linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH 0/2] perf record: mmap output file - v3
@ 2013-11-04 18:31 Jiri Olsa
  2013-11-04 18:31 ` [PATCH 1/2] perf tools: Add perf_data__write implementation into perf_data_file object Jiri Olsa
                   ` (2 more replies)
  0 siblings, 3 replies; 6+ messages in thread
From: Jiri Olsa @ 2013-11-04 18:31 UTC (permalink / raw)
  To: linux-kernel
  Cc: Jiri Olsa, Ingo Molnar, Frederic Weisbecker, Peter Zijlstra,
	Namhyung Kim, Mike Galbraith, Stephane Eranian, David Ahern,
	Adrian Hunter

hi,
after discussion with David, sending his change
updated into the new data file object.

David's original post:
  http://marc.info/?l=linux-kernel&m=138180575328977&w=2

Here's the change:
  git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
  perf/mmap1

thanks,
jirka

Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Stephane Eranian <eranian@google.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
---
Jiri Olsa (2):
      perf tools: Add perf_data__write implementation into perf_data_file object
      perf tools: Add perf_data_file__write mmap support

 tools/perf/builtin-record.c |  49 ++++++++++++++++++--------------------------
 tools/perf/util/data.c      | 116 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/data.h      |  23 ++++++++++++++-------
 3 files changed, 152 insertions(+), 36 deletions(-)

^ permalink raw reply	[flat|nested] 6+ messages in thread
* [PATCH 1/2] perf tools: Add perf_data__write implementation into perf_data_file object
  2013-11-04 18:31 [PATCH 0/2] perf record: mmap output file - v3 Jiri Olsa
@ 2013-11-04 18:31 ` Jiri Olsa
  2013-11-04 18:31 ` [PATCH 2/2] perf tools: Add perf_data_file__write mmap support Jiri Olsa
  2013-11-06 17:51 ` [PATCH 0/2] perf record: mmap output file - v3 David Ahern
  2 siblings, 0 replies; 6+ messages in thread
From: Jiri Olsa @ 2013-11-04 18:31 UTC (permalink / raw)
  To: linux-kernel
  Cc: Jiri Olsa, Ingo Molnar, Frederic Weisbecker, Peter Zijlstra,
	Namhyung Kim, Mike Galbraith, Stephane Eranian, David Ahern,
	Adrian Hunter

Adding perf_data__write implementation into perf_data_file
object. This interface is now used within record command
to store data.

Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Stephane Eranian <eranian@google.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
---
 tools/perf/builtin-record.c | 38 ++++++++++++++------------------------
 tools/perf/util/data.c      | 20 ++++++++++++++++++++
 tools/perf/util/data.h      | 15 ++++++++-------
 3 files changed, 42 insertions(+), 31 deletions(-)

diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index ab8d15e..5201677 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -82,24 +82,17 @@ static void advance_output(struct perf_record *rec, size_t size)
 	rec->bytes_written += size;
 }
 
-static int write_output(struct perf_record *rec, void *buf, size_t size)
+static ssize_t perf_record__write(struct perf_record *rec,
+				  void *buf, size_t size)
 {
-	struct perf_data_file *file = &rec->file;
-
-	while (size) {
-		int ret = write(file->fd, buf, size);
-
-		if (ret < 0) {
-			pr_err("failed to write perf data, error: %m\n");
-			return -1;
-		}
-
-		size -= ret;
-		buf += ret;
+	struct perf_session *session = rec->session;
+	ssize_t ret;
 
-		rec->bytes_written += ret;
-	}
+	ret = perf_data_file__write(session->file, buf, size);
+	if (ret < 0)
+		return -1;
 
+	rec->bytes_written += ret;
 	return 0;
 }
 
@@ -109,10 +102,7 @@ static int process_synthesized_event(struct perf_tool *tool,
 				     struct machine *machine __maybe_unused)
 {
 	struct perf_record *rec = container_of(tool, struct perf_record, tool);
-	if (write_output(rec, event, event->header.size) < 0)
-		return -1;
-
-	return 0;
+	return perf_record__write(rec, event, event->header.size);
 }
 
 static int perf_record__mmap_read(struct perf_record *rec,
@@ -137,7 +127,7 @@ static int perf_record__mmap_read(struct perf_record *rec,
 		size = md->mask + 1 - (old & md->mask);
 		old += size;
 
-		if (write_output(rec, buf, size) < 0) {
+		if (perf_record__write(rec, buf, size) < 0) {
 			rc = -1;
 			goto out;
 		}
@@ -147,7 +137,7 @@ static int perf_record__mmap_read(struct perf_record *rec,
 	size = head - old;
 	old += size;
 
-	if (write_output(rec, buf, size) < 0) {
+	if (perf_record__write(rec, buf, size) < 0) {
 		rc = -1;
 		goto out;
 	}
@@ -322,8 +312,8 @@ static struct perf_event_header finished_round_event = {
 
 static int perf_record__mmap_read_all(struct perf_record *rec)
 {
-	int i;
-	int rc = 0;
+	struct perf_session *session = rec->session;
+	int i, rc = 0;
 
 	for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 		if (rec->evlist->mmap[i].base) {
@@ -335,7 +325,7 @@ static int perf_record__mmap_read_all(struct perf_record *rec)
 	}
 
 	if (perf_header__has_feat(&rec->session->header, HEADER_TRACING_DATA))
-		rc = write_output(rec, &finished_round_event,
+		rc = perf_data_file__write(session->file, &finished_round_event,
 				  sizeof(finished_round_event));
 
 out:
diff --git a/tools/perf/util/data.c b/tools/perf/util/data.c
index 7d09faf..cce1256 100644
--- a/tools/perf/util/data.c
+++ b/tools/perf/util/data.c
@@ -118,3 +118,23 @@ void perf_data_file__close(struct perf_data_file *file)
 {
 	close(file->fd);
 }
+
+ssize_t perf_data_file__write(struct perf_data_file *file,
+			      void *buf, size_t size)
+{
+	ssize_t total = size;
+
+	while (size) {
+		ssize_t ret = write(file->fd, buf, size);
+
+		if (ret < 0) {
+			pr_err("failed to write perf data, error: %m\n");
+			return -1;
+		}
+
+		size -= ret;
+		buf  += ret;
+	}
+
+	return total;
+}
diff --git a/tools/perf/util/data.h b/tools/perf/util/data.h
index 8c2df80..02c53dc 100644
--- a/tools/perf/util/data.h
+++ b/tools/perf/util/data.h
@@ -9,12 +9,12 @@ enum perf_data_mode {
 };
 
 struct perf_data_file {
-	const char *path;
-	int fd;
-	bool is_pipe;
-	bool force;
-	unsigned long size;
-	enum perf_data_mode mode;
+	const char		*path;
+	int			 fd;
+	bool			 is_pipe;
+	bool			 force;
+	unsigned long		 size;
+	enum perf_data_mode	 mode;
 };
 
 static inline bool perf_data_file__is_read(struct perf_data_file *file)
@@ -44,5 +44,6 @@ static inline unsigned long perf_data_file__size(struct perf_data_file *file)
 
 int perf_data_file__open(struct perf_data_file *file);
 void perf_data_file__close(struct perf_data_file *file);
-
+ssize_t perf_data_file__write(struct perf_data_file *file,
+			      void *buf, size_t size);
 #endif /* __PERF_DATA_H */
-- 
1.7.11.7


^ permalink raw reply related	[flat|nested] 6+ messages in thread
* [PATCH 2/2] perf tools: Add perf_data_file__write mmap support
  2013-11-04 18:31 [PATCH 0/2] perf record: mmap output file - v3 Jiri Olsa
  2013-11-04 18:31 ` [PATCH 1/2] perf tools: Add perf_data__write implementation into perf_data_file object Jiri Olsa
@ 2013-11-04 18:31 ` Jiri Olsa
  2013-11-06 17:51 ` [PATCH 0/2] perf record: mmap output file - v3 David Ahern
  2 siblings, 0 replies; 6+ messages in thread
From: Jiri Olsa @ 2013-11-04 18:31 UTC (permalink / raw)
  To: linux-kernel
  Cc: Jiri Olsa, Ingo Molnar, Frederic Weisbecker, Peter Zijlstra,
	Namhyung Kim, Mike Galbraith, Stephane Eranian, David Ahern,
	Adrian Hunter

When recording raw_syscalls for the entire system, e.g.,
  perf record -e raw_syscalls:*,sched:sched_switch -a -- sleep 10

you end up with a negative feedback loop as perf itself calls
write() fairly often. This patch handles the problem by mmap'ing the
file in chunks of 64M at a time and copies events from the event buffers
to the file avoiding write system calls.

Before (with write syscall):
  # time ./perf.old record -e raw_syscalls:*,sched:sched_switch -a -- sleep 10
  [ perf record: Woken up 0 times to write data ]
  [ perf record: Captured and wrote 914.717 MB perf.data (~39964591 samples) ]

  real    0m11.390s
  user    0m2.029s
  sys     0m9.311s

After (using mmap):
  # time ./perf record -e raw_syscalls:*,sched:sched_switch -a -- sleep 10
  [ perf record: Woken up 74 times to write data ]
  [ perf record: Captured and wrote 19.231 MB perf.data (~840219 samples) ]

  real    0m10.182s
  user    0m0.067s
  sys     0m0.121s

In addition to perf-trace benefits using mmap lowers the overhead of
perf-record.

v3: moved David's code into perf_data_file object, also used
    most of his changelog

Original-patch-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Stephane Eranian <eranian@google.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
---
 tools/perf/builtin-record.c |  11 ++---
 tools/perf/util/data.c      | 100 +++++++++++++++++++++++++++++++++++++++++++-
 tools/perf/util/data.h      |   8 ++++
 3 files changed, 112 insertions(+), 7 deletions(-)

diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index 5201677..45722fc 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -240,12 +240,8 @@ out:
 
 static int process_buildids(struct perf_record *rec)
 {
-	struct perf_data_file *file  = &rec->file;
 	struct perf_session *session = rec->session;
-
-	u64 size = lseek(file->fd, 0, SEEK_CUR);
-	if (size == 0)
-		return 0;
+	u64 size = perf_data_file__size(&rec->file);
 
 	return __perf_session__process_events(session, rec->post_processing_offset,
 					      size - rec->post_processing_offset,
@@ -535,6 +531,11 @@ static int __cmd_record(struct perf_record *rec, int argc, const char **argv)
 	if (quiet || signr == SIGUSR1)
 		return 0;
 
+	if (perf_data_file__munmap(file)) {
+		pr_err("data file unmap failed\n");
+		goto out_delete_session;
+	}
+
 	fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 
 	/*
diff --git a/tools/perf/util/data.c b/tools/perf/util/data.c
index cce1256..af5d644 100644
--- a/tools/perf/util/data.c
+++ b/tools/perf/util/data.c
@@ -4,10 +4,13 @@
 #include <sys/stat.h>
 #include <unistd.h>
 #include <string.h>
+#include <sys/mman.h>
 
 #include "data.h"
 #include "util.h"
 
+#define MMAP_WRITE_SIZE   (64*1024*1024)
+
 static bool check_pipe(struct perf_data_file *file)
 {
 	struct stat st;
@@ -111,6 +114,9 @@ int perf_data_file__open(struct perf_data_file *file)
 	if (!file->path)
 		file->path = "perf.data";
 
+	if (!file->mmap_size)
+		file->mmap_size = MMAP_WRITE_SIZE;
+
 	return open_file(file);
 }
 
@@ -119,8 +125,70 @@ void perf_data_file__close(struct perf_data_file *file)
 	close(file->fd);
 }
 
-ssize_t perf_data_file__write(struct perf_data_file *file,
-			      void *buf, size_t size)
+static int do_mmap(struct perf_data_file *file, u64 offset)
+{
+	u64 mmap_size = file->mmap_size;
+
+	file->mmap_off  = offset % mmap_size;
+	file->mmap_foff = (offset / mmap_size) * mmap_size;
+
+	file->mmap_addr = mmap(NULL, mmap_size,
+			       PROT_WRITE | PROT_READ,
+			       MAP_SHARED,
+			       file->fd,
+			       file->mmap_foff);
+
+	if (file->mmap_addr == MAP_FAILED) {
+		pr_err("mmap failed: %d: %s\n", errno, strerror(errno));
+		return -1;
+	}
+
+	/* Expand file to include this mmap segment. */
+	if (ftruncate(file->fd, file->mmap_foff + file->mmap_size) != 0) {
+		pr_err("ftruncate failed: %d: %s\n", errno, strerror(errno));
+		return -1;
+	}
+
+	return 0;
+}
+
+static ssize_t write_mmap(struct perf_data_file *file,
+			  void *buf, size_t size)
+{
+	ssize_t total = size;
+
+	if (!file->mmap_addr) {
+		off_t offset = lseek(file->fd, 0, SEEK_CUR);
+		if (offset < 0)
+			return -1;
+
+		if (do_mmap(file, offset))
+			return -1;
+	}
+
+	while (size) {
+		u64 remain = file->mmap_size - file->mmap_off;
+
+		if (size > remain) {
+			memcpy(file->mmap_addr + file->mmap_off, buf, remain);
+			size -= remain;
+			buf  += remain;
+
+			munmap(file->mmap_addr, file->mmap_size);
+			if (do_mmap(file, file->mmap_foff + file->mmap_size))
+				return -1;
+		} else {
+			memcpy(file->mmap_addr + file->mmap_off, buf, size);
+			file->mmap_off += size;
+			size = 0;
+		}
+	}
+
+	return total;
+}
+
+static ssize_t write_raw(struct perf_data_file *file,
+			 void *buf, size_t size)
 {
 	ssize_t total = size;
 
@@ -138,3 +206,31 @@ ssize_t perf_data_file__write(struct perf_data_file *file,
 
 	return total;
 }
+
+ssize_t perf_data_file__write(struct perf_data_file *file,
+			      void *buf, size_t size)
+{
+	return file->is_pipe ? write_raw(file, buf, size) :
+			       write_mmap(file, buf, size);
+}
+
+int perf_data_file__munmap(struct perf_data_file *file)
+{
+	if (file->mmap_addr) {
+		int ret;
+
+		munmap(file->mmap_addr, file->mmap_size);
+
+		file->mmap_addr = NULL;
+		file->size = file->mmap_foff + file->mmap_off;
+
+		ret = ftruncate(file->fd, file->size);
+		if (ret)
+			pr_err("ftruncate failed: %d: %s\n", errno,
+			       strerror(errno));
+
+		return ret;
+	}
+
+	return 0;
+}
diff --git a/tools/perf/util/data.h b/tools/perf/util/data.h
index 02c53dc..de59ee0 100644
--- a/tools/perf/util/data.h
+++ b/tools/perf/util/data.h
@@ -2,6 +2,7 @@
 #define __PERF_DATA_H
 
 #include <stdbool.h>
+#include "types.h"
 
 enum perf_data_mode {
 	PERF_DATA_MODE_WRITE,
@@ -15,6 +16,12 @@ struct perf_data_file {
 	bool			 force;
 	unsigned long		 size;
 	enum perf_data_mode	 mode;
+
+	/* for MMAP based file writes */
+	void			*mmap_addr;
+	u64			 mmap_off;
+	u64			 mmap_foff;
+	u64			 mmap_size;
 };
 
 static inline bool perf_data_file__is_read(struct perf_data_file *file)
@@ -46,4 +53,5 @@ int perf_data_file__open(struct perf_data_file *file);
 void perf_data_file__close(struct perf_data_file *file);
 ssize_t perf_data_file__write(struct perf_data_file *file,
 			      void *buf, size_t size);
+int perf_data_file__munmap(struct perf_data_file *file);
 #endif /* __PERF_DATA_H */
-- 
1.7.11.7


^ permalink raw reply related	[flat|nested] 6+ messages in thread
* Re: [PATCH 0/2] perf record: mmap output file - v3
  2013-11-04 18:31 [PATCH 0/2] perf record: mmap output file - v3 Jiri Olsa
  2013-11-04 18:31 ` [PATCH 1/2] perf tools: Add perf_data__write implementation into perf_data_file object Jiri Olsa
  2013-11-04 18:31 ` [PATCH 2/2] perf tools: Add perf_data_file__write mmap support Jiri Olsa
@ 2013-11-06 17:51 ` David Ahern
  2013-11-07 10:06   ` Jiri Olsa
  2 siblings, 1 reply; 6+ messages in thread
From: David Ahern @ 2013-11-06 17:51 UTC (permalink / raw)
  To: Jiri Olsa, linux-kernel
  Cc: Ingo Molnar, Frederic Weisbecker, Peter Zijlstra, Namhyung Kim,
	Mike Galbraith, Stephane Eranian, Adrian Hunter

On 11/4/13, 11:31 AM, Jiri Olsa wrote:
> hi,
> after discussion with David, sending his change
> updated into the new data file object.
>
> David's original post:
>    http://marc.info/?l=linux-kernel&m=138180575328977&w=2
>
> Here's the change:
>    git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
>    perf/mmap1
>

Not having any luck with these patches. Basic use case:

perf trace record -a -- sleep 5
perf trace --summary

With your patches I get a parse error. With my latest versions 
everything works clean.

David


^ permalink raw reply	[flat|nested] 6+ messages in thread
* Re: [PATCH 0/2] perf record: mmap output file - v3
  2013-11-06 17:51 ` [PATCH 0/2] perf record: mmap output file - v3 David Ahern
@ 2013-11-07 10:06   ` Jiri Olsa
  2013-11-07 14:32     ` David Ahern
  0 siblings, 1 reply; 6+ messages in thread
From: Jiri Olsa @ 2013-11-07 10:06 UTC (permalink / raw)
  To: David Ahern
  Cc: linux-kernel, Ingo Molnar, Frederic Weisbecker, Peter Zijlstra,
	Namhyung Kim, Mike Galbraith, Stephane Eranian, Adrian Hunter

On Wed, Nov 06, 2013 at 10:51:52AM -0700, David Ahern wrote:
> On 11/4/13, 11:31 AM, Jiri Olsa wrote:
> >hi,
> >after discussion with David, sending his change
> >updated into the new data file object.
> >
> >David's original post:
> >   http://marc.info/?l=linux-kernel&m=138180575328977&w=2
> >
> >Here's the change:
> >   git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
> >   perf/mmap1
> >
> 
> Not having any luck with these patches. Basic use case:
> 
> perf trace record -a -- sleep 5
> perf trace --summary
> 
> With your patches I get a parse error. With my latest versions
> everything works clean.

hum, 
perf trace --summary seems stuck in a loop, but that's no
different from current acme's perf/core.

Anyway I'll postpone the change till your v3 goes in
and go from there.

jirka

^ permalink raw reply	[flat|nested] 6+ messages in thread
* Re: [PATCH 0/2] perf record: mmap output file - v3
  2013-11-07 10:06   ` Jiri Olsa
@ 2013-11-07 14:32     ` David Ahern
  0 siblings, 0 replies; 6+ messages in thread
From: David Ahern @ 2013-11-07 14:32 UTC (permalink / raw)
  To: Jiri Olsa
  Cc: linux-kernel, Ingo Molnar, Frederic Weisbecker, Peter Zijlstra,
	Namhyung Kim, Mike Galbraith, Stephane Eranian, Adrian Hunter

On 11/7/13, 3:06 AM, Jiri Olsa wrote:
> On Wed, Nov 06, 2013 at 10:51:52AM -0700, David Ahern wrote:
>> On 11/4/13, 11:31 AM, Jiri Olsa wrote:
>>> hi,
>>> after discussion with David, sending his change
>>> updated into the new data file object.
>>>
>>> David's original post:
>>>    http://marc.info/?l=linux-kernel&m=138180575328977&w=2
>>>
>>> Here's the change:
>>>    git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
>>>    perf/mmap1
>>>
>>
>> Not having any luck with these patches. Basic use case:
>>
>> perf trace record -a -- sleep 5
>> perf trace --summary
>>
>> With your patches I get a parse error. With my latest versions
>> everything works clean.
>
> hum,
> perf trace --summary seems stuck in a loop, but that's no
> different from current acme's perf/core.

D'oh. When writing the email I was trying to keep the command line small 
and in the process removed the -i /tmp/perf.data from the second 
command. Without it perf-trace runs in live mode; with it events are 
read from a file.

David


^ permalink raw reply	[flat|nested] 6+ messages in thread
end of thread, other threads:[~2013-11-07 14:32 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2013-11-04 18:31 [PATCH 0/2] perf record: mmap output file - v3 Jiri Olsa
2013-11-04 18:31 ` [PATCH 1/2] perf tools: Add perf_data__write implementation into perf_data_file object Jiri Olsa
2013-11-04 18:31 ` [PATCH 2/2] perf tools: Add perf_data_file__write mmap support Jiri Olsa
2013-11-06 17:51 ` [PATCH 0/2] perf record: mmap output file - v3 David Ahern
2013-11-07 10:06   ` Jiri Olsa
2013-11-07 14:32     ` David Ahern

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).