Groups > linux.kernel > #1273741 > unrolled thread

[RFC/PATCH] perf tools: Introduce perf_thread for backtrace

Started by	Namhyung Kim <namhyung@kernel.org>
First post	2015-11-20 07:10 +0100
Last post	2015-11-24 08:40 +0100
Articles	8 — 4 participants

Back to article view | Back to linux.kernel

  [RFC/PATCH] perf tools: Introduce perf_thread for backtrace Namhyung Kim <namhyung@kernel.org> - 2015-11-20 07:10 +0100
    RE: [RFC/PATCH] perf tools: Introduce perf_thread for backtrace 平松雅巳 / HIRAMATU，MASAMI   <masami.hiramatsu.pt@hitachi.com> - 2015-11-20 10:10 +0100
      Re: [RFC/PATCH] perf tools: Introduce perf_thread for backtrace Arnaldo Carvalho de Melo <acme@kernel.org> - 2015-11-20 13:20 +0100
        Re: [RFC/PATCH] perf tools: Introduce perf_thread for backtrace Namhyung Kim <namhyung@kernel.org> - 2015-11-24 08:30 +0100
    Re: [RFC/PATCH] perf tools: Introduce perf_thread for backtrace Jiri Olsa <jolsa@redhat.com> - 2015-11-20 10:40 +0100
      Re: [RFC/PATCH] perf tools: Introduce perf_thread for backtrace Namhyung Kim <namhyung@kernel.org> - 2015-11-24 08:40 +0100
    Re: [RFC/PATCH] perf tools: Introduce perf_thread for backtrace Arnaldo Carvalho de Melo <acme@kernel.org> - 2015-11-23 22:40 +0100
      Re: [RFC/PATCH] perf tools: Introduce perf_thread for backtrace Namhyung Kim <namhyung@kernel.org> - 2015-11-24 08:40 +0100

#1273741 — [RFC/PATCH] perf tools: Introduce perf_thread for backtrace

From	Namhyung Kim <namhyung@kernel.org>
Date	2015-11-20 07:10 +0100
Subject	[RFC/PATCH] perf tools: Introduce perf_thread for backtrace
Message-ID	<qwN3r-3oE-3@gated-at.bofh.it>

Backtrace is a crucial info for debugging.  And upcoming refcnt
tracking facility also wants to use it.

So instead of relying on glibc's backtrace_symbols[_fd] which misses
some (static) functions , use our own symbol searching mechanism.  To
do that, add perf_thread global variable to keep its maps and symbols.

The backtrace output from TUI is changed like below.  (I made a key
action to generate a segfault for testing):

Before:
  perf: Segmentation fault
  -------- backtrace --------
  perf[0x544a8b]
  /usr/lib/libc.so.6(+0x33680)[0x7fc46420b680]
  perf[0x54041b]
  perf(perf_evlist__tui_browse_hists+0x91)[0x5432e1]
  perf(cmd_report+0x1d20)[0x43cb10]
  perf[0x487073]
  perf(main+0x62f)[0x42cb1f]
  /usr/lib/libc.so.6(__libc_start_main+0xf0)[0x7fc4641f8610]
  perf(_start+0x29)[0x42cc39]
  [0x0]

After:
  perf: Segmentation fault
  -------- backtrace --------
  perf_evsel__hists_browse(+0x43b) in perf [0x54066b]
  perf_evlist__tui_browse_hists(+0x91) in perf [0x543531]
  cmd_report(+0x1d20) in perf [0x43cb50]
  run_builtin(+0x53) in perf [0x4870b3]
  main(+0x634) in perf [0x42cb54]
  __libc_start_main(+0xf0) in libc-2.22.so [0x7fea3577c610]
  _start(+0x29) in perf [0x42cc79]
  [0x0]

Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
---
 tools/perf/perf.c         |  7 +++++-
 tools/perf/ui/tui/setup.c | 21 ++++++++++++++--
 tools/perf/util/util.c    | 62 +++++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/util.h    |  5 ++++
 4 files changed, 92 insertions(+), 3 deletions(-)

diff --git a/tools/perf/perf.c b/tools/perf/perf.c
index 4bee53c3f796..f77eb440b05c 100644
--- a/tools/perf/perf.c
+++ b/tools/perf/perf.c
@@ -604,6 +604,8 @@ int main(int argc, const char **argv)
 	 */
 	pthread__block_sigwinch();
 
+	create_perf_thread();
+
 	while (1) {
 		static int done_help;
 		int was_alias = run_argv(&argc, &argv);
@@ -615,7 +617,7 @@ int main(int argc, const char **argv)
 			fprintf(stderr, "Expansion of alias '%s' failed; "
 				"'%s' is not a perf-command\n",
 				cmd, argv[0]);
-			goto out;
+			goto out_destroy;
 		}
 		if (!done_help) {
 			cmd = argv[0] = help_unknown_cmd(cmd);
@@ -626,6 +628,9 @@ int main(int argc, const char **argv)
 
 	fprintf(stderr, "Failed to run command '%s': %s\n",
 		cmd, strerror_r(errno, sbuf, sizeof(sbuf)));
+
+out_destroy:
+	destroy_perf_thread();
 out:
 	return 1;
 }
diff --git a/tools/perf/ui/tui/setup.c b/tools/perf/ui/tui/setup.c
index 7dfeba0a91f3..bc2da884e65a 100644
--- a/tools/perf/ui/tui/setup.c
+++ b/tools/perf/ui/tui/setup.c
@@ -13,6 +13,8 @@
 #include "../libslang.h"
 #include "../keysyms.h"
 #include "tui.h"
+#include "../../util/symbol.h"
+#include "../../util/thread.h"
 
 static volatile int ui__need_resize;
 
@@ -96,14 +98,29 @@ int ui__getch(int delay_secs)
 static void ui__signal_backtrace(int sig)
 {
 	void *stackdump[32];
-	size_t size;
+	size_t size, i;
 
 	ui__exit(false);
 	psignal(sig, "perf");
 
 	printf("-------- backtrace --------\n");
 	size = backtrace(stackdump, ARRAY_SIZE(stackdump));
-	backtrace_symbols_fd(stackdump, size, STDOUT_FILENO);
+	/* skip first two stack frame (for this function and signal stack) */
+	for (i = 2; i < size; i++) {
+		struct addr_location al = {
+			.sym = NULL,
+		};
+
+		thread__find_addr_location(perf_thread, PERF_RECORD_MISC_USER,
+					   MAP__FUNCTION, (long)stackdump[i], &al);
+
+		if (al.sym)
+			printf("%s(+0x%"PRIx64") in ", al.sym->name,
+			       map__map_ip(al.map, (u64)stackdump[i]) - al.sym->start);
+		if (al.map)
+			printf("%s ", al.map->dso->short_name);
+		printf("[0x%lx]\n", (unsigned long)stackdump[i]);
+	}
 
 	exit(0);
 }
diff --git a/tools/perf/util/util.c b/tools/perf/util/util.c
index 75759aebc7b8..f1a26ea14053 100644
--- a/tools/perf/util/util.c
+++ b/tools/perf/util/util.c
@@ -16,6 +16,9 @@
 #include <linux/kernel.h>
 #include <unistd.h>
 #include "callchain.h"
+#include "machine.h"
+#include "thread.h"
+#include "thread_map.h"
 
 struct callchain_param	callchain_param = {
 	.mode	= CHAIN_GRAPH_ABS,
@@ -696,3 +699,62 @@ fetch_kernel_version(unsigned int *puint, char *str,
 		*puint = (version << 16) + (patchlevel << 8) + sublevel;
 	return 0;
 }
+
+
+static int process_event(struct perf_tool *tool, union perf_event *event,
+			 struct perf_sample *sample, struct machine *machine)
+{
+	switch (event->header.type) {
+	case PERF_RECORD_COMM:
+		return tool->comm(tool, event, sample, machine);
+	case PERF_RECORD_MMAP:
+		return tool->mmap(tool, event, sample, machine);
+	case PERF_RECORD_MMAP2:
+		return tool->mmap2(tool, event, sample, machine);
+	default:
+		break;
+	}
+	return 0;
+}
+
+struct thread *perf_thread;
+
+void create_perf_thread(void)
+{
+	struct perf_tool tool = {
+		.comm	= perf_event__process_comm,
+		.mmap	= perf_event__process_mmap,
+		.mmap2	= perf_event__process_mmap2,
+	};
+	struct thread_map *tm;
+	struct machine *machine;
+	int pid = getpid();
+
+	machine = machine__new_host();
+	if (machine == NULL)
+		return;
+
+	tm = thread_map__new_dummy();
+	if (tm == NULL) {
+		machine__delete(machine);
+		return;
+	}
+
+	thread_map__set_pid(tm, 0, pid);
+
+	perf_event__synthesize_thread_map(&tool, tm, process_event, machine,
+					  false, 500);
+
+	perf_thread = machine__find_thread(machine, pid, pid);
+	BUG_ON(perf_thread == NULL);
+
+	thread_map__put(tm);
+}
+
+void destroy_perf_thread(void)
+{
+	struct machine *machine = perf_thread->mg->machine;
+
+	machine__delete_threads(machine);
+	machine__delete(machine);
+}
diff --git a/tools/perf/util/util.h b/tools/perf/util/util.h
index dcc659017976..630e145049aa 100644
--- a/tools/perf/util/util.h
+++ b/tools/perf/util/util.h
@@ -358,4 +358,9 @@ int fetch_kernel_version(unsigned int *puint,
 #define KVER_FMT	"%d.%d.%d"
 #define KVER_PARAM(x)	KVER_VERSION(x), KVER_PATCHLEVEL(x), KVER_SUBLEVEL(x)
 
+extern struct thread *perf_thread;
+
+void create_perf_thread(void);
+void destroy_perf_thread(void);
+
 #endif /* GIT_COMPAT_UTIL_H */
-- 
2.6.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [next] | [standalone]

#1273881

From	平松雅巳 / HIRAMATU，MASAMI <masami.hiramatsu.pt@hitachi.com>
Date	2015-11-20 10:10 +0100
Message-ID	<qwPRF-5cO-27@gated-at.bofh.it>
In reply to	#1273741

PkZyb206IE5hbWh5dW5nIEtpbSBbbWFpbHRvOm5hbWh5dW5nQGtlcm5lbC5vcmddDQo+DQo+QmFj
a3RyYWNlIGlzIGEgY3J1Y2lhbCBpbmZvIGZvciBkZWJ1Z2dpbmcuICBBbmQgdXBjb21pbmcgcmVm
Y250DQo+dHJhY2tpbmcgZmFjaWxpdHkgYWxzbyB3YW50cyB0byB1c2UgaXQuDQo+DQo+U28gaW5z
dGVhZCBvZiByZWx5aW5nIG9uIGdsaWJjJ3MgYmFja3RyYWNlX3N5bWJvbHNbX2ZkXSB3aGljaCBt
aXNzZXMNCj5zb21lIChzdGF0aWMpIGZ1bmN0aW9ucyAsIHVzZSBvdXIgb3duIHN5bWJvbCBzZWFy
Y2hpbmcgbWVjaGFuaXNtLiAgVG8NCj5kbyB0aGF0LCBhZGQgcGVyZl90aHJlYWQgZ2xvYmFsIHZh
cmlhYmxlIHRvIGtlZXAgaXRzIG1hcHMgYW5kIHN5bWJvbHMuDQoNCkhtbSwgSSBkb3VidCB0aGF0
IHRoaXMgY2FuIHdvcmsgZm9yIGRlYnVnZ2luZyBzaXR1YXRpb24sIGJlY2F1c2UNCnNvbWV0aW1l
cyBiYWNrdHJhY2UgZmFjaWxpdGllcyBoYXMgdG8gZGVidWcgaXRzZWxmIGJ5IGl0c2VsZi4NCg0K
Rm9yIHRoZSBzb21lIChzdGF0aWMpIGZ1bmN0aW9ucywgSSdkIHJhdGhlciBsaWtlIHRvIHVzZSBn
bGliYydzIA0KYmFja3RyYWNlX3N5bWJvbHMgYW5kIGFkZHIybGluZSBvciBldmVuIHdpdGggcmF3
IGFkZHJlc3MgZm9yDQpyZWxpYWJpbGl0eS4uLg0KDQpUaGFuayB5b3UsDQoNCj4NCj5UaGUgYmFj
a3RyYWNlIG91dHB1dCBmcm9tIFRVSSBpcyBjaGFuZ2VkIGxpa2UgYmVsb3cuICAoSSBtYWRlIGEg
a2V5DQo+YWN0aW9uIHRvIGdlbmVyYXRlIGEgc2VnZmF1bHQgZm9yIHRlc3RpbmcpOg0KPg0KPkJl
Zm9yZToNCj4gIHBlcmY6IFNlZ21lbnRhdGlvbiBmYXVsdA0KPiAgLS0tLS0tLS0gYmFja3RyYWNl
IC0tLS0tLS0tDQo+ICBwZXJmWzB4NTQ0YThiXQ0KPiAgL3Vzci9saWIvbGliYy5zby42KCsweDMz
NjgwKVsweDdmYzQ2NDIwYjY4MF0NCj4gIHBlcmZbMHg1NDA0MWJdDQo+ICBwZXJmKHBlcmZfZXZs
aXN0X190dWlfYnJvd3NlX2hpc3RzKzB4OTEpWzB4NTQzMmUxXQ0KPiAgcGVyZihjbWRfcmVwb3J0
KzB4MWQyMClbMHg0M2NiMTBdDQo+ICBwZXJmWzB4NDg3MDczXQ0KPiAgcGVyZihtYWluKzB4NjJm
KVsweDQyY2IxZl0NCj4gIC91c3IvbGliL2xpYmMuc28uNihfX2xpYmNfc3RhcnRfbWFpbisweGYw
KVsweDdmYzQ2NDFmODYxMF0NCj4gIHBlcmYoX3N0YXJ0KzB4MjkpWzB4NDJjYzM5XQ0KPiAgWzB4
MF0NCj4NCj5BZnRlcjoNCj4gIHBlcmY6IFNlZ21lbnRhdGlvbiBmYXVsdA0KPiAgLS0tLS0tLS0g
YmFja3RyYWNlIC0tLS0tLS0tDQo+ICBwZXJmX2V2c2VsX19oaXN0c19icm93c2UoKzB4NDNiKSBp
biBwZXJmIFsweDU0MDY2Yl0NCj4gIHBlcmZfZXZsaXN0X190dWlfYnJvd3NlX2hpc3RzKCsweDkx
KSBpbiBwZXJmIFsweDU0MzUzMV0NCj4gIGNtZF9yZXBvcnQoKzB4MWQyMCkgaW4gcGVyZiBbMHg0
M2NiNTBdDQo+ICBydW5fYnVpbHRpbigrMHg1MykgaW4gcGVyZiBbMHg0ODcwYjNdDQo+ICBtYWlu
KCsweDYzNCkgaW4gcGVyZiBbMHg0MmNiNTRdDQo+ICBfX2xpYmNfc3RhcnRfbWFpbigrMHhmMCkg
aW4gbGliYy0yLjIyLnNvIFsweDdmZWEzNTc3YzYxMF0NCj4gIF9zdGFydCgrMHgyOSkgaW4gcGVy
ZiBbMHg0MmNjNzldDQo+ICBbMHgwXQ0KPg0KPkNjOiBGcmVkZXJpYyBXZWlzYmVja2VyIDxmd2Vp
c2JlY0BnbWFpbC5jb20+DQo+Q2M6IE1hc2FtaSBIaXJhbWF0c3UgPG1hc2FtaS5oaXJhbWF0c3Uu
cHRAaGl0YWNoaS5jb20+DQo+U2lnbmVkLW9mZi1ieTogTmFtaHl1bmcgS2ltIDxuYW1oeXVuZ0Br
ZXJuZWwub3JnPg0KPi0tLQ0KPiB0b29scy9wZXJmL3BlcmYuYyAgICAgICAgIHwgIDcgKysrKyst
DQo+IHRvb2xzL3BlcmYvdWkvdHVpL3NldHVwLmMgfCAyMSArKysrKysrKysrKysrKy0tDQo+IHRv
b2xzL3BlcmYvdXRpbC91dGlsLmMgICAgfCA2MiArKysrKysrKysrKysrKysrKysrKysrKysrKysr
KysrKysrKysrKysrKysrKysrKw0KPiB0b29scy9wZXJmL3V0aWwvdXRpbC5oICAgIHwgIDUgKysr
Kw0KPiA0IGZpbGVzIGNoYW5nZWQsIDkyIGluc2VydGlvbnMoKyksIDMgZGVsZXRpb25zKC0pDQo+
DQo+ZGlmZiAtLWdpdCBhL3Rvb2xzL3BlcmYvcGVyZi5jIGIvdG9vbHMvcGVyZi9wZXJmLmMNCj5p
bmRleCA0YmVlNTNjM2Y3OTYuLmY3N2ViNDQwYjA1YyAxMDA2NDQNCj4tLS0gYS90b29scy9wZXJm
L3BlcmYuYw0KPisrKyBiL3Rvb2xzL3BlcmYvcGVyZi5jDQo+QEAgLTYwNCw2ICs2MDQsOCBAQCBp
bnQgbWFpbihpbnQgYXJnYywgY29uc3QgY2hhciAqKmFyZ3YpDQo+IAkgKi8NCj4gCXB0aHJlYWRf
X2Jsb2NrX3NpZ3dpbmNoKCk7DQo+DQo+KwljcmVhdGVfcGVyZl90aHJlYWQoKTsNCj4rDQo+IAl3
aGlsZSAoMSkgew0KPiAJCXN0YXRpYyBpbnQgZG9uZV9oZWxwOw0KPiAJCWludCB3YXNfYWxpYXMg
PSBydW5fYXJndigmYXJnYywgJmFyZ3YpOw0KPkBAIC02MTUsNyArNjE3LDcgQEAgaW50IG1haW4o
aW50IGFyZ2MsIGNvbnN0IGNoYXIgKiphcmd2KQ0KPiAJCQlmcHJpbnRmKHN0ZGVyciwgIkV4cGFu
c2lvbiBvZiBhbGlhcyAnJXMnIGZhaWxlZDsgIg0KPiAJCQkJIiclcycgaXMgbm90IGEgcGVyZi1j
b21tYW5kXG4iLA0KPiAJCQkJY21kLCBhcmd2WzBdKTsNCj4tCQkJZ290byBvdXQ7DQo+KwkJCWdv
dG8gb3V0X2Rlc3Ryb3k7DQo+IAkJfQ0KPiAJCWlmICghZG9uZV9oZWxwKSB7DQo+IAkJCWNtZCA9
IGFyZ3ZbMF0gPSBoZWxwX3Vua25vd25fY21kKGNtZCk7DQo+QEAgLTYyNiw2ICs2MjgsOSBAQCBp
bnQgbWFpbihpbnQgYXJnYywgY29uc3QgY2hhciAqKmFyZ3YpDQo+DQo+IAlmcHJpbnRmKHN0ZGVy
ciwgIkZhaWxlZCB0byBydW4gY29tbWFuZCAnJXMnOiAlc1xuIiwNCj4gCQljbWQsIHN0cmVycm9y
X3IoZXJybm8sIHNidWYsIHNpemVvZihzYnVmKSkpOw0KPisNCj4rb3V0X2Rlc3Ryb3k6DQo+Kwlk
ZXN0cm95X3BlcmZfdGhyZWFkKCk7DQo+IG91dDoNCj4gCXJldHVybiAxOw0KPiB9DQo+ZGlmZiAt
LWdpdCBhL3Rvb2xzL3BlcmYvdWkvdHVpL3NldHVwLmMgYi90b29scy9wZXJmL3VpL3R1aS9zZXR1
cC5jDQo+aW5kZXggN2RmZWJhMGE5MWYzLi5iYzJkYTg4NGU2NWEgMTAwNjQ0DQo+LS0tIGEvdG9v
bHMvcGVyZi91aS90dWkvc2V0dXAuYw0KPisrKyBiL3Rvb2xzL3BlcmYvdWkvdHVpL3NldHVwLmMN
Cj5AQCAtMTMsNiArMTMsOCBAQA0KPiAjaW5jbHVkZSAiLi4vbGlic2xhbmcuaCINCj4gI2luY2x1
ZGUgIi4uL2tleXN5bXMuaCINCj4gI2luY2x1ZGUgInR1aS5oIg0KPisjaW5jbHVkZSAiLi4vLi4v
dXRpbC9zeW1ib2wuaCINCj4rI2luY2x1ZGUgIi4uLy4uL3V0aWwvdGhyZWFkLmgiDQo+DQo+IHN0
YXRpYyB2b2xhdGlsZSBpbnQgdWlfX25lZWRfcmVzaXplOw0KPg0KPkBAIC05NiwxNCArOTgsMjkg
QEAgaW50IHVpX19nZXRjaChpbnQgZGVsYXlfc2VjcykNCj4gc3RhdGljIHZvaWQgdWlfX3NpZ25h
bF9iYWNrdHJhY2UoaW50IHNpZykNCj4gew0KPiAJdm9pZCAqc3RhY2tkdW1wWzMyXTsNCj4tCXNp
emVfdCBzaXplOw0KPisJc2l6ZV90IHNpemUsIGk7DQo+DQo+IAl1aV9fZXhpdChmYWxzZSk7DQo+
IAlwc2lnbmFsKHNpZywgInBlcmYiKTsNCj4NCj4gCXByaW50ZigiLS0tLS0tLS0gYmFja3RyYWNl
IC0tLS0tLS0tXG4iKTsNCj4gCXNpemUgPSBiYWNrdHJhY2Uoc3RhY2tkdW1wLCBBUlJBWV9TSVpF
KHN0YWNrZHVtcCkpOw0KPi0JYmFja3RyYWNlX3N5bWJvbHNfZmQoc3RhY2tkdW1wLCBzaXplLCBT
VERPVVRfRklMRU5PKTsNCj4rCS8qIHNraXAgZmlyc3QgdHdvIHN0YWNrIGZyYW1lIChmb3IgdGhp
cyBmdW5jdGlvbiBhbmQgc2lnbmFsIHN0YWNrKSAqLw0KPisJZm9yIChpID0gMjsgaSA8IHNpemU7
IGkrKykgew0KPisJCXN0cnVjdCBhZGRyX2xvY2F0aW9uIGFsID0gew0KPisJCQkuc3ltID0gTlVM
TCwNCj4rCQl9Ow0KPisNCj4rCQl0aHJlYWRfX2ZpbmRfYWRkcl9sb2NhdGlvbihwZXJmX3RocmVh
ZCwgUEVSRl9SRUNPUkRfTUlTQ19VU0VSLA0KPisJCQkJCSAgIE1BUF9fRlVOQ1RJT04sIChsb25n
KXN0YWNrZHVtcFtpXSwgJmFsKTsNCj4rDQo+KwkJaWYgKGFsLnN5bSkNCj4rCQkJcHJpbnRmKCIl
cygrMHglIlBSSXg2NCIpIGluICIsIGFsLnN5bS0+bmFtZSwNCj4rCQkJICAgICAgIG1hcF9fbWFw
X2lwKGFsLm1hcCwgKHU2NClzdGFja2R1bXBbaV0pIC0gYWwuc3ltLT5zdGFydCk7DQo+KwkJaWYg
KGFsLm1hcCkNCj4rCQkJcHJpbnRmKCIlcyAiLCBhbC5tYXAtPmRzby0+c2hvcnRfbmFtZSk7DQo+
KwkJcHJpbnRmKCJbMHglbHhdXG4iLCAodW5zaWduZWQgbG9uZylzdGFja2R1bXBbaV0pOw0KPisJ
fQ0KPg0KPiAJZXhpdCgwKTsNCj4gfQ0KPmRpZmYgLS1naXQgYS90b29scy9wZXJmL3V0aWwvdXRp
bC5jIGIvdG9vbHMvcGVyZi91dGlsL3V0aWwuYw0KPmluZGV4IDc1NzU5YWViYzdiOC4uZjFhMjZl
YTE0MDUzIDEwMDY0NA0KPi0tLSBhL3Rvb2xzL3BlcmYvdXRpbC91dGlsLmMNCj4rKysgYi90b29s
cy9wZXJmL3V0aWwvdXRpbC5jDQo+QEAgLTE2LDYgKzE2LDkgQEANCj4gI2luY2x1ZGUgPGxpbnV4
L2tlcm5lbC5oPg0KPiAjaW5jbHVkZSA8dW5pc3RkLmg+DQo+ICNpbmNsdWRlICJjYWxsY2hhaW4u
aCINCj4rI2luY2x1ZGUgIm1hY2hpbmUuaCINCj4rI2luY2x1ZGUgInRocmVhZC5oIg0KPisjaW5j
bHVkZSAidGhyZWFkX21hcC5oIg0KPg0KPiBzdHJ1Y3QgY2FsbGNoYWluX3BhcmFtCWNhbGxjaGFp
bl9wYXJhbSA9IHsNCj4gCS5tb2RlCT0gQ0hBSU5fR1JBUEhfQUJTLA0KPkBAIC02OTYsMyArNjk5
LDYyIEBAIGZldGNoX2tlcm5lbF92ZXJzaW9uKHVuc2lnbmVkIGludCAqcHVpbnQsIGNoYXIgKnN0
ciwNCj4gCQkqcHVpbnQgPSAodmVyc2lvbiA8PCAxNikgKyAocGF0Y2hsZXZlbCA8PCA4KSArIHN1
YmxldmVsOw0KPiAJcmV0dXJuIDA7DQo+IH0NCj4rDQo+Kw0KPitzdGF0aWMgaW50IHByb2Nlc3Nf
ZXZlbnQoc3RydWN0IHBlcmZfdG9vbCAqdG9vbCwgdW5pb24gcGVyZl9ldmVudCAqZXZlbnQsDQo+
KwkJCSBzdHJ1Y3QgcGVyZl9zYW1wbGUgKnNhbXBsZSwgc3RydWN0IG1hY2hpbmUgKm1hY2hpbmUp
DQo+K3sNCj4rCXN3aXRjaCAoZXZlbnQtPmhlYWRlci50eXBlKSB7DQo+KwljYXNlIFBFUkZfUkVD
T1JEX0NPTU06DQo+KwkJcmV0dXJuIHRvb2wtPmNvbW0odG9vbCwgZXZlbnQsIHNhbXBsZSwgbWFj
aGluZSk7DQo+KwljYXNlIFBFUkZfUkVDT1JEX01NQVA6DQo+KwkJcmV0dXJuIHRvb2wtPm1tYXAo
dG9vbCwgZXZlbnQsIHNhbXBsZSwgbWFjaGluZSk7DQo+KwljYXNlIFBFUkZfUkVDT1JEX01NQVAy
Og0KPisJCXJldHVybiB0b29sLT5tbWFwMih0b29sLCBldmVudCwgc2FtcGxlLCBtYWNoaW5lKTsN
Cj4rCWRlZmF1bHQ6DQo+KwkJYnJlYWs7DQo+Kwl9DQo+KwlyZXR1cm4gMDsNCj4rfQ0KPisNCj4r
c3RydWN0IHRocmVhZCAqcGVyZl90aHJlYWQ7DQo+Kw0KPit2b2lkIGNyZWF0ZV9wZXJmX3RocmVh
ZCh2b2lkKQ0KPit7DQo+KwlzdHJ1Y3QgcGVyZl90b29sIHRvb2wgPSB7DQo+KwkJLmNvbW0JPSBw
ZXJmX2V2ZW50X19wcm9jZXNzX2NvbW0sDQo+KwkJLm1tYXAJPSBwZXJmX2V2ZW50X19wcm9jZXNz
X21tYXAsDQo+KwkJLm1tYXAyCT0gcGVyZl9ldmVudF9fcHJvY2Vzc19tbWFwMiwNCj4rCX07DQo+
KwlzdHJ1Y3QgdGhyZWFkX21hcCAqdG07DQo+KwlzdHJ1Y3QgbWFjaGluZSAqbWFjaGluZTsNCj4r
CWludCBwaWQgPSBnZXRwaWQoKTsNCj4rDQo+KwltYWNoaW5lID0gbWFjaGluZV9fbmV3X2hvc3Qo
KTsNCj4rCWlmIChtYWNoaW5lID09IE5VTEwpDQo+KwkJcmV0dXJuOw0KPisNCj4rCXRtID0gdGhy
ZWFkX21hcF9fbmV3X2R1bW15KCk7DQo+KwlpZiAodG0gPT0gTlVMTCkgew0KPisJCW1hY2hpbmVf
X2RlbGV0ZShtYWNoaW5lKTsNCj4rCQlyZXR1cm47DQo+Kwl9DQo+Kw0KPisJdGhyZWFkX21hcF9f
c2V0X3BpZCh0bSwgMCwgcGlkKTsNCj4rDQo+KwlwZXJmX2V2ZW50X19zeW50aGVzaXplX3RocmVh
ZF9tYXAoJnRvb2wsIHRtLCBwcm9jZXNzX2V2ZW50LCBtYWNoaW5lLA0KPisJCQkJCSAgZmFsc2Us
IDUwMCk7DQo+Kw0KPisJcGVyZl90aHJlYWQgPSBtYWNoaW5lX19maW5kX3RocmVhZChtYWNoaW5l
LCBwaWQsIHBpZCk7DQo+KwlCVUdfT04ocGVyZl90aHJlYWQgPT0gTlVMTCk7DQo+Kw0KPisJdGhy
ZWFkX21hcF9fcHV0KHRtKTsNCj4rfQ0KPisNCj4rdm9pZCBkZXN0cm95X3BlcmZfdGhyZWFkKHZv
aWQpDQo+K3sNCj4rCXN0cnVjdCBtYWNoaW5lICptYWNoaW5lID0gcGVyZl90aHJlYWQtPm1nLT5t
YWNoaW5lOw0KPisNCj4rCW1hY2hpbmVfX2RlbGV0ZV90aHJlYWRzKG1hY2hpbmUpOw0KPisJbWFj
aGluZV9fZGVsZXRlKG1hY2hpbmUpOw0KPit9DQo+ZGlmZiAtLWdpdCBhL3Rvb2xzL3BlcmYvdXRp
bC91dGlsLmggYi90b29scy9wZXJmL3V0aWwvdXRpbC5oDQo+aW5kZXggZGNjNjU5MDE3OTc2Li42
MzBlMTQ1MDQ5YWEgMTAwNjQ0DQo+LS0tIGEvdG9vbHMvcGVyZi91dGlsL3V0aWwuaA0KPisrKyBi
L3Rvb2xzL3BlcmYvdXRpbC91dGlsLmgNCj5AQCAtMzU4LDQgKzM1OCw5IEBAIGludCBmZXRjaF9r
ZXJuZWxfdmVyc2lvbih1bnNpZ25lZCBpbnQgKnB1aW50LA0KPiAjZGVmaW5lIEtWRVJfRk1UCSIl
ZC4lZC4lZCINCj4gI2RlZmluZSBLVkVSX1BBUkFNKHgpCUtWRVJfVkVSU0lPTih4KSwgS1ZFUl9Q
QVRDSExFVkVMKHgpLCBLVkVSX1NVQkxFVkVMKHgpDQo+DQo+K2V4dGVybiBzdHJ1Y3QgdGhyZWFk
ICpwZXJmX3RocmVhZDsNCj4rDQo+K3ZvaWQgY3JlYXRlX3BlcmZfdGhyZWFkKHZvaWQpOw0KPit2
b2lkIGRlc3Ryb3lfcGVyZl90aHJlYWQodm9pZCk7DQo+Kw0KPiAjZW5kaWYgLyogR0lUX0NPTVBB
VF9VVElMX0ggKi8NCj4tLQ0KPjIuNi4yDQoNCg==
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1274021

From	Arnaldo Carvalho de Melo <acme@kernel.org>
Date	2015-11-20 13:20 +0100
Message-ID	<qwSPw-76t-27@gated-at.bofh.it>
In reply to	#1273881

Em Fri, Nov 20, 2015 at 09:05:23AM +0000, 平松雅巳 / HIRAMATU，MASAMI escreveu:
> >From: Namhyung Kim [mailto:namhyung@kernel.org]
> >
> >Backtrace is a crucial info for debugging.  And upcoming refcnt
> >tracking facility also wants to use it.
> >
> >So instead of relying on glibc's backtrace_symbols[_fd] which misses
> >some (static) functions , use our own symbol searching mechanism.  To
> >do that, add perf_thread global variable to keep its maps and symbols.
> 
> Hmm, I doubt that this can work for debugging situation, because
> sometimes backtrace facilities has to debug itself by itself.

That is a valid point, possibly we can have both and when we think that
the code we rely on for resolving symbols has issues, activate the
other, more expensive, binutils/elfutils spawned command line utilities
to do compare the results?
 
> For the some (static) functions, I'd rather like to use glibc's 
> backtrace_symbols and addr2line or even with raw address for
> reliability...

 
> Thank you,
> 
> >
> >The backtrace output from TUI is changed like below.  (I made a key
> >action to generate a segfault for testing):
> >
> >Before:
> >  perf: Segmentation fault
> >  -------- backtrace --------
> >  perf[0x544a8b]
> >  /usr/lib/libc.so.6(+0x33680)[0x7fc46420b680]
> >  perf[0x54041b]
> >  perf(perf_evlist__tui_browse_hists+0x91)[0x5432e1]
> >  perf(cmd_report+0x1d20)[0x43cb10]
> >  perf[0x487073]
> >  perf(main+0x62f)[0x42cb1f]
> >  /usr/lib/libc.so.6(__libc_start_main+0xf0)[0x7fc4641f8610]
> >  perf(_start+0x29)[0x42cc39]
> >  [0x0]
> >
> >After:
> >  perf: Segmentation fault
> >  -------- backtrace --------
> >  perf_evsel__hists_browse(+0x43b) in perf [0x54066b]
> >  perf_evlist__tui_browse_hists(+0x91) in perf [0x543531]
> >  cmd_report(+0x1d20) in perf [0x43cb50]
> >  run_builtin(+0x53) in perf [0x4870b3]
> >  main(+0x634) in perf [0x42cb54]
> >  __libc_start_main(+0xf0) in libc-2.22.so [0x7fea3577c610]
> >  _start(+0x29) in perf [0x42cc79]
> >  [0x0]
> >
> >Cc: Frederic Weisbecker <fweisbec@gmail.com>
> >Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
> >Signed-off-by: Namhyung Kim <namhyung@kernel.org>
> >---
> > tools/perf/perf.c         |  7 +++++-
> > tools/perf/ui/tui/setup.c | 21 ++++++++++++++--
> > tools/perf/util/util.c    | 62 +++++++++++++++++++++++++++++++++++++++++++++++
> > tools/perf/util/util.h    |  5 ++++
> > 4 files changed, 92 insertions(+), 3 deletions(-)
> >
> >diff --git a/tools/perf/perf.c b/tools/perf/perf.c
> >index 4bee53c3f796..f77eb440b05c 100644
> >--- a/tools/perf/perf.c
> >+++ b/tools/perf/perf.c
> >@@ -604,6 +604,8 @@ int main(int argc, const char **argv)
> > 	 */
> > 	pthread__block_sigwinch();
> >
> >+	create_perf_thread();
> >+
> > 	while (1) {
> > 		static int done_help;
> > 		int was_alias = run_argv(&argc, &argv);
> >@@ -615,7 +617,7 @@ int main(int argc, const char **argv)
> > 			fprintf(stderr, "Expansion of alias '%s' failed; "
> > 				"'%s' is not a perf-command\n",
> > 				cmd, argv[0]);
> >-			goto out;
> >+			goto out_destroy;
> > 		}
> > 		if (!done_help) {
> > 			cmd = argv[0] = help_unknown_cmd(cmd);
> >@@ -626,6 +628,9 @@ int main(int argc, const char **argv)
> >
> > 	fprintf(stderr, "Failed to run command '%s': %s\n",
> > 		cmd, strerror_r(errno, sbuf, sizeof(sbuf)));
> >+
> >+out_destroy:
> >+	destroy_perf_thread();
> > out:
> > 	return 1;
> > }
> >diff --git a/tools/perf/ui/tui/setup.c b/tools/perf/ui/tui/setup.c
> >index 7dfeba0a91f3..bc2da884e65a 100644
> >--- a/tools/perf/ui/tui/setup.c
> >+++ b/tools/perf/ui/tui/setup.c
> >@@ -13,6 +13,8 @@
> > #include "../libslang.h"
> > #include "../keysyms.h"
> > #include "tui.h"
> >+#include "../../util/symbol.h"
> >+#include "../../util/thread.h"
> >
> > static volatile int ui__need_resize;
> >
> >@@ -96,14 +98,29 @@ int ui__getch(int delay_secs)
> > static void ui__signal_backtrace(int sig)
> > {
> > 	void *stackdump[32];
> >-	size_t size;
> >+	size_t size, i;
> >
> > 	ui__exit(false);
> > 	psignal(sig, "perf");
> >
> > 	printf("-------- backtrace --------\n");
> > 	size = backtrace(stackdump, ARRAY_SIZE(stackdump));
> >-	backtrace_symbols_fd(stackdump, size, STDOUT_FILENO);
> >+	/* skip first two stack frame (for this function and signal stack) */
> >+	for (i = 2; i < size; i++) {
> >+		struct addr_location al = {
> >+			.sym = NULL,
> >+		};
> >+
> >+		thread__find_addr_location(perf_thread, PERF_RECORD_MISC_USER,
> >+					   MAP__FUNCTION, (long)stackdump[i], &al);
> >+
> >+		if (al.sym)
> >+			printf("%s(+0x%"PRIx64") in ", al.sym->name,
> >+			       map__map_ip(al.map, (u64)stackdump[i]) - al.sym->start);
> >+		if (al.map)
> >+			printf("%s ", al.map->dso->short_name);
> >+		printf("[0x%lx]\n", (unsigned long)stackdump[i]);
> >+	}
> >
> > 	exit(0);
> > }
> >diff --git a/tools/perf/util/util.c b/tools/perf/util/util.c
> >index 75759aebc7b8..f1a26ea14053 100644
> >--- a/tools/perf/util/util.c
> >+++ b/tools/perf/util/util.c
> >@@ -16,6 +16,9 @@
> > #include <linux/kernel.h>
> > #include <unistd.h>
> > #include "callchain.h"
> >+#include "machine.h"
> >+#include "thread.h"
> >+#include "thread_map.h"
> >
> > struct callchain_param	callchain_param = {
> > 	.mode	= CHAIN_GRAPH_ABS,
> >@@ -696,3 +699,62 @@ fetch_kernel_version(unsigned int *puint, char *str,
> > 		*puint = (version << 16) + (patchlevel << 8) + sublevel;
> > 	return 0;
> > }
> >+
> >+
> >+static int process_event(struct perf_tool *tool, union perf_event *event,
> >+			 struct perf_sample *sample, struct machine *machine)
> >+{
> >+	switch (event->header.type) {
> >+	case PERF_RECORD_COMM:
> >+		return tool->comm(tool, event, sample, machine);
> >+	case PERF_RECORD_MMAP:
> >+		return tool->mmap(tool, event, sample, machine);
> >+	case PERF_RECORD_MMAP2:
> >+		return tool->mmap2(tool, event, sample, machine);
> >+	default:
> >+		break;
> >+	}
> >+	return 0;
> >+}
> >+
> >+struct thread *perf_thread;
> >+
> >+void create_perf_thread(void)
> >+{
> >+	struct perf_tool tool = {
> >+		.comm	= perf_event__process_comm,
> >+		.mmap	= perf_event__process_mmap,
> >+		.mmap2	= perf_event__process_mmap2,
> >+	};
> >+	struct thread_map *tm;
> >+	struct machine *machine;
> >+	int pid = getpid();
> >+
> >+	machine = machine__new_host();
> >+	if (machine == NULL)
> >+		return;
> >+
> >+	tm = thread_map__new_dummy();
> >+	if (tm == NULL) {
> >+		machine__delete(machine);
> >+		return;
> >+	}
> >+
> >+	thread_map__set_pid(tm, 0, pid);
> >+
> >+	perf_event__synthesize_thread_map(&tool, tm, process_event, machine,
> >+					  false, 500);
> >+
> >+	perf_thread = machine__find_thread(machine, pid, pid);
> >+	BUG_ON(perf_thread == NULL);
> >+
> >+	thread_map__put(tm);
> >+}
> >+
> >+void destroy_perf_thread(void)
> >+{
> >+	struct machine *machine = perf_thread->mg->machine;
> >+
> >+	machine__delete_threads(machine);
> >+	machine__delete(machine);
> >+}
> >diff --git a/tools/perf/util/util.h b/tools/perf/util/util.h
> >index dcc659017976..630e145049aa 100644
> >--- a/tools/perf/util/util.h
> >+++ b/tools/perf/util/util.h
> >@@ -358,4 +358,9 @@ int fetch_kernel_version(unsigned int *puint,
> > #define KVER_FMT	"%d.%d.%d"
> > #define KVER_PARAM(x)	KVER_VERSION(x), KVER_PATCHLEVEL(x), KVER_SUBLEVEL(x)
> >
> >+extern struct thread *perf_thread;
> >+
> >+void create_perf_thread(void);
> >+void destroy_perf_thread(void);
> >+
> > #endif /* GIT_COMPAT_UTIL_H */
> >--
> >2.6.2
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1276103

From	Namhyung Kim <namhyung@kernel.org>
Date	2015-11-24 08:30 +0100
Message-ID	<qygd4-51K-3@gated-at.bofh.it>
In reply to	#1274021

Hi Arnaldo and Masami,

On Fri, Nov 20, 2015 at 09:10:44AM -0300, Arnaldo Carvalho de Melo wrote:
> Em Fri, Nov 20, 2015 at 09:05:23AM +0000, 平松雅巳 / HIRAMATU，MASAMI escreveu:
> > >From: Namhyung Kim [mailto:namhyung@kernel.org]
> > >
> > >Backtrace is a crucial info for debugging.  And upcoming refcnt
> > >tracking facility also wants to use it.
> > >
> > >So instead of relying on glibc's backtrace_symbols[_fd] which misses
> > >some (static) functions , use our own symbol searching mechanism.  To
> > >do that, add perf_thread global variable to keep its maps and symbols.
> > 
> > Hmm, I doubt that this can work for debugging situation, because
> > sometimes backtrace facilities has to debug itself by itself.
> 
> That is a valid point, possibly we can have both and when we think that
> the code we rely on for resolving symbols has issues, activate the
> other, more expensive, binutils/elfutils spawned command line utilities
> to do compare the results?

Yeah, that's a possible solution.  We can start by using our own, and
if there's a certain amount of failure in symbol resolving, then
fallback to glibc's backtrace_symbols + addr2line.

>  
> > For the some (static) functions, I'd rather like to use glibc's 
> > backtrace_symbols and addr2line or even with raw address for
> > reliability...

I also printed the raw addresses in case of doubts, so you could
verify its correctness. :)  And IMHO, if something is severely broken,
we might not rely on glibc too.

Having said that, I agree with your concern and it needs the fallback
method for possible malfunction.  But I guess it'd work quite well for
most cases so it's worth trying to convert using it.  I'll work on the
fallback method then..

Thanks,
Namhyung
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1273892

From	Jiri Olsa <jolsa@redhat.com>
Date	2015-11-20 10:40 +0100
Message-ID	<qwQkF-5nF-1@gated-at.bofh.it>
In reply to	#1273741

On Fri, Nov 20, 2015 at 03:03:03PM +0900, Namhyung Kim wrote:
> Backtrace is a crucial info for debugging.  And upcoming refcnt
> tracking facility also wants to use it.
> 
> So instead of relying on glibc's backtrace_symbols[_fd] which misses
> some (static) functions , use our own symbol searching mechanism.  To
> do that, add perf_thread global variable to keep its maps and symbols.
> 
> The backtrace output from TUI is changed like below.  (I made a key
> action to generate a segfault for testing):
> 
> Before:
>   perf: Segmentation fault
>   -------- backtrace --------
>   perf[0x544a8b]
>   /usr/lib/libc.so.6(+0x33680)[0x7fc46420b680]
>   perf[0x54041b]
>   perf(perf_evlist__tui_browse_hists+0x91)[0x5432e1]
>   perf(cmd_report+0x1d20)[0x43cb10]
>   perf[0x487073]
>   perf(main+0x62f)[0x42cb1f]
>   /usr/lib/libc.so.6(__libc_start_main+0xf0)[0x7fc4641f8610]
>   perf(_start+0x29)[0x42cc39]
>   [0x0]
> 
> After:
>   perf: Segmentation fault
>   -------- backtrace --------
>   perf_evsel__hists_browse(+0x43b) in perf [0x54066b]
>   perf_evlist__tui_browse_hists(+0x91) in perf [0x543531]
>   cmd_report(+0x1d20) in perf [0x43cb50]
>   run_builtin(+0x53) in perf [0x4870b3]
>   main(+0x634) in perf [0x42cb54]
>   __libc_start_main(+0xf0) in libc-2.22.so [0x7fea3577c610]
>   _start(+0x29) in perf [0x42cc79]
>   [0x0]

nice idea! 

SNIP

> +
> +void create_perf_thread(void)
> +{
> +	struct perf_tool tool = {
> +		.comm	= perf_event__process_comm,
> +		.mmap	= perf_event__process_mmap,
> +		.mmap2	= perf_event__process_mmap2,
> +	};
> +	struct thread_map *tm;
> +	struct machine *machine;
> +	int pid = getpid();
> +
> +	machine = machine__new_host();
> +	if (machine == NULL)
> +		return;
> +
> +	tm = thread_map__new_dummy();
> +	if (tm == NULL) {
> +		machine__delete(machine);
> +		return;
> +	}

I think we could treat errors the usual way in here..
if fail to alloc this early, something is terribly wrong anyway

jirka
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1276109

From	Namhyung Kim <namhyung@kernel.org>
Date	2015-11-24 08:40 +0100
Message-ID	<qygmK-558-11@gated-at.bofh.it>
In reply to	#1273892

Hi Jiri,

On Fri, Nov 20, 2015 at 10:29:48AM +0100, Jiri Olsa wrote:
> On Fri, Nov 20, 2015 at 03:03:03PM +0900, Namhyung Kim wrote:
> > Backtrace is a crucial info for debugging.  And upcoming refcnt
> > tracking facility also wants to use it.
> > 
> > So instead of relying on glibc's backtrace_symbols[_fd] which misses
> > some (static) functions , use our own symbol searching mechanism.  To
> > do that, add perf_thread global variable to keep its maps and symbols.
> > 
> > The backtrace output from TUI is changed like below.  (I made a key
> > action to generate a segfault for testing):
> > 
> > Before:
> >   perf: Segmentation fault
> >   -------- backtrace --------
> >   perf[0x544a8b]
> >   /usr/lib/libc.so.6(+0x33680)[0x7fc46420b680]
> >   perf[0x54041b]
> >   perf(perf_evlist__tui_browse_hists+0x91)[0x5432e1]
> >   perf(cmd_report+0x1d20)[0x43cb10]
> >   perf[0x487073]
> >   perf(main+0x62f)[0x42cb1f]
> >   /usr/lib/libc.so.6(__libc_start_main+0xf0)[0x7fc4641f8610]
> >   perf(_start+0x29)[0x42cc39]
> >   [0x0]
> > 
> > After:
> >   perf: Segmentation fault
> >   -------- backtrace --------
> >   perf_evsel__hists_browse(+0x43b) in perf [0x54066b]
> >   perf_evlist__tui_browse_hists(+0x91) in perf [0x543531]
> >   cmd_report(+0x1d20) in perf [0x43cb50]
> >   run_builtin(+0x53) in perf [0x4870b3]
> >   main(+0x634) in perf [0x42cb54]
> >   __libc_start_main(+0xf0) in libc-2.22.so [0x7fea3577c610]
> >   _start(+0x29) in perf [0x42cc79]
> >   [0x0]
> 
> nice idea! 
> 
> SNIP
> 
> > +
> > +void create_perf_thread(void)
> > +{
> > +	struct perf_tool tool = {
> > +		.comm	= perf_event__process_comm,
> > +		.mmap	= perf_event__process_mmap,
> > +		.mmap2	= perf_event__process_mmap2,
> > +	};
> > +	struct thread_map *tm;
> > +	struct machine *machine;
> > +	int pid = getpid();
> > +
> > +	machine = machine__new_host();
> > +	if (machine == NULL)
> > +		return;
> > +
> > +	tm = thread_map__new_dummy();
> > +	if (tm == NULL) {
> > +		machine__delete(machine);
> > +		return;
> > +	}
> 
> I think we could treat errors the usual way in here..
> if fail to alloc this early, something is terribly wrong anyway

OK, I'll change it to return error code and let perf fail with it.

Thanks,
Namhyung
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1275890

From	Arnaldo Carvalho de Melo <acme@kernel.org>
Date	2015-11-23 22:40 +0100
Message-ID	<qy708-7p3-61@gated-at.bofh.it>
In reply to	#1273741

Em Fri, Nov 20, 2015 at 03:03:03PM +0900, Namhyung Kim escreveu:
> Backtrace is a crucial info for debugging.  And upcoming refcnt
> tracking facility also wants to use it.
> 
> So instead of relying on glibc's backtrace_symbols[_fd] which misses
> some (static) functions , use our own symbol searching mechanism.  To
> do that, add perf_thread global variable to keep its maps and symbols.
> 
> The backtrace output from TUI is changed like below.  (I made a key
> action to generate a segfault for testing):

This is a really nice use of what we have, I guess we could simplify
things further, not having to create a struct machine, as we have just
one thread, and we're not interested in kernel addresses, so no need for
kmaps, etc.

But as-is it already looks better than what we were using :-)

I'll try testing it further and will probably switch to using it if
nobody voices any problem we haven't realised with such approach.

- Arnaldo
 
> Before:
>   perf: Segmentation fault
>   -------- backtrace --------
>   perf[0x544a8b]
>   /usr/lib/libc.so.6(+0x33680)[0x7fc46420b680]
>   perf[0x54041b]
>   perf(perf_evlist__tui_browse_hists+0x91)[0x5432e1]
>   perf(cmd_report+0x1d20)[0x43cb10]
>   perf[0x487073]
>   perf(main+0x62f)[0x42cb1f]
>   /usr/lib/libc.so.6(__libc_start_main+0xf0)[0x7fc4641f8610]
>   perf(_start+0x29)[0x42cc39]
>   [0x0]
> 
> After:
>   perf: Segmentation fault
>   -------- backtrace --------
>   perf_evsel__hists_browse(+0x43b) in perf [0x54066b]
>   perf_evlist__tui_browse_hists(+0x91) in perf [0x543531]
>   cmd_report(+0x1d20) in perf [0x43cb50]
>   run_builtin(+0x53) in perf [0x4870b3]
>   main(+0x634) in perf [0x42cb54]
>   __libc_start_main(+0xf0) in libc-2.22.so [0x7fea3577c610]
>   _start(+0x29) in perf [0x42cc79]
>   [0x0]
> 
> Cc: Frederic Weisbecker <fweisbec@gmail.com>
> Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
> ---
>  tools/perf/perf.c         |  7 +++++-
>  tools/perf/ui/tui/setup.c | 21 ++++++++++++++--
>  tools/perf/util/util.c    | 62 +++++++++++++++++++++++++++++++++++++++++++++++
>  tools/perf/util/util.h    |  5 ++++
>  4 files changed, 92 insertions(+), 3 deletions(-)
> 
> diff --git a/tools/perf/perf.c b/tools/perf/perf.c
> index 4bee53c3f796..f77eb440b05c 100644
> --- a/tools/perf/perf.c
> +++ b/tools/perf/perf.c
> @@ -604,6 +604,8 @@ int main(int argc, const char **argv)
>  	 */
>  	pthread__block_sigwinch();
>  
> +	create_perf_thread();
> +
>  	while (1) {
>  		static int done_help;
>  		int was_alias = run_argv(&argc, &argv);
> @@ -615,7 +617,7 @@ int main(int argc, const char **argv)
>  			fprintf(stderr, "Expansion of alias '%s' failed; "
>  				"'%s' is not a perf-command\n",
>  				cmd, argv[0]);
> -			goto out;
> +			goto out_destroy;
>  		}
>  		if (!done_help) {
>  			cmd = argv[0] = help_unknown_cmd(cmd);
> @@ -626,6 +628,9 @@ int main(int argc, const char **argv)
>  
>  	fprintf(stderr, "Failed to run command '%s': %s\n",
>  		cmd, strerror_r(errno, sbuf, sizeof(sbuf)));
> +
> +out_destroy:
> +	destroy_perf_thread();
>  out:
>  	return 1;
>  }
> diff --git a/tools/perf/ui/tui/setup.c b/tools/perf/ui/tui/setup.c
> index 7dfeba0a91f3..bc2da884e65a 100644
> --- a/tools/perf/ui/tui/setup.c
> +++ b/tools/perf/ui/tui/setup.c
> @@ -13,6 +13,8 @@
>  #include "../libslang.h"
>  #include "../keysyms.h"
>  #include "tui.h"
> +#include "../../util/symbol.h"
> +#include "../../util/thread.h"
>  
>  static volatile int ui__need_resize;
>  
> @@ -96,14 +98,29 @@ int ui__getch(int delay_secs)
>  static void ui__signal_backtrace(int sig)
>  {
>  	void *stackdump[32];
> -	size_t size;
> +	size_t size, i;
>  
>  	ui__exit(false);
>  	psignal(sig, "perf");
>  
>  	printf("-------- backtrace --------\n");
>  	size = backtrace(stackdump, ARRAY_SIZE(stackdump));
> -	backtrace_symbols_fd(stackdump, size, STDOUT_FILENO);
> +	/* skip first two stack frame (for this function and signal stack) */
> +	for (i = 2; i < size; i++) {
> +		struct addr_location al = {
> +			.sym = NULL,
> +		};
> +
> +		thread__find_addr_location(perf_thread, PERF_RECORD_MISC_USER,
> +					   MAP__FUNCTION, (long)stackdump[i], &al);
> +
> +		if (al.sym)
> +			printf("%s(+0x%"PRIx64") in ", al.sym->name,
> +			       map__map_ip(al.map, (u64)stackdump[i]) - al.sym->start);
> +		if (al.map)
> +			printf("%s ", al.map->dso->short_name);
> +		printf("[0x%lx]\n", (unsigned long)stackdump[i]);
> +	}
>  
>  	exit(0);
>  }
> diff --git a/tools/perf/util/util.c b/tools/perf/util/util.c
> index 75759aebc7b8..f1a26ea14053 100644
> --- a/tools/perf/util/util.c
> +++ b/tools/perf/util/util.c
> @@ -16,6 +16,9 @@
>  #include <linux/kernel.h>
>  #include <unistd.h>
>  #include "callchain.h"
> +#include "machine.h"
> +#include "thread.h"
> +#include "thread_map.h"
>  
>  struct callchain_param	callchain_param = {
>  	.mode	= CHAIN_GRAPH_ABS,
> @@ -696,3 +699,62 @@ fetch_kernel_version(unsigned int *puint, char *str,
>  		*puint = (version << 16) + (patchlevel << 8) + sublevel;
>  	return 0;
>  }
> +
> +
> +static int process_event(struct perf_tool *tool, union perf_event *event,
> +			 struct perf_sample *sample, struct machine *machine)
> +{
> +	switch (event->header.type) {
> +	case PERF_RECORD_COMM:
> +		return tool->comm(tool, event, sample, machine);
> +	case PERF_RECORD_MMAP:
> +		return tool->mmap(tool, event, sample, machine);
> +	case PERF_RECORD_MMAP2:
> +		return tool->mmap2(tool, event, sample, machine);
> +	default:
> +		break;
> +	}
> +	return 0;
> +}
> +
> +struct thread *perf_thread;
> +
> +void create_perf_thread(void)
> +{
> +	struct perf_tool tool = {
> +		.comm	= perf_event__process_comm,
> +		.mmap	= perf_event__process_mmap,
> +		.mmap2	= perf_event__process_mmap2,
> +	};
> +	struct thread_map *tm;
> +	struct machine *machine;
> +	int pid = getpid();
> +
> +	machine = machine__new_host();
> +	if (machine == NULL)
> +		return;
> +
> +	tm = thread_map__new_dummy();
> +	if (tm == NULL) {
> +		machine__delete(machine);
> +		return;
> +	}
> +
> +	thread_map__set_pid(tm, 0, pid);
> +
> +	perf_event__synthesize_thread_map(&tool, tm, process_event, machine,
> +					  false, 500);
> +
> +	perf_thread = machine__find_thread(machine, pid, pid);
> +	BUG_ON(perf_thread == NULL);
> +
> +	thread_map__put(tm);
> +}
> +
> +void destroy_perf_thread(void)
> +{
> +	struct machine *machine = perf_thread->mg->machine;
> +
> +	machine__delete_threads(machine);
> +	machine__delete(machine);
> +}
> diff --git a/tools/perf/util/util.h b/tools/perf/util/util.h
> index dcc659017976..630e145049aa 100644
> --- a/tools/perf/util/util.h
> +++ b/tools/perf/util/util.h
> @@ -358,4 +358,9 @@ int fetch_kernel_version(unsigned int *puint,
>  #define KVER_FMT	"%d.%d.%d"
>  #define KVER_PARAM(x)	KVER_VERSION(x), KVER_PATCHLEVEL(x), KVER_SUBLEVEL(x)
>  
> +extern struct thread *perf_thread;
> +
> +void create_perf_thread(void);
> +void destroy_perf_thread(void);
> +
>  #endif /* GIT_COMPAT_UTIL_H */
> -- 
> 2.6.2
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1276104

From	Namhyung Kim <namhyung@kernel.org>
Date	2015-11-24 08:40 +0100
Message-ID	<qygmJ-558-1@gated-at.bofh.it>
In reply to	#1275890

On Mon, Nov 23, 2015 at 06:39:36PM -0300, Arnaldo Carvalho de Melo wrote:
> Em Fri, Nov 20, 2015 at 03:03:03PM +0900, Namhyung Kim escreveu:
> > Backtrace is a crucial info for debugging.  And upcoming refcnt
> > tracking facility also wants to use it.
> > 
> > So instead of relying on glibc's backtrace_symbols[_fd] which misses
> > some (static) functions , use our own symbol searching mechanism.  To
> > do that, add perf_thread global variable to keep its maps and symbols.
> > 
> > The backtrace output from TUI is changed like below.  (I made a key
> > action to generate a segfault for testing):
> 
> This is a really nice use of what we have, I guess we could simplify
> things further, not having to create a struct machine, as we have just
> one thread, and we're not interested in kernel addresses, so no need for
> kmaps, etc.

Yes, I thought about it.  But as adding the perf thread to an existing
machine can affect other thread(s), I didn't do it.  Maybe we can set
the perf threads' machine pointer in a hacky way without adding the
thread into the machine, but I'd rather not doing that too because
it's fragile and current code is simple enough IMHO.


> 
> But as-is it already looks better than what we were using :-)
> 
> I'll try testing it further and will probably switch to using it if
> nobody voices any problem we haven't realised with such approach.

Thank you!
Namhyung
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [standalone]

csiph-web

[RFC/PATCH] perf tools: Introduce perf_thread for backtrace

Contents

#1273741 — [RFC/PATCH] perf tools: Introduce perf_thread for backtrace

#1273881

#1274021

#1276103

#1273892

#1276109

#1275890

#1276104