perf annotate: Handle samples not at objdump output addr boundaries
[linux-2.6.git] / tools / perf / builtin-annotate.c
index 377cb7c..5ec5de9 100644 (file)
 #include "perf.h"
 #include "util/debug.h"
 
+#include "util/event.h"
 #include "util/parse-options.h"
 #include "util/parse-events.h"
 #include "util/thread.h"
 #include "util/sort.h"
 #include "util/hist.h"
-#include "util/process_events.h"
+#include "util/session.h"
 
 static char            const *input_name = "perf.data";
 
 static int             force;
-static int             input;
 
 static int             full_paths;
 
 static int             print_line;
 
-static unsigned long   page_size;
-static unsigned long   mmap_window = 32;
-
 struct sym_hist {
        u64             sum;
        u64             ip[0];
@@ -54,39 +51,22 @@ struct sym_priv {
        struct sym_ext  *ext;
 };
 
-static struct symbol_conf symbol_conf = {
-       .priv_size        = sizeof(struct sym_priv),
-       .try_vmlinux_path = true,
-};
-
 static const char *sym_hist_filter;
 
-static int symbol_filter(struct map *map __used, struct symbol *sym)
+static int sym__alloc_hist(struct symbol *self)
 {
-       if (sym_hist_filter == NULL ||
-           strcmp(sym->name, sym_hist_filter) == 0) {
-               struct sym_priv *priv = symbol__priv(sym);
-               const int size = (sizeof(*priv->hist) +
-                                 (sym->end - sym->start) * sizeof(u64));
+       struct sym_priv *priv = symbol__priv(self);
+       const int size = (sizeof(*priv->hist) +
+                         (self->end - self->start) * sizeof(u64));
 
-               priv->hist = malloc(size);
-               if (priv->hist)
-                       memset(priv->hist, 0, size);
-               return 0;
-       }
-       /*
-        * FIXME: We should really filter it out, as we don't want to go thru symbols
-        * we're not interested, and if a DSO ends up with no symbols, delete it too,
-        * but right now the kernel loading routines in symbol.c bail out if no symbols
-        * are found, fix it later.
-        */
-       return 0;
+       priv->hist = zalloc(size);
+       return priv->hist == NULL ? -1 : 0;
 }
 
 /*
  * collect histogram counts
  */
-static void hist_hit(struct hist_entry *he, u64 ip)
+static int annotate__hist_hit(struct hist_entry *he, u64 ip)
 {
        unsigned int sym_size, offset;
        struct symbol *sym = he->sym;
@@ -96,180 +76,127 @@ static void hist_hit(struct hist_entry *he, u64 ip)
        he->count++;
 
        if (!sym || !he->map)
-               return;
+               return 0;
 
        priv = symbol__priv(sym);
-       if (!priv->hist)
-               return;
+       if (priv->hist == NULL && sym__alloc_hist(sym) < 0)
+               return -ENOMEM;
 
        sym_size = sym->end - sym->start;
        offset = ip - sym->start;
 
-       if (verbose)
-               fprintf(stderr, "%s: ip=%Lx\n", __func__,
-                       he->map->unmap_ip(he->map, ip));
+       pr_debug3("%s: ip=%#Lx\n", __func__, he->map->unmap_ip(he->map, ip));
 
        if (offset >= sym_size)
-               return;
+               return 0;
 
        h = priv->hist;
        h->sum++;
        h->ip[offset]++;
 
-       if (verbose >= 3)
-               printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
-                       (void *)(unsigned long)he->sym->start,
-                       he->sym->name,
-                       (void *)(unsigned long)ip, ip - he->sym->start,
-                       h->ip[offset]);
+       pr_debug3("%#Lx %s: count++ [ip: %#Lx, %#Lx] => %Ld\n", he->sym->start,
+                 he->sym->name, ip, ip - he->sym->start, h->ip[offset]);
+       return 0;
 }
 
-static int hist_entry__add(struct thread *thread, struct map *map,
-                          struct symbol *sym, u64 ip, u64 count, char level)
+static int perf_session__add_hist_entry(struct perf_session *self,
+                                       struct addr_location *al, u64 count)
 {
        bool hit;
-       struct hist_entry *he = __hist_entry__add(thread, map, sym, NULL, ip,
-                                                 count, level, &hit);
+       struct hist_entry *he;
+
+       if (sym_hist_filter != NULL &&
+           (al->sym == NULL || strcmp(sym_hist_filter, al->sym->name) != 0)) {
+               /* We're only interested in a symbol named sym_hist_filter */
+               if (al->sym != NULL) {
+                       rb_erase(&al->sym->rb_node,
+                                &al->map->dso->symbols[al->map->type]);
+                       symbol__delete(al->sym);
+               }
+               return 0;
+       }
+
+       he = __perf_session__add_hist_entry(self, al, NULL, count, &hit);
        if (he == NULL)
                return -ENOMEM;
-       hist_hit(he, ip);
-       return 0;
+
+       return annotate__hist_hit(he, al->addr);
 }
 
-static int
-process_sample_event(event_t *event, unsigned long offset, unsigned long head)
+static int process_sample_event(event_t *event, struct perf_session *session)
 {
-       char level;
-       u64 ip = event->ip.ip;
-       struct map *map = NULL;
-       struct symbol *sym = NULL;
-       struct thread *thread = threads__findnew(event->ip.pid);
-
-       dump_printf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
-               (void *)(offset + head),
-               (void *)(long)(event->header.size),
-               event->header.misc,
-               event->ip.pid,
-               (void *)(long)ip);
-
-       if (thread == NULL) {
-               fprintf(stderr, "problem processing %d event, skipping it.\n",
-                       event->header.type);
-               return -1;
-       }
+       struct addr_location al;
 
-       dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
-
-       if (event->header.misc & PERF_RECORD_MISC_KERNEL) {
-               level = 'k';
-               sym = kernel_maps__find_function(ip, &map, symbol_filter);
-               dump_printf(" ...... dso: %s\n",
-                           map ? map->dso->long_name : "<not found>");
-       } else if (event->header.misc & PERF_RECORD_MISC_USER) {
-               level = '.';
-               map = thread__find_map(thread, ip);
-               if (map != NULL) {
-got_map:
-                       ip = map->map_ip(map, ip);
-                       sym = map__find_function(map, ip, symbol_filter);
-               } else {
-                       /*
-                        * If this is outside of all known maps,
-                        * and is a negative address, try to look it
-                        * up in the kernel dso, as it might be a
-                        * vsyscall or vdso (which executes in user-mode).
-                        *
-                        * XXX This is nasty, we should have a symbol list in
-                        * the "[vdso]" dso, but for now lets use the old
-                        * trick of looking in the whole kernel symbol list.
-                        */
-                       if ((long long)ip < 0) {
-                               map = kernel_map__functions;
-                               goto got_map;
-                       }
-               }
-               dump_printf(" ...... dso: %s\n",
-                           map ? map->dso->long_name : "<not found>");
-       } else {
-               level = 'H';
-               dump_printf(" ...... dso: [hypervisor]\n");
+       dump_printf("(IP, %d): %d: %#Lx\n", event->header.misc,
+                   event->ip.pid, event->ip.ip);
+
+       if (event__preprocess_sample(event, session, &al, NULL) < 0) {
+               pr_warning("problem processing %d event, skipping it.\n",
+                          event->header.type);
+               return -1;
        }
 
-       if (hist_entry__add(thread, map, sym, ip, 1, level)) {
-               fprintf(stderr, "problem incrementing symbol count, "
-                               "skipping event\n");
+       if (!al.filtered && perf_session__add_hist_entry(session, &al, 1)) {
+               pr_warning("problem incrementing symbol count, "
+                          "skipping event\n");
                return -1;
        }
-       total++;
 
        return 0;
 }
 
-static int
-process_comm_event(event_t *event, unsigned long offset, unsigned long head)
-{
-       struct thread *thread = threads__findnew(event->comm.pid);
+struct objdump_line {
+       struct list_head node;
+       s64              offset;
+       char             *line;
+};
 
-       dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
-               (void *)(offset + head),
-               (void *)(long)(event->header.size),
-               event->comm.comm, event->comm.pid);
+static struct objdump_line *objdump_line__new(s64 offset, char *line)
+{
+       struct objdump_line *self = malloc(sizeof(*self));
 
-       if (thread == NULL ||
-           thread__set_comm(thread, event->comm.comm)) {
-               dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
-               return -1;
+       if (self != NULL) {
+               self->offset = offset;
+               self->line = line;
        }
-       total_comm++;
 
-       return 0;
+       return self;
 }
 
-static int
-process_event(event_t *event, unsigned long offset, unsigned long head)
+static void objdump_line__free(struct objdump_line *self)
 {
-       switch (event->header.type) {
-       case PERF_RECORD_SAMPLE:
-               return process_sample_event(event, offset, head);
-
-       case PERF_RECORD_MMAP:
-               return process_mmap_event(event, offset, head);
-
-       case PERF_RECORD_COMM:
-               return process_comm_event(event, offset, head);
-
-       case PERF_RECORD_FORK:
-               return process_task_event(event, offset, head);
-       /*
-        * We dont process them right now but they are fine:
-        */
+       free(self->line);
+       free(self);
+}
 
-       case PERF_RECORD_THROTTLE:
-       case PERF_RECORD_UNTHROTTLE:
-               return 0;
+static void objdump__add_line(struct list_head *head, struct objdump_line *line)
+{
+       list_add_tail(&line->node, head);
+}
 
-       default:
-               return -1;
-       }
+static struct objdump_line *objdump__get_next_ip_line(struct list_head *head,
+                                                     struct objdump_line *pos)
+{
+       list_for_each_entry_continue(pos, head, node)
+               if (pos->offset >= 0)
+                       return pos;
 
-       return 0;
+       return NULL;
 }
 
-static int parse_line(FILE *file, struct hist_entry *he, u64 len)
+static int parse_line(FILE *file, struct hist_entry *he,
+                     struct list_head *head)
 {
        struct symbol *sym = he->sym;
+       struct objdump_line *objdump_line;
        char *line = NULL, *tmp, *tmp2;
-       static const char *prev_line;
-       static const char *prev_color;
-       unsigned int offset;
        size_t line_len;
-       u64 start;
-       s64 line_ip;
-       int ret;
+       s64 line_ip, offset = -1;
        char *c;
 
        if (getline(&line, &line_len, file) < 0)
                return -1;
+
        if (!line)
                return -1;
 
@@ -278,8 +205,6 @@ static int parse_line(FILE *file, struct hist_entry *he, u64 len)
                *c = 0;
 
        line_ip = -1;
-       offset = 0;
-       ret = -2;
 
        /*
         * Strip leading spaces:
@@ -300,9 +225,30 @@ static int parse_line(FILE *file, struct hist_entry *he, u64 len)
                        line_ip = -1;
        }
 
-       start = he->map->unmap_ip(he->map, sym->start);
-
        if (line_ip != -1) {
+               u64 start = map__rip_2objdump(he->map, sym->start);
+               offset = line_ip - start;
+       }
+
+       objdump_line = objdump_line__new(offset, line);
+       if (objdump_line == NULL) {
+               free(line);
+               return -1;
+       }
+       objdump__add_line(head, objdump_line);
+
+       return 0;
+}
+
+static int objdump_line__print(struct objdump_line *self,
+                              struct list_head *head,
+                              struct hist_entry *he, u64 len)
+{
+       struct symbol *sym = he->sym;
+       static const char *prev_line;
+       static const char *prev_color;
+
+       if (self->offset != -1) {
                const char *path = NULL;
                unsigned int hits = 0;
                double percent = 0.0;
@@ -310,15 +256,22 @@ static int parse_line(FILE *file, struct hist_entry *he, u64 len)
                struct sym_priv *priv = symbol__priv(sym);
                struct sym_ext *sym_ext = priv->ext;
                struct sym_hist *h = priv->hist;
+               s64 offset = self->offset;
+               struct objdump_line *next = objdump__get_next_ip_line(head, self);
+
+               while (offset < (s64)len &&
+                      (next == NULL || offset < next->offset)) {
+                       if (sym_ext) {
+                               if (path == NULL)
+                                       path = sym_ext[offset].path;
+                               percent += sym_ext[offset].percent;
+                       } else
+                               hits += h->ip[offset];
+
+                       ++offset;
+               }
 
-               offset = line_ip - start;
-               if (offset < len)
-                       hits = h->ip[offset];
-
-               if (offset < len && sym_ext) {
-                       path = sym_ext[offset].path;
-                       percent = sym_ext[offset].percent;
-               } else if (h->sum)
+               if (sym_ext == NULL && h->sum)
                        percent = 100.0 * hits / h->sum;
 
                color = get_percent_color(percent);
@@ -339,12 +292,12 @@ static int parse_line(FILE *file, struct hist_entry *he, u64 len)
 
                color_fprintf(stdout, color, " %7.2f", percent);
                printf(" :      ");
-               color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
+               color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", self->line);
        } else {
-               if (!*line)
+               if (!*self->line)
                        printf("         :\n");
                else
-                       printf("         :      %s\n", line);
+                       printf("         :      %s\n", self->line);
        }
 
        return 0;
@@ -470,6 +423,20 @@ static void print_summary(const char *filename)
        }
 }
 
+static void hist_entry__print_hits(struct hist_entry *self)
+{
+       struct symbol *sym = self->sym;
+       struct sym_priv *priv = symbol__priv(sym);
+       struct sym_hist *h = priv->hist;
+       u64 len = sym->end - sym->start, offset;
+
+       for (offset = 0; offset < len; ++offset)
+               if (h->ip[offset] != 0)
+                       printf("%*Lx: %Lu\n", BITS_PER_LONG / 2,
+                              sym->start + offset, h->ip[offset]);
+       printf("%*s: %Lu\n", BITS_PER_LONG / 2, "h->sum", h->sum);
+}
+
 static void annotate_sym(struct hist_entry *he)
 {
        struct map *map = he->map;
@@ -479,15 +446,15 @@ static void annotate_sym(struct hist_entry *he)
        u64 len;
        char command[PATH_MAX*2];
        FILE *file;
+       LIST_HEAD(head);
+       struct objdump_line *pos, *n;
 
        if (!filename)
                return;
 
-       if (verbose)
-               fprintf(stderr, "%s: filename=%s, sym=%s, start=%Lx, end=%Lx\n",
-                       __func__, filename, sym->name,
-                       map->unmap_ip(map, sym->start),
-                       map->unmap_ip(map, sym->end));
+       pr_debug("%s: filename=%s, sym=%s, start=%#Lx, end=%#Lx\n", __func__,
+                filename, sym->name, map->unmap_ip(map, sym->start),
+                map->unmap_ip(map, sym->end));
 
        if (full_paths)
                d_filename = filename;
@@ -510,7 +477,8 @@ static void annotate_sym(struct hist_entry *he)
                       dso, dso->long_name, sym, sym->name);
 
        sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
-               map->unmap_ip(map, sym->start), map->unmap_ip(map, sym->end),
+               map__rip_2objdump(map, sym->start),
+               map__rip_2objdump(map, sym->end),
                filename, filename);
 
        if (verbose >= 3)
@@ -521,20 +489,30 @@ static void annotate_sym(struct hist_entry *he)
                return;
 
        while (!feof(file)) {
-               if (parse_line(file, he, len) < 0)
+               if (parse_line(file, he, &head) < 0)
                        break;
        }
 
        pclose(file);
+
+       if (verbose)
+               hist_entry__print_hits(he);
+
+       list_for_each_entry_safe(pos, n, &head, node) {
+               objdump_line__print(pos, &head, he, len);
+               list_del(&pos->node);
+               objdump_line__free(pos);
+       }
+
        if (print_line)
                free_source_line(he, len);
 }
 
-static void find_annotations(void)
+static void perf_session__find_annotations(struct perf_session *self)
 {
        struct rb_node *nd;
 
-       for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
+       for (nd = rb_first(&self->hists); nd; nd = rb_next(nd)) {
                struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
                struct sym_priv *priv;
 
@@ -555,123 +533,44 @@ static void find_annotations(void)
        }
 }
 
+static struct perf_event_ops event_ops = {
+       .sample = process_sample_event,
+       .mmap   = event__process_mmap,
+       .comm   = event__process_comm,
+       .fork   = event__process_task,
+};
+
 static int __cmd_annotate(void)
 {
-       int ret, rc = EXIT_FAILURE;
-       unsigned long offset = 0;
-       unsigned long head = 0;
-       struct stat input_stat;
-       event_t *event;
-       uint32_t size;
-       char *buf;
-
-       register_idle_thread();
-
-       input = open(input_name, O_RDONLY);
-       if (input < 0) {
-               perror("failed to open file");
-               exit(-1);
-       }
-
-       ret = fstat(input, &input_stat);
-       if (ret < 0) {
-               perror("failed to stat file");
-               exit(-1);
-       }
-
-       if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
-               fprintf(stderr, "file: %s not owned by current user or root\n", input_name);
-               exit(-1);
-       }
-
-       if (!input_stat.st_size) {
-               fprintf(stderr, "zero-sized file, nothing to do!\n");
-               exit(0);
-       }
-
-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);
-
-       size = event->header.size;
-       if (!size)
-               size = 8;
-
-       if (head + event->header.size >= page_size * mmap_window) {
-               unsigned long shift = page_size * (head / page_size);
-               int munmap_ret;
-
-               munmap_ret = munmap(buf, page_size * mmap_window);
-               assert(munmap_ret == 0);
-
-               offset += shift;
-               head -= shift;
-               goto remap;
-       }
-
-       size = event->header.size;
-
-       dump_printf("%p [%p]: event: %d\n",
-                       (void *)(offset + head),
-                       (void *)(long)event->header.size,
-                       event->header.type);
-
-       if (!size || process_event(event, offset, head) < 0) {
-
-               dump_printf("%p [%p]: skipping unknown header type: %d\n",
-                       (void *)(offset + head),
-                       (void *)(long)(event->header.size),
-                       event->header.type);
-
-               total_unknown++;
+       int ret;
+       struct perf_session *session;
 
-               /*
-                * assume we lost track of the stream, check alignment, and
-                * increment a single u64 in the hope to catch on again 'soon'.
-                */
+       session = perf_session__new(input_name, O_RDONLY, force);
+       if (session == NULL)
+               return -ENOMEM;
 
-               if (unlikely(head & 7))
-                       head &= ~7ULL;
+       ret = perf_session__process_events(session, &event_ops);
+       if (ret)
+               goto out_delete;
 
-               size = 8;
+       if (dump_trace) {
+               event__print_totals();
+               goto out_delete;
        }
 
-       head += size;
-
-       if (offset + head < (unsigned long)input_stat.st_size)
-               goto more;
-
-       rc = EXIT_SUCCESS;
-       close(input);
-
-       dump_printf("      IP events: %10ld\n", total);
-       dump_printf("    mmap events: %10ld\n", total_mmap);
-       dump_printf("    comm events: %10ld\n", total_comm);
-       dump_printf("    fork events: %10ld\n", total_fork);
-       dump_printf(" unknown events: %10ld\n", total_unknown);
-
-       if (dump_trace)
-               return 0;
-
        if (verbose > 3)
-               threads__fprintf(stdout);
+               perf_session__fprintf(session, stdout);
 
        if (verbose > 2)
                dsos__fprintf(stdout);
 
-       collapse__resort();
-       output__resort(total);
+       perf_session__collapse_resort(session);
+       perf_session__output_resort(session, session->event_total[0]);
+       perf_session__find_annotations(session);
+out_delete:
+       perf_session__delete(session);
 
-       find_annotations();
-
-       return rc;
+       return ret;
 }
 
 static const char * const annotate_usage[] = {
@@ -700,31 +599,17 @@ static const struct option options[] = {
        OPT_END()
 };
 
-static void setup_sorting(void)
-{
-       char *tmp, *tok, *str = strdup(sort_order);
-
-       for (tok = strtok_r(str, ", ", &tmp);
-                       tok; tok = strtok_r(NULL, ", ", &tmp)) {
-               if (sort_dimension__add(tok) < 0) {
-                       error("Unknown --sort key: `%s'", tok);
-                       usage_with_options(annotate_usage, options);
-               }
-       }
-
-       free(str);
-}
-
 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
 {
-       if (symbol__init(&symbol_conf) < 0)
-               return -1;
+       argc = parse_options(argc, argv, options, annotate_usage, 0);
 
-       page_size = getpagesize();
+       symbol_conf.priv_size = sizeof(struct sym_priv);
+       symbol_conf.try_vmlinux_path = true;
 
-       argc = parse_options(argc, argv, options, annotate_usage, 0);
+       if (symbol__init() < 0)
+               return -1;
 
-       setup_sorting();
+       setup_sorting(annotate_usage, options);
 
        if (argc) {
                /*
@@ -740,9 +625,8 @@ int cmd_annotate(int argc, const char **argv, const char *prefix __used)
        setup_pager();
 
        if (field_sep && *field_sep == '.') {
-               fputs("'.' is the only non valid --field-separator argument\n",
-                               stderr);
-               exit(129);
+               pr_err("'.' is the only non valid --field-separator argument\n");
+               return -1;
        }
 
        return __cmd_annotate();