When annotating source/disasm lines the perf tools parse the output of
objdump, trying to provide augmented output that allows navigating
jumps, calls, etc.
But when a line output by objdump can't be parsed the annotation code
falls back to just presenting the unparsed line.
When fixing a leak in the 0fb9f2aab7 commit ("perf annotate: Fix
memory leaks in LOCK handling") we failed to take that into account and
instead tried to free one of the data structures that should be freed
only when successfully allocated, oops, segfault.
There was a change in the way the objdump output for lock prefixed
instructions is formatted that lead the relevant parser to fail to grok
it.
At least RHEL7 works ok, but Fedora 20 segfaults.
Fix it by making the ins__delete() destructor work like the most basic
destructor: free().
Namely make it accept a NULL pointer and when handling it just do
nothing.
Further investigation is needed to figure out the nature of the objdump
output change so as to make the parser grok it.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Rabin Vincent <rabin@rab.in>
Link: http://lkml.kernel.org/n/tip-7wsy0zo292pif0yjoqpfryrz@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull alternative instructions framework improvements from Borislav Petkov:
"A more involved rework of the alternatives framework to be able to
pad instructions and thus make using the alternatives macros more
straightforward and without having to figure out old and new instruction
sizes but have the toolchain figure that out for us.
Furthermore, it optimizes JMPs used so that fetch and decode can be
relieved with smaller versions of the JMPs, where possible.
Some stats:
x86_64 defconfig:
Alternatives sites total: 2478
Total padding added (in Bytes): 6051
The padding is currently done for:
X86_FEATURE_ALWAYS
X86_FEATURE_ERMS
X86_FEATURE_LFENCE_RDTSC
X86_FEATURE_MFENCE_RDTSC
X86_FEATURE_SMAP
This is with the latest version of the patchset. Of course, on each
machine the alternatives sites actually being patched are a proper
subset of the total number."
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Adjust perf bench to the new changes in the alternatives code for
memcpy/memset.
Reviewed-by: Hitoshi Mitake <mitake.hitoshi@lab.ntt.co.jp>
Signed-off-by: Borislav Petkov <bp@suse.de>
Remove bias offset to find probe point by address.
Without this patch, probe points on kernel and executables are shown
correctly, but do not work with libraries:
# ./perf probe -l
probe:do_fork (on do_fork@kernel/fork.c)
probe_libc:malloc (on malloc in /usr/lib64/libc-2.17.so)
probe_perf:strlist__new (on strlist__new@util/strlist.c in /home/mhiramat/ksrc/linux-3/tools/perf/perf)
Removing bias allows it to show it as real place:
# ./perf probe -l
probe:do_fork (on do_fork@kernel/fork.c)
probe_libc:malloc (on __libc_malloc@malloc/malloc.c in /usr/lib64/libc-2.17.so)
probe_perf:strlist__new (on strlist__new@util/strlist.c in /home/mhiramat/ksrc/linux-3/tools/perf/perf)
Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Naohiro Aota <naota@elisp.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20150302124946.9191.64085.stgit@localhost.localdomain
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Warn if given uprobe event accesses memory on older kernel.
Until 3.14, uprobe event only supports accessing registers so this warns
to upgrade kernel if uprobe-event returns -EINVAL and an argument of the
event accesses memory ($stack, @+offset, and +|-offs() symtax).
With this patch (on 3.10.0-123.13.2.el7.x86_64);
-----
# ./perf probe -x ./perf warn_uprobe_event_compat stack=-0\(%sp\)
Added new event:
Failed to write event: Invalid argument
Please upgrade your kernel to at least 3.14 to have access to feature -0(%sp)
Error: Failed to add events.
-----
Suggested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lkml.kernel.org/r/20150228025329.32106.70581.stgit@localhost.localdomain
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In this commit:
commit 363b785f38
Author: Don Zickus <dzickus@redhat.com>
Date: Fri Mar 14 10:43:44 2014 -0400
perf tools: Speed up thread map generation
We ended up emitting PERF_RECORD_FORK events after their corresponding
PERF_RECORD_COMM, so the code below will remove the "existing thread"
and then recreates it, unnecessarily:
[root@ssdandy ~]# perf probe -x ~/bin/perf -L machine__process_fork_event
<machine__process_fork_event@/home/acme/git/linux/tools/perf/util/machine.c:0>
0 int machine__process_fork_event(struct machine *machine, union perf_event *event,
struct perf_sample *sample)
2 {
3 struct thread *thread = machine__find_thread(machine,
event->fork.pid,
event->fork.tid);
6 struct thread *parent = machine__findnew_thread(machine,
event->fork.ppid,
event->fork.ptid);
/* if a thread currently exists for the thread id remove it */
if (thread != NULL)
12 machine__remove_thread(machine, thread);
14 thread = machine__findnew_thread(machine, event->fork.pid,
event->fork.tid);
16 if (dump_trace)
17 perf_event__fprintf_task(event, stdout);
19 if (thread == NULL || parent == NULL ||
20 thread__fork(thread, parent, sample->time) < 0) {
21 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
22 return -1;
}
25 return 0;
26 }
[root@ssdandy ~]# perf probe -x ~/bin/perf fork_after_comm=machine__process_fork_event:12
Added new event:
probe_perf:fork_after_comm (on machine__process_fork_event:12 in /home/acme/bin/perf)
You can now use it in all perf tools, such as:
perf record -e probe_perf:fork_after_comm -aR sleep 1
[root@ssdandy ~]#
[root@ssdandy ~]# perf record -g -e probe_perf:* trace -o /tmp/bla
^C[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.021 MB perf.data (30 samples) ]
Terminated
[root@ssdandy ~]#
[root@ssdandy ~]# perf report --no-children --show-total-period --stdio
# To display the perf.data header info, please use --header/--header-only options.
#
# Samples: 30 of event 'probe_perf:fork_after_comm'
# Event count (approx.): 30
#
# Overhead Period Command Shared Object Symbol
# ........ ............ ....... ............. ...............................
#
100.00% 30 trace trace [.] machine__process_fork_event
|
---machine__process_fork_event
__event__synthesize_thread.part.2
perf_event__synthesize_threads
cmd_trace
main
__libc_start_main
[root@ssdandy ~]#
And Looking at 'perf report -D' output we see it:
0 0 0x8698 [0x30]: PERF_RECORD_COMM: auditd:703/707
0 0 0x86c8 [0x38]: PERF_RECORD_FORK(703:707):(703:703)
Fix it by more closely mimicking how the kernel generates those records
when a new fork happens, i.e. first a PERF_RECORD_FORK, then a
PERF_RECORD_COMM.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: David Ahern <dsahern@gmail.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/n/tip-h0emvymi2t3mw8dlqd6d6z73@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add --purge FILE to remove all caches of FILE.
Since the current --remove FILE removes a cache which has
same build-id of given FILE. Since the command takes a
FILE path, it can confuse user who tries to remove cache
about FILE path.
-----
# ./perf buildid-cache -v --add ./perf
Adding 133b7b5486d987a5ab5c3ebf4ea14941f45d4d4f ./perf: Ok
# (update the ./perf binary)
# ./perf buildid-cache -v --remove ./perf
Removing 305bbd1be68f66eca7e2d78db294653031edfa79 ./perf: FAIL
./perf wasn't in the cache
-----
Actually, the --remove's FAIL is not shown, it just silently fails.
So, this patch adds --purge FILE action for such usecase.
perf buildid-cache --purge FILE removes all caches which has same FILE
path.
In other words, it removes all caches including old binaries.
-----
# ./perf buildid-cache -v --add ./perf
Adding 133b7b5486d987a5ab5c3ebf4ea14941f45d4d4f ./perf: Ok
# (update the ./perf binary)
# ./perf buildid-cache -v --purge ./perf
Removing 133b7b5486d987a5ab5c3ebf4ea14941f45d4d4f ./perf: Ok
-----
BTW, if you want to purge all the caches, remove ~/.debug/* .
Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Hemant Kumar <hemant@linux.vnet.ibm.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20150227045026.1999.64084.stgit@localhost.localdomain
[ s/dirname/dir_name/g to fix build on fedora14, where dirname is a global ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
If the long_name of a 'struct option' is defined as NULL, --list-opts
will incorrectly print '--(null)' in its output. As a result, '--(null)'
will finally appear in the case of bash completion, e.g. 'perf record
--'.
Example:
Before this patch:
$ perf record --list-opts
--event --filter --pid --tid --realtime --no-buffering --raw-samples
--all-cpus --cpu --count --output --no-inherit --freq --mmap-pages
--group --(null) --call-graph --verbose --quiet --stat --data
--timestamp --period --no-samples --no-buildid-cache --no-buildid
--cgroup --delay --uid --branch-any --branch-filter --weight
--transaction --per-thread --intr-regs
After this patch:
$ perf record --list-opts
--event --filter --pid --tid --realtime --no-buffering --raw-samples
--all-cpus --cpu --count --output --no-inherit --freq --mmap-pages
--group --call-graph --verbose --quiet --stat --data --timestamp
--period --no-samples --no-buildid-cache --no-buildid --cgroup --delay
--uid --branch-any --branch-filter --weight --transaction --per-thread
--intr-regs
Signed-off-by: Yunlong Song <yunlong.song@huawei.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1425032491-20224-7-git-send-email-yunlong.song@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The minus operator has higher precedence than ?: Add parentheses around
?: fix this.
Before this patch:
$ echo 'p:myprobe do_sys_open' > /sys/kernel/debug/tracing/kprobe_events
$ perf probe -l -k ../vmlinux
kprobes:myprobe (on do_sys_open)
After this patch:
$ echo 'p:myprobe do_sys_open' > /sys/kernel/debug/tracing/kprobe_events
$ perf probe -l -k ../vmlinux
kprobes:myprobe (on do_sys_open@linux.git/fs/open.c)
Signed-off-by: He Kuang <hekuang@huawei.com>
Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1425034373-14511-1-git-send-email-hekuang@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently, the perf diff only works with same binaries. That's because
it compares the symbol start address. It doesn't work if the perf.data
comes from different binaries. This patch matches the symbol names.
Actually, perf diff once intended to compare the symbol names. The
commit as below can look for a pair by name.
604c5c9297 (perf diff: Change the default sort order to "dso,symbol")
However, at that time, perf diff used a global list of dsos. That means
the binaries which has same name can only be loaded once. That's a
problem for comparing different binaries.
For example, we have an old binary and an updated binary. They very
likely have same name and most of the functions, so only dsos from old
binary will be loaded. When processing the data from updated binary,
perf still use the symbol information from old binary. That's wrong.
Then the commit as below used IP to replace symbol name.
9c443dfdd3 ("perf diff: Fix support for all --sort combinations")
>From that time, perf diff starts to compare the symbol address.
The global dsos is discarded from a patch in 2010.
a1645ce12a ("perf: 'perf kvm' tool for monitoring guest performance
from host")
However, at that time, perf diff already compared by address. So perf
diff cannot work for different binaries as well.
This patch actually rolls back the perf diff to original design. The
document is also changed, so everybody knows the original design is to
compare the symbol names.
Here are some examples:
The only difference between example_v1.c and example_v2.c is the
location of f2 and f3. There is no change in behavior, but the previous
perf diff display the wrong differential profile.
example_v1.c
noinline void f3(void)
{
volatile int i;
for (i = 0; i < 10000;) {
if(i%2)
i++;
else
i++;
}
}
noinline void f2(void)
{
volatile int a = 100, b, c;
for (b = 0; b < 10000; b++)
c = a * b;
}
noinline void f1(void)
{
f2();
f3();
}
int main()
{
int i;
for (i = 0; i < 100000; i++)
f1();
}
example_v2.c
noinline void f2(void)
{
volatile int a = 100, b, c;
for (b = 0; b < 10000; b++)
c = a * b;
}
noinline void f3(void)
{
volatile int i;
for (i = 0; i < 10000;) {
if(i%2)
i++;
else
i++;
}
}
noinline void f1(void)
{
f2();
f3();
}
int main()
{
int i;
for (i = 0; i < 100000; i++)
f1();
}
[lk@localhost perf_diff]$ gcc example_v1.c -o example
[lk@localhost perf_diff]$ perf record -o example_v1.data ./example
[ perf record: Woken up 4 times to write data ]
[ perf record: Captured and wrote 0.813 MB example_v1.data (~35522 samples) ]
[lk@localhost perf_diff]$ gcc example_v2.c -o example
[lk@localhost perf_diff]$ perf record -o example_v2.data ./example
[ perf record: Woken up 4 times to write data ]
[ perf record: Captured and wrote 0.824 MB example_v2.data (~36015 samples) ]
Old perf diff result:
[lk@localhost perf_diff]$ perf diff example_v1.data example_v2.data
Event 'cycles'
Baseline Delta Shared Object Symbol
........ ....... ................ ...............................
[kernel.vmlinux] [k] __perf_event_task_sched_out
0.00% [kernel.vmlinux] [k] apic_timer_interrupt
[kernel.vmlinux] [k] idle_cpu
[kernel.vmlinux] [k] intel_pstate_timer_func
[kernel.vmlinux] [k] native_read_msr_safe
0.00% [kernel.vmlinux] [k] native_read_tsc
0.00% [kernel.vmlinux] [k] native_write_msr_safe
[kernel.vmlinux] [k] ntp_tick_length
0.00% [kernel.vmlinux] [k] rb_erase
0.00% [kernel.vmlinux] [k] tick_sched_timer
0.00% [kernel.vmlinux] [k] unmap_single_vma
0.00% [kernel.vmlinux] [k] update_wall_time
0.00% example [.] f1
46.24% example [.] f2
53.71% -7.55% example [.] f3
+53.81% example [.] f3
0.02% example [.] main
New perf diff result:
[lk@localhost perf_diff]$ perf diff example_v1.data example_v2.data
[kernel.vmlinux] [k] __perf_event_task_sched_out
0.00% [kernel.vmlinux] [k] apic_timer_interrupt
[kernel.vmlinux] [k] idle_cpu
[kernel.vmlinux] [k] intel_pstate_timer_func
[kernel.vmlinux] [k] native_read_msr_safe
0.00% [kernel.vmlinux] [k] native_read_tsc
0.00% [kernel.vmlinux] [k] native_write_msr_safe
[kernel.vmlinux] [k] ntp_tick_length
0.00% [kernel.vmlinux] [k] rb_erase
0.00% [kernel.vmlinux] [k] tick_sched_timer
0.00% [kernel.vmlinux] [k] unmap_single_vma
0.00% [kernel.vmlinux] [k] update_wall_time
0.00% example [.] f1
46.24% -0.08% example [.] f2
53.71% +0.11% example [.] f3
0.02% example [.] main
Signed-off-by: Kan Liang <kan.liang@intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Andi Kleen <ak@linux.intel.com>
Link: http://lkml.kernel.org/r/1423460384-11645-1-git-send-email-kan.liang@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add new buildid cache if the update target file is not cached.
This can happen when an old binary is replaced by new one after caching
the old one. In this case, user sees his operation just failed.
But it does not look straight, since user just pass the binary "path",
not "build-id".
----
# ./perf buildid-cache --add ./perf
(update ./perf to new binary)
# ./perf buildid-cache --update ./perf
./perf wasn't in the cache
#
----
This patch adds given new binary to cache if the new binary is
not cached. So we'll not see the above error.
----
# ./perf buildid-cache --add ./perf
(update ./perf to new binary)
# ./perf buildid-cache --update ./perf
#
----
Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Hemant Kumar <hemant@linux.vnet.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20150226065440.23912.1494.stgit@localhost.localdomain
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Recent linux kernel provides a blacklist of the functions which can not
be probed. perf probe can now check this blacklist before setting new
events and indicate better error message for users.
Without this patch,
----
# perf probe --add vmalloc_fault
Added new event:
Failed to write event: Invalid argument
Error: Failed to add events.
----
With this patch
----
# perf probe --add vmalloc_fault
Added new event:
Warning: Skipped probing on blacklisted function: vmalloc_fault
----
Reported-by: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20150219143113.14434.5387.stgit@localhost.localdomain
Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
perf-top is terminating due to SIGBUS on sparc64. git bisect points to:
commit 8239698603
Author: Arnaldo Carvalho de Melo <acme@redhat.com>
Date: Mon Sep 8 13:26:35 2014 -0300
perf evlist: Refcount mmaps
We need to know how many fds are using a perf mmap via
PERF_EVENT_IOC_SET_OUTPUT, so that we can know when to ditch an mmap,
refcount it.
This commit added 'int refcnt' to struct perf_mmap and the addition makes the
event_copy element no longer 8-byte aligned.
Fix by adding __attribute__((aligned(8))) to the event_copy struct
member.
Signed-off-by: David Ahern <david.ahern@oracle.com>
Link: http://lkml.kernel.org/r/1424304198-92028-1-git-send-email-david.ahern@oracle.com
[ Switched from 'int pad;' to using __attribute__, David tested/acked that ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Commit f6edb53c49 converted the probe to
a CPU wide event first (pid == -1). For kernels that do not support
the PERF_FLAG_FD_CLOEXEC flag the probe fails with EINVAL. Since this
errno is not handled pid is not reset to 0 and the subsequent use of
pid = -1 as an argument brings in an additional failure path if
perf_event_paranoid > 0:
$ perf record -- sleep 1
perf_event_open(..., 0) failed unexpectedly with error 13 (Permission denied)
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.007 MB /tmp/perf.data (11 samples) ]
Also, ensure the fd of the confirmation check is closed and comment why
pid = -1 is used.
Needs to go to 3.18 stable tree as well.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Based-on-patch-by: David Ahern <david.ahern@oracle.com>
Acked-by: David Ahern <david.ahern@oracle.com>
Cc: David Ahern <dsahern@gmail.com>
Link: http://lkml.kernel.org/r/54EC610C.8000403@intel.com
Cc: stable@vger.kernel.org # v3.18+
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Some of the tracers bring their own id or pid fields and we can end up
having two of them. This patch adds a "perf_" prefix to the 'generic'
fields so we avoid a clash of the member names.
The change is visible in the babeltrace output:
Before:
$ babeltrace ./ctf-data/
[03:19:13.962131936] (+0.000001935) cycles: { }, { ip = 0xFFFFFFFF8105443A, tid = 20714, pid = 20714, period = 8 }
[03:19:13.962133732] (+0.000001796) cycles: { }, { ip = 0xFFFFFFFF8105443A, tid = 20714, pid = 20714, period = 114 }
...
Now:
$ babeltrace ./ctf-data/
[03:19:13.962131936] (+0.000001935) cycles: { }, { perf_ip = 0xFFFFFFFF8105443A, perf_tid = 20714, perf_pid = 20714, perf_period = 8 }
[03:19:13.962133732] (+0.000001796) cycles: { }, { perf_ip = 0xFFFFFFFF8105443A, perf_tid = 20714, perf_pid = 20714, perf_period = 114 }
...
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Reviewed-by: David Ahern <dsahern@gmail.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jeremie Galarneau <jgalar@efficios.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tom Zanussi <tzanussi@gmail.com>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1424470628-5969-5-git-send-email-jolsa@kernel.org
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Adding 'perf data convert' to convert perf data file into different
format. This patch adds support for CTF format conversion.
To convert perf.data into CTF run:
$ perf data convert --to-ctf=./ctf-data/
[ perf data convert: Converted 'perf.data' into CTF data './ctf-data/' ]
[ perf data convert: Converted and wrote 11.268 MB (100230 samples) ]
The command will create CTF metadata out of perf.data file (or one
specified via -i option) and then convert all sample events into single
CTF stream.
Each sample_type bit is translated into separated CTF event field apart
from following exceptions:
PERF_SAMPLE_RAW - added in next patch
PERF_SAMPLE_READ - TODO
PERF_SAMPLE_CALLCHAIN - TODO
PERF_SAMPLE_BRANCH_STACK - TODO
PERF_SAMPLE_REGS_USER - TODO
PERF_SAMPLE_STACK_USER - TODO
$ perf --debug=data-convert=2 data convert ...
The converted CTF data could be analyzed by CTF tools, like babletrace
or tracecompass [1].
$ babeltrace ./ctf-data/
[03:19:13.962125533] (+?.?????????) cycles: { }, { ip = 0xFFFFFFFF8105443A, tid = 20714, pid = 20714, period = 1 }
[03:19:13.962130001] (+0.000004468) cycles: { }, { ip = 0xFFFFFFFF8105443A, tid = 20714, pid = 20714, period = 1 }
[03:19:13.962131936] (+0.000001935) cycles: { }, { ip = 0xFFFFFFFF8105443A, tid = 20714, pid = 20714, period = 8 }
[03:19:13.962133732] (+0.000001796) cycles: { }, { ip = 0xFFFFFFFF8105443A, tid = 20714, pid = 20714, period = 114 }
[03:19:13.962135557] (+0.000001825) cycles: { }, { ip = 0xFFFFFFFF8105443A, tid = 20714, pid = 20714, period = 2087 }
[03:19:13.962137627] (+0.000002070) cycles: { }, { ip = 0xFFFFFFFF81361938, tid = 20714, pid = 20714, period = 37582 }
[03:19:13.962161091] (+0.000023464) cycles: { }, { ip = 0xFFFFFFFF8124218F, tid = 20714, pid = 20714, period = 600246 }
[03:19:13.962517569] (+0.000356478) cycles: { }, { ip = 0xFFFFFFFF811A75DB, tid = 20714, pid = 20714, period = 1325731 }
[03:19:13.969518008] (+0.007000439) cycles: { }, { ip = 0x34080917B2, tid = 20714, pid = 20714, period = 1144298 }
The following members to the ctf-environment were decided to be added to
distinguish and specify perf CTF data:
- domain
It says "kernel" because it contains a kernel trace (not to be
confused with a user space like lttng-ust does)
- tracer_name
It says perf. This can be used to distinguish between lttng and perf
CTF based trace.
- version
The kernel version from stream. In addition to release, this is what
it looks like on a Debian kernel:
release = "3.14-1-amd64";
version = "3.14.0";
[1] http://projects.eclipse.org/projects/tools.tracecompass
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Reviewed-by: David Ahern <dsahern@gmail.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jeremie Galarneau <jgalar@efficios.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Tom Zanussi <tzanussi@gmail.com>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1424470628-5969-4-git-send-email-jolsa@kernel.org
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
User visible changes:
- No need to explicitely enable evsels for workload started from perf, let it
be enabled via perf_event_attr.enable_on_exec, removing some events that take
place in the 'perf trace' before a workload is really started by it.
(Arnaldo Carvalho de Melo)
- Fix to handle optimized not-inlined functions in 'perf probe' (Masami Hiramatsu)
- Update 'perf probe' man page (Masami Hiramatsu)
- 'perf trace': Allow mixing with tracepoints and suppressing plain syscalls
(Arnaldo Carvalho de Melo)
Infrastructure changes:
- Introduce {trace_seq_do,event_format_}_fprintf functions to allow
a default tracepoint field list printer to be used in tools that allows
redirecting output to a file. (Arnaldo Carvalho de Melo)
- The man page for pthread_attr_set_affinity_np states that _GNU_SOURCE
must be defined before pthread.h, do it to fix the build in some
systems (Josh Boyer)
- Cleanups in 'perf buildid-cache' (Masami Hiramatsu)
- Fix dso cache test case (Namhyung Kim)
- Do Not rely on dso__data_read_offset() to open DSO (Namhyung Kim)
- Make perf aware of tracefs (Steven Rostedt).
- Fix build by defining STT_GNU_IFUNC for glibc 2.9 and older (Vinson Lee)
- AArch64 symbol resolution fixes (Victor Kamensky)
- Kconfig beachhead (Jiri Olsa)
- Simplify nr_pages validity (Kaixu Xia)
- Fixup header positioning in 'perf list' (Yunlong Song)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
LBR call stack only has user-space callchains. It is output in the
PERF_SAMPLE_BRANCH_STACK data format. For kernel callchains, it's
still in the form of PERF_SAMPLE_CALLCHAIN.
The perf tool has to handle both data sources to construct a
complete callstack.
For the "perf report -D" option, both lbr and fp information will be
displayed.
A new call chain recording option "lbr" is introduced into the perf
tool for LBR call stack. The user can use --call-graph lbr to get
the call stack information from hardware.
Here are some examples.
When profiling bc(1) on Fedora 19:
echo 'scale=2000; 4*a(1)' > cmd; perf record --call-graph lbr bc -l < cmd
If enabling LBR, perf report output looks like:
50.36% bc bc [.] bc_divide
|
--- bc_divide
execute
run_code
yyparse
main
__libc_start_main
_start
33.66% bc bc [.] _one_mult
|
--- _one_mult
bc_divide
execute
run_code
yyparse
main
__libc_start_main
_start
7.62% bc bc [.] _bc_do_add
|
--- _bc_do_add
|
|--99.89%-- 0x2000186a8
--0.11%-- [...]
6.83% bc bc [.] _bc_do_sub
|
--- _bc_do_sub
|
|--99.94%-- bc_add
| execute
| run_code
| yyparse
| main
| __libc_start_main
| _start
--0.06%-- [...]
0.46% bc libc-2.17.so [.] __memset_sse2
|
--- __memset_sse2
|
|--54.13%-- bc_new_num
| |
| |--51.00%-- bc_divide
| | execute
| | run_code
| | yyparse
| | main
| | __libc_start_main
| | _start
| |
| |--30.46%-- _bc_do_sub
| | bc_add
| | execute
| | run_code
| | yyparse
| | main
| | __libc_start_main
| | _start
| |
| --18.55%-- _bc_do_add
| bc_add
| execute
| run_code
| yyparse
| main
| __libc_start_main
| _start
|
--45.87%-- bc_divide
execute
run_code
yyparse
main
__libc_start_main
_start
If using FP, perf report output looks like:
echo 'scale=2000; 4*a(1)' > cmd; perf record --call-graph fp bc -l < cmd
50.49% bc bc [.] bc_divide
|
--- bc_divide
33.57% bc bc [.] _one_mult
|
--- _one_mult
7.61% bc bc [.] _bc_do_add
|
--- _bc_do_add
0x2000186a8
6.88% bc bc [.] _bc_do_sub
|
--- _bc_do_sub
0.42% bc libc-2.17.so [.] __memcpy_ssse3_back
|
--- __memcpy_ssse3_back
If using LBR, perf report -D output looks like:
3458145275743 0x2fd750 [0xd8]: PERF_RECORD_SAMPLE(IP, 0x2): 9748/9748: 0x408ea8 period: 609644 addr: 0
... LBR call chain: nr:8
..... 0: fffffffffffffe00
..... 1: 0000000000408e50
..... 2: 000000000040a458
..... 3: 000000000040562e
..... 4: 0000000000408590
..... 5: 00000000004022c0
..... 6: 00000000004015dd
..... 7: 0000003d1cc21b43
... FP chain: nr:2
..... 0: fffffffffffffe00
..... 1: 0000000000408ea8
... thread: bc:9748
...... dso: /usr/bin/bc
The LBR call stack has the following known limitations:
- Zero length calls are not filtered out by the hardware
- Exception handing such as setjmp/longjmp will have calls/returns not
match
- Pushing different return address onto the stack will have
calls/returns not match
- If callstack is deeper than the LBR, only the last entries are
captured
Tested-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Borislav Petkov <bp@suse.de>
Cc: David Ahern <dsahern@gmail.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Simon Que <sque@chromium.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1420482185-29830-3-git-send-email-kan.liang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The hearer text 'List of pre-defined events (to be used in -e):' is
placed in an improper function, which causes an abnormal output, e.g.
'perf list hw' shows no guiding text at all, and 'perf list hw
L1-dcache*' shows the guiding text incorrectly in the middle of the
output.
Example
Before this patch:
$ perf list hw L1-dcache*
branch-instructions OR branches [Hardware event]
branch-misses [Hardware event]
bus-cycles [Hardware event]
cache-misses [Hardware event]
cache-references [Hardware event]
cpu-cycles OR cycles [Hardware event]
instructions [Hardware event]
stalled-cycles-backend OR idle-cycles-backend [Hardware event]
stalled-cycles-frontend OR idle-cycles-frontend [Hardware event]
List of pre-defined events (to be used in -e): <-- incorrect position
L1-dcache-load-misses [Hardware cache event]
L1-dcache-loads [Hardware cache event]
L1-dcache-prefetch-misses [Hardware cache event]
L1-dcache-prefetches [Hardware cache event]
L1-dcache-store-misses [Hardware cache event]
L1-dcache-stores [Hardware cache event]
After this patch:
$ perf list hw L1-dcache*
List of pre-defined events (to be used in -e): <-- correct position
branch-instructions OR branches [Hardware event]
branch-misses [Hardware event]
bus-cycles [Hardware event]
cache-misses [Hardware event]
cache-references [Hardware event]
cpu-cycles OR cycles [Hardware event]
instructions [Hardware event]
stalled-cycles-backend OR idle-cycles-backend [Hardware event]
stalled-cycles-frontend OR idle-cycles-frontend [Hardware event]
L1-dcache-load-misses [Hardware cache event]
L1-dcache-loads [Hardware cache event]
L1-dcache-prefetch-misses [Hardware cache event]
L1-dcache-prefetches [Hardware cache event]
L1-dcache-store-misses [Hardware cache event]
L1-dcache-stores [Hardware cache event]
Signed-off-by: Yunlong Song <yunlong.song@huawei.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1423833115-11199-8-git-send-email-yunlong.song@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Fix the 'segmentation fault' bug of 'perf list --list-cmds', which also
happens in other cases (e.g. record, report ...). This bug happens when
there are no cmds to list at all.
Example:
Before this patch:
$ perf list --list-cmds
Segmentation fault
$
After this patch:
$ perf list --list-cmds
$
As shown above, the result prints nothing rather than a segmentation
fault. The null result means 'perf list' has no cmds to display at this
time.
Signed-off-by: Yunlong Song <yunlong.song@huawei.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1423833115-11199-5-git-send-email-yunlong.song@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
