perf_counter: some simple userspace profiling
# perf-record make -j4 kernel/
# perf-report | tail -15
0.39 cc1 [kernel] lock_acquired
0.42 cc1 [kernel] lock_acquire
0.51 cc1 [ user ] /lib64/libc-2.8.90.so: _int_free
0.51 as [kernel] clear_page_c
0.53 cc1 [ user ] /lib64/libc-2.8.90.so: memcpy
0.56 cc1 [ user ] /lib64/libc-2.8.90.so: _IO_vfprintf
0.63 cc1 [kernel] lock_release
0.67 cc1 [ user ] /lib64/libc-2.8.90.so: strlen
0.68 cc1 [kernel] debug_smp_processor_id
1.38 cc1 [ user ] /lib64/libc-2.8.90.so: _int_malloc
1.55 cc1 [ user ] /lib64/libc-2.8.90.so: memset
1.77 cc1 [kernel] __lock_acquire
1.88 cc1 [kernel] clear_page_c
3.61 as [ user ] /usr/bin/as: <unknown>
59.16 cc1 [ user ] /usr/libexec/gcc/x86_64-redhat-linux/4.3.2/cc1: <unknown>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
LKML-Reference: <20090408130409.220518450@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
diff --git a/Documentation/perf_counter/perf-record.c b/Documentation/perf_counter/perf-record.c
new file mode 100644
index 0000000..614de7c
--- /dev/null
+++ b/Documentation/perf_counter/perf-record.c
@@ -0,0 +1,530 @@
+
+
+#define _GNU_SOURCE
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+#include <unistd.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include <getopt.h>
+#include <assert.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <errno.h>
+#include <ctype.h>
+#include <time.h>
+#include <sched.h>
+#include <pthread.h>
+
+#include <sys/syscall.h>
+#include <sys/ioctl.h>
+#include <sys/poll.h>
+#include <sys/prctl.h>
+#include <sys/wait.h>
+#include <sys/uio.h>
+#include <sys/mman.h>
+
+#include <linux/unistd.h>
+#include <linux/types.h>
+
+#include "../../include/linux/perf_counter.h"
+
+
+/*
+ * prctl(PR_TASK_PERF_COUNTERS_DISABLE) will (cheaply) disable all
+ * counters in the current task.
+ */
+#define PR_TASK_PERF_COUNTERS_DISABLE 31
+#define PR_TASK_PERF_COUNTERS_ENABLE 32
+
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
+
+#define rdclock() \
+({ \
+ struct timespec ts; \
+ \
+ clock_gettime(CLOCK_MONOTONIC, &ts); \
+ ts.tv_sec * 1000000000ULL + ts.tv_nsec; \
+})
+
+/*
+ * Pick up some kernel type conventions:
+ */
+#define __user
+#define asmlinkage
+
+#ifdef __x86_64__
+#define __NR_perf_counter_open 295
+#define rmb() asm volatile("lfence" ::: "memory")
+#define cpu_relax() asm volatile("rep; nop" ::: "memory");
+#endif
+
+#ifdef __i386__
+#define __NR_perf_counter_open 333
+#define rmb() asm volatile("lfence" ::: "memory")
+#define cpu_relax() asm volatile("rep; nop" ::: "memory");
+#endif
+
+#ifdef __powerpc__
+#define __NR_perf_counter_open 319
+#define rmb() asm volatile ("sync" ::: "memory")
+#define cpu_relax() asm volatile ("" ::: "memory");
+#endif
+
+#define unlikely(x) __builtin_expect(!!(x), 0)
+#define min(x, y) ({ \
+ typeof(x) _min1 = (x); \
+ typeof(y) _min2 = (y); \
+ (void) (&_min1 == &_min2); \
+ _min1 < _min2 ? _min1 : _min2; })
+
+asmlinkage int sys_perf_counter_open(
+ struct perf_counter_hw_event *hw_event_uptr __user,
+ pid_t pid,
+ int cpu,
+ int group_fd,
+ unsigned long flags)
+{
+ return syscall(
+ __NR_perf_counter_open, hw_event_uptr, pid, cpu, group_fd, flags);
+}
+
+#define MAX_COUNTERS 64
+#define MAX_NR_CPUS 256
+
+#define EID(type, id) (((__u64)(type) << PERF_COUNTER_TYPE_SHIFT) | (id))
+
+static int nr_counters = 0;
+static __u64 event_id[MAX_COUNTERS] = { };
+static int default_interval = 100000;
+static int event_count[MAX_COUNTERS];
+static int fd[MAX_NR_CPUS][MAX_COUNTERS];
+static int nr_cpus = 0;
+static unsigned int page_size;
+static unsigned int mmap_pages = 16;
+static int output;
+static char *output_name = "output.perf";
+static int group = 0;
+static unsigned int realtime_prio = 0;
+
+const unsigned int default_count[] = {
+ 1000000,
+ 1000000,
+ 10000,
+ 10000,
+ 1000000,
+ 10000,
+};
+
+static char *hw_event_names[] = {
+ "CPU cycles",
+ "instructions",
+ "cache references",
+ "cache misses",
+ "branches",
+ "branch misses",
+ "bus cycles",
+};
+
+static char *sw_event_names[] = {
+ "cpu clock ticks",
+ "task clock ticks",
+ "pagefaults",
+ "context switches",
+ "CPU migrations",
+ "minor faults",
+ "major faults",
+};
+
+struct event_symbol {
+ __u64 event;
+ char *symbol;
+};
+
+static struct event_symbol event_symbols[] = {
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), "cpu-cycles", },
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), "cycles", },
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS), "instructions", },
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES), "cache-references", },
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES), "cache-misses", },
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_INSTRUCTIONS), "branch-instructions", },
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_INSTRUCTIONS), "branches", },
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_MISSES), "branch-misses", },
+ {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BUS_CYCLES), "bus-cycles", },
+
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_CLOCK), "cpu-clock", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK), "task-clock", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), "page-faults", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), "faults", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS_MIN), "minor-faults", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS_MAJ), "major-faults", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), "context-switches", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), "cs", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), "cpu-migrations", },
+ {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), "migrations", },
+};
+
+/*
+ * Each event can have multiple symbolic names.
+ * Symbolic names are (almost) exactly matched.
+ */
+static __u64 match_event_symbols(char *str)
+{
+ __u64 config, id;
+ int type;
+ unsigned int i;
+
+ if (sscanf(str, "r%llx", &config) == 1)
+ return config | PERF_COUNTER_RAW_MASK;
+
+ if (sscanf(str, "%d:%llu", &type, &id) == 2)
+ return EID(type, id);
+
+ for (i = 0; i < ARRAY_SIZE(event_symbols); i++) {
+ if (!strncmp(str, event_symbols[i].symbol,
+ strlen(event_symbols[i].symbol)))
+ return event_symbols[i].event;
+ }
+
+ return ~0ULL;
+}
+
+static int parse_events(char *str)
+{
+ __u64 config;
+
+again:
+ if (nr_counters == MAX_COUNTERS)
+ return -1;
+
+ config = match_event_symbols(str);
+ if (config == ~0ULL)
+ return -1;
+
+ event_id[nr_counters] = config;
+ nr_counters++;
+
+ str = strstr(str, ",");
+ if (str) {
+ str++;
+ goto again;
+ }
+
+ return 0;
+}
+
+#define __PERF_COUNTER_FIELD(config, name) \
+ ((config & PERF_COUNTER_##name##_MASK) >> PERF_COUNTER_##name##_SHIFT)
+
+#define PERF_COUNTER_RAW(config) __PERF_COUNTER_FIELD(config, RAW)
+#define PERF_COUNTER_CONFIG(config) __PERF_COUNTER_FIELD(config, CONFIG)
+#define PERF_COUNTER_TYPE(config) __PERF_COUNTER_FIELD(config, TYPE)
+#define PERF_COUNTER_ID(config) __PERF_COUNTER_FIELD(config, EVENT)
+
+static void display_events_help(void)
+{
+ unsigned int i;
+ __u64 e;
+
+ printf(
+ " -e EVENT --event=EVENT # symbolic-name abbreviations");
+
+ for (i = 0; i < ARRAY_SIZE(event_symbols); i++) {
+ int type, id;
+
+ e = event_symbols[i].event;
+ type = PERF_COUNTER_TYPE(e);
+ id = PERF_COUNTER_ID(e);
+
+ printf("\n %d:%d: %-20s",
+ type, id, event_symbols[i].symbol);
+ }
+
+ printf("\n"
+ " rNNN: raw PMU events (eventsel+umask)\n\n");
+}
+
+static void display_help(void)
+{
+ printf(
+ "Usage: perf-record [<options>]\n"
+ "perf-record Options (up to %d event types can be specified at once):\n\n",
+ MAX_COUNTERS);
+
+ display_events_help();
+
+ printf(
+ " -c CNT --count=CNT # event period to sample\n"
+ " -m pages --mmap_pages=<pages> # number of mmap data pages\n"
+ " -o file --output=<file> # output file\n"
+ " -r prio --realtime=<prio> # use RT prio\n"
+ );
+
+ exit(0);
+}
+
+static void process_options(int argc, char *argv[])
+{
+ int error = 0, counter;
+
+ for (;;) {
+ int option_index = 0;
+ /** Options for getopt */
+ static struct option long_options[] = {
+ {"count", required_argument, NULL, 'c'},
+ {"event", required_argument, NULL, 'e'},
+ {"mmap_pages", required_argument, NULL, 'm'},
+ {"output", required_argument, NULL, 'o'},
+ {"realtime", required_argument, NULL, 'r'},
+ {NULL, 0, NULL, 0 }
+ };
+ int c = getopt_long(argc, argv, "+:c:e:m:o:r:",
+ long_options, &option_index);
+ if (c == -1)
+ break;
+
+ switch (c) {
+ case 'c': default_interval = atoi(optarg); break;
+ case 'e': error = parse_events(optarg); break;
+ case 'm': mmap_pages = atoi(optarg); break;
+ case 'o': output_name = strdup(optarg); break;
+ case 'r': realtime_prio = atoi(optarg); break;
+ default: error = 1; break;
+ }
+ }
+ if (error)
+ display_help();
+
+ if (!nr_counters) {
+ nr_counters = 1;
+ event_id[0] = 0;
+ }
+
+ for (counter = 0; counter < nr_counters; counter++) {
+ if (event_count[counter])
+ continue;
+
+ event_count[counter] = default_interval;
+ }
+}
+
+struct mmap_data {
+ int counter;
+ void *base;
+ unsigned int mask;
+ unsigned int prev;
+};
+
+static unsigned int mmap_read_head(struct mmap_data *md)
+{
+ struct perf_counter_mmap_page *pc = md->base;
+ int head;
+
+ head = pc->data_head;
+ rmb();
+
+ return head;
+}
+
+static long events;
+static struct timeval last_read, this_read;
+
+static void mmap_read(struct mmap_data *md)
+{
+ unsigned int head = mmap_read_head(md);
+ unsigned int old = md->prev;
+ unsigned char *data = md->base + page_size;
+ unsigned long size;
+ void *buf;
+ int diff;
+
+ gettimeofday(&this_read, NULL);
+
+ /*
+ * If we're further behind than half the buffer, there's a chance
+ * the writer will bite our tail and screw up the events under us.
+ *
+ * If we somehow ended up ahead of the head, we got messed up.
+ *
+ * In either case, truncate and restart at head.
+ */
+ diff = head - old;
+ if (diff > md->mask / 2 || diff < 0) {
+ struct timeval iv;
+ unsigned long msecs;
+
+ timersub(&this_read, &last_read, &iv);
+ msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
+
+ fprintf(stderr, "WARNING: failed to keep up with mmap data."
+ " Last read %lu msecs ago.\n", msecs);
+
+ /*
+ * head points to a known good entry, start there.
+ */
+ old = head;
+ }
+
+ last_read = this_read;
+
+ if (old != head)
+ events++;
+
+ size = head - old;
+
+ if ((old & md->mask) + size != (head & md->mask)) {
+ buf = &data[old & md->mask];
+ size = md->mask + 1 - (old & md->mask);
+ old += size;
+ while (size) {
+ int ret = write(output, buf, size);
+ if (ret < 0) {
+ perror("failed to write");
+ exit(-1);
+ }
+ size -= ret;
+ buf += ret;
+ }
+ }
+
+ buf = &data[old & md->mask];
+ size = head - old;
+ old += size;
+ while (size) {
+ int ret = write(output, buf, size);
+ if (ret < 0) {
+ perror("failed to write");
+ exit(-1);
+ }
+ size -= ret;
+ buf += ret;
+ }
+
+ md->prev = old;
+}
+
+static volatile int done = 0;
+
+static void sigchld_handler(int sig)
+{
+ if (sig == SIGCHLD)
+ done = 1;
+}
+
+int main(int argc, char *argv[])
+{
+ struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
+ struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
+ struct perf_counter_hw_event hw_event;
+ int i, counter, group_fd, nr_poll = 0;
+ pid_t pid;
+ int ret;
+
+ page_size = sysconf(_SC_PAGE_SIZE);
+
+ process_options(argc, argv);
+
+ nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+ assert(nr_cpus <= MAX_NR_CPUS);
+ assert(nr_cpus >= 0);
+
+ output = open(output_name, O_CREAT|O_RDWR, S_IRWXU);
+ if (output < 0) {
+ perror("failed to create output file");
+ exit(-1);
+ }
+
+ argc -= optind;
+ argv += optind;
+
+ for (i = 0; i < nr_cpus; i++) {
+ group_fd = -1;
+ for (counter = 0; counter < nr_counters; counter++) {
+
+ memset(&hw_event, 0, sizeof(hw_event));
+ hw_event.config = event_id[counter];
+ hw_event.irq_period = event_count[counter];
+ hw_event.record_type = PERF_RECORD_IP | PERF_RECORD_TID;
+ hw_event.nmi = 1;
+ hw_event.mmap = 1;
+ hw_event.comm = 1;
+
+ fd[i][counter] = sys_perf_counter_open(&hw_event, -1, i, group_fd, 0);
+ if (fd[i][counter] < 0) {
+ int err = errno;
+ printf("kerneltop error: syscall returned with %d (%s)\n",
+ fd[i][counter], strerror(err));
+ if (err == EPERM)
+ printf("Are you root?\n");
+ exit(-1);
+ }
+ assert(fd[i][counter] >= 0);
+ fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
+
+ /*
+ * First counter acts as the group leader:
+ */
+ if (group && group_fd == -1)
+ group_fd = fd[i][counter];
+
+ event_array[nr_poll].fd = fd[i][counter];
+ event_array[nr_poll].events = POLLIN;
+ nr_poll++;
+
+ mmap_array[i][counter].counter = counter;
+ mmap_array[i][counter].prev = 0;
+ mmap_array[i][counter].mask = mmap_pages*page_size - 1;
+ mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
+ PROT_READ, MAP_SHARED, fd[i][counter], 0);
+ if (mmap_array[i][counter].base == MAP_FAILED) {
+ printf("kerneltop error: failed to mmap with %d (%s)\n",
+ errno, strerror(errno));
+ exit(-1);
+ }
+ }
+ }
+
+ signal(SIGCHLD, sigchld_handler);
+
+ pid = fork();
+ if (pid < 0)
+ perror("failed to fork");
+
+ if (!pid) {
+ if (execvp(argv[0], argv)) {
+ perror(argv[0]);
+ exit(-1);
+ }
+ }
+
+ if (realtime_prio) {
+ struct sched_param param;
+
+ param.sched_priority = realtime_prio;
+ if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
+ printf("Could not set realtime priority.\n");
+ exit(-1);
+ }
+ }
+
+ /*
+ * TODO: store the current /proc/$/maps information somewhere
+ */
+
+ while (!done) {
+ int hits = events;
+
+ for (i = 0; i < nr_cpus; i++) {
+ for (counter = 0; counter < nr_counters; counter++)
+ mmap_read(&mmap_array[i][counter]);
+ }
+
+ if (hits == events)
+ ret = poll(event_array, nr_poll, 100);
+ }
+
+ return 0;
+}
