721-1-git-send-email-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Change Intel PT and BTS to pass up the length and the instruction
bytes of the decoded or sampled instruction in the perf sample.
The decoder already knows this information, we just need to pass it
up. Since it is only a couple of movs it is not very expensive.
Handle instruction cache too. Make sure ilen is always initialized.
Used in the next patch.
[Adrian: re-base on top (and adjust for) instruction buffer size tidy-up]
[Adrian: add BTS support and adjust commit message accordingly]
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Link: http://lkml.kernel.org/r/1475847747-30994-3-git-send-email-adrian.hunter@intel.com
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tidy instruction buffer size usage in preparation for copying the
instruction bytes onto samples.
The instruction buffer is presently used for debugging, so rename its
size macro from INTEL_PT_INSN_DBG_BUF_SZ to INTEL_PT_INSN_BUF_SZ, and
use it everywhere.
Note that the maximum instruction size is 15 which is a less efficient size
to copy than 16, which is why a separate buffer size is used.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Link: http://lkml.kernel.org/r/1475847747-30994-2-git-send-email-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
