Skip to contentSkip to conversation

Benchmark and improve event synthesis performance

PATCH Superseded Ian Rogers 2 participants 10 messages Apr 1, 2020 – Apr 2, 2020 perf events

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Ian Rogers

author reviewer
#0 lore raw

cover Benchmark and improve event synthesis performance

Event synthesis is performance critical in common tasks using perf. For example, when perf record starts in system wide mode the /proc file system is scanned with events synthesized for each process and all executable mmaps. With large machines and lots of processes, we have seen O(seconds) of wall clock time while synthesis is occurring.

This patch set adds a benchmark for synthesis performance in a new benchmark collection called 'internals'. The benchmark uses the machine__synthesize_threads function, single threaded on the perf process with a 'tool' that just drops the events, to measure how long synthesis takes.

By profiling this benchmark 2 performance bottlenecks were identified, hugetlbfs_mountpoint and stdio. The impact of theses changes are:

Before:
Average synthesis took: 167.616800 usec
Average data synthesis took: 208.655600 usec

After hugetlbfs_mountpoint scalability fix:
Average synthesis took: 120.195100 usec
Average data synthesis took: 156.582300 usec

After removal of stdio in /proc/pid/maps code:
Average synthesis took: 67.189100 usec
Average data synthesis took: 102.451600 usec

Time was measured on an Intel Xeon 6154 compiling with Debian gcc 9.2.1.

Two patches in the set were sent to LKML previously but are included here for context around the benchmark performance impact:
https://lore.kernel.org/lkml/[email protected]/T/#u
https://lore.kernel.org/lkml/[email protected]/T/#u

A future area of improvement could be to add the perf top num-thread-synthesize option more widely to other perf commands, and also to benchmark its effectiveness.

Ian Rogers (4):
perf bench: add event synthesis benchmark
perf synthetic-events: save 4kb from 2 stack frames tools api: add a lightweight buffered reading api perf synthetic events: Remove use of sscanf from /proc reading

Stephane Eranian (1):
tools api fs: make xxx__mountpoint() more scalable

 tools/lib/api/fs/fs.c              |  17 +++
 tools/lib/api/fs/fs.h              |  12 ++
 tools/lib/api/io.h                 | 103 +++++++++++++++++
 tools/perf/bench/Build             |   2 +-
 tools/perf/bench/bench.h           |   2 +-
 tools/perf/bench/synthesize.c      | 101 ++++++++++++++++
 tools/perf/builtin-bench.c         |   6 +
 tools/perf/util/synthetic-events.c | 177 +++++++++++++++++++----------
 8 files changed, 355 insertions(+), 65 deletions(-)
 create mode 100644 tools/lib/api/io.h
 create mode 100644 tools/perf/bench/synthesize.c
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog
author reviewer
#1 lore raw

PATCH 1/5 perf bench: add event synthesis benchmark

Event synthesis may occur at the start or end (tail) of a perf command. In system-wide mode it can scan every process in /proc, which may add seconds of latency before event recording. Add a new benchmark that times how long event synthesis takes with and without data synthesis.

---
 tools/perf/bench/Build        |   2 +-
 tools/perf/bench/bench.h      |   2 +-
 tools/perf/bench/synthesize.c | 101 ++++++++++++++++++++++++++++++++++
 tools/perf/builtin-bench.c    |   6 ++
 4 files changed, 109 insertions(+), 2 deletions(-)
 create mode 100644 tools/perf/bench/synthesize.c
tools/perf/bench/Build +1 -1
index e4e321b6f883..042827385c87 100644
@@ -6,9 +6,9 @@ perf-y += futex-wake.o
 perf-y += futex-wake-parallel.o
 perf-y += futex-requeue.o
 perf-y += futex-lock-pi.o
-
 perf-y += epoll-wait.o
 perf-y += epoll-ctl.o
+perf-y += synthesize.o
 
 perf-$(CONFIG_X86_64) += mem-memcpy-x86-64-lib.o
 perf-$(CONFIG_X86_64) += mem-memcpy-x86-64-asm.o
tools/perf/bench/bench.h +1 -1
index 4aa6de1aa67d..4d669c803237 100644
@@ -41,9 +41,9 @@ int bench_futex_wake_parallel(int argc, const char **argv);
 int bench_futex_requeue(int argc, const char **argv);
 /* pi futexes */
 int bench_futex_lock_pi(int argc, const char **argv);
-
 int bench_epoll_wait(int argc, const char **argv);
 int bench_epoll_ctl(int argc, const char **argv);
+int bench_synthesize(int argc, const char **argv);
 
 #define BENCH_FORMAT_DEFAULT_STR	"default"
 #define BENCH_FORMAT_DEFAULT		0
tools/perf/bench/synthesize.c new file +101 -0

101 lines not on this page · raw message

tools/perf/builtin-bench.c +6 -0
index c06fe21c8613..11c79a8d85d6 100644
@@ -76,6 +76,11 @@ static struct bench epoll_benchmarks[] = {
 };
 #endif // HAVE_EVENTFD
 
+static struct bench internals_benchmarks[] = {
+	{ "synthesize", "Benchmark perf event synthesis",	bench_synthesize	},
+	{ NULL,		NULL,					NULL			}
+};
+
 struct collection {
 	const char	*name;
 	const char	*summary;
@@ -92,6 +97,7 @@ static struct collection collections[] = {
 #ifdef HAVE_EVENTFD
 	{"epoll",       "Epoll stressing benchmarks",                   epoll_benchmarks        },
 #endif
+	{ "internals",	"Perf-internals benchmarks",			internals_benchmarks	},
 	{ "all",	"All benchmarks",				NULL			},
 	{ NULL,		NULL,						NULL			}
 };
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog
#2 lore raw

Re: [PATCH 1/5] perf bench: add event synthesis benchmark

Quoted 6 lines · Ian Rogers

Event synthesis may occur at the start or end (tail) of a perf command. In system-wide mode it can scan every process in /proc, which may add seconds of latency before event recording. Add a new benchmark that times how long event synthesis takes with and without data synthesis.

awesome, please put the command line example in changelog

jirka

author reviewer
#3 lore raw

PATCH 2/5 tools api fs: make xxx__mountpoint() more scalable

In reply to Ian Rogers

From: Stephane Eranian <[email protected]>

The xxx_mountpoint() interface provided by fs.c finds mount points for common pseudo filesystems. The first time xxx_mountpoint() is invoked, it scans the mount table (/proc/mounts) looking for a match. If found, it is cached. The price to scan /proc/mounts is paid once if the mount is found.

When the mount point is not found, subsequent calls to xxx_mountpoint() scan /proc/mounts over and over again. There is no caching.

This causes a scaling issue in perf record with hugeltbfs__mountpoint(). The function is called for each process found in synthesize__mmap_events(). If the machine has thousands of processes and if the /proc/mounts has many entries this could cause major overhead in perf record. We have observed multi-second slowdowns on some configurations.

As an example on a laptop:

Before:

$ sudo umount /dev/hugepages
$ strace -e trace=openat -o /tmp/tt perf record -a ls
$ fgrep mounts /tmp/tt
285

After:

$ sudo umount /dev/hugepages
$ strace -e trace=openat -o /tmp/tt perf record -a ls
$ fgrep mounts /tmp/tt
1

One could argue that the non-caching in case the moint point is not found is intentional. That way subsequent calls may discover a moint point if the sysadmin mounts the filesystem. But the same argument could be made against caching the mount point. It could be unmounted causing errors. It all depends on the intent of the interface. This patch assumes it is expected to scan /proc/mounts once. The patch documents the caching behavior in the fs.h header file.

An alternative would be to just fix perf record. But it would solve the problem with hugetlbs__mountpoint() but there could be similar issues (possibly down the line) with other xxx_mountpoint() calls in perf or other tools.

---
 tools/lib/api/fs/fs.c | 17 +++++++++++++++++
 tools/lib/api/fs/fs.h | 12 ++++++++++++
 2 files changed, 29 insertions(+)
tools/lib/api/fs/fs.c +17 -0
index 027b18f7ed8c..82f53d81a7a7 100644
@@ -90,6 +90,7 @@ struct fs {
 	const char * const	*mounts;
 	char			 path[PATH_MAX];
 	bool			 found;
+	bool			 checked;
 	long			 magic;
 };
 
@@ -111,31 +112,37 @@ static struct fs fs__entries[] = {
 		.name	= "sysfs",
 		.mounts	= sysfs__fs_known_mountpoints,
 		.magic	= SYSFS_MAGIC,
+		.checked = false,
 	},
 	[FS__PROCFS] = {
 		.name	= "proc",
 		.mounts	= procfs__known_mountpoints,
 		.magic	= PROC_SUPER_MAGIC,
+		.checked = false,
 	},
 	[FS__DEBUGFS] = {
 		.name	= "debugfs",
 		.mounts	= debugfs__known_mountpoints,
 		.magic	= DEBUGFS_MAGIC,
+		.checked = false,
 	},
 	[FS__TRACEFS] = {
 		.name	= "tracefs",
 		.mounts	= tracefs__known_mountpoints,
 		.magic	= TRACEFS_MAGIC,
+		.checked = false,
 	},
 	[FS__HUGETLBFS] = {
 		.name	= "hugetlbfs",
 		.mounts = hugetlbfs__known_mountpoints,
 		.magic	= HUGETLBFS_MAGIC,
+		.checked = false,
 	},
 	[FS__BPF_FS] = {
 		.name	= "bpf",
 		.mounts = bpf_fs__known_mountpoints,
 		.magic	= BPF_FS_MAGIC,
+		.checked = false,
 	},
 };
 
@@ -158,6 +165,7 @@ static bool fs__read_mounts(struct fs *fs)
 	}
 
 	fclose(fp);
+	fs->checked = true;
 	return fs->found = found;
 }
 
@@ -220,6 +228,7 @@ static bool fs__env_override(struct fs *fs)
 		return false;
 
 	fs->found = true;
+	fs->checked = true;
 	strncpy(fs->path, override_path, sizeof(fs->path) - 1);
 	fs->path[sizeof(fs->path) - 1] = '\0';
 	return true;
@@ -246,6 +255,14 @@ static const char *fs__mountpoint(int idx)
 	if (fs->found)
 		return (const char *)fs->path;
 
+	/* the mount point was already checked for the mount point
+	 * but and did not exist, so return NULL to avoid scanning again.
+	 * This makes the found and not found paths cost equivalent
+	 * in case of multiple calls.
+	 */
+	if (fs->checked)
+		return NULL;
+
 	return fs__get_mountpoint(fs);
 }
 
tools/lib/api/fs/fs.h +12 -0
index 936edb95e1f3..aa222ca30311 100644
@@ -18,6 +18,18 @@
 	const char *name##__mount(void);	\
 	bool name##__configured(void);		\
 
+/*
+ * The xxxx__mountpoint() entry points find the first match mount point for each
+ * filesystems listed below, where xxxx is the filesystem type.
+ *
+ * The interface is as follows:
+ *
+ * - If a mount point is found on first call, it is cached and used for all
+ *   subsequent calls.
+ *
+ * - If a mount point is not found, NULL is returned on first call and all
+ *   subsequent calls.
+ */
 FS(sysfs)
 FS(procfs)
 FS(debugfs)
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog
author reviewer
#4 lore raw

PATCH 3/5 perf synthetic-events: save 4kb from 2 stack frames

In reply to Ian Rogers

Reuse an existing char buffer to avoid two PATH_MAX sized char buffers. Reduces stack frame sizes by 4kb.

perf_event__synthesize_mmap_events before 'sub $0x45b8,%rsp' after 'sub $0x35b8,%rsp'.

perf_event__get_comm_ids before 'sub $0x2028,%rsp' after 'sub $0x1028,%rsp'.

The performance impact of this change is negligible.

---
 tools/perf/util/synthetic-events.c | 22 ++++++++++------------
 1 file changed, 10 insertions(+), 12 deletions(-)
tools/perf/util/synthetic-events.c +10 -12
index 3f28af39f9c6..1f3d8d4bb879 100644
@@ -70,7 +70,6 @@ int perf_tool__process_synth_event(struct perf_tool *tool,
 static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len,
 				    pid_t *tgid, pid_t *ppid)
 {
-	char filename[PATH_MAX];
 	char bf[4096];
 	int fd;
 	size_t size = 0;
@@ -80,11 +79,11 @@ static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len,
 	*tgid = -1;
 	*ppid = -1;
 
-	snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
+	snprintf(bf, sizeof(bf), "/proc/%d/status", pid);
 
-	fd = open(filename, O_RDONLY);
+	fd = open(bf, O_RDONLY);
 	if (fd < 0) {
-		pr_debug("couldn't open %s\n", filename);
+		pr_debug("couldn't open %s\n", bf);
 		return -1;
 	}
 
@@ -280,9 +279,9 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 				       struct machine *machine,
 				       bool mmap_data)
 {
-	char filename[PATH_MAX];
 	FILE *fp;
 	unsigned long long t;
+	char bf[BUFSIZ];
 	bool truncation = false;
 	unsigned long long timeout = proc_map_timeout * 1000000ULL;
 	int rc = 0;
@@ -292,15 +291,15 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 	if (machine__is_default_guest(machine))
 		return 0;
 
-	snprintf(filename, sizeof(filename), "%s/proc/%d/task/%d/maps",
-		 machine->root_dir, pid, pid);
+	snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps",
+		machine->root_dir, pid, pid);
 
-	fp = fopen(filename, "r");
+	fp = fopen(bf, "r");
 	if (fp == NULL) {
 		/*
 		 * We raced with a task exiting - just return:
 		 */
-		pr_debug("couldn't open %s\n", filename);
+		pr_debug("couldn't open %s\n", bf);
 		return -1;
 	}
 
@@ -308,7 +307,6 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 	t = rdclock();
 
 	while (1) {
-		char bf[BUFSIZ];
 		char prot[5];
 		char execname[PATH_MAX];
 		char anonstr[] = "//anon";
@@ -320,10 +318,10 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 			break;
 
 		if ((rdclock() - t) > timeout) {
-			pr_warning("Reading %s time out. "
+			pr_warning("Reading %s/proc/%d/task/%d/maps time out. "
 				   "You may want to increase "
 				   "the time limit by --proc-map-timeout\n",
-				   filename);
+				   machine->root_dir, pid, pid);
 			truncation = true;
 			goto out;
 		}
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog
author reviewer
#5 lore raw

PATCH 4/5 tools api: add a lightweight buffered reading api

In reply to Ian Rogers

The synthesize benchmark shows the majority of execution time going to fgets and sscanf, necessary to parse /proc/pid/maps. Add a new buffered reading library that will be used to replace these calls in a follow-up CL.

---
 tools/lib/api/io.h | 103 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 103 insertions(+)
 create mode 100644 tools/lib/api/io.h
tools/lib/api/io.h new file +103 -0
new file mode 100644
index 000000000000..e88d76c59440
@@ -0,0 +1,103 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Lightweight buffered reading library.
+ *
+ * Copyright 2019 Google LLC.
+ */
+
+struct io {
+	/* File descriptor being read/ */
+	int fd;
+	/* Size of the read buffer. */
+	unsigned int buf_len;
+	/* Pointer to storage for buffering read. */
+	char *buf;
+	/* End of the storage. */
+	char *end;
+	/* Currently accessed data pointer. */
+	char *data;
+	/* Set true on when the end of file on read error. */
+	bool eof;
+};
+
+static inline void init_io(struct io *io, int fd,
+			   char *buf, unsigned int buf_len)
+{
+	io->fd = fd;
+	io->buf_len = buf_len;
+	io->buf = buf;
+	io->end = buf;
+	io->data = buf;
+	io->eof = false;
+}
+
+/* Reads one character from the "io" file with similar semantics to fgetc. */
+static inline int get_char(struct io *io)
+{
+	char *ptr = io->data;
+
+	if (ptr == io->end) {
+		ssize_t n = read(io->fd, io->buf, io->buf_len);
+
+		if (n <= 0) {
+			io->eof = true;
+			return -1;
+		}
+		ptr = &io->buf[0];
+		io->end = &io->buf[n];
+	}
+	io->data = ptr + 1;
+	return *ptr;
+}
+
+/* Read a hexadecimal value with no 0x prefix into the out argument hex.
+ * Returns -1 on error or if nothing is read, otherwise returns the character
+ * after the hexadecimal value.
+ */
+static inline int get_hex(struct io *io, __u64 *hex)
+{
+	bool first_read = true;
+
+	*hex = 0;
+	while (true) {
+		char ch = get_char(io);
+
+		if (ch < 0)
+			return ch;
+		if (ch >= '0' && ch <= '9')
+			*hex = (*hex << 4) | (ch - '0');
+		else if (ch >= 'a' && ch <= 'f')
+			*hex = (*hex << 4) | (ch - 'a' + 10);
+		else if (ch >= 'A' && ch <= 'F')
+			*hex = (*hex << 4) | (ch - 'A' + 10);
+		else if (first_read)
+			return -1;
+		else
+			return ch;
+		first_read = false;
+	}
+}
+
+/* Read a decimal value into the out argument dec.
+ * Returns -1 on error or if nothing is read, otherwise returns the character
+ * after the decimal value.
+ */
+static inline int get_dec(struct io *io, __u64 *dec)
+{
+	bool first_read = true;
+
+	*dec = 0;
+	while (true) {
+		char ch = get_char(io);
+
+		if (ch < 0)
+			return ch;
+		if (ch >= '0' && ch <= '9')
+			*dec = (*dec * 10) + ch - '0';
+		else if (first_read)
+			return -1;
+		else
+			return ch;
+		first_read = false;
+	}
+}
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog
#6 lore raw

Re: [PATCH 4/5] tools api: add a lightweight buffered reading api

Quoted 17 lines · Ian Rogers

The synthesize benchmark shows the majority of execution time going to fgets and sscanf, necessary to parse /proc/pid/maps. Add a new buffered reading library that will be used to replace these calls in a follow-up CL.

---
 tools/lib/api/io.h | 103 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 103 insertions(+)
 create mode 100644 tools/lib/api/io.h
tools/lib/api/io.h new file +0 -0
new file mode 100644
index 000000000000..e88d76c59440

tools/lib/api/io.h line 6

Quoted 6 lines
tools/lib/api/io.h new file +6 -0
@@ -0,0 +1,103 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Lightweight buffered reading library.
+ *
+ * Copyright 2019 Google LLC.
+ */

missing the recursion header check

#ifndef __PERF_IO_H
#define __PERF_IO_H

Quoted 16 lines
tools/lib/api/io.h new file +16 -0
+
+struct io {
+	/* File descriptor being read/ */
+	int fd;
+	/* Size of the read buffer. */
+	unsigned int buf_len;
+	/* Pointer to storage for buffering read. */
+	char *buf;
+	/* End of the storage. */
+	char *end;
+	/* Currently accessed data pointer. */
+	char *data;
+	/* Set true on when the end of file on read error. */
+	bool eof;
+};
+

tools/lib/api/io.h line 32

Quoted 10 lines
tools/lib/api/io.h new file +10 -0
+static inline void init_io(struct io *io, int fd,
+			   char *buf, unsigned int buf_len)
+{
+	io->fd = fd;
+	io->buf_len = buf_len;
+	io->buf = buf;
+	io->end = buf;
+	io->data = buf;
+	io->eof = false;
+}

looks good, could you please add io__ prefix for all functions?

thanks,
jirka

Quoted 74 lines
tools/lib/api/io.h new file +71 -0
+
+/* Reads one character from the "io" file with similar semantics to fgetc. */
+static inline int get_char(struct io *io)
+{
+	char *ptr = io->data;
+
+	if (ptr == io->end) {
+		ssize_t n = read(io->fd, io->buf, io->buf_len);
+
+		if (n <= 0) {
+			io->eof = true;
+			return -1;
+		}
+		ptr = &io->buf[0];
+		io->end = &io->buf[n];
+	}
+	io->data = ptr + 1;
+	return *ptr;
+}
+
+/* Read a hexadecimal value with no 0x prefix into the out argument hex.
+ * Returns -1 on error or if nothing is read, otherwise returns the character
+ * after the hexadecimal value.
+ */
+static inline int get_hex(struct io *io, __u64 *hex)
+{
+	bool first_read = true;
+
+	*hex = 0;
+	while (true) {
+		char ch = get_char(io);
+
+		if (ch < 0)
+			return ch;
+		if (ch >= '0' && ch <= '9')
+			*hex = (*hex << 4) | (ch - '0');
+		else if (ch >= 'a' && ch <= 'f')
+			*hex = (*hex << 4) | (ch - 'a' + 10);
+		else if (ch >= 'A' && ch <= 'F')
+			*hex = (*hex << 4) | (ch - 'A' + 10);
+		else if (first_read)
+			return -1;
+		else
+			return ch;
+		first_read = false;
+	}
+}
+
+/* Read a decimal value into the out argument dec.
+ * Returns -1 on error or if nothing is read, otherwise returns the character
+ * after the decimal value.
+ */
+static inline int get_dec(struct io *io, __u64 *dec)
+{
+	bool first_read = true;
+
+	*dec = 0;
+	while (true) {
+		char ch = get_char(io);
+
+		if (ch < 0)
+			return ch;
+		if (ch >= '0' && ch <= '9')
+			*dec = (*dec * 10) + ch - '0';
+		else if (first_read)
+			return -1;
+		else
+			return ch;
+		first_read = false;
+	}
+}
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog
author reviewer
#7 lore raw

PATCH 5/5 perf synthetic events: Remove use of sscanf from /proc reading

In reply to Ian Rogers

The synthesize benchmark, run on a single process and thread, shows perf_event__synthesize_mmap_events as the hottest function with fgets and sscanf taking the majority of execution time. fscanf performs similarly well. Replace the scanf call with manual reading of each field of the /proc/pid/maps line, and remove some unnecessary buffering. This change also addresses potential, but unlikely, buffer overruns for the string values read by scanf.

Performance before is:
Average synthesis took: 120.195100 usec
Average data synthesis took: 156.582300 usec

And after is:
Average synthesis took: 67.189100 usec
Average data synthesis took: 102.451600 usec

On a Intel Xeon 6154 compiling with Debian gcc 9.2.1.

---
 tools/perf/util/synthetic-events.c | 157 +++++++++++++++++++----------
 1 file changed, 105 insertions(+), 52 deletions(-)
tools/perf/util/synthetic-events.c +105 -52
index 1f3d8d4bb879..64e010689a5f 100644
@@ -36,6 +36,7 @@
 #include <string.h>
 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */
 #include <api/fs/fs.h>
+#include <api/io.h>
 #include <sys/types.h>
 #include <sys/stat.h>
 #include <fcntl.h>
@@ -272,6 +273,79 @@ static int perf_event__synthesize_fork(struct perf_tool *tool,
 	return 0;
 }
 
+static bool read_proc_maps_line(struct io *io, __u64 *start, __u64 *end,
+				u32 *prot, u32 *flags, __u64 *offset,
+				u32 *maj, u32 *min,
+				__u64 *inode,
+				ssize_t pathname_size, char *pathname)
+{
+	__u64 temp;
+	int ch;
+	char *start_pathname = pathname;
+
+	if (get_hex(io, start) != '-')
+		return false;
+	if (get_hex(io, end) != ' ')
+		return false;
+
+	/* map protection and flags bits */
+	*prot = 0;
+	ch = get_char(io);
+	if (ch == 'r')
+		*prot |= PROT_READ;
+	else if (ch != '-')
+		return false;
+	ch = get_char(io);
+	if (ch == 'w')
+		*prot |= PROT_WRITE;
+	else if (ch != '-')
+		return false;
+	ch = get_char(io);
+	if (ch == 'x')
+		*prot |= PROT_EXEC;
+	else if (ch != '-')
+		return false;
+	ch = get_char(io);
+	if (ch == 's')
+		*flags = MAP_SHARED;
+	else if (ch == 'p')
+		*flags = MAP_PRIVATE;
+	else
+		return false;
+	if (get_char(io) != ' ')
+		return false;
+
+	if (get_hex(io, offset) != ' ')
+		return false;
+
+	if (get_hex(io, &temp) != ':')
+		return false;
+	*maj = temp;
+	if (get_hex(io, &temp) != ' ')
+		return false;
+	*min = temp;
+
+	ch = get_dec(io, inode);
+	if (ch != ' ') {
+		*pathname = '\0';
+		return ch == '\n';
+	}
+	do {
+		ch = get_char(io);
+	} while (ch == ' ');
+	while (true) {
+		if (ch < 0)
+			return false;
+		if (ch == '\0' || ch == '\n' ||
+		    (pathname + 1 - start_pathname) >= pathname_size) {
+			*pathname = '\0';
+			return true;
+		}
+		*pathname++ = ch;
+		ch = get_char(io);
+	}
+}
+
 int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 				       union perf_event *event,
 				       pid_t pid, pid_t tgid,
@@ -279,9 +353,9 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 				       struct machine *machine,
 				       bool mmap_data)
 {
-	FILE *fp;
 	unsigned long long t;
 	char bf[BUFSIZ];
+	struct io io;
 	bool truncation = false;
 	unsigned long long timeout = proc_map_timeout * 1000000ULL;
 	int rc = 0;
@@ -294,28 +368,39 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 	snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps",
 		machine->root_dir, pid, pid);
 
-	fp = fopen(bf, "r");
-	if (fp == NULL) {
+	io.fd = open(bf, O_RDONLY, 0);
+	if (io.fd < 0) {
 		/*
 		 * We raced with a task exiting - just return:
 		 */
 		pr_debug("couldn't open %s\n", bf);
 		return -1;
 	}
+	init_io(&io, io.fd, bf, sizeof(bf));
 
 	event->header.type = PERF_RECORD_MMAP2;
 	t = rdclock();
 
-	while (1) {
-		char prot[5];
-		char execname[PATH_MAX];
-		char anonstr[] = "//anon";
-		unsigned int ino;
+	while (!io.eof) {
+		static const char anonstr[] = "//anon";
 		size_t size;
-		ssize_t n;
 
-		if (fgets(bf, sizeof(bf), fp) == NULL)
-			break;
+		/* ensure null termination since stack will be reused. */
+		strcpy(event->mmap2.filename, "");
+
+		/* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
+		if (!read_proc_maps_line(&io,
+					&event->mmap2.start,
+					&event->mmap2.len,
+					&event->mmap2.prot,
+					&event->mmap2.flags,
+					&event->mmap2.pgoff,
+					&event->mmap2.maj,
+					&event->mmap2.min,
+					&event->mmap2.ino,
+					sizeof(event->mmap2.filename),
+					event->mmap2.filename))
+			continue;
 
 		if ((rdclock() - t) > timeout) {
 			pr_warning("Reading %s/proc/%d/task/%d/maps time out. "
@@ -326,23 +411,6 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 			goto out;
 		}
 
-		/* ensure null termination since stack will be reused. */
-		strcpy(execname, "");
-
-		/* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
-		n = sscanf(bf, "%"PRI_lx64"-%"PRI_lx64" %s %"PRI_lx64" %x:%x %u %[^\n]\n",
-		       &event->mmap2.start, &event->mmap2.len, prot,
-		       &event->mmap2.pgoff, &event->mmap2.maj,
-		       &event->mmap2.min,
-		       &ino, execname);
-
-		/*
- 		 * Anon maps don't have the execname.
- 		 */
-		if (n < 7)
-			continue;
-
-		event->mmap2.ino = (u64)ino;
 		event->mmap2.ino_generation = 0;
 
 		/*
@@ -353,23 +421,8 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 		else
 			event->header.misc = PERF_RECORD_MISC_GUEST_USER;
 
-		/* map protection and flags bits */
-		event->mmap2.prot = 0;
-		event->mmap2.flags = 0;
-		if (prot[0] == 'r')
-			event->mmap2.prot |= PROT_READ;
-		if (prot[1] == 'w')
-			event->mmap2.prot |= PROT_WRITE;
-		if (prot[2] == 'x')
-			event->mmap2.prot |= PROT_EXEC;
-
-		if (prot[3] == 's')
-			event->mmap2.flags |= MAP_SHARED;
-		else
-			event->mmap2.flags |= MAP_PRIVATE;
-
-		if (prot[2] != 'x') {
-			if (!mmap_data || prot[0] != 'r')
+		if ((event->mmap2.prot & PROT_EXEC) == 0) {
+			if (!mmap_data || (event->mmap2.prot & PROT_READ) == 0)
 				continue;
 
 			event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
@@ -379,17 +432,17 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 		if (truncation)
 			event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT;
 
-		if (!strcmp(execname, ""))
-			strcpy(execname, anonstr);
+		if (!strcmp(event->mmap2.filename, ""))
+			strcpy(event->mmap2.filename, anonstr);
 
 		if (hugetlbfs_mnt_len &&
-		    !strncmp(execname, hugetlbfs_mnt, hugetlbfs_mnt_len)) {
-			strcpy(execname, anonstr);
+		    !strncmp(event->mmap2.filename, hugetlbfs_mnt,
+			     hugetlbfs_mnt_len)) {
+			strcpy(event->mmap2.filename, anonstr);
 			event->mmap2.flags |= MAP_HUGETLB;
 		}
 
-		size = strlen(execname) + 1;
-		memcpy(event->mmap2.filename, execname, size);
+		size = strlen(event->mmap2.filename) + 1;
 		size = PERF_ALIGN(size, sizeof(u64));
 		event->mmap2.len -= event->mmap.start;
 		event->mmap2.header.size = (sizeof(event->mmap2) -
@@ -408,7 +461,7 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 			break;
 	}
 
-	fclose(fp);
+	close(io.fd);
 	return rc;
 }
 
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog
#8 lore raw

Re: [PATCH 5/5] perf synthetic events: Remove use of sscanf from /proc reading

Quoted 12 lines · Ian Rogers

The synthesize benchmark, run on a single process and thread, shows perf_event__synthesize_mmap_events as the hottest function with fgets and sscanf taking the majority of execution time. fscanf performs similarly well. Replace the scanf call with manual reading of each field of the /proc/pid/maps line, and remove some unnecessary buffering. This change also addresses potential, but unlikely, buffer overruns for the string values read by scanf.

Performance before is:
Average synthesis took: 120.195100 usec
Average data synthesis took: 156.582300 usec

Quoted 5 lines

And after is:
Average synthesis took: 67.189100 usec
Average data synthesis took: 102.451600 usec

On a Intel Xeon 6154 compiling with Debian gcc 9.2.1.

I can see the speedup as well, thanks!

jirka

#9 lore raw

Re: [PATCH 5/5] perf synthetic events: Remove use of sscanf from /proc reading

In reply to Ian Rogers

On Wed, Apr 01, 2020 at 04:39:45PM -0700, Ian Rogers wrote:

SNIP

Quoted 34 lines
tools/perf/util/synthetic-events.c +4 -7
@@ -279,9 +353,9 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 				       struct machine *machine,
 				       bool mmap_data)
 {
-	FILE *fp;
 	unsigned long long t;
 	char bf[BUFSIZ];
+	struct io io;
 	bool truncation = false;
 	unsigned long long timeout = proc_map_timeout * 1000000ULL;
 	int rc = 0;
@@ -294,28 +368,39 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 	snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps",
 		machine->root_dir, pid, pid);
 
-	fp = fopen(bf, "r");
-	if (fp == NULL) {
+	io.fd = open(bf, O_RDONLY, 0);
+	if (io.fd < 0) {
 		/*
 		 * We raced with a task exiting - just return:
 		 */
 		pr_debug("couldn't open %s\n", bf);
 		return -1;
 	}
+	init_io(&io, io.fd, bf, sizeof(bf));
 
 	event->header.type = PERF_RECORD_MMAP2;
 	t = rdclock();
 
-	while (1) {
-		char prot[5];
-		char execname[PATH_MAX];
-		char anonstr[] = "//anon";

tools/perf/util/synthetic-events.c int perf_event__synthesize_mmap_events(struct perf_tool *tool, · line 389

Quoted 10 lines
tools/perf/util/synthetic-events.c +4 -4
-		unsigned int ino;
+	while (!io.eof) {
+		static const char anonstr[] = "//anon";
 		size_t size;
-		ssize_t n;
 
-		if (fgets(bf, sizeof(bf), fp) == NULL)
-			break;
+		/* ensure null termination since stack will be reused. */
+		strcpy(event->mmap2.filename, "");

could you just do 'event->mmap2.filename[0] = 0x0' instad ?

jirka

Quoted 15 lines
tools/perf/util/synthetic-events.c +14 -0
+
+		/* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
+		if (!read_proc_maps_line(&io,
+					&event->mmap2.start,
+					&event->mmap2.len,
+					&event->mmap2.prot,
+					&event->mmap2.flags,
+					&event->mmap2.pgoff,
+					&event->mmap2.maj,
+					&event->mmap2.min,
+					&event->mmap2.ino,
+					sizeof(event->mmap2.filename),
+					event->mmap2.filename))
+			continue;
 

SNIP