We have defined YY_USER_ACTION to keep trace of the column location
during events parsing, but we need to clean it up when we call REJECT.
When REJECT is called, the lexer shrinks the text and re-runs the
matching, so we need to address it in resuming the previous location
value to keep it correct for error display, like:
Before:
$ perf stat -e 'cpu/uops_executed.core,krava/' true
event syntax error: '..38;5;9:mi=01;05;37;41:su=48;5;196;38;5;15:sg=48;5;1\
1;38;5;16:ca=48;5;196;38;5;226:tw=48;5;10;38;5;16:ow=48;5;10;38;5;21:st=48;5;\
21;38;50
�'
\___ unknown term
After:
$ ./perf stat -e 'cpu/uops_executed.core,krava/' true
event syntax error: '..cuted.core,krava/'
\___ unknown term
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Reported-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Andi Kleen <ak@linux.intel.com>
Cc: Changbin Du <changbin.du@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/n/tip-vug2hchlny30jfsfrumbym26@git.kernel.org
Link: http://lkml.kernel.org/r/20171009140944.GD28623@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Perf top is often crashing at very random locations on powerpc. After
investigating, I found the crash only happens when sample is of zero
length symbol. Powerpc kernel has many such symbols which does not
contain length details in vmlinux binary and thus start and end
addresses of such symbols are same.
Structure
struct sym_hist {
u64 nr_samples;
u64 period;
struct sym_hist_entry addr[0];
};
has last member 'addr[]' of size zero. 'addr[]' is an array of addresses
that belongs to one symbol (function). If function consist of 100
instructions, 'addr' points to an array of 100 'struct sym_hist_entry'
elements. For zero length symbol, it points to the *empty* array, i.e.
no members in the array and thus offset 0 is also invalid for such
array.
static int __symbol__inc_addr_samples(...)
{
...
offset = addr - sym->start;
h = annotation__histogram(notes, evidx);
h->nr_samples++;
h->addr[offset].nr_samples++;
h->period += sample->period;
h->addr[offset].period += sample->period;
...
}
Here, when 'addr' is same as 'sym->start', 'offset' becomes 0, which is
valid for normal symbols but *invalid* for zero length symbols and thus
updating h->addr[offset] causes memory corruption.
Fix this by adding one dummy element for zero length symbols.
Link: https://lkml.org/lkml/2016/10/10/148
Fixes: edee44be59 ("perf annotate: Don't throw error for zero length symbols")
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Kim Phillips <kim.phillips@arm.com>
Cc: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Taeung Song <treeze.taeung@gmail.com>
Link: http://lkml.kernel.org/r/1508854806-10542-1-git-send-email-ravi.bangoria@linux.vnet.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
On one hand this ensures that the memory is properly freed when the DSO
gets freed. On the other hand this significantly speeds up the
processing of the callchain nodes when lots of srclines are requested.
For one of my data files e.g.:
Before:
Performance counter stats for 'perf report -s srcline -g srcline --stdio':
52496.495043 task-clock (msec) # 0.999 CPUs utilized
634 context-switches # 0.012 K/sec
2 cpu-migrations # 0.000 K/sec
191,561 page-faults # 0.004 M/sec
165,074,498,235 cycles # 3.144 GHz
334,170,832,408 instructions # 2.02 insn per cycle
90,220,029,745 branches # 1718.591 M/sec
654,525,177 branch-misses # 0.73% of all branches
52.533273822 seconds time elapsedProcessed 236605 events and lost 40 chunks!
After:
Performance counter stats for 'perf report -s srcline -g srcline --stdio':
22606.323706 task-clock (msec) # 1.000 CPUs utilized
31 context-switches # 0.001 K/sec
0 cpu-migrations # 0.000 K/sec
185,471 page-faults # 0.008 M/sec
71,188,113,681 cycles # 3.149 GHz
133,204,943,083 instructions # 1.87 insn per cycle
34,886,384,979 branches # 1543.214 M/sec
278,214,495 branch-misses # 0.80% of all branches
22.609857253 seconds time elapsed
Note that the difference is only this large when `--inline` is not
passed. In such situations, we would use the inliner cache and thus do
not run this code path that often.
I think that this cache should actually be used in other places, too.
When looking at the valgrind leak report for perf report, we see tons of
srclines being leaked, most notably from calls to
hist_entry__get_srcline. The problem is that get_srcline has many
different formatting options (show_sym, show_addr, potentially even
unwind_inlines when calling __get_srcline directly). As such, the
srcline cannot easily be cached for all calls, or we'd have to add
caches for all formatting combinations (6 so far). An alternative would
be to remove the formatting options and handle that on a different level
- i.e. print the sym/addr on demand wherever we actually output
something. And the unwind_inlines could be moved into a separate
function that does not return the srcline.
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20171019113836.5548-4-milian.wolff@kdab.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When no inlined frames could be found for a given address, we did not
store this information anywhere. That means we potentially do the costly
inliner lookup repeatedly for cases where we know it can never succeed.
This patch makes dso__parse_addr_inlines always return a valid
inline_node. It will be empty when no inliners are found. This enables
us to cache the empty list in the DSO, thereby improving the performance
when many addresses fail to find the inliners.
For my trivial example, the performance impact is already quite
significant:
Before:
~~~~~
Performance counter stats for 'perf report --stdio --inline -g srcline -s srcline' (5 runs):
594.804032 task-clock (msec) # 0.998 CPUs utilized ( +- 0.07% )
53 context-switches # 0.089 K/sec ( +- 4.09% )
0 cpu-migrations # 0.000 K/sec ( +-100.00% )
5,687 page-faults # 0.010 M/sec ( +- 0.02% )
2,300,918,213 cycles # 3.868 GHz ( +- 0.09% )
4,395,839,080 instructions # 1.91 insn per cycle ( +- 0.00% )
939,177,205 branches # 1578.969 M/sec ( +- 0.00% )
11,824,633 branch-misses # 1.26% of all branches ( +- 0.10% )
0.596246531 seconds time elapsed ( +- 0.07% )
~~~~~
After:
~~~~~
Performance counter stats for 'perf report --stdio --inline -g srcline -s srcline' (5 runs):
113.111405 task-clock (msec) # 0.990 CPUs utilized ( +- 0.89% )
29 context-switches # 0.255 K/sec ( +- 54.25% )
0 cpu-migrations # 0.000 K/sec
5,380 page-faults # 0.048 M/sec ( +- 0.01% )
432,378,779 cycles # 3.823 GHz ( +- 0.75% )
670,057,633 instructions # 1.55 insn per cycle ( +- 0.01% )
141,001,247 branches # 1246.570 M/sec ( +- 0.01% )
2,346,845 branch-misses # 1.66% of all branches ( +- 0.19% )
0.114222393 seconds time elapsed ( +- 1.19% )
~~~~~
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20171019113836.5548-3-milian.wolff@kdab.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Please do not apply this to mainline directly, instead please re-run the
coccinelle script shown below and apply its output.
For several reasons, it is desirable to use {READ,WRITE}_ONCE() in
preference to ACCESS_ONCE(), and new code is expected to use one of the
former. So far, there's been no reason to change most existing uses of
ACCESS_ONCE(), as these aren't harmful, and changing them results in
churn.
However, for some features, the read/write distinction is critical to
correct operation. To distinguish these cases, separate read/write
accessors must be used. This patch migrates (most) remaining
ACCESS_ONCE() instances to {READ,WRITE}_ONCE(), using the following
coccinelle script:
----
// Convert trivial ACCESS_ONCE() uses to equivalent READ_ONCE() and
// WRITE_ONCE()
// $ make coccicheck COCCI=/home/mark/once.cocci SPFLAGS="--include-headers" MODE=patch
virtual patch
@ depends on patch @
expression E1, E2;
@@
- ACCESS_ONCE(E1) = E2
+ WRITE_ONCE(E1, E2)
@ depends on patch @
expression E;
@@
- ACCESS_ONCE(E)
+ READ_ONCE(E)
----
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: davem@davemloft.net
Cc: linux-arch@vger.kernel.org
Cc: mpe@ellerman.id.au
Cc: shuah@kernel.org
Cc: snitzer@redhat.com
Cc: thor.thayer@linux.intel.com
Cc: tj@kernel.org
Cc: viro@zeniv.linux.org.uk
Cc: will.deacon@arm.com
Link: http://lkml.kernel.org/r/1508792849-3115-19-git-send-email-paulmck@linux.vnet.ibm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The fake symbols we create for inlined frames will represent different
functions but can use the symbol start address. This leads to issues
when different inline branches all lead to the same function.
Before:
~~~~~
$ perf report -s sym -i perf.inlining.data --inline --stdio -g function
...
--38.86%--_start
__libc_start_main
main
|
--37.57%--std::norm<double> (inlined)
std::_Norm_helper<true>::_S_do_it<double> (inlined)
|
--36.36%--std::abs<double> (inlined)
std::__complex_abs (inlined)
|
--12.24%--std::linear_congruential_engine<unsigned long, 16807ul, 0ul, 2147483647ul>::operator() (inlined)
std::__detail::__mod<unsigned long, 2147483647ul, 16807ul, 0ul> (inlined)
std::__detail::_Mod<unsigned long, 2147483647ul, 16807ul, 0ul, true, true>::__calc (inlined)
~~~~~
Note that this backtrace representation is completely bogus.
Complex abs does not call the linear congruential engine! It
is just a side-effect of a longer inlined stack being appended
to a shorter, different inlined stack, both of which originate
in the same function (main).
This patch fixes the issue:
~~~~~
$ perf report -s sym -i perf.inlining.data --inline --stdio -g function
...
--38.86%--_start
__libc_start_main
main
|
|--35.59%--std::uniform_real_distribution<double>::operator()<std::linear_congruential_engine<unsigned long, 16807ul, 0ul, 2147483647ul> > (inlined)
| std::uniform_real_distribution<double>::operator()<std::linear_congruential_engine<unsigned long, 16807ul, 0ul, 2147483647ul> > (inlined)
| |
| --34.37%--std::__detail::_Adaptor<std::linear_congruential_engine<unsigned long, 16807ul, 0ul, 2147483647ul>, double>::operator() (inlined)
| std::generate_canonical<double, 53ul, std::linear_congruential_engine<unsigned long, 16807ul, 0ul, 2147483647ul> > (inlined)
| |
| --12.24%--std::linear_congruential_engine<unsigned long, 16807ul, 0ul, 2147483647ul>::operator() (inlined)
| std::__detail::__mod<unsigned long, 2147483647ul, 16807ul, 0ul> (inlined)
| std::__detail::_Mod<unsigned long, 2147483647ul, 16807ul, 0ul, true, true>::__calc (inlined)
|
--1.99%--std::norm<double> (inlined)
std::_Norm_helper<true>::_S_do_it<double> (inlined)
std::abs<double> (inlined)
std::__complex_abs (inlined)
~~~~~
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Reviewed-by: Jiri Olsa <jolsa@redhat.com>
Reviewed-by: Namhyung Kim <namhyung@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
Cc: Yao Jin <yao.jin@linux.intel.com>
Link: http://lkml.kernel.org/r/20171009203310.17362-10-milian.wolff@kdab.com
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
[ Fix up conflict with c1fbc0cf81 ("perf callchain: Compare dsos (as well) for CCKEY_FUNCTION"), remove unneeded hunk ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The original patch that introduced inline frame output in the various
browsers used this suffix already. The new centralized approach that
uses fake symbols for inlined frames was missing this approach so far.
Instead of changing the symbol name itself, we only print the suffix
where needed. This allows us to efficiently lookup the symbol for a
given name without first having to append the suffix before the lookup.
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Reviewed-by: Jiri Olsa <jolsa@redhat.com>
Reviewed-by: Namhyung Kim <namhyung@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Yao Jin <yao.jin@linux.intel.com>
Link: http://lkml.kernel.org/r/20171009203310.17362-8-milian.wolff@kdab.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When a callchain entry has no srcline available, we ended up comparing
the instruction pointer. I consider this to be not too useful. Rather, I
think we should group the entries by function name, which this patch
adds. For people who want to split the data on the IP boundary, using
`-g address` is the correct choice.
Before:
~~~~~
100.00% 38.86% [.] main
|
|--61.14%--main inlining.cpp:14
| std::norm<double> complex:664
| std::_Norm_helper<true>::_S_do_it<double> complex:654
| std::abs<double> complex:597
| std::__complex_abs complex:589
| |
| |--56.03%--hypot
| | |
| | |--8.45%--__hypot_finite
| | |
| | |--7.62%--__hypot_finite
| | |
| | |--2.29%--__hypot_finite
| | |
| | |--2.24%--__hypot_finite
| | |
| | |--2.06%--__hypot_finite
| | |
| | |--1.81%--__hypot_finite
...
~~~~~
After:
~~~~~
100.00% 38.86% [.] main
|
|--61.14%--main inlining.cpp:14
| std::norm<double> complex:664
| std::_Norm_helper<true>::_S_do_it<double> complex:654
| std::abs<double> complex:597
| std::__complex_abs complex:589
| |
| |--60.29%--hypot
| | |
| | --56.03%--__hypot_finite
| |
| --0.85%--cabs
~~~~~
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Reviewed-by: Jiri Olsa <jolsa@redhat.com>
Reviewed-by: Namhyung Kim <namhyung@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Yao Jin <yao.jin@linux.intel.com>
Link: http://lkml.kernel.org/r/20171009203310.17362-7-milian.wolff@kdab.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The inline_node structs are maintained by the new dso->inlines tree.
This in turn keeps ownership of the fake symbols and srcline string
representing an inline frame.
This tree is sorted by address to allow quick lookups. All other entries
of the symbol beside the function name are unused for inline frames. The
advantage of this approach is that all existing users of the callchain
API can now transparently display inlined frames without having to patch
their code.
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Reviewed-by: Jiri Olsa <jolsa@redhat.com>
Reviewed-by: Namhyung Kim <namhyung@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Yao Jin <yao.jin@linux.intel.com>
Link: http://lkml.kernel.org/r/20171009203310.17362-6-milian.wolff@kdab.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This is a requirement to create real callchain entries for inlined
frames.
Since the list of inlines usually contains the target symbol too, i.e.
the location where the frames get inlined to, we alias that symbol and
reuse it as-is is. This ensures that other dependent functionality keeps
working, most notably annotation of the target frames.
For all other entries in the inline_list, a fake symbol is created.
These are marked by new 'inlined' member which is set to true. Only
those symbols are managed by the inline_list and get freed when the
inline_list is deleted from within inline_node__delete.
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Reviewed-by: Jiri Olsa <jolsa@redhat.com>
Reviewed-by: Namhyung Kim <namhyung@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Yao Jin <yao.jin@linux.intel.com>
Link: http://lkml.kernel.org/r/20171009203310.17362-4-milian.wolff@kdab.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This is mostly a preparation to enable the creation of full callchain
nodes for inline frames. Such frames will reference the IP of the
non-inlined frame, but hold the symbol and srcline for an inlined
location. As such, we won't be able to query the srcline on-demand based
on the IP alone. Instead, we will leverage the functionality provided by
this patch here, and store the srcline for the inlined nodes in the new
srcline member of callchain_cursor_node.
Note that this patch on its own leaks the srcline, as there is no
free_callchain_cursor_node or similar. A future patch will add caching
of the srcline and handle deletion properly.
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Reviewed-by: Jiri Olsa <jolsa@redhat.com>
Reviewed-by: Namhyung Kim <namhyung@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Yao Jin <yao.jin@linux.intel.com>
Link: http://lkml.kernel.org/r/20171009203310.17362-3-milian.wolff@kdab.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In current xyarray code, xyarray__max_x() returns max_y, and xyarray__max_y()
returns max_x.
It's confusing and for code logic it looks not correct.
Error happens when closing evsel fd. Let's see this scenario:
1. Allocate an fd (pseudo-code)
perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
{
evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
}
xyarray__new(int xlen, int ylen, size_t entry_size)
{
size_t row_size = ylen * entry_size;
struct xyarray *xy = zalloc(sizeof(*xy) + xlen * row_size);
xy->entry_size = entry_size;
xy->row_size = row_size;
xy->entries = xlen * ylen;
xy->max_x = xlen;
xy->max_y = ylen;
......
}
So max_x is ncpus, max_y is nthreads and row_size = nthreads * 4.
2. Use perf syscall and get the fd
int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
struct thread_map *threads)
{
for (cpu = 0; cpu < cpus->nr; cpu++) {
for (thread = 0; thread < nthreads; thread++) {
int fd, group_fd;
fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
group_fd, flags);
FD(evsel, cpu, thread) = fd;
}
}
static inline void *xyarray__entry(struct xyarray *xy, int x, int y)
{
return &xy->contents[x * xy->row_size + y * xy->entry_size];
}
These codes don't have issues. The issue happens in the closing of fd.
3. Close fd.
void perf_evsel__close_fd(struct perf_evsel *evsel)
{
int cpu, thread;
for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
close(FD(evsel, cpu, thread));
FD(evsel, cpu, thread) = -1;
}
}
Since xyarray__max_x() returns max_y (nthreads) and xyarry__max_y()
returns max_x (ncpus), so above code is actually to be:
for (cpu = 0; cpu < nthreads; cpu++)
for (thread = 0; thread < ncpus; ++thread) {
close(FD(evsel, cpu, thread));
FD(evsel, cpu, thread) = -1;
}
It's not correct!
This change is introduced by "475fb533fb7d" ("perf evsel: Fix buffer overflow
while freeing events")
This fix is to let xyarray__max_x() return max_x (ncpus) and
let xyarry__max_y() return max_y (nthreads)
Committer note:
This was also fixed by Ravi Bangoria, who provided the same patch,
noticing the problem with 'perf record':
<quote Ravi>
I see 'perf record -p <pid>' crashes with following log:
*** Error in `./perf': free(): invalid next size (normal): 0x000000000298b340 ***
======= Backtrace: =========
/lib/x86_64-linux-gnu/libc.so.6(+0x777e5)[0x7f7fd85c87e5]
/lib/x86_64-linux-gnu/libc.so.6(+0x8037a)[0x7f7fd85d137a]
/lib/x86_64-linux-gnu/libc.so.6(cfree+0x4c)[0x7f7fd85d553c]
./perf(perf_evsel__close+0xb4)[0x4b7614]
./perf(perf_evlist__delete+0x100)[0x4ab180]
./perf(cmd_record+0x1d9)[0x43a5a9]
./perf[0x49aa2f]
./perf(main+0x631)[0x427841]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf0)[0x7f7fd8571830]
./perf(_start+0x29)[0x427a59]
</>
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
Fixes: d74be47673 ("perf xyarray: Save max_x, max_y")
Link: http://lkml.kernel.org/r/1508339478-26674-1-git-send-email-yao.jin@linux.intel.com
Link: http://lkml.kernel.org/r/1508327446-15302-1-git-send-email-ravi.bangoria@linux.vnet.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Thomas reported that 'perf buildid-list' gets a SEGFAULT due to NULL
pointer deref when he ran it on a data with namespace events. It was
because the buildid_id__mark_dso_hit_ops lacks the namespace event
handler and perf_too__fill_default() didn't set it.
Program received signal SIGSEGV, Segmentation fault.
0x0000000000000000 in ?? ()
Missing separate debuginfos, use: dnf debuginfo-install audit-libs-2.7.7-1.fc25.s390x bzip2-libs-1.0.6-21.fc25.s390x elfutils-libelf-0.169-1.fc25.s390x
+elfutils-libs-0.169-1.fc25.s390x libcap-ng-0.7.8-1.fc25.s390x numactl-libs-2.0.11-2.ibm.fc25.s390x openssl-libs-1.1.0e-1.1.ibm.fc25.s390x perl-libs-5.24.1-386.fc25.s390x
+python-libs-2.7.13-2.fc25.s390x slang-2.3.0-7.fc25.s390x xz-libs-5.2.3-2.fc25.s390x zlib-1.2.8-10.fc25.s390x
(gdb) where
#0 0x0000000000000000 in ?? ()
#1 0x00000000010fad6a in machines__deliver_event (machines=<optimized out>, machines@entry=0x2c6fd18,
evlist=<optimized out>, event=event@entry=0x3fffdf00470, sample=0x3ffffffe880, sample@entry=0x3ffffffe888,
tool=tool@entry=0x1312968 <build_id.mark_dso_hit_ops>, file_offset=1136) at util/session.c:1287
#2 0x00000000010fbf4e in perf_session__deliver_event (file_offset=1136, tool=0x1312968 <build_id.mark_dso_hit_ops>,
sample=0x3ffffffe888, event=0x3fffdf00470, session=0x2c6fc30) at util/session.c:1340
#3 perf_session__process_event (session=0x2c6fc30, session@entry=0x0, event=event@entry=0x3fffdf00470,
file_offset=file_offset@entry=1136) at util/session.c:1522
#4 0x00000000010fddde in __perf_session__process_events (file_size=11880, data_size=<optimized out>,
data_offset=<optimized out>, session=0x0) at util/session.c:1899
#5 perf_session__process_events (session=0x0, session@entry=0x2c6fc30) at util/session.c:1953
#6 0x000000000103b2ac in perf_session__list_build_ids (with_hits=<optimized out>, force=<optimized out>)
at builtin-buildid-list.c:83
#7 cmd_buildid_list (argc=<optimized out>, argv=<optimized out>) at builtin-buildid-list.c:115
#8 0x00000000010a026c in run_builtin (p=0x1311f78 <commands+24>, argc=argc@entry=2, argv=argv@entry=0x3fffffff3c0)
at perf.c:296
#9 0x000000000102bc00 in handle_internal_command (argv=<optimized out>, argc=2) at perf.c:348
#10 run_argv (argcp=<synthetic pointer>, argv=<synthetic pointer>) at perf.c:392
#11 main (argc=<optimized out>, argv=0x3fffffff3c0) at perf.c:536
(gdb)
Fix it by adding a stub event handler for namespace event.
Committer testing:
Further clarifying, plain using 'perf buildid-list' will not end up in a
SEGFAULT when processing a perf.data file with namespace info:
# perf record -a --namespaces sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 2.024 MB perf.data (1058 samples) ]
# perf buildid-list | wc -l
38
# perf buildid-list | head -5
e2a171c7b905826fc8494f0711ba76ab6abbd604 /lib/modules/4.14.0-rc3+/build/vmlinux
874840a02d8f8a31cedd605d0b8653145472ced3 /lib/modules/4.14.0-rc3+/kernel/arch/x86/kvm/kvm-intel.ko
ea7223776730cd8a22f320040aae4d54312984bc /lib/modules/4.14.0-rc3+/kernel/drivers/gpu/drm/i915/i915.ko
5961535e6732a8edb7f22b3f148bb2fa2e0be4b9 /lib/modules/4.14.0-rc3+/kernel/drivers/gpu/drm/drm.ko
f045f54aa78cf1931cc893f78b6cbc52c72a8cb1 /usr/lib64/libc-2.25.so
#
It is only when one asks for checking what of those entries actually had
samples, i.e. when we use either -H or --with-hits, that we will process
all the PERF_RECORD_ events, and since tools/perf/builtin-buildid-list.c
neither explicitely set a perf_tool.namespaces() callback nor the
default stub was set that we end up, when processing a
PERF_RECORD_NAMESPACE record, causing a SEGFAULT:
# perf buildid-list -H
Segmentation fault (core dumped)
^C
#
Reported-and-Tested-by: Thomas-Mich Richter <tmricht@linux.vnet.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Hari Bathini <hbathini@linux.vnet.ibm.com>
Cc: Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas-Mich Richter <tmricht@linux.vnet.ibm.com>
Fixes: f3b3614a28 ("perf tools: Add PERF_RECORD_NAMESPACES to include namespaces related info")
Link: http://lkml.kernel.org/r/20171017132900.11043-1-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Adding the check wether the eBPF file exists, to consider it
as eBPF input file. This way we can differentiate eBPF events
from events that end up with same suffix as eBPF file.
Before:
$ perf stat -e 'cpu/uops_executed.core/' true
bpf: builtin compilation failed: -95, try external compiler
WARNING: unable to get correct kernel building directory.
Hint: Set correct kbuild directory using 'kbuild-dir' option in [llvm]
section of ~/.perfconfig or set it to "" to suppress kbuild
detection.
event syntax error: 'cpu/uops_executed.core/'
\___ Failed to load cpu/uops_executed.c from source: 'version' section incorrect or lost
After:
$ perf stat -e 'cpu/uops_executed.core/' true
Performance counter stats for 'true':
181,533 cpu/uops_executed.core/:u
0.002795447 seconds time elapsed
If user makes type in the eBPF file, we prioritize the event syntax
and show following warning:
$ perf stat -e 'krava.c//' true
event syntax error: 'krava.c//'
\___ Cannot find PMU `krava.c'. Missing kernel support?
Reported-and-Tested-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Changbin Du <changbin.du@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/20171013083736.15037-9-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently, perf record is broken on arm/arm64 systems when the PMU is
specified explicitly as part of the event, e.g.
$ ./perf record -e armv8_cortex_a53/cpu_cycles/u true
In such cases, perf record fails to open events unless
perf_event_paranoid is set to -1, even if the PMU in question supports
mode exclusion. Further, even when perf_event_paranoid is toggled, no
samples are recorded.
This is an unintended side effect of commit:
e3ba76deef ("perf tools: Force uncore events to system wide monitoring)
... which assumes that if a PMU has an associated cpu_map, it is an
uncore PMU, and forces events for such PMUs to be system-wide.
This is not true for arm/arm64 systems, which can have heterogeneous
CPUs. To account for this, multiple CPU PMUs are exposed, each with a
"cpus" field under sysfs, which the perf tool parses into a cpu_map. ARM
PMUs do not have a "cpumask" file, and only have a "cpus" file. For the
gory details as to why, see commit:
7e3fcffe95 ("perf pmu: Support alternative sysfs cpumask")
Given all of this, we can instead identify uncore PMUs by explicitly
checking for a "cpumask" file, and restore arm/arm64 PMU support back to
a working state. This patch does so, adding a new perf_pmu::is_uncore
field, and splitting the existing cpumask parsing so that it can be
reused.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Tested-by Will Deacon <will.deacon@arm.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: 4.12+ <stable@vger.kernel.org>
Fixes: e3ba76deef ("perf tools: Force uncore events to system wide monitoring)
Link: http://lkml.kernel.org/r/1507315102-5942-1-git-send-email-mark.rutland@arm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The proc files which is sorted with alphabetical order are evenly
assigned to several synthesize threads to be processed in parallel.
For 'perf top', the threads number hard code to online CPU number. The
following patch will introduce an option to set it.
For other perf tools, the thread number is 1. Because the process
function is not ready for multithreading, e.g.
process_synthesized_event.
This patch series only support event synthesize multithreading for 'perf
top'. For other tools, it can be done separately later.
With multithread applied, the total processing time can get up to 1.56x
speedup on Knights Mill for 'perf top'.
For specific single event processing, the processing time could increase
because of the lock contention. So proc_map_timeout may need to be
increased. Otherwise some proc maps will be truncated.
Based on my test, increasing the proc_map_timeout has small impact
on the total processing time. The total processing time still get 1.49x
speedup on Knights Mill after increasing the proc_map_timeout.
The patch itself doesn't increase the proc_map_timeout.
Doesn't need to implement multithreading for per task monitoring,
perf_event__synthesize_thread_map. It doesn't have performance issue.
Committer testing:
# getconf _NPROCESSORS_ONLN
4
# perf trace --no-inherit -e clone -o /tmp/output perf top
# tail -4 /tmp/bla
0.124 ( 0.041 ms): clone(flags: VM|FS|FILES|SIGHAND|THREAD|SYSVSEM|SETTLS|PARENT_SETTID|CHILD_CLEARTID, child_stack: 0x7fc3eb3a8f30, parent_tidptr: 0x7fc3eb3a99d0, child_tidptr: 0x7fc3eb3a99d0, tls: 0x7fc3eb3a9700) = 9548 (perf)
0.246 ( 0.023 ms): clone(flags: VM|FS|FILES|SIGHAND|THREAD|SYSVSEM|SETTLS|PARENT_SETTID|CHILD_CLEARTID, child_stack: 0x7fc3eaba7f30, parent_tidptr: 0x7fc3eaba89d0, child_tidptr: 0x7fc3eaba89d0, tls: 0x7fc3eaba8700) = 9549 (perf)
0.286 ( 0.019 ms): clone(flags: VM|FS|FILES|SIGHAND|THREAD|SYSVSEM|SETTLS|PARENT_SETTID|CHILD_CLEARTID, child_stack: 0x7fc3ea3a6f30, parent_tidptr: 0x7fc3ea3a79d0, child_tidptr: 0x7fc3ea3a79d0, tls: 0x7fc3ea3a7700) = 9550 (perf)
246.540 ( 0.047 ms): clone(flags: VM|FS|FILES|SIGHAND|THREAD|SYSVSEM|SETTLS|PARENT_SETTID|CHILD_CLEARTID, child_stack: 0x7fc3ea3a6f30, parent_tidptr: 0x7fc3ea3a79d0, child_tidptr: 0x7fc3ea3a79d0, tls: 0x7fc3ea3a7700) = 9551 (perf)
#
Signed-off-by: Kan Liang <kan.liang@intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: He Kuang <hekuang@huawei.com>
Cc: Lukasz Odzioba <lukasz.odzioba@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1506696477-146932-4-git-send-email-kan.liang@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The build of kernel v4.14-rc1 for i686 fails on RHEL 6 with the error
in tools/perf:
util/syscalltbl.c:157: error: expected ';', ',' or ')' before '__maybe_unused'
mv: cannot stat `util/.syscalltbl.o.tmp': No such file or directory
Fix it by placing/moving:
#include <linux/compiler.h>
outside of #ifdef HAVE_SYSCALL_TABLE block.
Signed-off-by: Akemi Yagi <toracat@elrepo.org>
Cc: Alan Bartlett <ajb@elrepo.org>
Link: http://lkml.kernel.org/r/oq41r8$1v9$1@blaine.gmane.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
With --call-graph option, perf report can display call chains using
type, min percent threshold, optional print limit and order. And the
default call-graph parameter is 'graph,0.5,caller,function,percent'.
Before this patch, 'perf report --call-graph' shows incorrect debug
messages as below:
# perf report --call-graph
Invalid callchain mode: 0.5
Invalid callchain order: 0.5
Invalid callchain sort key: 0.5
Invalid callchain config key: 0.5
Invalid callchain mode: caller
Invalid callchain mode: function
Invalid callchain order: function
Invalid callchain mode: percent
Invalid callchain order: percent
Invalid callchain sort key: percent
That is because in function __parse_callchain_report_opt(),each field of
the call-graph parameter is passed to parse_callchain_{mode,order,
sort_key,value} in turn until it meets the matching value.
For example, the order field "caller" is passed to
parse_callchain_mode() firstly and obviously it doesn't match any mode
field. Therefore parse_callchain_mode() will shows the debug message
"Invalid callchain mode: caller", which could confuse users.
The patch fixes this issue by moving the warning out of the function
parse_callchain_{mode,order,sort_key,value}.
Signed-off-by: Mengting Zhang <zhangmengting@huawei.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Krister Johansen <kjlx@templeofstupid.com>
Cc: Li Bin <huawei.libin@huawei.com>
Cc: Milian Wolff <milian.wolff@kdab.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Cc: Yao Jin <yao.jin@linux.intel.com>
Link: http://lkml.kernel.org/r/1506154694-39691-1-git-send-email-zhangmengting@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Yet another fix for probing the max attr.precise_ip setting: it is not
enough settting attr.exclude_kernel for !root users, as they _can_
profile the kernel if the kernel.perf_event_paranoid sysctl is set to
-1, so check that as well.
Testing it:
As non root:
$ sysctl kernel.perf_event_paranoid
kernel.perf_event_paranoid = 2
$ perf record sleep 1
$ perf evlist -v
cycles:uppp: ..., exclude_kernel: 1, ... precise_ip: 3, ...
Now as non-root, but with kernel.perf_event_paranoid set set to the
most permissive value, -1:
$ sysctl kernel.perf_event_paranoid
kernel.perf_event_paranoid = -1
$ perf record sleep 1
$ perf evlist -v
cycles:ppp: ..., exclude_kernel: 0, ... precise_ip: 3, ...
$
I.e. non-root, default kernel.perf_event_paranoid: :uppp modifier = not allowed to sample the kernel,
non-root, most permissible kernel.perf_event_paranoid: :ppp = allowed to sample the kernel.
In both cases, use the highest available precision: attr.precise_ip = 3.
Reported-and-Tested-by: Ingo Molnar <mingo@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Fixes: d37a369790 ("perf evsel: Fix attr.exclude_kernel setting for default cycles:p")
Link: http://lkml.kernel.org/n/tip-nj2qkf75xsd6pw6hhjzfqqdx@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Andi reported a performance drop in single threaded perf tools such as
'perf script' due to the growing number of locks being put in place to
allow for multithreaded tools, so wrap the POSIX threads rwlock routines
with the names used for such kinds of locks in the Linux kernel and then
allow for tools to ask for those locks to be used or not.
I.e. a tool may have a multithreaded phase and then switch to single
threaded, like the upcoming patches for the synthesizing of
PERF_RECORD_{FORK,MMAP,etc} for pre-existing processes to then switch to
single threaded mode in 'perf top'.
The init routines will not be conditional, this way starting as single
threaded to then move to multi threaded mode should be possible.
Reported-by: Andi Kleen <ak@linux.intel.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/20170404161739.GH12903@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The -M metric group parser threw away the events of earlier groups when
multiple groups were specified. Fix this here by not overwriting the
string incorrectly.
Now this works correctly:
% perf stat -M Summary,SMT --metric-only -a sleep 1
Performance counter stats for 'system wide':
Instructions CPI CLKS CPU_Utilization GFLOPs SMT_2T_Utilization SMT_2T_Utilization Kernel_Utilization CoreIPC CORE_CLKS
900907376.0 2.7 2398954144.0 0.1 0.0 0.2 0.2 0.1 0.4 2080822855.5
while previously it would only show the SMT metrics.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170914205735.18431-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When a PMU is missing print a better error message mentioning
the missing PMU.
% mkdir empty
% mount --bind empty /sys/devices/msr
% perf stat -M Summary true
event syntax error: '{inst_retired.any,cycles}:W,{cpu_clk_unhalted.thread}:W,{inst_retired.any}:W,{cpu_clk_unhalted.ref_tsc,msr/tsc/}:W,{fp_comp_ops_exe.sse_scalar..'
\___ Cannot find PMU `msr'. Missing kernel support?
It still cannot find the right column for aliases, but it's already a vast improvement.
v2: Check asprintf
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170913215006.32222-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To process any events, it needs to find the thread in the machine first.
The machine maintains a rb tree to store all threads. The rb tree is
protected by a rw lock.
It is not a problem for current perf which serially processing events.
However, it will have scalability performance issue to process events in
parallel, especially on a heavy load system which have many threads.
Introduce a hashtable to divide the big rb tree into many samll rb tree
for threads. The index is thread id % hashtable size. It can reduce the
lock contention.
Committer notes:
Renamed some variables and function names to reduce semantic confusion:
'struct threads' pointers: thread -> threads
threads hastable index: tid -> hash_bucket
struct threads *machine__thread() -> machine__threads()
Cast tid to (unsigned int) to handle -1 in machine__threads() (Kan Liang)
Signed-off-by: Kan Liang <kan.liang@intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Lukasz Odzioba <lukasz.odzioba@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1505096603-215017-2-git-send-email-kan.liang@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When there isn't a config file (e.g. ~/.perfconfig) or it has nothing,
the config set wasn't created.
If the config set does not exist, a config file can't be autogenerated.
So allow creating a empty config set in the above case,
then we can support the config file autogeneration.
Before:
$ rm -f ~/.perfconfig
$ perf config --user report.children=false
$ cat ~/.perfconfig
cat: /root/.perfconfig: No such file or directory
But I think it should work even if there isn't a config file.
After:
$ rm -f ~/.perfconfig
$ perf config --user report.children=false
$ cat ~/.perfconfig
# this file is auto-generated.
[report]
children = false
NOTE:
As a result, if perf_config_set__init() fails, it looks as if the config
set isn't freed. But it isn't a problem. Because the config set will be
freed by perf_config_set__delete() at the end of cmd_config().
Signed-off-by: Taeung Song <treeze.taeung@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1504754336-9824-1-git-send-email-treeze.taeung@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
