Skip to contentSkip to conversation

perf tools: Fix pyrf_evlist__read_on_cpu interface

PATCH Applied Jiri Olsa 3 participants 6 messages Aug 17, 2018 – Aug 23, 2018 perf events

lore
Share & tools
lore

Jiri Olsa

author reviewer
#0 lore raw

cover perf tools: Fix pyrf_evlist__read_on_cpu interface

hi,
Jaroslav reported errors from valgrind over perf python script, which led us to discover serious issue with python interface.

It affects tuned daemon, which depends on this interface, and makes it hang.

thanks,
jirka

---
Jiri Olsa (2):
perf tools: Store real cpu number in the struct perf_mmap perf python: Fix pyrf_evlist__read_on_cpu interface

 tools/perf/util/evlist.c |  2 +-
 tools/perf/util/mmap.c   |  4 +++-
 tools/perf/util/mmap.h   |  4 +++-
 tools/perf/util/python.c | 20 +++++++++++++++++++-
 4 files changed, 26 insertions(+), 4 deletions(-)
author reviewer
#1 lore raw

PATCH 1/2 perf tools: Store real cpu number in the struct perf_mmap

Storing the real cpu number in the struct perf_mmap, which be used by python interface that allows user to read particular memory map for given cpu.

---
 tools/perf/util/evlist.c | 2 +-
 tools/perf/util/mmap.c   | 4 +++-
 tools/perf/util/mmap.h   | 4 +++-
 3 files changed, 7 insertions(+), 3 deletions(-)
tools/perf/util/evlist.c +1 -1
index e7a4b31a84fb..26e1c02440f9 100644
@@ -803,7 +803,7 @@ static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 		if (*output == -1) {
 			*output = fd;
 
-			if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
+			if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu)  < 0)
 				return -1;
 		} else {
 			if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
tools/perf/util/mmap.c +3 -1
index fc832676a798..2eac6dee0b9c 100644
@@ -164,7 +164,8 @@ void perf_mmap__munmap(struct perf_mmap *map)
 	auxtrace_mmap__munmap(&map->auxtrace_mmap);
 }
 
-int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd)
+int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp,
+		    int fd, int cpu)
 {
 	/*
 	 * The last one will be done at perf_mmap__consume(), so that we
@@ -191,6 +192,7 @@ int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd)
 		return -1;
 	}
 	map->fd = fd;
+	map->cpu = cpu;
 
 	if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
 				&mp->auxtrace_mp, map->base, fd))
tools/perf/util/mmap.h +3 -1
index d82294db1295..cc4fc8cc33ca 100644
@@ -18,6 +18,7 @@ struct perf_mmap {
 	void		 *base;
 	int		 mask;
 	int		 fd;
+	int		 cpu;
 	refcount_t	 refcnt;
 	u64		 prev;
 	u64		 start;
@@ -60,7 +61,8 @@ struct mmap_params {
 	struct auxtrace_mmap_params auxtrace_mp;
 };
 
-int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd);
+int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp,
+		    int fd, int cpu);
 void perf_mmap__munmap(struct perf_mmap *map);
 
 void perf_mmap__get(struct perf_mmap *map);
Signature · 1 line
2.17.1
#2 lore raw

[tip:perf/urgent] perf mmap: Store real cpu number in 'struct perf_mmap'

perf mmap: Store real cpu number in 'struct perf_mmap'
Commit-ID:  31fb4c0d7b88f036edb96a6a3bd791289ea2f931
Gitweb:     https://git.kernel.org/tip/31fb4c0d7b88f036edb96a6a3bd791289ea2f931
Author:     Jiri Olsa <[email protected]>
AuthorDate: Fri, 17 Aug 2018 13:45:55 +0200
Committer:  Arnaldo Carvalho de Melo <[email protected]>
CommitDate: Mon, 20 Aug 2018 08:54:59 -0300

perf mmap: Store real cpu number in 'struct perf_mmap'

Store the real cpu number in 'struct perf_mmap', which will be used by python interface that allows user to read a particular memory map for given cpu.

---
 tools/perf/util/evlist.c | 2 +-
 tools/perf/util/mmap.c   | 3 ++-
 tools/perf/util/mmap.h   | 3 ++-
 3 files changed, 5 insertions(+), 3 deletions(-)
tools/perf/util/evlist.c +1 -1
index e7a4b31a84fb..be440df29615 100644
@@ -803,7 +803,7 @@ static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 		if (*output == -1) {
 			*output = fd;
 
-			if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
+			if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
 				return -1;
 		} else {
 			if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
tools/perf/util/mmap.c +2 -1
index fc832676a798..215f69f41672 100644
@@ -164,7 +164,7 @@ void perf_mmap__munmap(struct perf_mmap *map)
 	auxtrace_mmap__munmap(&map->auxtrace_mmap);
 }
 
-int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd)
+int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd, int cpu)
 {
 	/*
 	 * The last one will be done at perf_mmap__consume(), so that we
@@ -191,6 +191,7 @@ int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd)
 		return -1;
 	}
 	map->fd = fd;
+	map->cpu = cpu;
 
 	if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
 				&mp->auxtrace_mp, map->base, fd))
tools/perf/util/mmap.h +2 -1
index d82294db1295..05a6d47c7956 100644
@@ -18,6 +18,7 @@ struct perf_mmap {
 	void		 *base;
 	int		 mask;
 	int		 fd;
+	int		 cpu;
 	refcount_t	 refcnt;
 	u64		 prev;
 	u64		 start;
@@ -60,7 +61,7 @@ struct mmap_params {
 	struct auxtrace_mmap_params auxtrace_mp;
 };
 
-int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd);
+int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd, int cpu);
 void perf_mmap__munmap(struct perf_mmap *map);
 
 void perf_mmap__get(struct perf_mmap *map);
author reviewer
#3 lore raw

PATCH 2/2 perf python: Fix pyrf_evlist__read_on_cpu interface

In reply to Jiri Olsa

Jaroslav reported errors from valgrind over perf python script:

  # echo 0 > /sys/devices/system/cpu/cpu4/online
  # valgrind ./test.py
  ==7524== Memcheck, a memory error detector
  ...
  ==7524== Command: ./test.py
  ==7524==
  pid 7526 exited
  ==7524== Invalid read of size 8
  ==7524==    at 0xCC2C2B3: perf_mmap__read_forward (evlist.c:780)
  ==7524==    by 0xCC2A681: pyrf_evlist__read_on_cpu (python.c:959)
  ...
  ==7524==  Address 0x65c4868 is 16 bytes after a block of size 459,36..
  ==7524==    at 0x4C2B955: calloc (vg_replace_malloc.c:711)
  ==7524==    by 0xCC2F484: zalloc (util.h:35)
  ==7524==    by 0xCC2F484: perf_evlist__alloc_mmap (evlist.c:978)
  ...

The reason for this is in the python interface, that allows script to pass arbitrary cpu number, which is then used to access struct perf_evlist::mmap array. That's obviously wrong and works only when if all cpus are available and fails if some cpu is missing, like in the example above.

This patch makes the pyrf_evlist__read_on_cpu search the evlist's maps array for the proper map to access.

It's linear search at the moment. Based on the way how is the read_on_cpu used, I don't think we need to be fast in here. But we could add some hash in the middle to make it fast/er.

We don't allow python interface to set write_backward event attribute, so it's safe to check only evlist's mmaps.

---
 tools/perf/util/python.c | 20 +++++++++++++++++++-
 1 file changed, 19 insertions(+), 1 deletion(-)
tools/perf/util/python.c +19 -1
index f74fbb652a4f..ce501ba14b08 100644
@@ -11,6 +11,7 @@
 #include "cpumap.h"
 #include "print_binary.h"
 #include "thread_map.h"
+#include "mmap.h"
 
 #if PY_MAJOR_VERSION < 3
 #define _PyUnicode_FromString(arg) \
@@ -976,6 +977,20 @@ static PyObject *pyrf_evlist__add(struct pyrf_evlist *pevlist,
 	return Py_BuildValue("i", evlist->nr_entries);
 }
 
+static struct perf_mmap *get_md(struct perf_evlist *evlist, int cpu)
+{
+	int i;
+
+	for (i = 0; i < evlist->nr_mmaps; i++) {
+		struct perf_mmap *md = &evlist->mmap[i];
+
+		if (md->cpu == cpu)
+			return md;
+	}
+
+	return NULL;
+}
+
 static PyObject *pyrf_evlist__read_on_cpu(struct pyrf_evlist *pevlist,
 					  PyObject *args, PyObject *kwargs)
 {
@@ -990,7 +1005,10 @@ static PyObject *pyrf_evlist__read_on_cpu(struct pyrf_evlist *pevlist,
 					 &cpu, &sample_id_all))
 		return NULL;
 
-	md = &evlist->mmap[cpu];
+	md = get_md(evlist, cpu);
+	if (!md)
+		return NULL;
+
 	if (perf_mmap__read_init(md) < 0)
 		goto end;
 
Signature · 1 line
2.17.1
#4 lore raw

[tip:perf/urgent] perf python: Fix pyrf_evlist__read_on_cpu() interface

perf python: Fix pyrf_evlist__read_on_cpu() interface
Commit-ID:  721f0dfc3ce821c6a32820ab63edfb48ed4af075
Gitweb:     https://git.kernel.org/tip/721f0dfc3ce821c6a32820ab63edfb48ed4af075
Author:     Jiri Olsa <[email protected]>
AuthorDate: Fri, 17 Aug 2018 13:45:56 +0200
Committer:  Arnaldo Carvalho de Melo <[email protected]>
CommitDate: Mon, 20 Aug 2018 08:54:59 -0300

perf python: Fix pyrf_evlist__read_on_cpu() interface

Jaroslav reported errors from valgrind over perf python script:

  # echo 0 > /sys/devices/system/cpu/cpu4/online
  # valgrind ./test.py
  ==7524== Memcheck, a memory error detector
  ...
  ==7524== Command: ./test.py
  ==7524==
  pid 7526 exited
  ==7524== Invalid read of size 8
  ==7524==    at 0xCC2C2B3: perf_mmap__read_forward (evlist.c:780)
  ==7524==    by 0xCC2A681: pyrf_evlist__read_on_cpu (python.c:959)
  ...
  ==7524==  Address 0x65c4868 is 16 bytes after a block of size 459,36..
  ==7524==    at 0x4C2B955: calloc (vg_replace_malloc.c:711)
  ==7524==    by 0xCC2F484: zalloc (util.h:35)
  ==7524==    by 0xCC2F484: perf_evlist__alloc_mmap (evlist.c:978)
  ...

The reason for this is in the python interface, that allows a script to pass arbitrary cpu number, which is then used to access struct perf_evlist::mmap array. That's obviously wrong and works only when if all cpus are available and fails if some cpu is missing, like in the example above.

This patch makes pyrf_evlist__read_on_cpu() search the evlist's maps array for the proper map to access.

It's linear search at the moment. Based on the way how is the read_on_cpu used, I don't think we need to be fast in here. But we could add some hash in the middle to make it fast/er.

We don't allow python interface to set write_backward event attribute, so it's safe to check only evlist's mmaps.

---
 tools/perf/util/python.c | 20 +++++++++++++++++++-
 1 file changed, 19 insertions(+), 1 deletion(-)
tools/perf/util/python.c +19 -1
index f74fbb652a4f..ce501ba14b08 100644
@@ -11,6 +11,7 @@
 #include "cpumap.h"
 #include "print_binary.h"
 #include "thread_map.h"
+#include "mmap.h"
 
 #if PY_MAJOR_VERSION < 3
 #define _PyUnicode_FromString(arg) \
@@ -976,6 +977,20 @@ static PyObject *pyrf_evlist__add(struct pyrf_evlist *pevlist,
 	return Py_BuildValue("i", evlist->nr_entries);
 }
 
+static struct perf_mmap *get_md(struct perf_evlist *evlist, int cpu)
+{
+	int i;
+
+	for (i = 0; i < evlist->nr_mmaps; i++) {
+		struct perf_mmap *md = &evlist->mmap[i];
+
+		if (md->cpu == cpu)
+			return md;
+	}
+
+	return NULL;
+}
+
 static PyObject *pyrf_evlist__read_on_cpu(struct pyrf_evlist *pevlist,
 					  PyObject *args, PyObject *kwargs)
 {
@@ -990,7 +1005,10 @@ static PyObject *pyrf_evlist__read_on_cpu(struct pyrf_evlist *pevlist,
 					 &cpu, &sample_id_all))
 		return NULL;
 
-	md = &evlist->mmap[cpu];
+	md = get_md(evlist, cpu);
+	if (!md)
+		return NULL;
+
 	if (perf_mmap__read_init(md) < 0)
 		goto end;
 
maintainer Applied
#5 lore raw
In reply to Jiri Olsa

Em Fri, Aug 17, 2018 at 01:45:54PM +0200, Jiri Olsa escreveu:

Quoted 6 lines

hi,
Jaroslav reported errors from valgrind over perf python script, which led us to discover serious issue with python interface.

It affects tuned daemon, which depends on this interface, and makes it hang.

Thanks, applied.

- Arnaldo