Skip to contentSkip to conversation

perf tools: fix buffer allocation

PATCH Applied Eric Dumazet 3 participants 10 messages Sep 24, 2009 – Sep 26, 2009 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf tools: fix buffer allocation

Hi Ingo

Here is a patch for perf.

BTW, use of openat() is a nuisance, since many machines have old glibc (RHEL 4 Update 5 -> glibc-2.3.4-2.36 on my dev machine for example)

Thanks

[PATCH] perf tools: fix buffer allocation

"perf top" cores dump on my dev machine, if run from a directory where vmlinux is present.

*** glibc detected *** malloc(): memory corruption: 0x085670d0 ***

---
 tools/perf/util/module.c |    2 +-
 1 files changed, 1 insertion(+), 1 deletion(-)
tools/perf/util/module.c +1 -1
index 8f81622..0d8c85d 100644
@@ -423,7 +423,7 @@ static int mod_dso__load_module_paths(struct mod_dso *self)
 	len += strlen(uts.release);
 	len += strlen("/modules.dep");
 
-	dpath = calloc(1, len);
+	dpath = calloc(1, len + 1);
 	if (dpath == NULL)
 		return err;
 
Applied
#1 lore raw
Quoted 3 lines · Eric Dumazet

Hi Ingo

Here is a patch for perf.

Applied, thanks Eric!

Quoted 2 lines

BTW, use of openat() is a nuisance, since many machines have old glibc (RHEL 4 Update 5 -> glibc-2.3.4-2.36 on my dev machine for example)

We can certainly remove that reliance - wanna send a patch for it?

Ingo

#2 lore raw
Quoted 4 lines · Ingo Molnar
Quoted 2 lines · Eric Dumazet

Hi Ingo

Quoted 8 lines
Quoted 1 line

Here is a patch for perf.

Applied, thanks Eric!

Quoted 2 lines

BTW, use of openat() is a nuisance, since many machines have old glibc (RHEL 4 Update 5 -> glibc-2.3.4-2.36 on my dev machine for example)

We can certainly remove that reliance - wanna send a patch for it?

Yes I'll do it, thanks.

#9 lore raw

[tip:perf/urgent] perf tools: Fix buffer allocation

In reply to Eric Dumazet
Commit-ID:  a255a9981a8566a1efabec983b7811e937e662d2
Gitweb:     http://git.kernel.org/tip/a255a9981a8566a1efabec983b7811e937e662d2
Author:     Eric Dumazet <[email protected]>
AuthorDate: Thu, 24 Sep 2009 15:05:59 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Thu, 24 Sep 2009 15:11:23 +0200

perf tools: Fix buffer allocation

"perf top" cores dump on my dev machine, if run from a directory where vmlinux is present:

*** glibc detected *** malloc(): memory corruption: 0x085670d0 ***

LKML-Reference: <[email protected]>

---
 tools/perf/util/module.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
tools/perf/util/module.c +1 -1
index 8f81622..0d8c85d 100644
@@ -423,7 +423,7 @@ static int mod_dso__load_module_paths(struct mod_dso *self)
 	len += strlen(uts.release);
 	len += strlen("/modules.dep");
 
-	dpath = calloc(1, len);
+	dpath = calloc(1, len + 1);
 	if (dpath == NULL)
 		return err;
 
#3 lore raw

PATCH 1/1 perf tools: Dont use openat()

In reply to Ingo Molnar
Quoted 4 lines · Ingo Molnar
Quoted 2 lines · Eric Dumazet

Hi Ingo

Quoted 8 lines
Quoted 1 line

Here is a patch for perf.

Applied, thanks Eric!

Quoted 2 lines

BTW, use of openat() is a nuisance, since many machines have old glibc (RHEL 4 Update 5 -> glibc-2.3.4-2.36 on my dev machine for example)

We can certainly remove that reliance - wanna send a patch for it?

Here it is

Thanks

[PATCH] perf tools: Dont use openat()

openat() is still a young glibc facility, better to not use it in a non performance critical program (perf list)

---
 tools/perf/util/parse-events.c |   49 ++++++++++++-------------------
 1 files changed, 20 insertions(+), 29 deletions(-)
tools/perf/util/parse-events.c +20 -29
index 13ab4b8..865208d 100644
@@ -165,33 +165,31 @@ struct tracepoint_path *tracepoint_id_to_path(u64 config)
 	DIR *sys_dir, *evt_dir;
 	struct dirent *sys_next, *evt_next, sys_dirent, evt_dirent;
 	char id_buf[4];
-	int sys_dir_fd, fd;
+	int fd;
 	u64 id;
 	char evt_path[MAXPATHLEN];
+	char dir_path[MAXPATHLEN];
 
 	if (valid_debugfs_mount(debugfs_path))
 		return NULL;
 
 	sys_dir = opendir(debugfs_path);
 	if (!sys_dir)
-		goto cleanup;
-	sys_dir_fd = dirfd(sys_dir);
+		return NULL;
 
 	for_each_subsystem(sys_dir, sys_dirent, sys_next) {
-		int dfd = openat(sys_dir_fd, sys_dirent.d_name,
-				 O_RDONLY|O_DIRECTORY), evt_dir_fd;
-		if (dfd == -1)
-			continue;
-		evt_dir = fdopendir(dfd);
-		if (!evt_dir) {
-			close(dfd);
+
+		snprintf(dir_path, MAXPATHLEN, "%s/%s", debugfs_path,
+			 sys_dirent.d_name);
+		evt_dir = opendir(dir_path);
+		if (!evt_dir)
 			continue;
-		}
-		evt_dir_fd = dirfd(evt_dir);
+
 		for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next) {
-			snprintf(evt_path, MAXPATHLEN, "%s/id",
+			
+			snprintf(evt_path, MAXPATHLEN, "%s/%s/id", dir_path,
 				 evt_dirent.d_name);
-			fd = openat(evt_dir_fd, evt_path, O_RDONLY);
+			fd = open(evt_path, O_RDONLY);
 			if (fd < 0)
 				continue;
 			if (read(fd, id_buf, sizeof(id_buf)) < 0) {
@@ -225,7 +223,6 @@ struct tracepoint_path *tracepoint_id_to_path(u64 config)
 		closedir(evt_dir);
 	}
 
-cleanup:
 	closedir(sys_dir);
 	return NULL;
 }
@@ -761,28 +758,24 @@ static void print_tracepoint_events(void)
 {
 	DIR *sys_dir, *evt_dir;
 	struct dirent *sys_next, *evt_next, sys_dirent, evt_dirent;
-	int sys_dir_fd;
 	char evt_path[MAXPATHLEN];
+	char dir_path[MAXPATHLEN];
 
 	if (valid_debugfs_mount(debugfs_path))
 		return;
 
 	sys_dir = opendir(debugfs_path);
 	if (!sys_dir)
-		goto cleanup;
-	sys_dir_fd = dirfd(sys_dir);
+		return;
 
 	for_each_subsystem(sys_dir, sys_dirent, sys_next) {
-		int dfd = openat(sys_dir_fd, sys_dirent.d_name,
-				 O_RDONLY|O_DIRECTORY), evt_dir_fd;
-		if (dfd == -1)
-			continue;
-		evt_dir = fdopendir(dfd);
-		if (!evt_dir) {
-			close(dfd);
+
+		snprintf(dir_path, MAXPATHLEN, "%s/%s", debugfs_path,
+			 sys_dirent.d_name);
+		evt_dir = opendir(dir_path);
+		if (!evt_dir)
 			continue;
-		}
-		evt_dir_fd = dirfd(evt_dir);
+
 		for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next) {
 			snprintf(evt_path, MAXPATHLEN, "%s:%s",
 				 sys_dirent.d_name, evt_dirent.d_name);
@@ -791,8 +784,6 @@ static void print_tracepoint_events(void)
 		}
 		closedir(evt_dir);
 	}
-
-cleanup:
 	closedir(sys_dir);
 }
 
Applied
#4 lore raw

Re: [PATCH] perf tools: Dont use openat()

Quoted 20 lines · Eric Dumazet

Ingo Molnar a ?crit :

Quoted 12 lines · Eric Dumazet
Quoted 4 lines

Hi Ingo

Here is a patch for perf.

Applied, thanks Eric!

Quoted 2 lines

BTW, use of openat() is a nuisance, since many machines have old glibc (RHEL 4 Update 5 -> glibc-2.3.4-2.36 on my dev machine for example)

We can certainly remove that reliance - wanna send a patch for it?

Here it is

Thanks

[PATCH] perf tools: Dont use openat()

Quoted 7 lines

openat() is still a young glibc facility, better to not use it in a non performance critical program (perf list)

---
 tools/perf/util/parse-events.c |   49 ++++++++++++-------------------
 1 files changed, 20 insertions(+), 29 deletions(-)

Applied, thanks Eric! I think it's also significant enough to be pushed to 2.6.32.

Ingo

#5 lore raw

[tip:perf/urgent] perf tools: Dont use openat()

In reply to Eric Dumazet
Commit-ID:  725b13685c61168f71825b3fb67d96d2d7aa3b0f
Gitweb:     http://git.kernel.org/tip/725b13685c61168f71825b3fb67d96d2d7aa3b0f
Author:     Eric Dumazet <[email protected]>
AuthorDate: Thu, 24 Sep 2009 15:39:09 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Thu, 24 Sep 2009 21:20:08 +0200

perf tools: Dont use openat()

openat() is still a young glibc facility, better to not use it in a non performance critical program (perf list)

Many machines have older glibc (RHEL 4 Update 5 -> glibc-2.3.4-2.36 on my dev machine for example).

Cc Peter Zijlstra <[email protected]>
Cc Ulrich Drepper <[email protected]>

LKML-Reference: <[email protected]>

---
 tools/perf/util/parse-events.c |   49 ++++++++++++++++-----------------------
 1 files changed, 20 insertions(+), 29 deletions(-)
tools/perf/util/parse-events.c +20 -29
index 13ab4b8..87c424d 100644
@@ -165,33 +165,31 @@ struct tracepoint_path *tracepoint_id_to_path(u64 config)
 	DIR *sys_dir, *evt_dir;
 	struct dirent *sys_next, *evt_next, sys_dirent, evt_dirent;
 	char id_buf[4];
-	int sys_dir_fd, fd;
+	int fd;
 	u64 id;
 	char evt_path[MAXPATHLEN];
+	char dir_path[MAXPATHLEN];
 
 	if (valid_debugfs_mount(debugfs_path))
 		return NULL;
 
 	sys_dir = opendir(debugfs_path);
 	if (!sys_dir)
-		goto cleanup;
-	sys_dir_fd = dirfd(sys_dir);
+		return NULL;
 
 	for_each_subsystem(sys_dir, sys_dirent, sys_next) {
-		int dfd = openat(sys_dir_fd, sys_dirent.d_name,
-				 O_RDONLY|O_DIRECTORY), evt_dir_fd;
-		if (dfd == -1)
-			continue;
-		evt_dir = fdopendir(dfd);
-		if (!evt_dir) {
-			close(dfd);
+
+		snprintf(dir_path, MAXPATHLEN, "%s/%s", debugfs_path,
+			 sys_dirent.d_name);
+		evt_dir = opendir(dir_path);
+		if (!evt_dir)
 			continue;
-		}
-		evt_dir_fd = dirfd(evt_dir);
+
 		for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next) {
-			snprintf(evt_path, MAXPATHLEN, "%s/id",
+
+			snprintf(evt_path, MAXPATHLEN, "%s/%s/id", dir_path,
 				 evt_dirent.d_name);
-			fd = openat(evt_dir_fd, evt_path, O_RDONLY);
+			fd = open(evt_path, O_RDONLY);
 			if (fd < 0)
 				continue;
 			if (read(fd, id_buf, sizeof(id_buf)) < 0) {
@@ -225,7 +223,6 @@ struct tracepoint_path *tracepoint_id_to_path(u64 config)
 		closedir(evt_dir);
 	}
 
-cleanup:
 	closedir(sys_dir);
 	return NULL;
 }
@@ -761,28 +758,24 @@ static void print_tracepoint_events(void)
 {
 	DIR *sys_dir, *evt_dir;
 	struct dirent *sys_next, *evt_next, sys_dirent, evt_dirent;
-	int sys_dir_fd;
 	char evt_path[MAXPATHLEN];
+	char dir_path[MAXPATHLEN];
 
 	if (valid_debugfs_mount(debugfs_path))
 		return;
 
 	sys_dir = opendir(debugfs_path);
 	if (!sys_dir)
-		goto cleanup;
-	sys_dir_fd = dirfd(sys_dir);
+		return;
 
 	for_each_subsystem(sys_dir, sys_dirent, sys_next) {
-		int dfd = openat(sys_dir_fd, sys_dirent.d_name,
-				 O_RDONLY|O_DIRECTORY), evt_dir_fd;
-		if (dfd == -1)
-			continue;
-		evt_dir = fdopendir(dfd);
-		if (!evt_dir) {
-			close(dfd);
+
+		snprintf(dir_path, MAXPATHLEN, "%s/%s", debugfs_path,
+			 sys_dirent.d_name);
+		evt_dir = opendir(dir_path);
+		if (!evt_dir)
 			continue;
-		}
-		evt_dir_fd = dirfd(evt_dir);
+
 		for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next) {
 			snprintf(evt_path, MAXPATHLEN, "%s:%s",
 				 sys_dirent.d_name, evt_dirent.d_name);
@@ -791,8 +784,6 @@ static void print_tracepoint_events(void)
 		}
 		closedir(evt_dir);
 	}
-
-cleanup:
 	closedir(sys_dir);
 }
 
#6 lore raw

Re: [PATCH] perf tools: Dont use openat()

In reply to Eric Dumazet

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 1 line · Eric Dumazet
Quoted 1 line

We can certainly remove that reliance - wanna send a patch for it?

Come on, the silliness has to stop. The kernel must be recent and to use it adequately the C library also must be recent. And "recent" is not even correct anymore: the functions are available for more then two years. Removing the use of the modern interfaces makes everything slower and might even re-introduce race conditions my patch fixed.

- --
➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖ -----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)
Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/

iEYEARECAAYFAkq733cACgkQ2ijCOnn/RHQHzQCgsmre/xVEVghMVRWDCIevMPwc +ecAmwaSOSLuSySZXRzSY/S64jf7K0Bd
=y6MN
-----END PGP SIGNATURE-----

#7 lore raw

Re: [PATCH] perf tools: Dont use openat()

Quoted 12 lines · Ulrich Drepper

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 1 line · Eric Dumazet
Quoted 1 line

We can certainly remove that reliance - wanna send a patch for it?

Come on, the silliness has to stop. The kernel must be recent and to use it adequately the C library also must be recent. And "recent" is not even correct anymore: the functions are available for more then two years. Removing the use of the modern interfaces makes everything slower and might even re-introduce race conditions my patch fixed.

First time I ear that C library *must* be recent. This was never mentionned in Documentation/Changes, "Minimal Requirements"

"perf list" can be 100x slower, and even racy, I dont mind. At all.

$ time perf list >/dev/null 2>/dev/null
real    0m0.001s
user    0m0.000s
sys     0m0.001s

With openat(), I cannot use "perf" on machines I can only change kernel, since changing glibc is too risky for legacy apps. I could use static and private glibc, but last time I tried this I lost few hours and failed.

If modern interfaces means : "Upgrade glibc or die, silly you..." I prefer to be silly, and keep increasing linux usability, even if I dont have the chance to have a modern lab with up2date distros.

BTW, oprofile compiles perfectly on RHEL4 and 'old' glibc.

Thanks

#8 lore raw

Re: [PATCH] perf tools: Dont use openat()

Quoted 17 lines · Eric Dumazet

Ulrich Drepper a ??crit :

Quoted 12 lines

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 1 line · Eric Dumazet
Quoted 1 line

We can certainly remove that reliance - wanna send a patch for it?

Come on, the silliness has to stop. The kernel must be recent and to use it adequately the C library also must be recent. And "recent" is not even correct anymore: the functions are available for more then two years. Removing the use of the modern interfaces makes everything slower and might even re-introduce race conditions my patch fixed.

First time I ear that C library *must* be recent. This was never mentionned in Documentation/Changes, "Minimal Requirements"

Quoted 7 lines

"perf list" can be 100x slower, and even racy, I dont mind. At all.

$ time perf list >/dev/null 2>/dev/null
real    0m0.001s
user    0m0.000s
sys     0m0.001s

Yep, doesnt look problematic from a performance POV.

Quoted 4 lines

With openat(), I cannot use "perf" on machines I can only change kernel, since changing glibc is too risky for legacy apps. I could use static and private glibc, but last time I tried this I lost few hours and failed.

Yeah, we generally dont want to force an upgrade of other components if possible. The fact that you ran into it in a pre -rc1 kernel means that there's a thousand others out there who will run into similar problems.

Ingo