In LBR call stack mode, the depth of reconstructed LBR call stack limits
to the number of LBR registers.
For example, on skylake, the depth of reconstructed LBR call stack is
always <= 32.
# To display the perf.data header info, please use
# --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 6K of event 'cycles'
# Event count (approx.): 6487119731
#
# Children Self Command Shared Object Symbol
# ........ ........ ............... ..................
# ................................
99.97% 99.97% tchain_edit tchain_edit [.] f43
|
--99.64%--f11
f12
f13
f14
f15
f16
f17
f18
f19
f20
f21
f22
f23
f24
f25
f26
f27
f28
f29
f30
f31
f32
f33
f34
f35
f36
f37
f38
f39
f40
f41
f42
f43
For a call stack which is deeper than LBR limit, HW will overwrite the
LBR register with oldest branch. Only partial call stacks can be
reconstructed.
However, the overwritten LBRs may still be retrieved from previous
sample. At that moment, HW hasn't overwritten the LBR registers yet.
Perf tools can stitch those overwritten LBRs on current call stacks to
get a more complete call stack.
To determine if LBRs can be stitched, perf tools need to compare current
sample with previous sample.
- They should have identical LBR records (Same from, to and flags
values, and the same physical index of LBR registers).
- The searching starts from the base-of-stack of current sample.
Once perf determines to stitch the previous LBRs, the corresponding LBR
cursor nodes will be copied to 'lists'. The 'lists' is to track the LBR
cursor nodes which are going to be stitched.
When the stitching is over, the nodes will not be freed immediately.
They will be moved to 'free_lists'. Next stitching may reuse the space.
Both 'lists' and 'free_lists' will be freed when all samples are
processed.
Committer notes:
Fix the intel-pt.c initialization of the union with 'struct
branch_flags', that breaks the build with its unnamed union on older gcc
versions.
Uninline thread__free_stitch_list(), as it grew big and started dragging
includes to thread.h, so move it to thread.c where what it needs in
terms of headers are already there.
This fixes the build in several systems such as debian:experimental when
cross building to the MIPS32 architecture, i.e. in the other cases what
was needed was being included by sheer luck.
In file included from builtin-sched.c:11:
util/thread.h: In function 'thread__free_stitch_list':
util/thread.h:169:3: error: implicit declaration of function 'free' [-Werror=implicit-function-declaration]
169 | free(pos);
| ^~~~
util/thread.h:169:3: error: incompatible implicit declaration of built-in function 'free' [-Werror]
util/thread.h:19:1: note: include '<stdlib.h>' or provide a declaration of 'free'
18 | #include "callchain.h"
+++ |+#include <stdlib.h>
19 |
util/thread.h:174:3: error: incompatible implicit declaration of built-in function 'free' [-Werror]
174 | free(pos);
| ^~~~
util/thread.h:174:3: note: include '<stdlib.h>' or provide a declaration of 'free'
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Pavel Gerasimov <pavel.gerasimov@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Vitaly Slobodskoy <vitaly.slobodskoy@intel.com>
Link: http://lore.kernel.org/lkml/20200319202517.23423-13-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
And pick the shortest name: 'struct maps'.
The split existed because we used to have two groups of maps, one for
functions and one for variables, but that only complicated things,
sometimes we needed to figure out what was at some address and then had
to first try it on the functions group and if that failed, fall back to
the variables one.
That split is long gone, so for quite a while we had only one struct
maps per struct map_groups, simplify things by combining those structs.
First patch is the minimum needed to merge both, follow up patches will
rename 'thread->mg' to 'thread->maps', etc.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lkml.kernel.org/n/tip-hom6639ro7020o708trhxh59@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Commit e5adfc3e7e ("perf map: Synthesize maps only for thread group
leader") changed the recording side so that we no longer get mmap events
for threads other than the thread group leader (when synthesising these
events for threads which exist before perf is started).
When a file recorded after this change is loaded, the lack of mmap
records mean that unwinding is not set up for any other threads.
This can be seen in a simple record/report scenario:
perf record --call-graph=dwarf -t $TID
perf report
If $TID is a process ID then the report will show call graphs, but if
$TID is a secondary thread the output is as if --call-graph=none was
specified.
Following the rationale in that commit, move the libunwind fields into
struct map_groups and update the libunwind functions to take this
instead of the struct thread. This is only required for
unwind__finish_access which must now be called from map_groups__delete
and the others are changed for symmetry.
Note that unwind__get_entries keeps the thread argument since it is
required for symbol lookup and the libdw unwind provider uses the thread
ID.
Signed-off-by: John Keeping <john@metanate.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Fixes: e5adfc3e7e ("perf map: Synthesize maps only for thread group leader")
Link: http://lkml.kernel.org/r/20190815100146.28842-2-john@metanate.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Threads are created when we either synthesize PERF_RECORD_FORK events
for pre-existing threads or when we receive PERF_RECORD_FORK events from
the kernel as new threads get created.
We then keep them in machine->threads[].entries rb trees till when we
receive a PERF_RECORD_EXIT, i.e. that thread terminated.
The thread object has a reference count that is grabbed when, for
instance, we keep that thread referenced in struct hist_entry, in 'perf
report' and 'perf top'.
When we receive a PERF_RECORD_EXIT we remove the thread object from the
rb tree and move it to the corresponding machine->threads[].dead list,
then we do a thread__put(), dropping the reference we had for keeping it
in the rb tree.
In thread__put() we were assuming that when the reference count hit zero
we should remove it from the dead list by simply doing a
list_del_init(&thread->node).
That works well when all the thread lifetime is during the machine that
has the list heads lifetime, since we know that we can do the
list_del_init() and it will update the 'dead' list_head.
But in 'perf sched lat' we were doing:
machine__new() (via perf_session__new)
process events, grabbing refcounts to keep those thread objects
in 'perf sched' local data structures.
machine__exit() (via perf_session__delete) which would delete the
'dead' list heads.
And then doing the final thread__put() for the refcounts 'perf sched'
rightfully obtained for keeping those thread object references.
b00m, since thread__put() would do the list_del_init() touching
a dead dead list head.
Fix it by removing all the dead threads from machine->threads[].dead at
machine__exit(), since whatever is there should have refcounts taken by
things like 'perf sched lat', and make thread__put() check if the thread
is in a linked list before removing it from that list.
Reported-by: Wei Li <liwei391@huawei.com>
Link: https://lkml.kernel.org/r/20190508143648.8153-1-liwei391@huawei.com
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Zhipeng Xie <xiezhipeng1@huawei.com>
Link: https://lkml.kernel.org/r/20190704194355.GI10740@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When compiled with libunwind, perf does some preparatory work when
processing side-band events. This is not needed when report actually
don't unwind dwarf callchains, so it's disabled with
dwarf_callchain_users bool.
However we could move that check to higher level and shield more
unwanted code for normal report processing, giving us following speed up
on kernel build profile:
Before:
$ perf record make -j40
...
$ ll ../../perf.data
-rw-------. 1 jolsa jolsa 461783932 Apr 26 09:11 perf.data
$ perf stat -e cycles:u,instructions:u perf report -i perf.data > out
Performance counter stats for 'perf report -i perf.data':
78,669,920,155 cycles:u
99,076,431,951 instructions:u # 1.26 insn per cycle
55.382823668 seconds time elapsed
27.512341000 seconds user
27.712871000 seconds sys
After:
$ perf stat -e cycles:u,instructions:u perf report -i perf.data > out
Performance counter stats for 'perf report -i perf.data':
59,626,798,904 cycles:u
88,583,575,849 instructions:u # 1.49 insn per cycle
21.296935559 seconds time elapsed
20.010191000 seconds user
1.202935000 seconds sys
The speed is higher with profile having many side-band events,
because these trigger libunwind preparatory code.
This does not apply for perf compiled with libdw for dwarf unwind,
only for build with libunwind.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20190426073804.17238-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a utility function to fetch executable code. Convert one
user over to it. There are more places doing that, but they
do significantly different actions, so they are not
easy to fit into a single library function.
Committer changes:
. No need to cast around, make 'buf' be a void pointer.
. Rename it to thread__memcpy() to reflect the fact it is about copying
a chunk of memory from a thread, i.e. from its address space.
. No need to have it in a separate object file, move it to thread.[ch]
. Check the return of map__load(), the original code didn't do it, but
since we're moving this around, check that as well.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lkml.kernel.org/r/20190305144758.12397-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When looking at PT or brstackinsn traces with 'perf script' it can be
very useful to see the source code. This adds a simple facility to print
them with 'perf script', if the information is available through dwarf
% perf record ...
% perf script -F insn,ip,sym,srccode
...
4004c6 main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004c6 main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004b3 main
6 v++;
% perf record -b ...
% perf script -F insn,ip,sym,srccode,brstackinsn
...
main+22:
0000000000400543 insn: e8 ca ff ff ff # PRED
|18 f1();
f1:
0000000000400512 insn: 55
|10 {
0000000000400513 insn: 48 89 e5
0000000000400516 insn: b8 00 00 00 00
|11 f2();
000000000040051b insn: e8 d6 ff ff ff # PRED
f2:
00000000004004f6 insn: 55
|5 {
00000000004004f7 insn: 48 89 e5
00000000004004fa insn: 8b 05 2c 0b 20 00
|6 c = a / b;
0000000000400500 insn: 8b 0d 2a 0b 20 00
0000000000400506 insn: 99
0000000000400507 insn: f7 f9
0000000000400509 insn: 89 05 29 0b 20 00
000000000040050f insn: 90
|7 }
0000000000400510 insn: 5d
0000000000400511 insn: c3 # PRED
f1+14:
0000000000400520 insn: b8 00 00 00 00
|12 f2();
0000000000400525 insn: e8 cc ff ff ff # PRED
f2:
00000000004004f6 insn: 55
|5 {
00000000004004f7 insn: 48 89 e5
00000000004004fa insn: 8b 05 2c 0b 20 00
|6 c = a / b;
Not supported for callchains currently, would need some layout changes
there.
Committer notes:
Fixed the build on Alpine Linux (3.4 .. 3.8) by addressing this
warning:
In file included from util/srccode.c:19:0:
/usr/include/sys/fcntl.h:1:2: error: #warning redirecting incorrect #include <sys/fcntl.h> to <fcntl.h> [-Werror=cpp]
#warning redirecting incorrect #include <sys/fcntl.h> to <fcntl.h>
^~~~~~~
cc1: all warnings being treated as errors
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20181204001848.24769-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When synthesizing FORK events, we are trying to create thread objects
for the already running tasks on the machine.
Normally, for a kernel FORK event, we want to clone the parent's maps
because that is what the kernel just did.
But when synthesizing, this should not be done. If we do, we end up
with overlapping maps as we process the sythesized MMAP2 events that
get delivered shortly thereafter.
Use the FORK event misc flags in an internal way to signal this
situation, so we can elide the map clone when appropriate.
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Joe Mario <jmario@redhat.com>
Link: http://lkml.kernel.org/r/20181030.222404.2085088822877051075.davem@davemloft.net
[ Added comment about flag use in machine__process_fork_event(),
use ternary op in thread__clone_map_groups() as suggested by Jiri ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Remove the split of symbol tables for data (MAP__VARIABLE) and for
functions (MAP__FUNCTION), its unneeded and there were various places
doing two lookups to find a symbol, so simplify this.
We still will consider only the symbols that matched the filters in
place, i.e. see the (elf_(sec,sym)|symbol_type)__filter() routines in
the patch, just so that we consider only the same symbols as before,
to reduce the possibility of regressions.
All the tests on 50-something build environments, in varios versions
of lots of distros and cross build environments were performed without
build regressions, as usual with all pull requests the other tests were
also performed: 'perf test' and 'make -C tools/perf build-test'.
Also this was done at a great granularity so that regressions can be
bisected more easily.
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: https://lkml.kernel.org/n/tip-hiq0fy2rsleupnqqwuojo1ne@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Out of thread__find_addr_location(..., MAP__FUNCTION, ...), idea here is to
continue removing references to MAP__{FUNCTION,VARIABLE} ahead of
getting both types of symbols in the same rbtree, as various places do
two lookups, looking first at MAP__FUNCTION, then at MAP__VARIABLE.
So thread__find_symbol() will eventually do just that, and 'struct
symbol' will have the symbol type, for code that cares about that.
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: https://lkml.kernel.org/n/tip-n7528en9e08yd3flzmb26tth@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
