perf_counter: add sample user-space to Documentation/perf_counter/
authorIngo Molnar <mingo@elte.hu>
Mon, 23 Mar 2009 20:29:59 +0000 (21:29 +0100)
committerIngo Molnar <mingo@elte.hu>
Mon, 6 Apr 2009 07:30:19 +0000 (09:30 +0200)
Initial version of kerneltop.c and perfstat.c.

Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Documentation/perf_counter/Makefile [new file with mode: 0644]
Documentation/perf_counter/kerneltop.c [new file with mode: 0644]
Documentation/perf_counter/perfstat.c [new file with mode: 0644]

diff --git a/Documentation/perf_counter/Makefile b/Documentation/perf_counter/Makefile
new file mode 100644 (file)
index 0000000..b457497
--- /dev/null
@@ -0,0 +1,12 @@
+BINS = kerneltop perfstat
+
+all: $(BINS)
+
+kerneltop: kerneltop.c perfcounters.h
+       cc -O6 -Wall -lrt `pkg-config --cflags --libs glib-2.0` -o $@ $<
+
+perfstat: kerneltop
+       ln -sf kerneltop perfstat
+
+clean:
+       rm $(BINS)
diff --git a/Documentation/perf_counter/kerneltop.c b/Documentation/perf_counter/kerneltop.c
new file mode 100644 (file)
index 0000000..cf0e30b
--- /dev/null
@@ -0,0 +1,956 @@
+/*
+ * kerneltop.c: show top kernel functions - performance counters showcase
+
+   Build with:
+
+     cc -O6 -Wall `pkg-config --cflags --libs glib-2.0` -o kerneltop kerneltop.c
+
+   Sample output:
+
+------------------------------------------------------------------------------
+ KernelTop:    2669 irqs/sec  [NMI, cache-misses/cache-refs],  (all, cpu: 2)
+------------------------------------------------------------------------------
+
+             weight         RIP          kernel function
+             ______   ________________   _______________
+
+              35.20 - ffffffff804ce74b : skb_copy_and_csum_dev
+              33.00 - ffffffff804cb740 : sock_alloc_send_skb
+              31.26 - ffffffff804ce808 : skb_push
+              22.43 - ffffffff80510004 : tcp_established_options
+              19.00 - ffffffff8027d250 : find_get_page
+              15.76 - ffffffff804e4fc9 : eth_type_trans
+              15.20 - ffffffff804d8baa : dst_release
+              14.86 - ffffffff804cf5d8 : skb_release_head_state
+              14.00 - ffffffff802217d5 : read_hpet
+              12.00 - ffffffff804ffb7f : __ip_local_out
+              11.97 - ffffffff804fc0c8 : ip_local_deliver_finish
+               8.54 - ffffffff805001a3 : ip_queue_xmit
+
+  Started by Ingo Molnar <mingo@redhat.com>
+
+  Improvements and fixes by:
+
+    Arjan van de Ven <arjan@linux.intel.com>
+    Yanmin Zhang <yanmin.zhang@intel.com>
+    Mike Galbraith <efault@gmx.de>
+
+  Released under the GPL v2. (and only v2, not any later version)
+
+ */
+#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 <getopt.h>
+#include <assert.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <errno.h>
+#include <ctype.h>
+#include <time.h>
+
+#include <glib.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 <linux/unistd.h>
+
+#ifdef __x86_64__
+# define __NR_perf_counter_open        295
+#endif
+
+#ifdef __i386__
+# define __NR_perf_counter_open 333
+#endif
+
+/*
+ * Pick up some kernel type conventions:
+ */
+#define __user
+#define asmlinkage
+
+typedef unsigned int           __u32;
+typedef unsigned long long     __u64;
+typedef long long              __s64;
+
+/*
+ * User-space ABI bits:
+ */
+
+/*
+ * Generalized performance counter event types, used by the hw_event.type
+ * parameter of the sys_perf_counter_open() syscall:
+ */
+enum hw_event_types {
+       /*
+        * Common hardware events, generalized by the kernel:
+        */
+       PERF_COUNT_CPU_CYCLES           =  0,
+       PERF_COUNT_INSTRUCTIONS         =  1,
+       PERF_COUNT_CACHE_REFERENCES     =  2,
+       PERF_COUNT_CACHE_MISSES         =  3,
+       PERF_COUNT_BRANCH_INSTRUCTIONS  =  4,
+       PERF_COUNT_BRANCH_MISSES        =  5,
+       PERF_COUNT_BUS_CYCLES           =  6,
+
+       PERF_HW_EVENTS_MAX              =  7,
+
+       /*
+        * Special "software" counters provided by the kernel, even if
+        * the hardware does not support performance counters. These
+        * counters measure various physical and sw events of the
+        * kernel (and allow the profiling of them as well):
+        */
+       PERF_COUNT_CPU_CLOCK            = -1,
+       PERF_COUNT_TASK_CLOCK           = -2,
+       PERF_COUNT_PAGE_FAULTS          = -3,
+       PERF_COUNT_CONTEXT_SWITCHES     = -4,
+       PERF_COUNT_CPU_MIGRATIONS       = -5,
+
+       PERF_SW_EVENTS_MIN              = -6,
+};
+
+/*
+ * IRQ-notification data record type:
+ */
+enum perf_counter_record_type {
+       PERF_RECORD_SIMPLE              =  0,
+       PERF_RECORD_IRQ                 =  1,
+       PERF_RECORD_GROUP               =  2,
+};
+
+/*
+ * Hardware event to monitor via a performance monitoring counter:
+ */
+struct perf_counter_hw_event {
+       __s64                   type;
+
+       __u64                   irq_period;
+       __u64                   record_type;
+       __u64                   read_format;
+
+       __u64                   disabled       :  1, /* off by default        */
+                               nmi            :  1, /* NMI sampling          */
+                               raw            :  1, /* raw event type        */
+                               inherit        :  1, /* children inherit it   */
+                               pinned         :  1, /* must always be on PMU */
+                               exclusive      :  1, /* only group on PMU     */
+                               exclude_user   :  1, /* don't count user      */
+                               exclude_kernel :  1, /* ditto kernel          */
+                               exclude_hv     :  1, /* ditto hypervisor      */
+                               exclude_idle   :  1, /* don't count when idle */
+
+                               __reserved_1   : 54;
+
+       __u32                   extra_config_len;
+       __u32                   __reserved_4;
+
+       __u64                   __reserved_2;
+       __u64                   __reserved_3;
+};
+
+/*
+ * Ioctls that can be done on a perf counter fd:
+ */
+#define PERF_COUNTER_IOC_ENABLE                _IO('$', 0)
+#define PERF_COUNTER_IOC_DISABLE       _IO('$', 1)
+
+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)
+{
+       int ret;
+
+       ret = syscall(
+               __NR_perf_counter_open, hw_event_uptr, pid, cpu, group_fd, flags);
+#if defined(__x86_64__) || defined(__i386__)
+       if (ret < 0 && ret > -4096) {
+               errno = -ret;
+               ret = -1;
+       }
+#endif
+       return ret;
+}
+
+const char *event_types [] = {
+       "CPU cycles",
+       "instructions",
+       "cache-refs",
+       "cache-misses",
+       "branches",
+       "branch-misses",
+       "bus cycles"
+};
+
+const unsigned int default_count[] = {
+       1000000,
+       1000000,
+         10000,
+         10000,
+       1000000,
+         10000,
+};
+
+/*
+ * 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 MAX_COUNTERS           8
+
+static int                     nr_counters                     = -1;
+
+static __u64                   count_filter                   = 100;
+
+#define MAX_NR_CPUS            256
+
+static int                     event_count[MAX_COUNTERS];
+static unsigned long           event_id[MAX_COUNTERS];
+static int                     event_raw[MAX_COUNTERS];
+
+static int                     tid                             = -1;
+static int                     profile_cpu                     = -1;
+static int                     nr_cpus                         =  0;
+static int                     nmi                             =  1;
+static int                     group                           =  0;
+
+static char                    *vmlinux;
+
+static char                    *sym_filter;
+static unsigned long           filter_start;
+static unsigned long           filter_end;
+
+static int                     delay_secs                      =  2;
+static int                     zero;
+static int                     dump_symtab;
+
+struct source_line {
+       uint64_t                EIP;
+       unsigned long           count;
+       char                    *line;
+};
+
+static GList                   *lines;
+
+static void display_help(void)
+{
+       printf(
+       "Usage: kerneltop [<options>]\n\n"
+       "KernelTop Options (up to %d event types can be specified at once):\n\n",
+                MAX_COUNTERS);
+       printf(
+       " -e EID    --event_id=EID     # event type ID                    [default:  0]\n"
+       "                                   0: CPU cycles\n"
+       "                                   1: instructions\n"
+       "                                   2: cache accesses\n"
+       "                                   3: cache misses\n"
+       "                                   4: branch instructions\n"
+       "                                   5: branch prediction misses\n"
+       "                                   6: bus cycles\n\n"
+       "                                rNNN: raw PMU events (eventsel+umask)\n\n"
+       " -c CNT    --count=CNT        # event period to sample\n\n"
+       " -C CPU    --cpu=CPU          # CPU (-1 for all)                 [default: -1]\n"
+       " -p PID    --pid=PID          # PID of sampled task (-1 for all) [default: -1]\n\n"
+       " -d delay  --delay=<seconds>  # sampling/display delay           [default:  2]\n"
+       " -f CNT --filter=CNT          # min-event-count filter          [default: 100]\n\n"
+       " -s symbol --symbol=<symbol>  # function to be showed annotated one-shot\n"
+       " -x path   --vmlinux=<path>   # the vmlinux binary, required for -s use:\n"
+       " -z        --zero             # zero counts after display\n"
+       " -D        --dump_symtab      # dump symbol table to stderr on startup\n"
+       "\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'},
+                       {"cpu",         required_argument,      NULL, 'C'},
+                       {"delay",       required_argument,      NULL, 'd'},
+                       {"dump_symtab", no_argument,            NULL, 'D'},
+                       {"event_id",    required_argument,      NULL, 'e'},
+                       {"filter",      required_argument,      NULL, 'f'},
+                       {"group",       required_argument,      NULL, 'g'},
+                       {"help",        no_argument,            NULL, 'h'},
+                       {"nmi",         required_argument,      NULL, 'n'},
+                       {"pid",         required_argument,      NULL, 'p'},
+                       {"vmlinux",     required_argument,      NULL, 'x'},
+                       {"symbol",      required_argument,      NULL, 's'},
+                       {"zero",        no_argument,            NULL, 'z'},
+                       {NULL,          0,                      NULL,  0 }
+               };
+               int c = getopt_long(argc, argv, "c:C:d:De:f:g:hn:p:s:x:z",
+                                   long_options, &option_index);
+               if (c == -1)
+                       break;
+
+               switch (c) {
+               case 'c':
+                       if (nr_counters == -1)
+                               nr_counters = 0;
+                       event_count[nr_counters]        =   atoi(optarg); break;
+               case 'C':
+                       /* CPU and PID are mutually exclusive */
+                       if (tid != -1) {
+                               printf("WARNING: CPU switch overriding PID\n");
+                               sleep(1);
+                               tid = -1;
+                       }
+                       profile_cpu                     =   atoi(optarg); break;
+               case 'd': delay_secs                    =   atoi(optarg); break;
+               case 'D': dump_symtab                   =              1; break;
+
+               case 'e':
+                       nr_counters++;
+                       if (nr_counters == MAX_COUNTERS) {
+                               error = 1;
+                               break;
+                       }
+                       if (*optarg == 'r') {
+                               event_raw[nr_counters] = 1;
+                               ++optarg;
+                       }
+                       event_id[nr_counters] = strtol(optarg, NULL, 16);
+                       break;
+
+               case 'f': count_filter                  =   atoi(optarg); break;
+               case 'g': group                         =   atoi(optarg); break;
+               case 'h':                                 display_help(); break;
+               case 'n': nmi                           =   atoi(optarg); break;
+               case 'p':
+                       /* CPU and PID are mutually exclusive */
+                       if (profile_cpu != -1) {
+                               printf("WARNING: PID switch overriding CPU\n");
+                               sleep(1);
+                               profile_cpu = -1;
+                       }
+                       tid                             =   atoi(optarg); break;
+               case 's': sym_filter                    = strdup(optarg); break;
+               case 'x': vmlinux                       = strdup(optarg); break;
+               case 'z': zero                          =              1; break;
+               default: error = 1; break;
+               }
+       }
+       if (error)
+               display_help();
+
+       nr_counters++;
+       if (nr_counters < 1)
+               nr_counters = 1;
+
+       for (counter = 0; counter < nr_counters; counter++) {
+               if (event_count[counter])
+                       continue;
+
+               if (event_id[counter] < PERF_HW_EVENTS_MAX)
+                       event_count[counter] = default_count[event_id[counter]];
+               else
+                       event_count[counter] = 100000;
+       }
+}
+
+static uint64_t                        min_ip;
+static uint64_t                        max_ip = -1ll;
+
+struct sym_entry {
+       unsigned long long      addr;
+       char                    *sym;
+       unsigned long           count[MAX_COUNTERS];
+       int                     skip;
+       GList                   *source;
+};
+
+#define MAX_SYMS               100000
+
+static int sym_table_count;
+
+struct sym_entry               *sym_filter_entry;
+
+static struct sym_entry                sym_table[MAX_SYMS];
+
+static void show_details(struct sym_entry *sym);
+
+/*
+ * Ordering weight: count-1 * count-1 * ... / count-n
+ */
+static double sym_weight(const struct sym_entry *sym)
+{
+       double weight;
+       int counter;
+
+       weight = sym->count[0];
+
+       for (counter = 1; counter < nr_counters-1; counter++)
+               weight *= sym->count[counter];
+
+       weight /= (sym->count[counter] + 1);
+
+       return weight;
+}
+
+static int compare(const void *__sym1, const void *__sym2)
+{
+       const struct sym_entry *sym1 = __sym1, *sym2 = __sym2;
+
+       return sym_weight(sym1) < sym_weight(sym2);
+}
+
+static time_t                  last_refresh;
+static long                    events;
+static long                    userspace_events;
+static const char              CONSOLE_CLEAR[] = "\e[H\e[2J";
+
+static struct sym_entry                tmp[MAX_SYMS];
+
+static void print_sym_table(void)
+{
+       int i, printed;
+       int counter;
+       float events_per_sec = events/delay_secs;
+       float kevents_per_sec = (events-userspace_events)/delay_secs;
+
+       memcpy(tmp, sym_table, sizeof(sym_table[0])*sym_table_count);
+       qsort(tmp, sym_table_count, sizeof(tmp[0]), compare);
+
+       write(1, CONSOLE_CLEAR, strlen(CONSOLE_CLEAR));
+
+       printf(
+"------------------------------------------------------------------------------\n");
+       printf( " KernelTop:%8.0f irqs/sec  kernel:%3.1f%% [%s, ",
+               events_per_sec,
+               100.0 - (100.0*((events_per_sec-kevents_per_sec)/events_per_sec)),
+               nmi ? "NMI" : "IRQ");
+
+       if (nr_counters == 1)
+               printf("%d ", event_count[0]);
+
+       for (counter = 0; counter < nr_counters; counter++) {
+               if (counter)
+                       printf("/");
+
+               if (event_id[counter] < PERF_HW_EVENTS_MAX)
+                       printf( "%s", event_types[event_id[counter]]);
+               else
+                       printf( "raw:%04lx", event_id[counter]);
+       }
+
+       printf( "], ");
+
+       if (tid != -1)
+               printf(" (tid: %d", tid);
+       else
+               printf(" (all");
+
+       if (profile_cpu != -1)
+               printf(", cpu: %d)\n", profile_cpu);
+       else {
+               if (tid != -1)
+                       printf(")\n");
+               else
+                       printf(", %d CPUs)\n", nr_cpus);
+       }
+
+       printf("------------------------------------------------------------------------------\n\n");
+
+       if (nr_counters == 1)
+               printf("             events");
+       else
+               printf("  weight     events");
+
+       printf("         RIP          kernel function\n"
+                      "  ______     ______   ________________   _______________\n\n"
+       );
+
+       printed = 0;
+       for (i = 0; i < sym_table_count; i++) {
+               int count;
+
+               if (nr_counters == 1) {
+                       if (printed <= 18 &&
+                                       tmp[i].count[0] >= count_filter) {
+                               printf("%19.2f - %016llx : %s\n",
+                                 sym_weight(tmp + i), tmp[i].addr, tmp[i].sym);
+                               printed++;
+                       }
+               } else {
+                       if (printed <= 18 &&
+                                       tmp[i].count[0] >= count_filter) {
+                               printf("%8.1f %10ld - %016llx : %s\n",
+                                 sym_weight(tmp + i),
+                                 tmp[i].count[0],
+                                 tmp[i].addr, tmp[i].sym);
+                               printed++;
+                       }
+               }
+               /*
+                * Add decay to the counts:
+                */
+               for (count = 0; count < nr_counters; count++)
+                       sym_table[i].count[count] = zero ? 0 : sym_table[i].count[count] * 7 / 8;
+       }
+
+       if (sym_filter_entry)
+               show_details(sym_filter_entry);
+
+       last_refresh = time(NULL);
+
+       {
+               struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
+
+               if (poll(&stdin_poll, 1, 0) == 1) {
+                       printf("key pressed - exiting.\n");
+                       exit(0);
+               }
+       }
+}
+
+static int read_symbol(FILE *in, struct sym_entry *s)
+{
+       static int filter_match = 0;
+       char *sym, stype;
+       char str[500];
+       int rc, pos;
+
+       rc = fscanf(in, "%llx %c %499s", &s->addr, &stype, str);
+       if (rc == EOF)
+               return -1;
+
+       assert(rc == 3);
+
+       /* skip until end of line: */
+       pos = strlen(str);
+       do {
+               rc = fgetc(in);
+               if (rc == '\n' || rc == EOF || pos >= 499)
+                       break;
+               str[pos] = rc;
+               pos++;
+       } while (1);
+       str[pos] = 0;
+
+       sym = str;
+
+       /* Filter out known duplicates and non-text symbols. */
+       if (!strcmp(sym, "_text"))
+               return 1;
+       if (!min_ip && !strcmp(sym, "_stext"))
+               return 1;
+       if (!strcmp(sym, "_etext") || !strcmp(sym, "_sinittext"))
+               return 1;
+       if (stype != 'T' && stype != 't')
+               return 1;
+       if (!strncmp("init_module", sym, 11) || !strncmp("cleanup_module", sym, 14))
+               return 1;
+       if (strstr(sym, "_text_start") || strstr(sym, "_text_end"))
+               return 1;
+
+       s->sym = malloc(strlen(str));
+       assert(s->sym);
+
+       strcpy((char *)s->sym, str);
+       s->skip = 0;
+
+       /* Tag events to be skipped. */
+       if (!strcmp("default_idle", s->sym) || !strcmp("cpu_idle", s->sym))
+               s->skip = 1;
+       if (!strcmp("enter_idle", s->sym) || !strcmp("exit_idle", s->sym))
+               s->skip = 1;
+
+       if (filter_match == 1) {
+               filter_end = s->addr;
+               filter_match = -1;
+               if (filter_end - filter_start > 10000) {
+                       printf("hm, too large filter symbol <%s> - skipping.\n",
+                               sym_filter);
+                       printf("symbol filter start: %016lx\n", filter_start);
+                       printf("                end: %016lx\n", filter_end);
+                       filter_end = filter_start = 0;
+                       sym_filter = NULL;
+                       sleep(1);
+               }
+       }
+       if (filter_match == 0 && sym_filter && !strcmp(s->sym, sym_filter)) {
+               filter_match = 1;
+               filter_start = s->addr;
+       }
+
+       return 0;
+}
+
+int compare_addr(const void *__sym1, const void *__sym2)
+{
+       const struct sym_entry *sym1 = __sym1, *sym2 = __sym2;
+
+       return sym1->addr > sym2->addr;
+}
+
+static void sort_symbol_table(void)
+{
+       int i, dups;
+
+       do {
+               qsort(sym_table, sym_table_count, sizeof(sym_table[0]), compare_addr);
+               for (i = 0, dups = 0; i < sym_table_count; i++) {
+                       if (sym_table[i].addr == sym_table[i+1].addr) {
+                               sym_table[i+1].addr = -1ll;
+                               dups++;
+                       }
+               }
+               sym_table_count -= dups;
+       } while(dups);
+}
+
+static void parse_symbols(void)
+{
+       struct sym_entry *last;
+
+       FILE *kallsyms = fopen("/proc/kallsyms", "r");
+
+       if (!kallsyms) {
+               printf("Could not open /proc/kallsyms - no CONFIG_KALLSYMS_ALL=y?\n");
+               exit(-1);
+       }
+
+       while (!feof(kallsyms)) {
+               if (read_symbol(kallsyms, &sym_table[sym_table_count]) == 0) {
+                       sym_table_count++;
+                       assert(sym_table_count <= MAX_SYMS);
+               }
+       }
+
+       sort_symbol_table();
+       min_ip = sym_table[0].addr;
+       max_ip = sym_table[sym_table_count-1].addr;
+       last = sym_table + sym_table_count++;
+
+       last->addr = -1ll;
+       last->sym = "<end>";
+
+       if (filter_end) {
+               int count;
+               for (count=0; count < sym_table_count; count ++) {
+                       if (!strcmp(sym_table[count].sym, sym_filter)) {
+                               sym_filter_entry = &sym_table[count];
+                               break;
+                       }
+               }
+       }
+       if (dump_symtab) {
+               int i;
+
+               for (i = 0; i < sym_table_count; i++)
+                       fprintf(stderr, "%llx %s\n",
+                               sym_table[i].addr, sym_table[i].sym);
+       }
+}
+
+
+static void parse_vmlinux(char *filename)
+{
+       FILE *file;
+       char command[PATH_MAX*2];
+       if (!filename)
+               return;
+
+       sprintf(command, "objdump --start-address=0x%016lx --stop-address=0x%016lx -dS %s", filter_start, filter_end, filename);
+
+       file = popen(command, "r");
+       if (!file)
+               return;
+
+       while (!feof(file)) {
+               struct source_line *src;
+               size_t dummy = 0;
+               char *c;
+
+               src = malloc(sizeof(struct source_line));
+               assert(src != NULL);    
+               memset(src, 0, sizeof(struct source_line));
+
+               if (getline(&src->line, &dummy, file) < 0)
+                       break;
+               if (!src->line)
+                       break;
+
+               c = strchr(src->line, '\n');
+               if (c)
+                       *c = 0;
+
+               lines = g_list_prepend(lines, src);
+
+               if (strlen(src->line)>8 && src->line[8] == ':')
+                       src->EIP = strtoull(src->line, NULL, 16);
+               if (strlen(src->line)>8 && src->line[16] == ':')
+                       src->EIP = strtoull(src->line, NULL, 16);
+       }
+       pclose(file);
+       lines = g_list_reverse(lines);
+}
+
+static void record_precise_ip(uint64_t ip)
+{
+       struct source_line *line;
+       GList *item;
+
+       item = g_list_first(lines);
+       while (item) {
+               line = item->data;
+               if (line->EIP == ip)
+                       line->count++;
+               if (line->EIP > ip)
+                       break;
+               item = g_list_next(item);
+       }
+}
+
+static void lookup_sym_in_vmlinux(struct sym_entry *sym)
+{
+       struct source_line *line;
+       GList *item;
+       char pattern[PATH_MAX];
+       sprintf(pattern, "<%s>:", sym->sym);
+
+       item = g_list_first(lines);
+       while (item) {
+               line = item->data;
+               if (strstr(line->line, pattern)) {
+                       sym->source = item;
+                       break;
+               }
+               item = g_list_next(item);
+       }
+}
+
+void show_lines(GList *item_queue, int item_queue_count)
+{
+       int i;
+       struct source_line *line;
+
+       for (i = 0; i < item_queue_count; i++) {
+               line = item_queue->data;
+               printf("%8li\t%s\n", line->count, line->line);
+               item_queue = g_list_next(item_queue);
+       }
+}
+
+#define TRACE_COUNT     3
+
+static void show_details(struct sym_entry *sym)
+{
+       struct source_line *line;
+       GList *item;
+       int displayed = 0;
+       GList *item_queue = NULL;
+       int item_queue_count = 0;
+
+       if (!sym->source)
+               lookup_sym_in_vmlinux(sym);
+       if (!sym->source)
+               return;
+
+       printf("Showing details for %s\n", sym->sym);
+
+       item = sym->source;
+       while (item) {
+               line = item->data;
+               if (displayed && strstr(line->line, ">:"))
+                       break;
+
+               if (!item_queue_count)
+                       item_queue = item;
+               item_queue_count ++;
+
+               if (line->count >= count_filter) {
+                       show_lines(item_queue, item_queue_count);
+                       item_queue_count = 0;
+                       item_queue = NULL;
+               } else if (item_queue_count > TRACE_COUNT) {
+                       item_queue = g_list_next(item_queue);
+                       item_queue_count --;
+               }
+
+               line->count = 0;
+               displayed++;
+               if (displayed > 300)
+                       break;
+               item = g_list_next(item);
+       }
+}
+
+/*
+ * Binary search in the histogram table and record the hit:
+ */
+static void record_ip(uint64_t ip, int counter)
+{
+       int left_idx, middle_idx, right_idx, idx;
+       unsigned long left, middle, right;
+
+       record_precise_ip(ip);
+
+       left_idx = 0;
+       right_idx = sym_table_count-1;
+       assert(ip <= max_ip && ip >= min_ip);
+
+       while (left_idx + 1 < right_idx) {
+               middle_idx = (left_idx + right_idx) / 2;
+
+               left   = sym_table[  left_idx].addr;
+               middle = sym_table[middle_idx].addr;
+               right  = sym_table[ right_idx].addr;
+
+               if (!(left <= middle && middle <= right)) {
+                       printf("%016lx...\n%016lx...\n%016lx\n", left, middle, right);
+                       printf("%d %d %d\n", left_idx, middle_idx, right_idx);
+               }
+               assert(left <= middle && middle <= right);
+               if (!(left <= ip && ip <= right)) {
+                       printf(" left: %016lx\n", left);
+                       printf("   ip: %016lx\n", ip);
+                       printf("right: %016lx\n", right);
+               }
+               assert(left <= ip && ip <= right);
+               /*
+                * [ left .... target .... middle .... right ]
+                *   => right := middle
+                */
+               if (ip < middle) {
+                       right_idx = middle_idx;
+                       continue;
+               }
+               /*
+                * [ left .... middle ... target ... right ]
+                *   => left := middle
+                */
+               left_idx = middle_idx;
+       }
+
+       idx = left_idx;
+
+       if (!sym_table[idx].skip)
+               sym_table[idx].count[counter]++;
+       else events--;
+}
+
+static void process_event(uint64_t ip, int counter)
+{
+       events++;
+
+       if (ip < min_ip || ip > max_ip) {
+               userspace_events++;
+               return;
+       }
+
+       record_ip(ip, counter);
+}
+
+int main(int argc, char *argv[])
+{
+       struct pollfd event_array[MAX_NR_CPUS][MAX_COUNTERS];
+       struct perf_counter_hw_event hw_event;
+       int fd[MAX_NR_CPUS][MAX_COUNTERS];
+       int i, counter, group_fd;
+       unsigned int cpu;
+       uint64_t ip;
+       ssize_t res;
+       int ret;
+
+       process_options(argc, argv);
+
+       nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+       if (tid != -1 || profile_cpu != -1)
+               nr_cpus = 1;
+
+       assert(nr_cpus <= MAX_NR_CPUS);
+
+       for (i = 0; i < nr_cpus; i++) {
+               group_fd = -1;
+               for (counter = 0; counter < nr_counters; counter++) {
+
+                       cpu     = profile_cpu;
+                       if (tid == -1 && profile_cpu == -1)
+                               cpu = i;
+
+                       memset(&hw_event, 0, sizeof(hw_event));
+                       hw_event.type           = event_id[counter];
+                       hw_event.raw            = event_raw[counter];
+                       hw_event.irq_period     = event_count[counter];
+                       hw_event.record_type    = PERF_RECORD_IRQ;
+                       hw_event.nmi            = nmi;
+
+                       fd[i][counter] = sys_perf_counter_open(&hw_event, tid, cpu, group_fd, 0);
+                       fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
+                       if (fd[i][counter] < 0) {
+                               printf("kerneltop error: syscall returned with %d (%s)\n",
+                                       fd[i][counter], strerror(-fd[i][counter]));
+                               if (fd[i][counter] == -1)
+                                       printf("Are you root?\n");
+                               exit(-1);
+                       }
+                       assert(fd[i][counter] >= 0);
+
+                       /*
+                        * First counter acts as the group leader:
+                        */
+                       if (group && group_fd == -1)
+                               group_fd = fd[i][counter];
+
+                       event_array[i][counter].fd = fd[i][counter];
+                       event_array[i][counter].events = POLLIN;
+               }
+       }
+
+       parse_symbols();
+       if (vmlinux && sym_filter_entry)
+               parse_vmlinux(vmlinux);
+
+       printf("KernelTop refresh period: %d seconds\n", delay_secs);
+       last_refresh = time(NULL);
+
+       while (1) {
+               int hits = events;
+
+               for (i = 0; i < nr_cpus; i++) {
+                       for (counter = 0; counter < nr_counters; counter++) {
+                               res = read(fd[i][counter], (char *) &ip, sizeof(ip));
+                               if (res > 0) {
+                                       assert(res == sizeof(ip));
+
+                                       process_event(ip, counter);
+                               }
+                       }
+               }
+
+               if (time(NULL) >= last_refresh + delay_secs) {
+                       print_sym_table();
+                       events = userspace_events = 0;
+               }
+
+               if (hits == events)
+                       ret = poll(event_array[0], nr_cpus, 1000);
+               hits = events;
+       }
+
+       return 0;
+}
diff --git a/Documentation/perf_counter/perfstat.c b/Documentation/perf_counter/perfstat.c
new file mode 100644 (file)
index 0000000..9a5808f
--- /dev/null
@@ -0,0 +1,521 @@
+/*
+ * perfstat:  /usr/bin/time -alike performance counter statistics utility
+ *
+ *        It summarizes the counter events of all tasks (and child tasks),
+ *        covering all CPUs that the command (or workload) executes on.
+ *        It only counts the per-task events of the workload started,
+ *        independent of how many other tasks run on those CPUs.
+ *
+ * Build with:       cc -O2 -g -lrt -Wall -W -o perfstat perfstat.c
+ *
+ * Sample output:
+ *
+
+   $ ./perfstat -e 1 -e 3 -e 5 ls -lR /usr/include/ >/dev/null
+
+   Performance counter stats for 'ls':
+
+           163516953 instructions
+                2295 cache-misses
+             2855182 branch-misses
+
+ *
+ * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
+ *
+ * Released under the GPLv2 (not later).
+ *
+ * Percpu counter support by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
+ * Symbolic event options by: Wu Fengguang <fengguang.wu@intel.com>
+ */
+#define _GNU_SOURCE
+
+#include <assert.h>
+#include <getopt.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <ctype.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <time.h>
+
+#include <sys/syscall.h>
+#include <sys/ioctl.h>
+#include <sys/prctl.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+#include <sys/wait.h>
+#include <sys/uio.h>
+
+#include <linux/unistd.h>
+
+#ifdef __x86_64__
+# define __NR_perf_counter_open        295
+#endif
+
+#ifdef __i386__
+# define __NR_perf_counter_open 333
+#endif
+
+#ifdef __powerpc__
+#define __NR_perf_counter_open 319
+#endif
+
+/*
+ * Pick up some kernel type conventions:
+ */
+#define __user
+#define asmlinkage
+
+typedef unsigned int           __u32;
+typedef unsigned long long     __u64;
+typedef long long              __s64;
+
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
+
+/*
+ * User-space ABI bits:
+ */
+
+/*
+ * Generalized performance counter event types, used by the hw_event.type
+ * parameter of the sys_perf_counter_open() syscall:
+ */
+enum hw_event_types {
+       /*
+        * Common hardware events, generalized by the kernel:
+        */
+       PERF_COUNT_CPU_CYCLES           =  0,
+       PERF_COUNT_INSTRUCTIONS         =  1,
+       PERF_COUNT_CACHE_REFERENCES     =  2,
+       PERF_COUNT_CACHE_MISSES         =  3,
+       PERF_COUNT_BRANCH_INSTRUCTIONS  =  4,
+       PERF_COUNT_BRANCH_MISSES        =  5,
+       PERF_COUNT_BUS_CYCLES           =  6,
+
+       PERF_HW_EVENTS_MAX              =  7,
+
+       /*
+        * Special "software" counters provided by the kernel, even if
+        * the hardware does not support performance counters. These
+        * counters measure various physical and sw events of the
+        * kernel (and allow the profiling of them as well):
+        */
+       PERF_COUNT_CPU_CLOCK            = -1,
+       PERF_COUNT_TASK_CLOCK           = -2,
+       PERF_COUNT_PAGE_FAULTS          = -3,
+       PERF_COUNT_CONTEXT_SWITCHES     = -4,
+       PERF_COUNT_CPU_MIGRATIONS       = -5,
+
+       PERF_SW_EVENTS_MIN              = -6,
+};
+
+/*
+ * IRQ-notification data record type:
+ */
+enum perf_counter_record_type {
+       PERF_RECORD_SIMPLE              =  0,
+       PERF_RECORD_IRQ                 =  1,
+       PERF_RECORD_GROUP               =  2,
+};
+
+/*
+ * Hardware event to monitor via a performance monitoring counter:
+ */
+struct perf_counter_hw_event {
+       __s64                   type;
+
+       __u64                   irq_period;
+       __u64                   record_type;
+       __u64                   read_format;
+
+       __u64                   disabled       :  1, /* off by default        */
+                               nmi            :  1, /* NMI sampling          */
+                               raw            :  1, /* raw event type        */
+                               inherit        :  1, /* children inherit it   */
+                               pinned         :  1, /* must always be on PMU */
+                               exclusive      :  1, /* only group on PMU     */
+                               exclude_user   :  1, /* don't count user      */
+                               exclude_kernel :  1, /* ditto kernel          */
+                               exclude_hv     :  1, /* ditto hypervisor      */
+                               exclude_idle   :  1, /* don't count when idle */
+
+                               __reserved_1   : 54;
+
+       __u32                   extra_config_len;
+       __u32                   __reserved_4;
+
+       __u64                   __reserved_2;
+       __u64                   __reserved_3;
+};
+
+/*
+ * Ioctls that can be done on a perf counter fd:
+ */
+#define PERF_COUNTER_IOC_ENABLE                _IO('$', 0)
+#define PERF_COUNTER_IOC_DISABLE       _IO('$', 1)
+
+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)
+{
+       int ret;
+
+       ret = syscall(
+               __NR_perf_counter_open, hw_event_uptr, pid, cpu, group_fd, flags);
+#if defined(__x86_64__) || defined(__i386__)
+       if (ret < 0 && ret > -4096) {
+               errno = -ret;
+               ret = -1;
+       }
+#endif
+       return ret;
+}
+
+
+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",
+};
+
+struct event_symbol {
+       int event;
+       char *symbol;
+};
+
+static struct event_symbol event_symbols [] = {
+       {PERF_COUNT_CPU_CYCLES,                 "cpu-cycles",           },
+       {PERF_COUNT_CPU_CYCLES,                 "cycles",               },
+       {PERF_COUNT_INSTRUCTIONS,               "instructions",         },
+       {PERF_COUNT_CACHE_REFERENCES,           "cache-references",     },
+       {PERF_COUNT_CACHE_MISSES,               "cache-misses",         },
+       {PERF_COUNT_BRANCH_INSTRUCTIONS,        "branch-instructions",  },
+       {PERF_COUNT_BRANCH_INSTRUCTIONS,        "branches",             },
+       {PERF_COUNT_BRANCH_MISSES,              "branch-misses",        },
+       {PERF_COUNT_BUS_CYCLES,                 "bus-cycles",           },
+       {PERF_COUNT_CPU_CLOCK,                  "cpu-ticks",            },
+       {PERF_COUNT_CPU_CLOCK,                  "ticks",                },
+       {PERF_COUNT_TASK_CLOCK,                 "task-ticks",           },
+       {PERF_COUNT_PAGE_FAULTS,                "page-faults",          },
+       {PERF_COUNT_PAGE_FAULTS,                "faults",               },
+       {PERF_COUNT_CONTEXT_SWITCHES,           "context-switches",     },
+       {PERF_COUNT_CONTEXT_SWITCHES,           "cs",                   },
+       {PERF_COUNT_CPU_MIGRATIONS,             "cpu-migrations",       },
+       {PERF_COUNT_CPU_MIGRATIONS,             "migrations",           },
+};
+
+#define MAX_COUNTERS                                   64
+#define MAX_NR_CPUS                                    256
+
+static int                     nr_counters             = 0;
+static int                     nr_cpus                 = 0;
+
+static int                     event_id[MAX_COUNTERS]  =
+                                        { -2, -5, -4, -3, 0, 1, 2, 3};
+
+static int                     event_raw[MAX_COUNTERS];
+
+static int                     system_wide             = 0;
+
+static void display_help(void)
+{
+       unsigned int i;
+       int e;
+
+       printf(
+       "Usage: perfstat [<events...>] <cmd...>\n\n"
+       "PerfStat Options (up to %d event types can be specified):\n\n",
+                MAX_COUNTERS);
+       printf(
+       " -e EVENT     --event=EVENT        #  symbolic-name        abbreviations");
+
+       for (i = 0, e = PERF_HW_EVENTS_MAX; i < ARRAY_SIZE(event_symbols); i++) {
+               if (e != event_symbols[i].event) {
+                       e = event_symbols[i].event;
+                       printf(
+       "\n                                  %2d: %-20s", e, event_symbols[i].symbol);
+               } else
+                       printf(" %s", event_symbols[i].symbol);
+       }
+
+       printf("\n"
+       "                                rNNN: raw event type\n\n"
+       " -s                                # system-wide collection\n\n"
+       " -c <cmd..>   --command=<cmd..>    # command+arguments to be timed.\n"
+       "\n");
+       exit(0);
+}
+
+static int type_valid(int type)
+{
+       if (type >= PERF_HW_EVENTS_MAX)
+               return 0;
+       if (type <= PERF_SW_EVENTS_MIN)
+               return 0;
+
+       return 1;
+}
+
+static char *event_name(int ctr)
+{
+       int type = event_id[ctr];
+       static char buf[32];
+
+       if (event_raw[ctr]) {
+               sprintf(buf, "raw 0x%x", type);
+               return buf;
+       }
+       if (!type_valid(type))
+               return "unknown";
+
+       if (type >= 0)
+               return hw_event_names[type];
+
+       return sw_event_names[-type-1];
+}
+
+/*
+ * Each event can have multiple symbolic names.
+ * Symbolic names are (almost) exactly matched.
+ */
+static int match_event_symbols(char *str)
+{
+       unsigned int i;
+
+       if (isdigit(str[0]) || str[0] == '-')
+               return atoi(str);
+
+       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 PERF_HW_EVENTS_MAX;
+}
+
+static void parse_events(char *str)
+{
+       int type, raw;
+
+again:
+       nr_counters++;
+       if (nr_counters == MAX_COUNTERS)
+               display_help();
+
+       raw = 0;
+       if (*str == 'r') {
+               raw = 1;
+               ++str;
+               type = strtol(str, NULL, 16);
+       } else {
+               type = match_event_symbols(str);
+               if (!type_valid(type))
+                       display_help();
+       }
+
+       event_id[nr_counters] = type;
+       event_raw[nr_counters] = raw;
+
+       str = strstr(str, ",");
+       if (str) {
+               str++;
+               goto again;
+       }
+}
+
+static void process_options(int argc, char *argv[])
+{
+       for (;;) {
+               int option_index = 0;
+               /** Options for getopt */
+               static struct option long_options[] = {
+                       {"event",       required_argument,      NULL, 'e'},
+                       {"help",        no_argument,            NULL, 'h'},
+                       {"command",     no_argument,            NULL, 'c'},
+                       {NULL,          0,                      NULL,  0 }
+               };
+               int c = getopt_long(argc, argv, "+:e:c:s",
+                                   long_options, &option_index);
+               if (c == -1)
+                       break;
+
+               switch (c) {
+               case 'c':
+                       break;
+               case 's':
+                       system_wide = 1;
+                       break;
+               case 'e':
+                       parse_events(optarg);
+                       break;
+               default:
+                       break;
+               }
+       }
+       if (optind == argc)
+               goto err;
+
+       if (!nr_counters)
+               nr_counters = 8;
+       else
+               nr_counters++;
+       return;
+
+err:
+       display_help();
+}
+
+char fault_here[1000000];
+
+#define PR_TASK_PERF_COUNTERS_DISABLE           31
+#define PR_TASK_PERF_COUNTERS_ENABLE            32
+
+static int fd[MAX_NR_CPUS][MAX_COUNTERS];
+
+static void create_counter(int counter)
+{
+       struct perf_counter_hw_event hw_event;
+
+       memset(&hw_event, 0, sizeof(hw_event));
+       hw_event.type           = event_id[counter];
+       hw_event.raw            = event_raw[counter];
+       hw_event.record_type    = PERF_RECORD_SIMPLE;
+       hw_event.nmi            = 0;
+
+       if (system_wide) {
+               int cpu;
+               for (cpu = 0; cpu < nr_cpus; cpu ++) {
+                       fd[cpu][counter] = sys_perf_counter_open(&hw_event, -1, cpu, -1, 0);
+                       if (fd[cpu][counter] < 0) {
+                               printf("perfstat error: syscall returned with %d (%s)\n",
+                                               fd[cpu][counter], strerror(errno));
+                               exit(-1);
+                       }
+                       
+               }
+       } else {
+               hw_event.inherit        = 1;
+               hw_event.disabled       = 1;
+
+               fd[0][counter] = sys_perf_counter_open(&hw_event, 0, -1, -1, 0);
+               if (fd[0][counter] < 0) {
+                       printf("perfstat error: syscall returned with %d (%s)\n",
+                                       fd[0][counter], strerror(errno));
+                       exit(-1);
+               }
+       }
+}
+
+
+#define rdclock()                                      \
+({                                                     \
+       struct timespec ts;                             \
+                                                       \
+       clock_gettime(CLOCK_MONOTONIC, &ts);            \
+       ts.tv_sec * 1000000000ULL + ts.tv_nsec;         \
+})
+
+int main(int argc, char *argv[])
+{
+       unsigned long long t0, t1;
+       int counter;
+       ssize_t res;
+       int status;
+       int pid;
+
+       process_options(argc, argv);
+
+       if (system_wide) {
+               nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+               assert(nr_cpus <= MAX_NR_CPUS);
+               assert(nr_cpus >= 0);
+       } else
+               nr_cpus = 1;
+
+       for (counter = 0; counter < nr_counters; counter++)
+               create_counter(counter);
+
+       argc -= optind;
+       argv += optind;
+
+       /*
+        * Enable counters and exec the command:
+        */
+       t0 = rdclock();
+       prctl(PR_TASK_PERF_COUNTERS_ENABLE);
+
+       if ((pid = fork()) < 0)
+               perror("failed to fork");
+       if (!pid) {
+               if (execvp(argv[0], argv)) {
+                       perror(argv[0]);
+                       exit(-1);
+               }
+       }
+       while (wait(&status) >= 0)
+               ;
+       prctl(PR_TASK_PERF_COUNTERS_DISABLE);
+       t1 = rdclock();
+
+       fflush(stdout);
+
+       fprintf(stderr, "\n");
+       fprintf(stderr, " Performance counter stats for \'%s\':\n",
+               argv[0]);
+       fprintf(stderr, "\n");
+
+       for (counter = 0; counter < nr_counters; counter++) {
+               int cpu;
+               __u64 count, single_count;
+
+               count = 0;
+               for (cpu = 0; cpu < nr_cpus; cpu ++) {
+                       res = read(fd[cpu][counter],
+                                       (char *) &single_count, sizeof(single_count));
+                       assert(res == sizeof(single_count));
+                       count += single_count;
+               }
+
+               if (!event_raw[counter] &&
+                   (event_id[counter] == PERF_COUNT_CPU_CLOCK ||
+                    event_id[counter] == PERF_COUNT_TASK_CLOCK)) {
+
+                       double msecs = (double)count / 1000000;
+
+                       fprintf(stderr, " %14.6f  %-20s (msecs)\n",
+                               msecs, event_name(counter));
+               } else {
+                       fprintf(stderr, " %14Ld  %-20s (events)\n",
+                               count, event_name(counter));
+               }
+               if (!counter)
+                       fprintf(stderr, "\n");
+       }
+       fprintf(stderr, "\n");
+       fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n",
+                       (double)(t1-t0)/1e6);
+       fprintf(stderr, "\n");
+
+       return 0;
+}