perf_counter: some simple userspace profiling
authorPeter Zijlstra <a.p.zijlstra@chello.nl>
Wed, 8 Apr 2009 13:01:31 +0000 (15:01 +0200)
committerIngo Molnar <mingo@elte.hu>
Wed, 8 Apr 2009 17:05:54 +0000 (19:05 +0200)
# 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>
Documentation/perf_counter/Makefile
Documentation/perf_counter/perf-record.c [new file with mode: 0644]
Documentation/perf_counter/perf-report.cc [new file with mode: 0644]

index 194b662..1dd37ee 100644 (file)
@@ -1,10 +1,16 @@
-BINS = kerneltop perfstat
+BINS = kerneltop perfstat perf-record perf-report
 
 all: $(BINS)
 
 kerneltop: kerneltop.c ../../include/linux/perf_counter.h
        cc -O6 -Wall -lrt -o $@ $<
 
+perf-record: perf-record.c ../../include/linux/perf_counter.h
+       cc -O6 -Wall -lrt -o $@ $<
+
+perf-report: perf-report.cc ../../include/linux/perf_counter.h
+       g++ -O6 -Wall -lrt -o $@ $<
+
 perfstat: kerneltop
        ln -sf kerneltop perfstat
 
diff --git a/Documentation/perf_counter/perf-record.c b/Documentation/perf_counter/perf-record.c
new file mode 100644 (file)
index 0000000..614de7c
--- /dev/null
@@ -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, &param)) {
+                       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;
+}
diff --git a/Documentation/perf_counter/perf-report.cc b/Documentation/perf_counter/perf-report.cc
new file mode 100644 (file)
index 0000000..09da0ba
--- /dev/null
@@ -0,0 +1,472 @@
+#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 <fcntl.h>
+#include <stdio.h>
+#include <errno.h>
+#include <ctype.h>
+#include <time.h>
+#include <getopt.h>
+
+#include <sys/ioctl.h>
+#include <sys/poll.h>
+#include <sys/prctl.h>
+#include <sys/wait.h>
+#include <sys/mman.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+
+#include <linux/unistd.h>
+#include <linux/types.h>
+
+#include "../../include/linux/perf_counter.h"
+
+#include <set>
+#include <map>
+#include <string>
+
+
+static char            const *input_name = "output.perf";
+static int             input;
+
+static unsigned long   page_size;
+static unsigned long   mmap_window = 32;
+
+struct ip_event {
+       struct perf_event_header header;
+       __u64 ip;
+       __u32 pid, tid;
+};
+struct mmap_event {
+       struct perf_event_header header;
+       __u32 pid, tid;
+       __u64 start;
+       __u64 len;
+       __u64 pgoff;
+       char filename[PATH_MAX];
+};
+struct comm_event {
+       struct perf_event_header header;
+       __u32 pid,tid;
+       char comm[16];
+};
+
+typedef union event_union {
+       struct perf_event_header header;
+       struct ip_event ip;
+       struct mmap_event mmap;
+       struct comm_event comm;
+} event_t;
+
+struct section {
+       uint64_t start;
+       uint64_t end;
+
+       uint64_t offset;
+
+       std::string name;
+
+       section() { };
+
+       section(uint64_t stab) : end(stab) { };
+
+       section(uint64_t start, uint64_t size, uint64_t offset, std::string name) :
+               start(start), end(start + size), offset(offset), name(name)
+       { };
+
+       bool operator < (const struct section &s) const {
+               return end < s.end;
+       };
+};
+
+typedef std::set<struct section> sections_t;
+
+struct symbol {
+       uint64_t start;
+       uint64_t end;
+
+       std::string name;
+
+       symbol() { };
+
+       symbol(uint64_t ip) : start(ip) { }
+
+       symbol(uint64_t start, uint64_t len, std::string name) :
+               start(start), end(start + len), name(name)
+       { };
+
+       bool operator < (const struct symbol &s) const {
+               return start < s.start;
+       };
+};
+
+typedef std::set<struct symbol> symbols_t;
+
+struct dso {
+       sections_t sections;
+       symbols_t syms;
+};
+
+static std::map<std::string, struct dso> dsos;
+
+static void load_dso_sections(std::string dso_name)
+{
+       struct dso &dso = dsos[dso_name];
+
+       std::string cmd = "readelf -DSW " + dso_name;
+
+       FILE *file = popen(cmd.c_str(), "r");
+       if (!file) {
+               perror("failed to open pipe");
+               exit(-1);
+       }
+
+       char *line = NULL;
+       size_t n = 0;
+
+       while (!feof(file)) {
+               uint64_t addr, off, size;
+               char name[32];
+
+               if (getline(&line, &n, file) < 0)
+                       break;
+               if (!line)
+                       break;
+
+               if (sscanf(line, "  [%*2d] %16s %*14s %Lx %Lx %Lx",
+                                       name, &addr, &off, &size) == 4) {
+
+                       dso.sections.insert(section(addr, size, addr - off, name));
+               }
+#if 0
+               /*
+                * for reading readelf symbols (-s), however these don't seem
+                * to include nearly everything, so use nm for that.
+                */
+               if (sscanf(line, " %*4d %*3d: %Lx %5Lu %*7s %*6s %*7s %3d %s",
+                          &start, &size, &section, sym) == 4) {
+
+                       start -= dso.section_offsets[section];
+
+                       dso.syms.insert(symbol(start, size, std::string(sym)));
+               }
+#endif
+       }
+       pclose(file);
+}
+
+static void load_dso_symbols(std::string dso_name, std::string args)
+{
+       struct dso &dso = dsos[dso_name];
+
+       std::string cmd = "nm -nSC " + args + " " + dso_name;
+
+       FILE *file = popen(cmd.c_str(), "r");
+       if (!file) {
+               perror("failed to open pipe");
+               exit(-1);
+       }
+
+       char *line = NULL;
+       size_t n = 0;
+
+       while (!feof(file)) {
+               uint64_t start, size;
+               char c;
+               char sym[1024];
+
+               if (getline(&line, &n, file) < 0)
+                       break;
+               if (!line)
+                       break;
+
+
+               if (sscanf(line, "%Lx %Lx %c %s", &start, &size, &c, sym) == 4) {
+                       sections_t::const_iterator si =
+                               dso.sections.upper_bound(section(start));
+                       if (si == dso.sections.end()) {
+                               printf("symbol in unknown section: %s\n", sym);
+                               continue;
+                       }
+
+                       start -= si->offset;
+
+                       dso.syms.insert(symbol(start, size, sym));
+               }
+       }
+       pclose(file);
+}
+
+static void load_dso(std::string dso_name)
+{
+       load_dso_sections(dso_name);
+       load_dso_symbols(dso_name, "-D"); /* dynamic symbols */
+       load_dso_symbols(dso_name, "");   /* regular ones */
+}
+
+void load_kallsyms(void)
+{
+       struct dso &dso = dsos["[kernel]"];
+
+       FILE *file = fopen("/proc/kallsyms", "r");
+       if (!file) {
+               perror("failed to open kallsyms");
+               exit(-1);
+       }
+
+       char *line;
+       size_t n;
+
+       while (!feof(file)) {
+               uint64_t start;
+               char c;
+               char sym[1024];
+
+               if (getline(&line, &n, file) < 0)
+                       break;
+               if (!line)
+                       break;
+
+               if (sscanf(line, "%Lx %c %s", &start, &c, sym) == 3)
+                       dso.syms.insert(symbol(start, 0x1000000, std::string(sym)));
+       }
+       fclose(file);
+}
+
+struct map {
+       uint64_t start;
+       uint64_t end;
+       uint64_t pgoff;
+
+       std::string dso;
+
+       map() { };
+
+       map(uint64_t ip) : end(ip) { }
+
+       map(mmap_event *mmap) {
+               start = mmap->start;
+               end = mmap->start + mmap->len;
+               pgoff = mmap->pgoff;
+
+               dso = std::string(mmap->filename);
+
+               if (dsos.find(dso) == dsos.end())
+                       load_dso(dso);
+       };
+
+       bool operator < (const struct map &m) const {
+               return end < m.end;
+       };
+};
+
+typedef std::set<struct map> maps_t;
+
+static std::map<int, maps_t> maps;
+
+static std::map<int, std::string> comms;
+
+static std::map<std::string, int> hist;
+static std::multimap<int, std::string> rev_hist;
+
+static std::string resolve_comm(int pid)
+{
+       std::string comm = "<unknown>";
+       std::map<int, std::string>::const_iterator ci = comms.find(pid);
+       if (ci != comms.end())
+               comm = ci->second;
+
+       return comm;
+}
+
+static std::string resolve_user_symbol(int pid, uint64_t ip)
+{
+       std::string sym = "<unknown>";
+
+       maps_t &m = maps[pid];
+       maps_t::const_iterator mi = m.upper_bound(map(ip));
+       if (mi == m.end())
+               return sym;
+
+       ip -= mi->start + mi->pgoff;
+
+       symbols_t &s = dsos[mi->dso].syms;
+       symbols_t::const_iterator si = s.upper_bound(symbol(ip));
+
+       sym = mi->dso + ": <unknown>";
+
+       if (si == s.begin())
+               return sym;
+       si--;
+
+       if (si->start <= ip && ip < si->end)
+               sym = mi->dso + ": " + si->name;
+#if 0
+       else if (si->start <= ip)
+               sym = mi->dso + ": ?" + si->name;
+#endif
+
+       return sym;
+}
+
+static std::string resolve_kernel_symbol(uint64_t ip)
+{
+       std::string sym = "<unknown>";
+
+       symbols_t &s = dsos["[kernel]"].syms;
+       symbols_t::const_iterator si = s.upper_bound(symbol(ip));
+
+       if (si == s.begin())
+               return sym;
+       si--;
+
+       if (si->start <= ip && ip < si->end)
+               sym = si->name;
+
+       return sym;
+}
+
+static void display_help(void)
+{
+       printf(
+       "Usage: perf-report [<options>]\n"
+       " -i file   --input=<file>      # input file\n"
+       );
+
+       exit(0);
+}
+
+static void process_options(int argc, char *argv[])
+{
+       int error = 0;
+
+       for (;;) {
+               int option_index = 0;
+               /** Options for getopt */
+               static struct option long_options[] = {
+                       {"input",       required_argument,      NULL, 'i'},
+                       {NULL,          0,                      NULL,  0 }
+               };
+               int c = getopt_long(argc, argv, "+:i:",
+                                   long_options, &option_index);
+               if (c == -1)
+                       break;
+
+               switch (c) {
+               case 'i': input_name                    = strdup(optarg); break;
+               default: error = 1; break;
+               }
+       }
+
+       if (error)
+               display_help();
+}
+
+int main(int argc, char *argv[])
+{
+       unsigned long offset = 0;
+       unsigned long head = 0;
+       struct stat stat;
+       char *buf;
+       event_t *event;
+       int ret;
+       unsigned long total = 0;
+
+       page_size = getpagesize();
+
+       process_options(argc, argv);
+
+       input = open(input_name, O_RDONLY);
+       if (input < 0) {
+               perror("failed to open file");
+               exit(-1);
+       }
+
+       ret = fstat(input, &stat);
+       if (ret < 0) {
+               perror("failed to stat file");
+               exit(-1);
+       }
+
+       load_kallsyms();
+
+remap:
+       buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
+                          MAP_SHARED, input, offset);
+       if (buf == MAP_FAILED) {
+               perror("failed to mmap file");
+               exit(-1);
+       }
+
+more:
+       event = (event_t *)(buf + head);
+
+       if (head + event->header.size >= page_size * mmap_window) {
+               unsigned long shift = page_size * (head / page_size);
+
+               munmap(buf, page_size * mmap_window);
+               offset += shift;
+               head -= shift;
+               goto remap;
+       }
+       head += event->header.size;
+
+       if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
+               std::string comm, sym, level;
+               char output[1024];
+
+               if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
+                       level = "[kernel]";
+                       sym = resolve_kernel_symbol(event->ip.ip);
+               } else if (event->header.misc & PERF_EVENT_MISC_USER) {
+                       level = "[ user ]";
+                       sym = resolve_user_symbol(event->ip.pid, event->ip.ip);
+               } else {
+                       level = "[  hv  ]";
+               }
+               comm = resolve_comm(event->ip.pid);
+
+               snprintf(output, sizeof(output), "%16s %s %s",
+                               comm.c_str(), level.c_str(), sym.c_str());
+               hist[output]++;
+
+               total++;
+
+       } else switch (event->header.type) {
+       case PERF_EVENT_MMAP:
+               maps[event->mmap.pid].insert(map(&event->mmap));
+               break;
+
+       case PERF_EVENT_COMM:
+               comms[event->comm.pid] = std::string(event->comm.comm);
+               break;
+       }
+
+       if (offset + head < stat.st_size)
+               goto more;
+
+       close(input);
+
+       std::map<std::string, int>::iterator hi = hist.begin();
+
+       while (hi != hist.end()) {
+               rev_hist.insert(std::pair<int, std::string>(hi->second, hi->first));
+               hist.erase(hi++);
+       }
+
+       std::multimap<int, std::string>::const_iterator ri = rev_hist.begin();
+
+       while (ri != rev_hist.end()) {
+               printf(" %5.2f %s\n", (100.0 * ri->first)/total, ri->second.c_str());
+               ri++;
+       }
+
+       return 0;
+}
+