perf/core improvements and fixes

Hi Ingo,

Please consider pulling,

- Arnaldo

The following changes since commit a41f3c8cd4e28dcbebd8ec27a9602c86cfa5f009:

perf/x86/intel/uncore: Add Broadwell-U uncore IMC PMU support (2015-05-11 11:57:47 +0200)

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git tags/perf-core-for-mingo

for you to fetch changes up to 70923bd26c732ac2d9e248d80214df6dfd75f78c:

perf tools: Make flex/bison calls honour V=1 (2015-05-14 19:27:47 -0300)

----------------------------------------------------------------
perf/core improvements and fixes:

User visible:

- Add --range option to show a variable's location range in 'perf probe', helping in collecting variables in probes when there is a mismatch between assembly and source code (He Kuang)

- Show better error message when failed to find variable in 'perf probe' (He Kuang)

- Fix 'perf report --thread' handling and document it better (Namhyung Kim)

Infrastructure:

- Fix to get negative exit codes in 'perf test' test routines (He Kuang)

- Make flex/bison calls honour V=1 (Jiri Olsa)

- Ignore tail calls to probed functions in 'perf probe' (Naveen N. Rao)

- Fix refcount expectations in map_group share 'perf test' (Arnaldo Carvalho de Melo)

Build Fixes:

- Fix 'perf kmem' build due to compiler thinking uninitialized var is being accessed (Arnaldo Carvalho de Melo)

- Provide le16toh if not defined, to fix the libtraceevent build on older distros (Arnaldo Carvalho de Melo)

- Fix 'perf trace' build on older distros by providing some CLOEXEC, NONBLOCK defines (Arnaldo Carvalho de Melo)

----------------------------------------------------------------
Arnaldo Carvalho de Melo (6):
perf kmem: Fix compiler warning about may be accessing uninitialized variable perf tests: Show refcounting broken expectations in thread-mg-share test perf machine: No need to keep a refcnt for last_match
perf tests: Fix map_groups refcount test
tools lib traceevent: Provide le16toh define for older systems perf trace: Fix the build on older distros

He Kuang (5):
perf trace: Removed duplicated NULL test
perf probe: Remove length limitation for showing available variables perf probe: Add --range option to show a variable's location range perf probe: Show better error message when failed to find variable perf tests: Fix to get negative exit codes

Jiri Olsa (1):
perf tools: Make flex/bison calls honour V=1

Namhyung Kim (4):
perf tools: Document relation of per-thread event count feature perf report: Force tty output if -T/--thread option is given perf report: Do not restrict -T option by other options perf report: Fix some option handling on --stdio

Naveen N. Rao (1):
perf probe: Ignore tail calls to probed functions

 tools/lib/traceevent/plugin_cfg80211.c   |  13 ++
 tools/perf/Documentation/perf-record.txt |   3 +-
 tools/perf/Documentation/perf-report.txt |   3 +-
 tools/perf/builtin-kmem.c                |   2 +-
 tools/perf/builtin-probe.c               |   2 +
 tools/perf/builtin-report.c              |  17 ++-
 tools/perf/builtin-trace.c               |  32 ++++-
 tools/perf/tests/builtin-test.c          |   2 +-
 tools/perf/tests/tests.h                 |   9 ++
 tools/perf/tests/thread-mg-share.c       |  29 +++--
 tools/perf/util/Build                    |   8 +-
 tools/perf/util/cache.h                  |   1 -
 tools/perf/util/dwarf-aux.c              | 204 +++++++++++++++++++++++++++----
 tools/perf/util/dwarf-aux.h              |  10 +-
 tools/perf/util/environment.c            |   1 -
 tools/perf/util/machine.c                |   8 +-
 tools/perf/util/pager.c                  |   5 -
 tools/perf/util/probe-event.h            |   1 +
 tools/perf/util/probe-finder.c           |  90 ++++++++++----
 19 files changed, 346 insertions(+), 94 deletions(-)

From: Arnaldo Carvalho de Melo <[email protected]>

The last argument to strtok_r doesn't need to be initialized, its just a placeholder to make this routine reentrant, but gcc doesn't know about that and complains, breaking the build, fix it by setting it to NULL.

---
 tools/perf/builtin-kmem.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index e0173c7f17b1..254614b10c4a 100644
@@ -713,7 +713,7 @@ static int parse_gfp_flags(struct perf_evsel *evsel, struct perf_sample *sample,
 		.size = sample->raw_size,
 	};
 	struct trace_seq seq;
-	char *str, *pos;
+	char *str, *pos = NULL;
 
 	if (nr_gfps) {
 		struct gfp_flag key = {

2.1.0

From: Namhyung Kim <[email protected]>

The 'perf record -s' and 'perf report -T' should be used together to see per-thread event counts. Document the relation of these commands.

---
 tools/perf/Documentation/perf-record.txt | 3 ++-
 tools/perf/Documentation/perf-report.txt | 3 ++-
 2 files changed, 4 insertions(+), 2 deletions(-)

index 57dd57bcef95..280533ebf9df 100644
@@ -147,7 +147,8 @@ OPTIONS
 
 -s::
 --stat::
-	Per thread counts.
+	Record per-thread event counts.  Use it with 'perf report -T' to see
+	the values.
 
 -d::
 --data::

index 27190ed06f9c..c33b69f3374f 100644
@@ -34,7 +34,8 @@ OPTIONS
 
 -T::
 --threads::
-	Show per-thread event counters
+	Show per-thread event counters.  The input data file should be recorded
+	with -s option.
 -c::
 --comms=::
 	Only consider symbols in these comms. CSV that understands

2.1.0

From: Namhyung Kim <[email protected]>

The -T/--thread option is supported only on --stdio mode (at least for now). So enforce the tty output if the option was requested.

---
 tools/perf/builtin-report.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

index 62b49ca0fc7b..4d642db642c5 100644
@@ -815,8 +815,8 @@ repeat:
 		goto error;
 	}
 
-	/* Force tty output for header output. */
-	if (report.header || report.header_only)
+	/* Force tty output for header output and per-thread stat. */
+	if (report.header || report.header_only || report.show_threads)
 		use_browser = 0;
 
 	if (strcmp(input_name, "-") != 0)

2.1.0

From: He Kuang <[email protected]>

No need to test trace.evlist against NULL twice.

---
 tools/perf/builtin-trace.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

index cbfdb9523868..96a2ebafc755 100644
@@ -2721,11 +2721,10 @@ int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
 	signal(SIGFPE, sighandler_dump_stack);
 
 	trace.evlist = perf_evlist__new();
-	if (trace.evlist == NULL)
-		return -ENOMEM;
 
 	if (trace.evlist == NULL) {
 		pr_err("Not enough memory to run!\n");
+		err = -ENOMEM;
 		goto out;
 	}
 

2.1.0

From: He Kuang <[email protected]>

Use struct strbuf instead of bare char[] to remove the length limitation of variables in variable_list, so they will not disappear due to overlength, and make preparation for adding more description for variables.

---
 tools/perf/util/dwarf-aux.c    | 50 +++++++++++++++++++-----------------------
 tools/perf/util/dwarf-aux.h    |  4 ++--
 tools/perf/util/probe-finder.c | 17 ++++++++------
 3 files changed, 34 insertions(+), 37 deletions(-)

index 16d46e26edac..737c9dbe5643 100644
@@ -848,19 +848,17 @@ Dwarf_Die *die_find_member(Dwarf_Die *st_die, const char *name,
 /**
  * die_get_typename - Get the name of given variable DIE
  * @vr_die: a variable DIE
- * @buf: a buffer for result type name
- * @len: a max-length of @buf
+ * @buf: a strbuf for result type name
  *
- * Get the name of @vr_die and stores it to @buf. Return the actual length
- * of type name if succeeded. Return -E2BIG if @len is not enough long, and
- * Return -ENOENT if failed to find type name.
+ * Get the name of @vr_die and stores it to @buf. Return 0 if succeeded.
+ * and Return -ENOENT if failed to find type name.
  * Note that the result will stores typedef name if possible, and stores
  * "*(function_type)" if the type is a function pointer.
  */
-int die_get_typename(Dwarf_Die *vr_die, char *buf, int len)
+int die_get_typename(Dwarf_Die *vr_die, struct strbuf *buf)
 {
 	Dwarf_Die type;
-	int tag, ret, ret2;
+	int tag, ret;
 	const char *tmp = "";
 
 	if (__die_get_real_type(vr_die, &type) == NULL)

@@ -871,8 +869,8 @@ int die_get_typename(Dwarf_Die *vr_die, char *buf, int len)
 		tmp = "*";
 	else if (tag == DW_TAG_subroutine_type) {
 		/* Function pointer */
-		ret = snprintf(buf, len, "(function_type)");
-		return (ret >= len) ? -E2BIG : ret;
+		strbuf_addf(buf, "(function_type)");
+		return 0;
 	} else {
 		if (!dwarf_diename(&type))
 			return -ENOENT;

@@ -883,39 +881,35 @@ int die_get_typename(Dwarf_Die *vr_die, char *buf, int len)
 		else if (tag == DW_TAG_enumeration_type)
 			tmp = "enum ";
 		/* Write a base name */
-		ret = snprintf(buf, len, "%s%s", tmp, dwarf_diename(&type));
-		return (ret >= len) ? -E2BIG : ret;
-	}
-	ret = die_get_typename(&type, buf, len);
-	if (ret > 0) {
-		ret2 = snprintf(buf + ret, len - ret, "%s", tmp);
-		ret = (ret2 >= len - ret) ? -E2BIG : ret2 + ret;
+		strbuf_addf(buf, "%s%s", tmp, dwarf_diename(&type));
+		return 0;
 	}
+	ret = die_get_typename(&type, buf);
+	if (ret == 0)
+		strbuf_addf(buf, "%s", tmp);
+
 	return ret;
 }
 
 /**
  * die_get_varname - Get the name and type of given variable DIE
  * @vr_die: a variable DIE
- * @buf: a buffer for type and variable name
- * @len: the max-length of @buf
+ * @buf: a strbuf for type and variable name
  *
  * Get the name and type of @vr_die and stores it in @buf as "type\tname".
  */
-int die_get_varname(Dwarf_Die *vr_die, char *buf, int len)
+int die_get_varname(Dwarf_Die *vr_die, struct strbuf *buf)
 {
-	int ret, ret2;
+	int ret;
 
-	ret = die_get_typename(vr_die, buf, len);
+	ret = die_get_typename(vr_die, buf);
 	if (ret < 0) {
 		pr_debug("Failed to get type, make it unknown.\n");
-		ret = snprintf(buf, len, "(unknown_type)");
-	}
-	if (ret > 0) {
-		ret2 = snprintf(buf + ret, len - ret, "\t%s",
-				dwarf_diename(vr_die));
-		ret = (ret2 >= len - ret) ? -E2BIG : ret2 + ret;
+		strbuf_addf(buf, "(unknown_type)");
 	}
-	return ret;
+
+	strbuf_addf(buf, "\t%s", dwarf_diename(vr_die));
+
+	return 0;
 }
 

index 50a3cdc55fd7..60676fda4824 100644
@@ -117,8 +117,8 @@ extern Dwarf_Die *die_find_member(Dwarf_Die *st_die, const char *name,
 				  Dwarf_Die *die_mem);
 
 /* Get the name of given variable DIE */
-extern int die_get_typename(Dwarf_Die *vr_die, char *buf, int len);
+extern int die_get_typename(Dwarf_Die *vr_die, struct strbuf *buf);
 
 /* Get the name and type of given variable DIE, stored as "type\tname" */
-extern int die_get_varname(Dwarf_Die *vr_die, char *buf, int len);
+extern int die_get_varname(Dwarf_Die *vr_die, struct strbuf *buf);
 #endif

index 8b9e274f940c..d7c2e90ab4ce 100644
@@ -1255,14 +1255,11 @@ int debuginfo__find_trace_events(struct debuginfo *dbg,
 	return (ret < 0) ? ret : tf.ntevs;
 }
 
-#define MAX_VAR_LEN 64
-
 /* Collect available variables in this scope */
 static int collect_variables_cb(Dwarf_Die *die_mem, void *data)
 {
 	struct available_var_finder *af = data;
 	struct variable_list *vl;
-	char buf[MAX_VAR_LEN];
 	int tag, ret;
 
 	vl = &af->vls[af->nvls - 1];

@@ -1274,10 +1271,16 @@ static int collect_variables_cb(Dwarf_Die *die_mem, void *data)
 						af->pf.fb_ops, &af->pf.sp_die,
 						NULL);
 		if (ret == 0) {
-			ret = die_get_varname(die_mem, buf, MAX_VAR_LEN);
-			pr_debug2("Add new var: %s\n", buf);
-			if (ret > 0)
-				strlist__add(vl->vars, buf);
+			struct strbuf buf;
+
+			strbuf_init(&buf, 64);
+			ret = die_get_varname(die_mem, &buf);
+			pr_debug2("Add new var: %s\n", buf.buf);
+			if (ret == 0) {
+				strlist__add(vl->vars,
+					strbuf_detach(&buf, NULL));
+			}
+			strbuf_release(&buf);
 		}
 	}
 

2.1.0

From: He Kuang <[email protected]>

It is not easy for users to get the accurate byte offset or the line number where a local variable can be probed.

With '--range' option, local variables in the scope of the probe point are showed with a byte offset range, and can be added according to this range information.

For example, there are some variables in the function generic_perform_write():

  <generic_perform_write@mm/filemap.c:0>
  0  ssize_t generic_perform_write(struct file *file,
  1                                 struct iov_iter *i, loff_t pos)
  2  {
  3          struct address_space *mapping = file->f_mapping;
  4          const struct address_space_operations *a_ops = mapping->a_ops;
  ...
  42                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
                                               &page, &fsdata);
  44                 if (unlikely(status < 0))

But we fail when we try to probe the variable 'a_ops' at line 42 or 44.

  $ perf probe --add 'generic_perform_write:42 a_ops'
  Failed to find the location of a_ops at this address.
    Perhaps, it has been optimized out.

This is because the source code do not match the assembly, so a variable may not be available in the source code line where it appears.

After this patch, we can lookup the accurate byte offset range of a variable, 'INV' indicates that this variable is not valid at the given point, but available in the scope:

  $ perf probe --vars 'generic_perform_write:42' --range
  Available variables at generic_perform_write:42
    @<generic_perform_write+141>
       [INV] ssize_t written @<generic_perform_write+[324-331]>
       [INV] struct address_space_operations*        a_ops   @<generic_perform_write+[55-61,170-176,223-246]>
       [VAL] (unknown_type)  fsdata  @<generic_perform_write+[70-307,346-411]>
       [VAL] loff_t  pos     @<generic_perform_write+[0-286,286-336,346-411]>
       [VAL] long int        status  @<generic_perform_write+[83-342,346-411]>
       [VAL] long unsigned int       bytes   @<generic_perform_write+[122-311,320-338,346-403,403-411]>
       [VAL] struct address_space*   mapping @<generic_perform_write+[35-344,346-411]>
       [VAL] struct iov_iter*        i       @<generic_perform_write+[0-340,346-411]>
       [VAL] struct page*    page    @<generic_perform_write+[70-307,346-411]>

Then it is more clear for us to add a probe with this variable:

  $ perf probe --add 'generic_perform_write+170 a_ops'
  Added new event:
    probe:generic_perform_write (on generic_perform_write+170 with a_ops)

---
 tools/perf/builtin-probe.c     |   2 +
 tools/perf/util/dwarf-aux.c    | 121 +++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/dwarf-aux.h    |   2 +
 tools/perf/util/probe-event.h  |   1 +
 tools/perf/util/probe-finder.c |  57 ++++++++++++++-----
 5 files changed, 170 insertions(+), 13 deletions(-)

index 7fa2c7a1086a..1272559fa22d 100644
@@ -372,6 +372,8 @@ __cmd_probe(int argc, const char **argv, const char *prefix __maybe_unused)
 		     "Show accessible variables on PROBEDEF", opt_show_vars),
 	OPT_BOOLEAN('\0', "externs", &probe_conf.show_ext_vars,
 		    "Show external variables too (with --vars only)"),
+	OPT_BOOLEAN('\0', "range", &probe_conf.show_location_range,
+		"Show variables location range in scope (with --vars only)"),
 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 		   "file", "vmlinux pathname"),
 	OPT_STRING('s', "source", &symbol_conf.source_prefix,

index 60676fda4824..c154c0b80880 100644
@@ -121,4 +121,6 @@ extern int die_get_typename(Dwarf_Die *vr_die, struct strbuf *buf);
 
 /* Get the name and type of given variable DIE, stored as "type\tname" */
 extern int die_get_varname(Dwarf_Die *vr_die, struct strbuf *buf);
+extern int die_get_var_range(Dwarf_Die *sp_die, Dwarf_Die *vr_die,
+			struct strbuf *buf);
 #endif

index 1e2faa3559d2..537eb329c2cf 100644
@@ -9,6 +9,7 @@
 /* Probe related configurations */
 struct probe_conf {
 	bool	show_ext_vars;
+	bool	show_location_range;
 	bool	force_add;
 	bool	no_inlines;
 	int	max_probes;

2.1.0

...

From: He Kuang <[email protected]>

Indicate to check variable location range in error message when we got failed to find the variable.

Before this patch:

  $ perf probe --add 'generic_perform_write+118 bytes'
  Failed to find the location of bytes at this address.
   Perhaps, it has been optimized out.
    Error: Failed to add events.

After this patch:

  $ perf probe --add 'generic_perform_write+118 bytes'
  Failed to find the location of the 'bytes' variable at this address.
   Perhaps it has been optimized out.
   Use -V with the --range option to show 'bytes' location range.
    Error: Failed to add events.

---
 tools/perf/util/probe-finder.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

index 5804086425aa..590a24a7f555 100644
@@ -527,10 +527,12 @@ static int convert_variable(Dwarf_Die *vr_die, struct probe_finder *pf)
 
 	ret = convert_variable_location(vr_die, pf->addr, pf->fb_ops,
 					&pf->sp_die, pf->tvar);
-	if (ret == -ENOENT || ret == -EINVAL)
-		pr_err("Failed to find the location of %s at this address.\n"
-		       " Perhaps, it has been optimized out.\n", pf->pvar->var);
-	else if (ret == -ENOTSUP)
+	if (ret == -ENOENT || ret == -EINVAL) {
+		pr_err("Failed to find the location of the '%s' variable at this address.\n"
+		       " Perhaps it has been optimized out.\n"
+		       " Use -V with the --range option to show '%s' location range.\n",
+		       pf->pvar->var, pf->pvar->var);
+	} else if (ret == -ENOTSUP)
 		pr_err("Sorry, we don't support this variable location yet.\n");
 	else if (ret == 0 && pf->pvar->field) {
 		ret = convert_variable_fields(vr_die, pf->pvar->var,

2.1.0

From: He Kuang <[email protected]>

WEXITSTATUS consists of the least significant 8 bits of the status argument, so we should convert the value to signed char if we have valid negative exit codes. And the return value of test->func() contains negative values:

  enum {
          TEST_OK   =  0,
          TEST_FAIL = -1,
          TEST_SKIP = -2,
  };

Before this patch:

  $ perf test -v 1
  ...
  test child finished with 254
  ---- end ----
  vmlinux symtab matches kallsyms: FAILED!

After this patch:

  $ perf test -v 1
  ...
  test child finished with -2
  ---- end ----
  vmlinux symtab matches kallsyms: Skip

---
 tools/perf/tests/builtin-test.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index 4f4098167112..f42af98a5c16 100644
@@ -219,7 +219,7 @@ static int run_test(struct test *test)
 	wait(&status);
 
 	if (WIFEXITED(status)) {
-		err = WEXITSTATUS(status);
+		err = (signed char)WEXITSTATUS(status);
 		pr_debug("test child finished with %d\n", err);
 	} else if (WIFSIGNALED(status)) {
 		err = -1;

2.1.0

From: Namhyung Kim <[email protected]>

It seems there's no reason to suppress per-thread event stat by -T option when -s or -p option is used. Make it work with those options.

---
 tools/perf/builtin-report.c | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)

index 4d642db642c5..92fca2157e5e 100644
@@ -333,15 +333,14 @@ static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
 	}
 
 	if (sort_order == NULL &&
-	    parent_pattern == default_parent_pattern) {
+	    parent_pattern == default_parent_pattern)
 		fprintf(stdout, "#\n# (%s)\n#\n", help);
 
-		if (rep->show_threads) {
-			bool style = !strcmp(rep->pretty_printing_style, "raw");
-			perf_read_values_display(stdout, &rep->show_threads_values,
-						 style);
-			perf_read_values_destroy(&rep->show_threads_values);
-		}
+	if (rep->show_threads) {
+		bool style = !strcmp(rep->pretty_printing_style, "raw");
+		perf_read_values_display(stdout, &rep->show_threads_values,
+					 style);
+		perf_read_values_destroy(&rep->show_threads_values);
 	}
 
 	return 0;

2.1.0

From: Arnaldo Carvalho de Melo <[email protected]>

To help understand the failure.

[acme@zoo linux]$ perf test -v 30
30: Test thread mg sharing :
--- start ---
test child forked, pid 12275

  FAILED tests/thread-mg-share.c:68 wrong refcnt (4 != 3)
  test child finished with -1
  ---- end ----
  Test thread mg sharing: FAILED!
  [acme@zoo linux]$

This is under investigation, the thread__delete() calls were replaced with thread__put(), and those cause mismatches because now we need to be more judicious with the thread lifetime management.

I.e. previously the thread__delete() would drop the map_group refcount, but now since thread__put doesn't call thread__delete() necessarily. because we have other refcount holders, the map_group refcount will not be as we expected when this test was implemented.

Will be fixed soon...

---
 tools/perf/tests/tests.h           |  9 +++++++++
 tools/perf/tests/thread-mg-share.c | 12 ++++++------
 2 files changed, 15 insertions(+), 6 deletions(-)

index 52758a33f64c..a10eaf5c4767 100644
@@ -9,6 +9,15 @@ do {									 \
 	}								 \
 } while (0)
 
+#define TEST_ASSERT_EQUAL(text, val, expected)				 \
+do {									 \
+	if (val != expected) {						 \
+		pr_debug("FAILED %s:%d %s (%d != %d)\n",		 \
+			 __FILE__, __LINE__, text, val, expected);	 \
+		return -1;						 \
+	}								 \
+} while (0)
+
 enum {
 	TEST_OK   =  0,
 	TEST_FAIL = -1,

index dc05bd62b4a3..04ecceed4e5b 100644
@@ -43,7 +43,7 @@ int test__thread_mg_share(void)
 			leader && t1 && t2 && t3 && other);
 
 	mg = leader->mg;
-	TEST_ASSERT_VAL("wrong refcnt", mg->refcnt == 4);
+	TEST_ASSERT_EQUAL("wrong refcnt", mg->refcnt, 4);
 
 	/* test the map groups pointer is shared */
 	TEST_ASSERT_VAL("map groups don't match", mg == t1->mg);

@@ -59,25 +59,25 @@ int test__thread_mg_share(void)
 	TEST_ASSERT_VAL("failed to find other leader", other_leader);
 
 	other_mg = other->mg;
-	TEST_ASSERT_VAL("wrong refcnt", other_mg->refcnt == 2);
+	TEST_ASSERT_EQUAL("wrong refcnt", other_mg->refcnt, 2);
 
 	TEST_ASSERT_VAL("map groups don't match", other_mg == other_leader->mg);
 
 	/* release thread group */
 	thread__put(leader);
-	TEST_ASSERT_VAL("wrong refcnt", mg->refcnt == 3);
+	TEST_ASSERT_EQUAL("wrong refcnt", mg->refcnt, 3);
 
 	thread__put(t1);
-	TEST_ASSERT_VAL("wrong refcnt", mg->refcnt == 2);
+	TEST_ASSERT_EQUAL("wrong refcnt", mg->refcnt, 2);
 
 	thread__put(t2);
-	TEST_ASSERT_VAL("wrong refcnt", mg->refcnt == 1);
+	TEST_ASSERT_EQUAL("wrong refcnt", mg->refcnt, 1);
 
 	thread__put(t3);
 
 	/* release other group  */
 	thread__put(other_leader);
-	TEST_ASSERT_VAL("wrong refcnt", other_mg->refcnt == 1);
+	TEST_ASSERT_EQUAL("wrong refcnt", other_mg->refcnt, 1);
 
 	thread__put(other);
 

2.1.0

From: Arnaldo Carvalho de Melo <[email protected]>

Since it is all associated with the refcount for keeping the thread in the rbtree, it is excessive and unecessarily complex to hold a refcont when changing machine->last_match.

---
 tools/perf/util/machine.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

index 8b0b307d91f4..34bf89f7f4f3 100644
@@ -364,7 +364,7 @@ static struct thread *____machine__findnew_thread(struct machine *machine,
 			return th;
 		}
 
-		thread__zput(machine->last_match);
+		machine->last_match = NULL;
 	}
 
 	while (*p != NULL) {

@@ -372,7 +372,7 @@ static struct thread *____machine__findnew_thread(struct machine *machine,
 		th = rb_entry(parent, struct thread, rb_node);
 
 		if (th->tid == tid) {
-			machine->last_match = thread__get(th);
+			machine->last_match = th;
 			machine__update_thread_pid(machine, th, pid);
 			return th;
 		}

@@ -409,7 +409,7 @@ static struct thread *____machine__findnew_thread(struct machine *machine,
 		 * It is now in the rbtree, get a ref
 		 */
 		thread__get(th);
-		machine->last_match = thread__get(th);
+		machine->last_match = th;
 	}
 
 	return th;

@@ -1309,7 +1309,7 @@ out_problem:
 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
 {
 	if (machine->last_match == th)
-		thread__zput(machine->last_match);
+		machine->last_match = NULL;
 
 	BUG_ON(th->refcnt.counter == 0);
 	if (lock)

2.1.0

From: Arnaldo Carvalho de Melo <[email protected]>

When introducing reference counting for struct thread instances I forgot to remove the synthetic threads from the machine's rbtree so that it then the threads would have just one reference and thus the thread__put() replacing the thread__delete() really turns into a thread__delete() (thread->refcnt == 1 at thread__put() time) and thus drop the thread->mg refcount, as expected by the this test.

Fix it by calling machine__remove_thread() (the counterpart of machine__findnew_thread()) on all the synthetic threads after the checks that involves the rbtree were done.

Before:

  # perf test -v mg
  30: Test thread mg sharing                                 :
  --- start ---
  test child forked, pid 26995
  FAILED tests/thread-mg-share.c:68 wrong refcnt (4 != 3)
  test child finished with -1
  ---- end ----
  Test thread mg sharing: FAILED!
  #

After:

  # perf test mg
  30: Test thread mg sharing: Ok
  #

---
 tools/perf/tests/thread-mg-share.c | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

index 04ecceed4e5b..c0ed56f7efc6 100644
@@ -58,6 +58,18 @@ int test__thread_mg_share(void)
 	other_leader = machine__find_thread(machine, 4, 4);
 	TEST_ASSERT_VAL("failed to find other leader", other_leader);
 
+	/*
+	 * Ok, now that all the rbtree related operations were done,
+	 * lets remove all of them from there so that we can do the
+	 * refcounting tests.
+	 */
+	machine__remove_thread(machine, leader);
+	machine__remove_thread(machine, t1);
+	machine__remove_thread(machine, t2);
+	machine__remove_thread(machine, t3);
+	machine__remove_thread(machine, other);
+	machine__remove_thread(machine, other_leader);
+
 	other_mg = other->mg;
 	TEST_ASSERT_EQUAL("wrong refcnt", other_mg->refcnt, 2);
 

@@ -81,11 +93,6 @@ int test__thread_mg_share(void)
 
 	thread__put(other);
 
-	/*
-	 * Cannot call machine__delete_threads(machine) now,
-	 * because we've already released all the threads.
-	 */
-
 	machines__exit(&machines);
 	return 0;
 }

2.1.0

From: "Naveen N. Rao" <[email protected]>

perf probe currently errors out if there are any tail calls to probed functions:

[root@rhel71be]# perf probe do_fork
Failed to find probe point in any functions.
  Error: Failed to add events.

Fix this by teaching perf to ignore tail calls.

Without patch:

[root@rhel71be perf]# ./perf probe -v do_fork probe-definition(0): do_fork symbol:do_fork file:(null) line:0 offset:0 return:0 lazy:(null)
0 arguments
Looking at the vmlinux_path (7 entries long)
symsrc__init: build id mismatch for /boot/vmlinux.
Using /usr/lib/debug/lib/modules/3.10.0-201.el7.ppc64/vmlinux for symbols Open Debuginfo file:
/usr/lib/debug/lib/modules/3.10.0-201.el7.ppc64/vmlinux
Try to find probe point from debuginfo.
found inline addr: 0xc0000000000bb9b0
Probe point found: do_fork+0
found inline addr: 0xc0000000000bbe20
Probe point found: kernel_thread+48
found inline addr: 0xc0000000000bbe5c
Probe point found: sys_fork+28
found inline addr: 0xc0000000000bbfac
Probe point found: sys_vfork+44
found inline addr: 0xc0000000000bc27c
Failed to find probe point in any functions.
An error occurred in debuginfo analysis (-2).
Error: Failed to add events. Reason: No such file or directory (Code: -2)

With patch:

[root@rhel71be perf]# ./perf probe -v do_fork probe-definition(0): do_fork symbol:do_fork file:(null) line:0 offset:0 return:0 lazy:(null)
0 arguments
Looking at the vmlinux_path (7 entries long)
symsrc__init: build id mismatch for /boot/vmlinux.
Using /usr/lib/debug/lib/modules/3.10.0-201.el7.ppc64/vmlinux for symbols Open Debuginfo file:
/usr/lib/debug/lib/modules/3.10.0-201.el7.ppc64/vmlinux
Try to find probe point from debuginfo.
found inline addr: 0xc0000000000bb9b0
Probe point found: do_fork+0
found inline addr: 0xc0000000000bbe20
Probe point found: kernel_thread+48
found inline addr: 0xc0000000000bbe5c
Probe point found: sys_fork+28
found inline addr: 0xc0000000000bbfac
Probe point found: sys_vfork+44
found inline addr: 0xc0000000000bc27c
Ignoring tail call from SyS_clone
Found 4 probe_trace_events.
Opening /sys/kernel/debug/tracing/kprobe_events write=1
No kprobe blacklist support, ignored
Added new events:
Writing event: p:probe/do_fork _text+768432
Failed to write event: Invalid argument
Error: Failed to add events. Reason: Invalid argument (Code: -22)

[Ignore the error about failure to write event - this kernel is missing a patch to resolve _text properly]

The reason to ignore tail calls is that the address does not belong to any function frame. In the example above, the address in SyS_clone is 0xc0000000000bc27c, but looking at the debug-info:

<1><830081>: Abbrev Number: 133 (DW_TAG_subprogram)

    <830083>   DW_AT_external    : 1
    <830083>   DW_AT_name        : (indirect string, offset: 0x3cea3): SyS_clone
    <830087>   DW_AT_decl_file   : 7
    <830088>   DW_AT_decl_line   : 1689
    <83008a>   DW_AT_prototyped  : 1
    <83008a>   DW_AT_type        : <0x8110eb>
    <83008e>   DW_AT_low_pc      : 0xc0000000000bc270
    <830096>   DW_AT_high_pc     : 0xc
    <83009e>   DW_AT_frame_base  : 1 byte block: 9c 	(DW_OP_call_frame_cfa)
    <8300a0>   DW_AT_GNU_all_call_sites: 1
    <8300a0>   DW_AT_sibling     : <0x830178>
<snip>
 <3><830147>: Abbrev Number: 125 (DW_TAG_GNU_call_site)
    <830148>   DW_AT_low_pc      : 0xc0000000000bc27c
    <830150>   DW_AT_GNU_tail_call: 1
    <830150>   DW_AT_abstract_origin: <0x82e7e1>

The frame ends at 0xc0000000000bc27c. I suppose this is why this particular call is a "tail" call. FWIW, systemtap seems to ignore these as well and requires users to explicitly place probes at these call sites if necessary. I print out the caller so that users know.

---
 tools/perf/util/dwarf-aux.c    | 37 +++++++++++++++++++++++++++++++++++++
 tools/perf/util/dwarf-aux.h    |  4 ++++
 tools/perf/util/probe-finder.c | 12 +++++++++---
 3 files changed, 50 insertions(+), 3 deletions(-)

index afa09719336b..eb47abdcf0ac 100644
@@ -433,6 +433,43 @@ struct __addr_die_search_param {
 	Dwarf_Die	*die_mem;
 };
 
+static int __die_search_func_tail_cb(Dwarf_Die *fn_die, void *data)
+{
+	struct __addr_die_search_param *ad = data;
+	Dwarf_Addr addr = 0;
+
+	if (dwarf_tag(fn_die) == DW_TAG_subprogram &&
+	    !dwarf_highpc(fn_die, &addr) &&
+	    addr == ad->addr) {
+		memcpy(ad->die_mem, fn_die, sizeof(Dwarf_Die));
+		return DWARF_CB_ABORT;
+	}
+	return DWARF_CB_OK;
+}
+
+/**
+ * die_find_tailfunc - Search for a non-inlined function with tail call at
+ * given address
+ * @cu_die: a CU DIE which including @addr
+ * @addr: target address
+ * @die_mem: a buffer for result DIE
+ *
+ * Search for a non-inlined function DIE with tail call at @addr. Stores the
+ * DIE to @die_mem and returns it if found. Returns NULL if failed.
+ */
+Dwarf_Die *die_find_tailfunc(Dwarf_Die *cu_die, Dwarf_Addr addr,
+				    Dwarf_Die *die_mem)
+{
+	struct __addr_die_search_param ad;
+	ad.addr = addr;
+	ad.die_mem = die_mem;
+	/* dwarf_getscopes can't find subprogram. */
+	if (!dwarf_getfuncs(cu_die, __die_search_func_tail_cb, &ad, 0))
+		return NULL;
+	else
+		return die_mem;
+}
+
 /* die_find callback for non-inlined function search */
 static int __die_search_func_cb(Dwarf_Die *fn_die, void *data)
 {

index c154c0b80880..c42ec366f2a7 100644
@@ -85,6 +85,10 @@ extern Dwarf_Die *die_find_child(Dwarf_Die *rt_die,
 extern Dwarf_Die *die_find_realfunc(Dwarf_Die *cu_die, Dwarf_Addr addr,
 				    Dwarf_Die *die_mem);
 
+/* Search a non-inlined function with tail call at given address */
+Dwarf_Die *die_find_tailfunc(Dwarf_Die *cu_die, Dwarf_Addr addr,
+				    Dwarf_Die *die_mem);
+
 /* Search the top inlined function including given address */
 extern Dwarf_Die *die_find_top_inlinefunc(Dwarf_Die *sp_die, Dwarf_Addr addr,
 					  Dwarf_Die *die_mem);

index 590a24a7f555..c50da392e256 100644
@@ -674,9 +674,15 @@ static int call_probe_finder(Dwarf_Die *sc_die, struct probe_finder *pf)
 	/* If not a real subprogram, find a real one */
 	if (!die_is_func_def(sc_die)) {
 		if (!die_find_realfunc(&pf->cu_die, pf->addr, &pf->sp_die)) {
-			pr_warning("Failed to find probe point in any "
-				   "functions.\n");
-			return -ENOENT;
+			if (die_find_tailfunc(&pf->cu_die, pf->addr, &pf->sp_die)) {
+				pr_warning("Ignoring tail call from %s\n",
+						dwarf_diename(&pf->sp_die));
+				return 0;
+			} else {
+				pr_warning("Failed to find probe point in any "
+					   "functions.\n");
+				return -ENOENT;
+			}
 		}
 	} else
 		memcpy(&pf->sp_die, sc_die, sizeof(Dwarf_Die));

2.1.0

From: Namhyung Kim <[email protected]>

There's a bug that perf report sometimes ignore some options on --stdio output. This bug is triggered only if a related config variable is set. For example, let's assume we have a following config file.

  $ cat ~/.perfconfig
  [call-graph]
    print-type = graph
  [hist]
    percentage = absolute

Then, following perf config will not honor some options.

  $ perf record -ag sleep 1
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.199 MB perf.data (77 samples) ]

  $ perf report -g none --stdio
  # To display the perf.data header info, please use --header/--header-only options.
  #
  # Samples: 77  of event 'cycles'
  # Event count (approx.): 25425383
  #
  # Overhead  Command          Shared Object            Symbol
  # ........  ...............  .......................  ..............
  #
      16.34%  swapper          [kernel.vmlinux]         [k] intel_idle
                      |
                      ---intel_idle
                         cpuidle_enter_state
                         cpuidle_enter
                         cpu_startup_entry
   ...

With '-g none' option, it should not show callchains, but it still shows callchains. However it works as expected on --tui output.

Similarly, '--percentage relative' option is not work and still shows a absolute percentage values.

Looking at the source, I found that those setting were overwritten by config variables when setup_pager() called. The setup_pager() is to start a pager process so that it can manage long lines of output on the stdio mode. But as it calls the perf_config() after parsing arguments, the settings were overwritten regardless of command line options.

The reason it calls perf_config() is to find the 'pager_program' which might be set by a config variable, I guess. However current perf code does not provide the config variable for it, so it's just meaningless IMHO. Eliminating the call makes the option working as expected.

---
 tools/perf/util/cache.h       | 1 -
 tools/perf/util/environment.c | 1 -
 tools/perf/util/pager.c       | 5 -----
 3 files changed, 7 deletions(-)

index fbcca21d66ab..c861373aaed3 100644
@@ -30,7 +30,6 @@ extern const char *perf_config_dirname(const char *, const char *);
 
 /* pager.c */
 extern void setup_pager(void);
-extern const char *pager_program;
 extern int pager_in_use(void);
 extern int pager_use_color;
 

index 275b0ee345f5..7405123692f1 100644
@@ -5,5 +5,4 @@
  */
 #include "cache.h"
 
-const char *pager_program;
 int pager_use_color = 1;

index 31ee02d4e988..53ef006a951c 100644
@@ -50,11 +50,6 @@ void setup_pager(void)
 
 	if (!isatty(1))
 		return;
-	if (!pager) {
-		if (!pager_program)
-			perf_config(perf_default_config, NULL);
-		pager = pager_program;
-	}
 	if (!pager)
 		pager = getenv("PAGER");
 	if (!(pager || access("/usr/bin/pager", X_OK)))

2.1.0

From: Arnaldo Carvalho de Melo <[email protected]>

Where such macro is not present, so just copy its definition from glibc's endian.h and define it if not already.

---
 tools/lib/traceevent/plugin_cfg80211.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

index 4592d8438318..ec57d0c1fbc2 100644
@@ -4,6 +4,19 @@
 #include <endian.h>
 #include "event-parse.h"
 
+/*
+ * From glibc endian.h, for older systems where it is not present, e.g.: RHEL5,
+ * Fedora6.
+ */
+#ifndef le16toh
+# if __BYTE_ORDER == __LITTLE_ENDIAN
+#  define le16toh(x) (x)
+# else
+#  define le16toh(x) __bswap_16 (x)
+# endif
+#endif
+
+
 static unsigned long long
 process___le16_to_cpup(struct trace_seq *s, unsigned long long *args)
 {

2.1.0

From: Arnaldo Carvalho de Melo <[email protected]>

Such as RHEL5, where CLOEXEC, NONBLOCK flags are not present, use a ifdef+define approach instead to make it build on all distros.

---
 tools/perf/builtin-trace.c | 29 ++++++++++++++++++++++++++++-
 1 file changed, 28 insertions(+), 1 deletion(-)

index 96a2ebafc755..a05490d06374 100644
@@ -16,7 +16,6 @@
 
 #include <libaudit.h>
 #include <stdlib.h>
-#include <sys/eventfd.h>
 #include <sys/mman.h>
 #include <linux/futex.h>
 

@@ -41,6 +40,34 @@
 # define EFD_SEMAPHORE		1
 #endif
 
+#ifndef EFD_NONBLOCK
+# define EFD_NONBLOCK		00004000
+#endif
+
+#ifndef EFD_CLOEXEC
+# define EFD_CLOEXEC		02000000
+#endif
+
+#ifndef O_CLOEXEC
+# define O_CLOEXEC		02000000
+#endif
+
+#ifndef SOCK_DCCP
+# define SOCK_DCCP		6
+#endif
+
+#ifndef SOCK_CLOEXEC
+# define SOCK_CLOEXEC		02000000
+#endif
+
+#ifndef SOCK_NONBLOCK
+# define SOCK_NONBLOCK		00004000
+#endif
+
+#ifndef MSG_CMSG_CLOEXEC
+# define MSG_CMSG_CLOEXEC	0x40000000
+#endif
+
 struct tp_field {
 	int offset;
 	union {

2.1.0

From: Jiri Olsa <[email protected]>

---
 tools/perf/util/Build | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

index d552203aead0..6966d0743bf7 100644
@@ -102,19 +102,19 @@ CFLAGS_exec_cmd.o += -DPERF_EXEC_PATH="BUILD_STR($(perfexecdir_SQ))" -DPREFIX="B
 
 $(OUTPUT)util/parse-events-flex.c: util/parse-events.l $(OUTPUT)util/parse-events-bison.c
 	$(call rule_mkdir)
-	@$(call echo-cmd,flex)$(FLEX) -o $@ --header-file=$(OUTPUT)util/parse-events-flex.h $(PARSER_DEBUG_FLEX) util/parse-events.l
+	$(Q)$(call echo-cmd,flex)$(FLEX) -o $@ --header-file=$(OUTPUT)util/parse-events-flex.h $(PARSER_DEBUG_FLEX) util/parse-events.l
 
 $(OUTPUT)util/parse-events-bison.c: util/parse-events.y
 	$(call rule_mkdir)
-	@$(call echo-cmd,bison)$(BISON) -v util/parse-events.y -d $(PARSER_DEBUG_BISON) -o $@ -p parse_events_
+	$(Q)$(call echo-cmd,bison)$(BISON) -v util/parse-events.y -d $(PARSER_DEBUG_BISON) -o $@ -p parse_events_
 
 $(OUTPUT)util/pmu-flex.c: util/pmu.l $(OUTPUT)util/pmu-bison.c
 	$(call rule_mkdir)
-	@$(call echo-cmd,flex)$(FLEX) -o $@ --header-file=$(OUTPUT)util/pmu-flex.h util/pmu.l
+	$(Q)$(call echo-cmd,flex)$(FLEX) -o $@ --header-file=$(OUTPUT)util/pmu-flex.h util/pmu.l
 
 $(OUTPUT)util/pmu-bison.c: util/pmu.y
 	$(call rule_mkdir)
-	@$(call echo-cmd,bison)$(BISON) -v util/pmu.y -d -o $@ -p perf_pmu_
+	$(Q)$(call echo-cmd,bison)$(BISON) -v util/pmu.y -d -o $@ -p perf_pmu_
 
 CFLAGS_parse-events-flex.o  += -w
 CFLAGS_pmu-flex.o           += -w

2.1.0

Pulled, thanks a lot Arnaldo!

Ingo