There's a problem on finding correct kernel symbols when perf report
runs on a different kernel. Although a part of the problem was solved
by the prior commit 0a7e6d1b68 ("perf tools: Check recorded kernel
version when finding vmlinux"), there's a remaining problem still.
When perf records samples, it synthesizes the kernel map using
machine__mmap_name() and ref_reloc_sym like "[kernel.kallsyms]_text".
You can easily see it using 'perf report -D' command.
After finishing record, it goes through the recorded events to find
maps/dsos actually used. And then record build-id info of them.
During this process, it needs to load symbols in a dso and it'd call
dso__load_vmlinux_path() since the default value of the symbol_conf.
try_vmlinux_path is true. However it changes dso->long_name to a real
path of the vmlinux file (e.g. /lib/modules/3.16.4/build/vmlinux) if one
is running on a custom kernel.
It resulted in that perf report reads the build-id of the vmlinux, but
cannot use it since it only knows about the [kernel.kallsyms] map. It
then falls back to possible vmlinux paths by using the recorded kernel
version (in case of a recent version) or a running kernel silently.
Even with the recent tools, this still has a possibility of breaking
the result. As the build directory is a symbolic link, if one built a
new kernel in the same directory with different source/config, the old
link to vmlinux will point the new file. So it's absolutely needed to
use build-id when finding a kernel image.
In this patch, it's now changed to try to search a kernel dso in the
existing dso list which was constructed during build-id table parsing
so it'll always have a build-id. If not found, search "[kernel.kallsyms]".
Before:
$ perf report
# Children Self Command Shared Object Symbol
# ........ ........ ....... ................. ...............................
#
72.15% 0.00% swapper [kernel.kallsyms] [k] set_curr_task_rt
72.15% 0.00% swapper [kernel.kallsyms] [k] native_calibrate_tsc
72.15% 0.00% swapper [kernel.kallsyms] [k] tsc_refine_calibration_work
71.87% 71.87% swapper [kernel.kallsyms] [k] module_finalize
...
After (for the same perf.data):
72.15% 0.00% swapper vmlinux [k] cpu_startup_entry
72.15% 0.00% swapper vmlinux [k] arch_cpu_idle
72.15% 0.00% swapper vmlinux [k] default_idle
71.87% 71.87% swapper vmlinux [k] native_safe_halt
...
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Ingo Molnar <mingo@kernel.org>
Link: http://lkml.kernel.org/r/20140924073356.GB1962@gmail.com
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1415063674-17206-8-git-send-email-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
A segfault happens on 'perf test hists_link' because we end up using a
struct machines on the stack, and then machines__init() was not
initializing the newly introduced rb_root, just the existing list_head.
When we introduced struct dsos, to group the two ways to store dsos,
i.e. the linked list and the rbtree, we didn't turned the initialization
done in:
machines__init(machines->host) ->
machine__init() ->
INIT_LIST_HEAD
into a dsos__init() to keep on initializing the list_head but _as well_
initializing the rb_root, oops.
All worked because outside perf-test we probably zalloc the whole thing
which ends up initializing it in to NULL.
So the problem looks contained to 'perf test' that uses it on stack,
etc.
Reported-by: Jiri Olsa <jolsa@redhat.com>
Acked-by: Waiman Long <Waiman.Long@hp.com>,
Cc: Adrian Hunter <adrian.hunter@intel.com>,
Cc: Don Zickus <dzickus@redhat.com>
Cc: Douglas Hatch <doug.hatch@hp.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 <peterz@infradead.org>
Cc: Scott J Norton <scott.norton@hp.com>
Cc: Waiman Long <Waiman.Long@hp.com>,
Link: http://lkml.kernel.org/r/20141014180353.GF3198@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
With workload that spawns and destroys many threads and processes, it
was found that perf-mem could took a long time to post-process the perf
data after the target workload had completed its operation.
The performance bottleneck was found to be the lookup and insertion of
the new DSO structures (thousands of them in this case).
In a dual-socket Ivy-Bridge E7-4890 v2 machine (30-core, 60-thread), the
perf profile below shows what perf was doing after the profiled AIM7
shared workload completed:
- 83.94% perf libc-2.11.3.so [.] __strcmp_sse42
- __strcmp_sse42
- 99.82% map__new
machine__process_mmap_event
perf_session_deliver_event
perf_session__process_event
__perf_session__process_events
cmd_record
cmd_mem
run_builtin
main
__libc_start_main
- 13.17% perf perf [.] __dsos__findnew
__dsos__findnew
map__new
machine__process_mmap_event
perf_session_deliver_event
perf_session__process_event
__perf_session__process_events
cmd_record
cmd_mem
run_builtin
main
__libc_start_main
So about 97% of CPU times were spent in the map__new() function trying
to insert new DSO entry into the DSO linked list. The whole
post-processing step took about 9 minutes.
The DSO structures are currently searched linearly. So the total
processing time will be proportional to n^2.
To overcome this performance problem, the DSO code is modified to also
put the DSO structures in a RB tree sorted by its long name in
additional to being in a simple linked list. With this change, the
processing time will become proportional to n*log(n) which will be much
quicker for large n. However, the short name will still be searched
using the old linear searching method. With that patch in place, the
same perf-mem post-processing step took less than 30 seconds to
complete.
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Douglas Hatch <doug.hatch@hp.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 <peterz@infradead.org>
Cc: Scott J Norton <scott.norton@hp.com>
Link: http://lkml.kernel.org/r/1412098575-27863-3-git-send-email-Waiman.Long@hp.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
As we run "perf c2c" on more applications, we noticed we're missing
significant samples from a common customer's application. Looking at
the /proc/<pid>/maps file for the app, we see "rwxs" and "rwxp"
permissions on many of the shared memory & heap regions, and on all the
thread stacks.
Because those regions have the "x" bit set, perf marks them with a
MAP_FUNCTION type. Hence ip_resolve_data() never finds load or store
events coming from them.
We fixed this by re-calling thread__find_addr_location with
MAP__FUNCTION in the case where map is NULL as a last ditch effort to
map the sample before giving up and dropping it.
Reported-by: Joe Mario <jmario@redhat.com>
Tested-by: Joe Mario <jmario@redhat.com>
Signed-off-by: Don Zickus <dzickus@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Joe Mario <jmario@redhat.com>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1408591511-57884-1-git-send-email-dzickus@redhat.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When saving the callchain on Power, the kernel conservatively saves excess
entries in the callchain. A few of these entries are needed in some cases
but not others. We should use the DWARF debug information to determine
when the entries are needed.
Eg: the value in the link register (LR) is needed only when it holds the
return address of a function. At other times it must be ignored.
If the unnecessary entries are not ignored, we end up with duplicate arcs
in the call-graphs.
Use the DWARF debug information to determine if any callchain entries
should be ignored when building call-graphs.
Callgraph before the patch:
14.67% 2234 sprintft libc-2.18.so [.] __random
|
--- __random
|
|--61.12%-- __random
| |
| |--97.15%-- rand
| | do_my_sprintf
| | main
| | generic_start_main.isra.0
| | __libc_start_main
| | 0x0
| |
| --2.85%-- do_my_sprintf
| main
| generic_start_main.isra.0
| __libc_start_main
| 0x0
|
--38.88%-- rand
|
|--94.01%-- rand
| do_my_sprintf
| main
| generic_start_main.isra.0
| __libc_start_main
| 0x0
|
--5.99%-- do_my_sprintf
main
generic_start_main.isra.0
__libc_start_main
0x0
Callgraph after the patch:
14.67% 2234 sprintft libc-2.18.so [.] __random
|
--- __random
|
|--95.93%-- rand
| do_my_sprintf
| main
| generic_start_main.isra.0
| __libc_start_main
| 0x0
|
--4.07%-- do_my_sprintf
main
generic_start_main.isra.0
__libc_start_main
0x0
TODO: For split-debug info objects like glibc, we can only determine
the call-frame-address only when both .eh_frame and .debug_info
sections are available. We should be able to determin the CFA
even without the .eh_frame section.
Fix suggested by Anton Blanchard.
Thanks to valuable input on DWARF debug information from Ulrich Weigand.
Reported-by: Maynard Johnson <maynard@us.ibm.com>
Tested-by: Maynard Johnson <maynard@us.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/20140625154903.GA29607@us.ibm.com
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
The function machine__get_kernel_start_addr() was taking the first symbol
of kallsyms as the start address. This is incorrect in certain cases
where the first symbol is something at 0, while the actual kernel
functions begin at a later point (e.g. 0x80200000).
This patch fixes machine__get_kernel_start_addr() to search for the
symbol "_text" or "_stext", which marks the beginning of kernel mapping.
This was already being done in machine__create_kernel_maps(). Thus, this
patch is just a refactor, to move that code into
machine__get_kernel_start_addr().
Signed-off-by: Simon Que <sque@chromium.org>
Link: http://lkml.kernel.org/r/1402943529-13244-1-git-send-email-sque@chromium.org
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Modules installed outside of the kernel's build system should go into
"%s/lib/modules/%s/extra", but at present, perf will only look at them
when they are in "%s/lib/modules/%s/kernel". Lets encourage good
citizenship by relaxing this requirement to "%s/lib/modules/%s". This
way open source modules that are out-of-tree have no incentive to start
populating a directory reserved for in-kernel modules and I can stop
hex-editing my system's perf binary when profiling OSS out-of-tree
modules.
Feedback from Namhyung Kim correctly revealed that the hex-edits that I
had been doing meant that perf was also traversing the build and source
symlinks in %s/lib/modules/%s. That is undesireable, so we explicitly
exclude them from traversal with a minor tweak to the traversal routine.
Signed-off-by: Richard Yao <ryao@gentoo.org>
Acked-by: Namhyung kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1398532675-13684-1-git-send-email-ryao@gentoo.org
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
When trying to map a bunch of instruction addresses to their respective
threads, I kept getting a lot of bogus entries [I forget the exact
reason as I patched my code months ago].
Looking through ip__resolve_ams, I noticed the check for
if (al.sym)
and realized, most times I have an al.map definition but sometimes an
al.sym is undefined. In the cases where al.sym is undefined, the loop
keeps going even though a valid al.map exists.
Modify this check to use the more reliable al.map. This fixed my bogus
entries.
Signed-off-by: Don Zickus <dzickus@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1393386227-149412-2-git-send-email-dzickus@redhat.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
