Merge branch 'perf/live' into perf/core
[safe/jmp/linux-2.6] / tools / perf / builtin-record.c
index 66979a5..a1b99ee 100644 (file)
@@ -5,49 +5,62 @@
  * (or a CPU, or a PID) into the perf.data output file - for
  * later analysis via perf report.
  */
+#define _FILE_OFFSET_BITS 64
+
 #include "builtin.h"
 
 #include "perf.h"
 
+#include "util/build-id.h"
 #include "util/util.h"
 #include "util/parse-options.h"
 #include "util/parse-events.h"
-#include "util/string.h"
 
 #include "util/header.h"
 #include "util/event.h"
 #include "util/debug.h"
 #include "util/session.h"
 #include "util/symbol.h"
+#include "util/cpumap.h"
 
 #include <unistd.h>
 #include <sched.h>
 
-static int                     fd[MAX_NR_CPUS][MAX_COUNTERS];
+enum write_mode_t {
+       WRITE_FORCE,
+       WRITE_APPEND
+};
+
+static int                     *fd[MAX_NR_CPUS][MAX_COUNTERS];
 
+static unsigned int            user_interval                   = UINT_MAX;
 static long                    default_interval                =      0;
 
 static int                     nr_cpus                         =      0;
 static unsigned int            page_size;
 static unsigned int            mmap_pages                      =    128;
+static unsigned int            user_freq                       = UINT_MAX;
 static int                     freq                            =   1000;
 static int                     output;
+static int                     pipe_output                     =      0;
 static const char              *output_name                    = "perf.data";
 static int                     group                           =      0;
 static unsigned int            realtime_prio                   =      0;
-static int                     raw_samples                     =      0;
-static int                     system_wide                     =      0;
+static bool                    raw_samples                     =  false;
+static bool                    system_wide                     =  false;
 static int                     profile_cpu                     =     -1;
 static pid_t                   target_pid                      =     -1;
+static pid_t                   target_tid                      =     -1;
+static pid_t                   *all_tids                       =      NULL;
+static int                     thread_num                      =      0;
 static pid_t                   child_pid                       =     -1;
-static int                     inherit                         =      1;
-static int                     force                           =      0;
-static int                     append_file                     =      0;
-static int                     call_graph                      =      0;
-static int                     inherit_stat                    =      0;
-static int                     no_samples                      =      0;
-static int                     sample_address                  =      0;
-static int                     multiplex                       =      0;
+static bool                    inherit                         =   true;
+static enum write_mode_t       write_mode                      = WRITE_FORCE;
+static bool                    call_graph                      =  false;
+static bool                    inherit_stat                    =  false;
+static bool                    no_samples                      =  false;
+static bool                    sample_address                  =  false;
+static bool                    multiplex                       =  false;
 static int                     multiplex_fd                    =     -1;
 
 static long                    samples                         =      0;
@@ -56,12 +69,13 @@ static struct timeval               this_read;
 
 static u64                     bytes_written                   =      0;
 
-static struct pollfd           event_array[MAX_NR_CPUS * MAX_COUNTERS];
+static struct pollfd           *event_array;
 
 static int                     nr_poll                         =      0;
 static int                     nr_cpu                          =      0;
 
 static int                     file_new                        =      1;
+static off_t                   post_processing_offset;
 
 static struct perf_session     *session;
 
@@ -72,7 +86,7 @@ struct mmap_data {
        unsigned int            prev;
 };
 
-static struct mmap_data                mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
+static struct mmap_data                *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
 
 static unsigned long mmap_read_head(struct mmap_data *md)
 {
@@ -96,6 +110,11 @@ static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
        pc->data_tail = tail;
 }
 
+static void advance_output(size_t size)
+{
+       bytes_written += size;
+}
+
 static void write_output(void *buf, size_t size)
 {
        while (size) {
@@ -111,22 +130,10 @@ static void write_output(void *buf, size_t size)
        }
 }
 
-static void write_event(event_t *buf, size_t size)
-{
-       /*
-       * Add it to the list of DSOs, so that when we finish this
-        * record session we can pick the available build-ids.
-        */
-       if (buf->header.type == PERF_RECORD_MMAP)
-               dsos__findnew(buf->mmap.filename);
-
-       write_output(buf, size);
-}
-
 static int process_synthesized_event(event_t *event,
                                     struct perf_session *self __used)
 {
-       write_event(event, event->header.size);
+       write_output(event, event->header.size);
        return 0;
 }
 
@@ -178,14 +185,14 @@ static void mmap_read(struct mmap_data *md)
                size = md->mask + 1 - (old & md->mask);
                old += size;
 
-               write_event(buf, size);
+               write_output(buf, size);
        }
 
        buf = &data[old & md->mask];
        size = head - old;
        old += size;
 
-       write_event(buf, size);
+       write_output(buf, size);
 
        md->prev = old;
        mmap_write_tail(md, old);
@@ -232,12 +239,13 @@ static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int n
        return h_attr;
 }
 
-static void create_counter(int counter, int cpu, pid_t pid)
+static void create_counter(int counter, int cpu)
 {
        char *filter = filters[counter];
        struct perf_event_attr *attr = attrs + counter;
        struct perf_header_attr *h_attr;
        int track = !counter; /* only the first counter needs these */
+       int thread_index;
        int ret;
        struct {
                u64 count;
@@ -252,10 +260,22 @@ static void create_counter(int counter, int cpu, pid_t pid)
 
        attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 
-       if (freq) {
-               attr->sample_type       |= PERF_SAMPLE_PERIOD;
-               attr->freq              = 1;
-               attr->sample_freq       = freq;
+       if (nr_counters > 1)
+               attr->sample_type |= PERF_SAMPLE_ID;
+
+       /*
+        * We default some events to a 1 default interval. But keep
+        * it a weak assumption overridable by the user.
+        */
+       if (!attr->sample_period || (user_freq != UINT_MAX &&
+                                    user_interval != UINT_MAX)) {
+               if (freq) {
+                       attr->sample_type       |= PERF_SAMPLE_PERIOD;
+                       attr->freq              = 1;
+                       attr->sample_freq       = freq;
+               } else {
+                       attr->sample_period = default_interval;
+               }
        }
 
        if (no_samples)
@@ -278,128 +298,162 @@ static void create_counter(int counter, int cpu, pid_t pid)
 
        attr->mmap              = track;
        attr->comm              = track;
-       attr->inherit           = (cpu < 0) && inherit;
-       attr->disabled          = 1;
+       attr->inherit           = inherit;
+       if (target_pid == -1 && !system_wide) {
+               attr->disabled = 1;
+               attr->enable_on_exec = 1;
+       }
 
+       for (thread_index = 0; thread_index < thread_num; thread_index++) {
 try_again:
-       fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
-
-       if (fd[nr_cpu][counter] < 0) {
-               int err = errno;
-
-               if (err == EPERM || err == EACCES)
-                       die("Permission error - are you root?\n");
-               else if (err ==  ENODEV && profile_cpu != -1)
-                       die("No such device - did you specify an out-of-range profile CPU?\n");
+               fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
+                               all_tids[thread_index], cpu, group_fd, 0);
+
+               if (fd[nr_cpu][counter][thread_index] < 0) {
+                       int err = errno;
+
+                       if (err == EPERM || err == EACCES)
+                               die("Permission error - are you root?\n"
+                                       "\t Consider tweaking"
+                                       " /proc/sys/kernel/perf_event_paranoid.\n");
+                       else if (err ==  ENODEV && profile_cpu != -1) {
+                               die("No such device - did you specify"
+                                       " an out-of-range profile CPU?\n");
+                       }
 
-               /*
-                * If it's cycles then fall back to hrtimer
-                * based cpu-clock-tick sw counter, which
-                * is always available even if no PMU support:
-                */
-               if (attr->type == PERF_TYPE_HARDWARE
-                       && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
-
-                       if (verbose)
-                               warning(" ... trying to fall back to cpu-clock-ticks\n");
-                       attr->type = PERF_TYPE_SOFTWARE;
-                       attr->config = PERF_COUNT_SW_CPU_CLOCK;
-                       goto try_again;
-               }
-               printf("\n");
-               error("perfcounter syscall returned with %d (%s)\n",
-                       fd[nr_cpu][counter], strerror(err));
+                       /*
+                        * If it's cycles then fall back to hrtimer
+                        * based cpu-clock-tick sw counter, which
+                        * is always available even if no PMU support:
+                        */
+                       if (attr->type == PERF_TYPE_HARDWARE
+                                       && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
+
+                               if (verbose)
+                                       warning(" ... trying to fall back to cpu-clock-ticks\n");
+                               attr->type = PERF_TYPE_SOFTWARE;
+                               attr->config = PERF_COUNT_SW_CPU_CLOCK;
+                               goto try_again;
+                       }
+                       printf("\n");
+                       error("perfcounter syscall returned with %d (%s)\n",
+                                       fd[nr_cpu][counter][thread_index], strerror(err));
 
 #if defined(__i386__) || defined(__x86_64__)
-               if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
-                       die("No hardware sampling interrupt available. No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.\n");
+                       if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
+                               die("No hardware sampling interrupt available."
+                                   " No APIC? If so then you can boot the kernel"
+                                   " with the \"lapic\" boot parameter to"
+                                   " force-enable it.\n");
 #endif
 
-               die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
-               exit(-1);
-       }
+                       die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
+                       exit(-1);
+               }
 
-       h_attr = get_header_attr(attr, counter);
-       if (h_attr == NULL)
-               die("nomem\n");
+               h_attr = get_header_attr(attr, counter);
+               if (h_attr == NULL)
+                       die("nomem\n");
 
-       if (!file_new) {
-               if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
-                       fprintf(stderr, "incompatible append\n");
-                       exit(-1);
+               if (!file_new) {
+                       if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
+                               fprintf(stderr, "incompatible append\n");
+                               exit(-1);
+                       }
                }
-       }
 
-       if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
-               perror("Unable to read perf file descriptor\n");
-               exit(-1);
-       }
+               if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
+                       perror("Unable to read perf file descriptor\n");
+                       exit(-1);
+               }
 
-       if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
-               pr_warning("Not enough memory to add id\n");
-               exit(-1);
-       }
+               if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
+                       pr_warning("Not enough memory to add id\n");
+                       exit(-1);
+               }
 
-       assert(fd[nr_cpu][counter] >= 0);
-       fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
+               assert(fd[nr_cpu][counter][thread_index] >= 0);
+               fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
 
-       /*
-        * First counter acts as the group leader:
-        */
-       if (group && group_fd == -1)
-               group_fd = fd[nr_cpu][counter];
-       if (multiplex && multiplex_fd == -1)
-               multiplex_fd = fd[nr_cpu][counter];
+               /*
+                * First counter acts as the group leader:
+                */
+               if (group && group_fd == -1)
+                       group_fd = fd[nr_cpu][counter][thread_index];
+               if (multiplex && multiplex_fd == -1)
+                       multiplex_fd = fd[nr_cpu][counter][thread_index];
 
-       if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
+               if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
 
-               ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
-               assert(ret != -1);
-       } else {
-               event_array[nr_poll].fd = fd[nr_cpu][counter];
-               event_array[nr_poll].events = POLLIN;
-               nr_poll++;
-
-               mmap_array[nr_cpu][counter].counter = counter;
-               mmap_array[nr_cpu][counter].prev = 0;
-               mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
-               mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
-                               PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
-               if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
-                       error("failed to mmap with %d (%s)\n", errno, strerror(errno));
-                       exit(-1);
+                       ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
+                       assert(ret != -1);
+               } else {
+                       event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
+                       event_array[nr_poll].events = POLLIN;
+                       nr_poll++;
+
+                       mmap_array[nr_cpu][counter][thread_index].counter = counter;
+                       mmap_array[nr_cpu][counter][thread_index].prev = 0;
+                       mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
+                       mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
+                               PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
+                       if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
+                               error("failed to mmap with %d (%s)\n", errno, strerror(errno));
+                               exit(-1);
+                       }
                }
-       }
 
-       if (filter != NULL) {
-               ret = ioctl(fd[nr_cpu][counter],
-                           PERF_EVENT_IOC_SET_FILTER, filter);
-               if (ret) {
-                       error("failed to set filter with %d (%s)\n", errno,
-                             strerror(errno));
-                       exit(-1);
+               if (filter != NULL) {
+                       ret = ioctl(fd[nr_cpu][counter][thread_index],
+                                       PERF_EVENT_IOC_SET_FILTER, filter);
+                       if (ret) {
+                               error("failed to set filter with %d (%s)\n", errno,
+                                               strerror(errno));
+                               exit(-1);
+                       }
                }
        }
-
-       ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
 }
 
-static void open_counters(int cpu, pid_t pid)
+static void open_counters(int cpu)
 {
        int counter;
 
        group_fd = -1;
        for (counter = 0; counter < nr_counters; counter++)
-               create_counter(counter, cpu, pid);
+               create_counter(counter, cpu);
 
        nr_cpu++;
 }
 
+static int process_buildids(void)
+{
+       u64 size = lseek(output, 0, SEEK_CUR);
+
+       if (size == 0)
+               return 0;
+
+       session->fd = output;
+       return __perf_session__process_events(session, post_processing_offset,
+                                             size - post_processing_offset,
+                                             size, &build_id__mark_dso_hit_ops);
+}
+
 static void atexit_header(void)
 {
-       session->header.data_size += bytes_written;
+       if (!pipe_output) {
+               session->header.data_size += bytes_written;
+
+               process_buildids();
+               perf_header__write(&session->header, output, true);
+       } else {
+               int err;
 
-       perf_header__write(&session->header, output, true);
+               err = event__synthesize_build_ids(process_synthesized_event,
+                                                 session);
+               if (err < 0)
+                       pr_err("Couldn't synthesize build ids.\n");
+       }
 }
 
 static int __cmd_record(int argc, const char **argv)
@@ -410,55 +464,59 @@ static int __cmd_record(int argc, const char **argv)
        int flags;
        int err;
        unsigned long waking = 0;
+       int child_ready_pipe[2], go_pipe[2];
+       const bool forks = argc > 0;
+       char buf;
 
        page_size = sysconf(_SC_PAGE_SIZE);
-       nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
-       assert(nr_cpus <= MAX_NR_CPUS);
-       assert(nr_cpus >= 0);
 
        atexit(sig_atexit);
        signal(SIGCHLD, sig_handler);
        signal(SIGINT, sig_handler);
 
-       if (!stat(output_name, &st) && st.st_size) {
-               if (!force) {
-                       if (!append_file) {
-                               pr_err("Error, output file %s exists, use -A "
-                                      "to append or -f to overwrite.\n",
-                                      output_name);
-                               exit(-1);
-                       }
-               } else {
+       if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
+               perror("failed to create pipes");
+               exit(-1);
+       }
+
+       if (!strcmp(output_name, "-"))
+               pipe_output = 1;
+       else if (!stat(output_name, &st) && st.st_size) {
+               if (write_mode == WRITE_FORCE) {
                        char oldname[PATH_MAX];
                        snprintf(oldname, sizeof(oldname), "%s.old",
                                 output_name);
                        unlink(oldname);
                        rename(output_name, oldname);
                }
-       } else {
-               append_file = 0;
+       } else if (write_mode == WRITE_APPEND) {
+               write_mode = WRITE_FORCE;
        }
 
        flags = O_CREAT|O_RDWR;
-       if (append_file)
+       if (write_mode == WRITE_APPEND)
                file_new = 0;
        else
                flags |= O_TRUNC;
 
-       output = open(output_name, flags, S_IRUSR|S_IWUSR);
+       if (pipe_output)
+               output = STDOUT_FILENO;
+       else
+               output = open(output_name, flags, S_IRUSR | S_IWUSR);
        if (output < 0) {
                perror("failed to create output file");
                exit(-1);
        }
 
-       session = perf_session__new(output_name, O_WRONLY, force, NULL);
+       session = perf_session__new(output_name, O_WRONLY,
+                                   write_mode == WRITE_FORCE);
        if (session == NULL) {
                pr_err("Not enough memory for reading perf file header\n");
                return -1;
        }
 
        if (!file_new) {
-               err = perf_header__read(&session->header, output);
+               err = perf_header__read(session, output);
                if (err < 0)
                        return err;
        }
@@ -476,58 +534,129 @@ static int __cmd_record(int argc, const char **argv)
 
        atexit(atexit_header);
 
-       if (!system_wide) {
-               pid = target_pid;
-               if (pid == -1)
-                       pid = getpid();
+       if (forks) {
+               child_pid = fork();
+               if (pid < 0) {
+                       perror("failed to fork");
+                       exit(-1);
+               }
 
-               open_counters(profile_cpu, pid);
-       } else {
-               if (profile_cpu != -1) {
-                       open_counters(profile_cpu, target_pid);
-               } else {
-                       for (i = 0; i < nr_cpus; i++)
-                               open_counters(i, target_pid);
+               if (!child_pid) {
+                       if (pipe_output)
+                               dup2(2, 1);
+                       close(child_ready_pipe[0]);
+                       close(go_pipe[1]);
+                       fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
+
+                       /*
+                        * Do a dummy execvp to get the PLT entry resolved,
+                        * so we avoid the resolver overhead on the real
+                        * execvp call.
+                        */
+                       execvp("", (char **)argv);
+
+                       /*
+                        * Tell the parent we're ready to go
+                        */
+                       close(child_ready_pipe[1]);
+
+                       /*
+                        * Wait until the parent tells us to go.
+                        */
+                       if (read(go_pipe[0], &buf, 1) == -1)
+                               perror("unable to read pipe");
+
+                       execvp(argv[0], (char **)argv);
+
+                       perror(argv[0]);
+                       exit(-1);
+               }
+
+               if (!system_wide && target_tid == -1 && target_pid == -1)
+                       all_tids[0] = child_pid;
+
+               close(child_ready_pipe[1]);
+               close(go_pipe[0]);
+               /*
+                * wait for child to settle
+                */
+               if (read(child_ready_pipe[0], &buf, 1) == -1) {
+                       perror("unable to read pipe");
+                       exit(-1);
                }
+               close(child_ready_pipe[0]);
        }
 
-       if (file_new) {
+       if ((!system_wide && !inherit) || profile_cpu != -1) {
+               open_counters(profile_cpu);
+       } else {
+               nr_cpus = read_cpu_map();
+               for (i = 0; i < nr_cpus; i++)
+                       open_counters(cpumap[i]);
+       }
+
+       if (pipe_output) {
+               err = perf_header__write_pipe(output);
+               if (err < 0)
+                       return err;
+       } else if (file_new) {
                err = perf_header__write(&session->header, output, false);
                if (err < 0)
                        return err;
        }
 
-       if (!system_wide)
-               event__synthesize_thread(pid, process_synthesized_event,
-                                        session);
-       else
-               event__synthesize_threads(process_synthesized_event, session);
+       post_processing_offset = lseek(output, 0, SEEK_CUR);
 
-       if (target_pid == -1 && argc) {
-               pid = fork();
-               if (pid < 0)
-                       die("failed to fork");
+       if (pipe_output) {
+               err = event__synthesize_attrs(&session->header,
+                                             process_synthesized_event,
+                                             session);
+               if (err < 0) {
+                       pr_err("Couldn't synthesize attrs.\n");
+                       return err;
+               }
 
-               if (!pid) {
-                       if (execvp(argv[0], (char **)argv)) {
-                               perror(argv[0]);
-                               exit(-1);
-                       }
-               } else {
-                       /*
-                        * Wait a bit for the execv'ed child to appear
-                        * and be updated in /proc
-                        * FIXME: Do you know a less heuristical solution?
-                        */
-                       usleep(1000);
-                       event__synthesize_thread(pid,
-                                                process_synthesized_event,
-                                                session);
+               err = event__synthesize_event_types(process_synthesized_event,
+                                                   session);
+               if (err < 0) {
+                       pr_err("Couldn't synthesize event_types.\n");
+                       return err;
+               }
+
+               err = event__synthesize_tracing_data(output, attrs,
+                                                    nr_counters,
+                                                    process_synthesized_event,
+                                                    session);
+               if (err <= 0) {
+                       pr_err("Couldn't record tracing data.\n");
+                       return err;
                }
 
-               child_pid = pid;
+               advance_output(err);
        }
 
+       err = event__synthesize_kernel_mmap(process_synthesized_event,
+                                           session, "_text");
+       if (err < 0)
+               err = event__synthesize_kernel_mmap(process_synthesized_event,
+                                                   session, "_stext");
+       if (err < 0) {
+               pr_err("Couldn't record kernel reference relocation symbol.\n");
+               return err;
+       }
+
+       err = event__synthesize_modules(process_synthesized_event, session);
+       if (err < 0) {
+               pr_err("Couldn't record kernel reference relocation symbol.\n");
+               return err;
+       }
+
+       if (!system_wide && profile_cpu == -1)
+               event__synthesize_thread(target_tid, process_synthesized_event,
+                                        session);
+       else
+               event__synthesize_threads(process_synthesized_event, session);
+
        if (realtime_prio) {
                struct sched_param param;
 
@@ -538,13 +667,24 @@ static int __cmd_record(int argc, const char **argv)
                }
        }
 
+       /*
+        * Let the child rip
+        */
+       if (forks)
+               close(go_pipe[1]);
+
        for (;;) {
                int hits = samples;
+               int thread;
 
                for (i = 0; i < nr_cpu; i++) {
                        for (counter = 0; counter < nr_counters; counter++) {
-                               if (mmap_array[i][counter].base)
-                                       mmap_read(&mmap_array[i][counter]);
+                               for (thread = 0;
+                                       thread < thread_num; thread++) {
+                                       if (mmap_array[i][counter][thread].base)
+                                               mmap_read(&mmap_array[i][counter][thread]);
+                               }
+
                        }
                }
 
@@ -557,8 +697,15 @@ static int __cmd_record(int argc, const char **argv)
 
                if (done) {
                        for (i = 0; i < nr_cpu; i++) {
-                               for (counter = 0; counter < nr_counters; counter++)
-                                       ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
+                               for (counter = 0;
+                                       counter < nr_counters;
+                                       counter++) {
+                                       for (thread = 0;
+                                               thread < thread_num;
+                                               thread++)
+                                               ioctl(fd[i][counter][thread],
+                                                       PERF_EVENT_IOC_DISABLE);
+                               }
                        }
                }
        }
@@ -583,6 +730,8 @@ static const char * const record_usage[] = {
        NULL
 };
 
+static bool force, append_file;
+
 static const struct option options[] = {
        OPT_CALLBACK('e', "event", NULL, "event",
                     "event selector. use 'perf list' to list available events",
@@ -590,7 +739,9 @@ static const struct option options[] = {
        OPT_CALLBACK(0, "filter", NULL, "filter",
                     "event filter", parse_filter),
        OPT_INTEGER('p', "pid", &target_pid,
-                   "record events on existing pid"),
+                   "record events on existing process id"),
+       OPT_INTEGER('t', "tid", &target_tid,
+                   "record events on existing thread id"),
        OPT_INTEGER('r', "realtime", &realtime_prio,
                    "collect data with this RT SCHED_FIFO priority"),
        OPT_BOOLEAN('R', "raw-samples", &raw_samples,
@@ -602,20 +753,20 @@ static const struct option options[] = {
        OPT_INTEGER('C', "profile_cpu", &profile_cpu,
                            "CPU to profile on"),
        OPT_BOOLEAN('f', "force", &force,
-                       "overwrite existing data file"),
-       OPT_LONG('c', "count", &default_interval,
+                       "overwrite existing data file (deprecated)"),
+       OPT_LONG('c', "count", &user_interval,
                    "event period to sample"),
        OPT_STRING('o', "output", &output_name, "file",
                    "output file name"),
        OPT_BOOLEAN('i', "inherit", &inherit,
                    "child tasks inherit counters"),
-       OPT_INTEGER('F', "freq", &freq,
+       OPT_INTEGER('F', "freq", &user_freq,
                    "profile at this frequency"),
        OPT_INTEGER('m', "mmap-pages", &mmap_pages,
                    "number of mmap data pages"),
        OPT_BOOLEAN('g', "call-graph", &call_graph,
                    "do call-graph (stack chain/backtrace) recording"),
-       OPT_BOOLEAN('v', "verbose", &verbose,
+       OPT_INCR('v', "verbose", &verbose,
                    "be more verbose (show counter open errors, etc)"),
        OPT_BOOLEAN('s', "stat", &inherit_stat,
                    "per thread counts"),
@@ -630,14 +781,25 @@ static const struct option options[] = {
 
 int cmd_record(int argc, const char **argv, const char *prefix __used)
 {
-       int counter;
-
-       symbol__init(0);
+       int i,j;
 
        argc = parse_options(argc, argv, options, record_usage,
-               PARSE_OPT_STOP_AT_NON_OPTION);
-       if (!argc && target_pid == -1 && !system_wide)
+                           PARSE_OPT_STOP_AT_NON_OPTION);
+       if (!argc && target_pid == -1 && target_tid == -1 &&
+               !system_wide && profile_cpu == -1)
+               usage_with_options(record_usage, options);
+
+       if (force && append_file) {
+               fprintf(stderr, "Can't overwrite and append at the same time."
+                               " You need to choose between -f and -A");
                usage_with_options(record_usage, options);
+       } else if (append_file) {
+               write_mode = WRITE_APPEND;
+       } else {
+               write_mode = WRITE_FORCE;
+       }
+
+       symbol__init();
 
        if (!nr_counters) {
                nr_counters     = 1;
@@ -645,6 +807,42 @@ int cmd_record(int argc, const char **argv, const char *prefix __used)
                attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
        }
 
+       if (target_pid != -1) {
+               target_tid = target_pid;
+               thread_num = find_all_tid(target_pid, &all_tids);
+               if (thread_num <= 0) {
+                       fprintf(stderr, "Can't find all threads of pid %d\n",
+                                       target_pid);
+                       usage_with_options(record_usage, options);
+               }
+       } else {
+               all_tids=malloc(sizeof(pid_t));
+               if (!all_tids)
+                       return -ENOMEM;
+
+               all_tids[0] = target_tid;
+               thread_num = 1;
+       }
+
+       for (i = 0; i < MAX_NR_CPUS; i++) {
+               for (j = 0; j < MAX_COUNTERS; j++) {
+                       fd[i][j] = malloc(sizeof(int)*thread_num);
+                       mmap_array[i][j] = zalloc(
+                               sizeof(struct mmap_data)*thread_num);
+                       if (!fd[i][j] || !mmap_array[i][j])
+                               return -ENOMEM;
+               }
+       }
+       event_array = malloc(
+               sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
+       if (!event_array)
+               return -ENOMEM;
+
+       if (user_interval != UINT_MAX)
+               default_interval = user_interval;
+       if (user_freq != UINT_MAX)
+               freq = user_freq;
+
        /*
         * User specified count overrides default frequency.
         */
@@ -657,12 +855,5 @@ int cmd_record(int argc, const char **argv, const char *prefix __used)
                exit(EXIT_FAILURE);
        }
 
-       for (counter = 0; counter < nr_counters; counter++) {
-               if (attrs[counter].sample_period)
-                       continue;
-
-               attrs[counter].sample_period = default_interval;
-       }
-
        return __cmd_record(argc, argv);
 }