Back to home page

OSCL-LXR
 
 

     

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * builtin-top.c
0004  *
0005  * Builtin top command: Display a continuously updated profile of
0006  * any workload, CPU or specific PID.
0007  *
0008  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
0009  *       2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
0010  *
0011  * Improvements and fixes by:
0012  *
0013  *   Arjan van de Ven <arjan@linux.intel.com>
0014  *   Yanmin Zhang <yanmin.zhang@intel.com>
0015  *   Wu Fengguang <fengguang.wu@intel.com>
0016  *   Mike Galbraith <efault@gmx.de>
0017  *   Paul Mackerras <paulus@samba.org>
0018  */
0019 #include "builtin.h"
0020 
0021 #include "perf.h"
0022 
0023 #include "util/annotate.h"
0024 #include "util/bpf-event.h"
0025 #include "util/cgroup.h"
0026 #include "util/config.h"
0027 #include "util/color.h"
0028 #include "util/dso.h"
0029 #include "util/evlist.h"
0030 #include "util/evsel.h"
0031 #include "util/evsel_config.h"
0032 #include "util/event.h"
0033 #include "util/machine.h"
0034 #include "util/map.h"
0035 #include "util/mmap.h"
0036 #include "util/session.h"
0037 #include "util/thread.h"
0038 #include "util/symbol.h"
0039 #include "util/synthetic-events.h"
0040 #include "util/top.h"
0041 #include "util/util.h"
0042 #include <linux/rbtree.h>
0043 #include <subcmd/parse-options.h>
0044 #include "util/parse-events.h"
0045 #include "util/callchain.h"
0046 #include "util/cpumap.h"
0047 #include "util/sort.h"
0048 #include "util/string2.h"
0049 #include "util/term.h"
0050 #include "util/intlist.h"
0051 #include "util/parse-branch-options.h"
0052 #include "arch/common.h"
0053 #include "ui/ui.h"
0054 
0055 #include "util/debug.h"
0056 #include "util/ordered-events.h"
0057 #include "util/pfm.h"
0058 
0059 #include <assert.h>
0060 #include <elf.h>
0061 #include <fcntl.h>
0062 
0063 #include <stdio.h>
0064 #include <termios.h>
0065 #include <unistd.h>
0066 #include <inttypes.h>
0067 
0068 #include <errno.h>
0069 #include <time.h>
0070 #include <sched.h>
0071 #include <signal.h>
0072 
0073 #include <sys/syscall.h>
0074 #include <sys/ioctl.h>
0075 #include <poll.h>
0076 #include <sys/prctl.h>
0077 #include <sys/wait.h>
0078 #include <sys/uio.h>
0079 #include <sys/utsname.h>
0080 #include <sys/mman.h>
0081 
0082 #include <linux/stringify.h>
0083 #include <linux/time64.h>
0084 #include <linux/types.h>
0085 #include <linux/err.h>
0086 
0087 #include <linux/ctype.h>
0088 #include <perf/mmap.h>
0089 
0090 static volatile int done;
0091 static volatile int resize;
0092 
0093 #define HEADER_LINE_NR  5
0094 
0095 static void perf_top__update_print_entries(struct perf_top *top)
0096 {
0097     top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
0098 }
0099 
0100 static void winch_sig(int sig __maybe_unused)
0101 {
0102     resize = 1;
0103 }
0104 
0105 static void perf_top__resize(struct perf_top *top)
0106 {
0107     get_term_dimensions(&top->winsize);
0108     perf_top__update_print_entries(top);
0109 }
0110 
0111 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
0112 {
0113     struct evsel *evsel;
0114     struct symbol *sym;
0115     struct annotation *notes;
0116     struct map *map;
0117     int err = -1;
0118 
0119     if (!he || !he->ms.sym)
0120         return -1;
0121 
0122     evsel = hists_to_evsel(he->hists);
0123 
0124     sym = he->ms.sym;
0125     map = he->ms.map;
0126 
0127     /*
0128      * We can't annotate with just /proc/kallsyms
0129      */
0130     if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
0131         !dso__is_kcore(map->dso)) {
0132         pr_err("Can't annotate %s: No vmlinux file was found in the "
0133                "path\n", sym->name);
0134         sleep(1);
0135         return -1;
0136     }
0137 
0138     notes = symbol__annotation(sym);
0139     pthread_mutex_lock(&notes->lock);
0140 
0141     if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
0142         pthread_mutex_unlock(&notes->lock);
0143         pr_err("Not enough memory for annotating '%s' symbol!\n",
0144                sym->name);
0145         sleep(1);
0146         return err;
0147     }
0148 
0149     err = symbol__annotate(&he->ms, evsel, &top->annotation_opts, NULL);
0150     if (err == 0) {
0151         top->sym_filter_entry = he;
0152     } else {
0153         char msg[BUFSIZ];
0154         symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
0155         pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
0156     }
0157 
0158     pthread_mutex_unlock(&notes->lock);
0159     return err;
0160 }
0161 
0162 static void __zero_source_counters(struct hist_entry *he)
0163 {
0164     struct symbol *sym = he->ms.sym;
0165     symbol__annotate_zero_histograms(sym);
0166 }
0167 
0168 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
0169 {
0170     struct utsname uts;
0171     int err = uname(&uts);
0172 
0173     ui__warning("Out of bounds address found:\n\n"
0174             "Addr:   %" PRIx64 "\n"
0175             "DSO:    %s %c\n"
0176             "Map:    %" PRIx64 "-%" PRIx64 "\n"
0177             "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
0178             "Arch:   %s\n"
0179             "Kernel: %s\n"
0180             "Tools:  %s\n\n"
0181             "Not all samples will be on the annotation output.\n\n"
0182             "Please report to linux-kernel@vger.kernel.org\n",
0183             ip, map->dso->long_name, dso__symtab_origin(map->dso),
0184             map->start, map->end, sym->start, sym->end,
0185             sym->binding == STB_GLOBAL ? 'g' :
0186             sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
0187             err ? "[unknown]" : uts.machine,
0188             err ? "[unknown]" : uts.release, perf_version_string);
0189     if (use_browser <= 0)
0190         sleep(5);
0191 
0192     map->erange_warned = true;
0193 }
0194 
0195 static void perf_top__record_precise_ip(struct perf_top *top,
0196                     struct hist_entry *he,
0197                     struct perf_sample *sample,
0198                     struct evsel *evsel, u64 ip)
0199 {
0200     struct annotation *notes;
0201     struct symbol *sym = he->ms.sym;
0202     int err = 0;
0203 
0204     if (sym == NULL || (use_browser == 0 &&
0205                 (top->sym_filter_entry == NULL ||
0206                  top->sym_filter_entry->ms.sym != sym)))
0207         return;
0208 
0209     notes = symbol__annotation(sym);
0210 
0211     if (pthread_mutex_trylock(&notes->lock))
0212         return;
0213 
0214     err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
0215 
0216     pthread_mutex_unlock(&notes->lock);
0217 
0218     if (unlikely(err)) {
0219         /*
0220          * This function is now called with he->hists->lock held.
0221          * Release it before going to sleep.
0222          */
0223         pthread_mutex_unlock(&he->hists->lock);
0224 
0225         if (err == -ERANGE && !he->ms.map->erange_warned)
0226             ui__warn_map_erange(he->ms.map, sym, ip);
0227         else if (err == -ENOMEM) {
0228             pr_err("Not enough memory for annotating '%s' symbol!\n",
0229                    sym->name);
0230             sleep(1);
0231         }
0232 
0233         pthread_mutex_lock(&he->hists->lock);
0234     }
0235 }
0236 
0237 static void perf_top__show_details(struct perf_top *top)
0238 {
0239     struct hist_entry *he = top->sym_filter_entry;
0240     struct evsel *evsel;
0241     struct annotation *notes;
0242     struct symbol *symbol;
0243     int more;
0244 
0245     if (!he)
0246         return;
0247 
0248     evsel = hists_to_evsel(he->hists);
0249 
0250     symbol = he->ms.sym;
0251     notes = symbol__annotation(symbol);
0252 
0253     pthread_mutex_lock(&notes->lock);
0254 
0255     symbol__calc_percent(symbol, evsel);
0256 
0257     if (notes->src == NULL)
0258         goto out_unlock;
0259 
0260     printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
0261     printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
0262 
0263     more = symbol__annotate_printf(&he->ms, top->sym_evsel, &top->annotation_opts);
0264 
0265     if (top->evlist->enabled) {
0266         if (top->zero)
0267             symbol__annotate_zero_histogram(symbol, top->sym_evsel->core.idx);
0268         else
0269             symbol__annotate_decay_histogram(symbol, top->sym_evsel->core.idx);
0270     }
0271     if (more != 0)
0272         printf("%d lines not displayed, maybe increase display entries [e]\n", more);
0273 out_unlock:
0274     pthread_mutex_unlock(&notes->lock);
0275 }
0276 
0277 static void perf_top__resort_hists(struct perf_top *t)
0278 {
0279     struct evlist *evlist = t->evlist;
0280     struct evsel *pos;
0281 
0282     evlist__for_each_entry(evlist, pos) {
0283         struct hists *hists = evsel__hists(pos);
0284 
0285         /*
0286          * unlink existing entries so that they can be linked
0287          * in a correct order in hists__match() below.
0288          */
0289         hists__unlink(hists);
0290 
0291         if (evlist->enabled) {
0292             if (t->zero) {
0293                 hists__delete_entries(hists);
0294             } else {
0295                 hists__decay_entries(hists, t->hide_user_symbols,
0296                              t->hide_kernel_symbols);
0297             }
0298         }
0299 
0300         hists__collapse_resort(hists, NULL);
0301 
0302         /* Non-group events are considered as leader */
0303         if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
0304             struct hists *leader_hists = evsel__hists(evsel__leader(pos));
0305 
0306             hists__match(leader_hists, hists);
0307             hists__link(leader_hists, hists);
0308         }
0309     }
0310 
0311     evlist__for_each_entry(evlist, pos) {
0312         evsel__output_resort(pos, NULL);
0313     }
0314 }
0315 
0316 static void perf_top__print_sym_table(struct perf_top *top)
0317 {
0318     char bf[160];
0319     int printed = 0;
0320     const int win_width = top->winsize.ws_col - 1;
0321     struct evsel *evsel = top->sym_evsel;
0322     struct hists *hists = evsel__hists(evsel);
0323 
0324     puts(CONSOLE_CLEAR);
0325 
0326     perf_top__header_snprintf(top, bf, sizeof(bf));
0327     printf("%s\n", bf);
0328 
0329     printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
0330 
0331     if (!top->record_opts.overwrite &&
0332         (top->evlist->stats.nr_lost_warned !=
0333          top->evlist->stats.nr_events[PERF_RECORD_LOST])) {
0334         top->evlist->stats.nr_lost_warned =
0335                   top->evlist->stats.nr_events[PERF_RECORD_LOST];
0336         color_fprintf(stdout, PERF_COLOR_RED,
0337                   "WARNING: LOST %d chunks, Check IO/CPU overload",
0338                   top->evlist->stats.nr_lost_warned);
0339         ++printed;
0340     }
0341 
0342     if (top->sym_filter_entry) {
0343         perf_top__show_details(top);
0344         return;
0345     }
0346 
0347     perf_top__resort_hists(top);
0348 
0349     hists__output_recalc_col_len(hists, top->print_entries - printed);
0350     putchar('\n');
0351     hists__fprintf(hists, false, top->print_entries - printed, win_width,
0352                top->min_percent, stdout, !symbol_conf.use_callchain);
0353 }
0354 
0355 static void prompt_integer(int *target, const char *msg)
0356 {
0357     char *buf = malloc(0), *p;
0358     size_t dummy = 0;
0359     int tmp;
0360 
0361     fprintf(stdout, "\n%s: ", msg);
0362     if (getline(&buf, &dummy, stdin) < 0)
0363         return;
0364 
0365     p = strchr(buf, '\n');
0366     if (p)
0367         *p = 0;
0368 
0369     p = buf;
0370     while(*p) {
0371         if (!isdigit(*p))
0372             goto out_free;
0373         p++;
0374     }
0375     tmp = strtoul(buf, NULL, 10);
0376     *target = tmp;
0377 out_free:
0378     free(buf);
0379 }
0380 
0381 static void prompt_percent(int *target, const char *msg)
0382 {
0383     int tmp = 0;
0384 
0385     prompt_integer(&tmp, msg);
0386     if (tmp >= 0 && tmp <= 100)
0387         *target = tmp;
0388 }
0389 
0390 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
0391 {
0392     char *buf = malloc(0), *p;
0393     struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
0394     struct hists *hists = evsel__hists(top->sym_evsel);
0395     struct rb_node *next;
0396     size_t dummy = 0;
0397 
0398     /* zero counters of active symbol */
0399     if (syme) {
0400         __zero_source_counters(syme);
0401         top->sym_filter_entry = NULL;
0402     }
0403 
0404     fprintf(stdout, "\n%s: ", msg);
0405     if (getline(&buf, &dummy, stdin) < 0)
0406         goto out_free;
0407 
0408     p = strchr(buf, '\n');
0409     if (p)
0410         *p = 0;
0411 
0412     next = rb_first_cached(&hists->entries);
0413     while (next) {
0414         n = rb_entry(next, struct hist_entry, rb_node);
0415         if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
0416             found = n;
0417             break;
0418         }
0419         next = rb_next(&n->rb_node);
0420     }
0421 
0422     if (!found) {
0423         fprintf(stderr, "Sorry, %s is not active.\n", buf);
0424         sleep(1);
0425     } else
0426         perf_top__parse_source(top, found);
0427 
0428 out_free:
0429     free(buf);
0430 }
0431 
0432 static void perf_top__print_mapped_keys(struct perf_top *top)
0433 {
0434     char *name = NULL;
0435 
0436     if (top->sym_filter_entry) {
0437         struct symbol *sym = top->sym_filter_entry->ms.sym;
0438         name = sym->name;
0439     }
0440 
0441     fprintf(stdout, "\nMapped keys:\n");
0442     fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
0443     fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
0444 
0445     if (top->evlist->core.nr_entries > 1)
0446         fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", evsel__name(top->sym_evsel));
0447 
0448     fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
0449 
0450     fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
0451     fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
0452     fprintf(stdout, "\t[S]     stop annotation.\n");
0453 
0454     fprintf(stdout,
0455         "\t[K]     hide kernel symbols.             \t(%s)\n",
0456         top->hide_kernel_symbols ? "yes" : "no");
0457     fprintf(stdout,
0458         "\t[U]     hide user symbols.               \t(%s)\n",
0459         top->hide_user_symbols ? "yes" : "no");
0460     fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
0461     fprintf(stdout, "\t[qQ]    quit.\n");
0462 }
0463 
0464 static int perf_top__key_mapped(struct perf_top *top, int c)
0465 {
0466     switch (c) {
0467         case 'd':
0468         case 'e':
0469         case 'f':
0470         case 'z':
0471         case 'q':
0472         case 'Q':
0473         case 'K':
0474         case 'U':
0475         case 'F':
0476         case 's':
0477         case 'S':
0478             return 1;
0479         case 'E':
0480             return top->evlist->core.nr_entries > 1 ? 1 : 0;
0481         default:
0482             break;
0483     }
0484 
0485     return 0;
0486 }
0487 
0488 static bool perf_top__handle_keypress(struct perf_top *top, int c)
0489 {
0490     bool ret = true;
0491 
0492     if (!perf_top__key_mapped(top, c)) {
0493         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
0494         struct termios save;
0495 
0496         perf_top__print_mapped_keys(top);
0497         fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
0498         fflush(stdout);
0499 
0500         set_term_quiet_input(&save);
0501 
0502         poll(&stdin_poll, 1, -1);
0503         c = getc(stdin);
0504 
0505         tcsetattr(0, TCSAFLUSH, &save);
0506         if (!perf_top__key_mapped(top, c))
0507             return ret;
0508     }
0509 
0510     switch (c) {
0511         case 'd':
0512             prompt_integer(&top->delay_secs, "Enter display delay");
0513             if (top->delay_secs < 1)
0514                 top->delay_secs = 1;
0515             break;
0516         case 'e':
0517             prompt_integer(&top->print_entries, "Enter display entries (lines)");
0518             if (top->print_entries == 0) {
0519                 perf_top__resize(top);
0520                 signal(SIGWINCH, winch_sig);
0521             } else {
0522                 signal(SIGWINCH, SIG_DFL);
0523             }
0524             break;
0525         case 'E':
0526             if (top->evlist->core.nr_entries > 1) {
0527                 /* Select 0 as the default event: */
0528                 int counter = 0;
0529 
0530                 fprintf(stderr, "\nAvailable events:");
0531 
0532                 evlist__for_each_entry(top->evlist, top->sym_evsel)
0533                     fprintf(stderr, "\n\t%d %s", top->sym_evsel->core.idx, evsel__name(top->sym_evsel));
0534 
0535                 prompt_integer(&counter, "Enter details event counter");
0536 
0537                 if (counter >= top->evlist->core.nr_entries) {
0538                     top->sym_evsel = evlist__first(top->evlist);
0539                     fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
0540                     sleep(1);
0541                     break;
0542                 }
0543                 evlist__for_each_entry(top->evlist, top->sym_evsel)
0544                     if (top->sym_evsel->core.idx == counter)
0545                         break;
0546             } else
0547                 top->sym_evsel = evlist__first(top->evlist);
0548             break;
0549         case 'f':
0550             prompt_integer(&top->count_filter, "Enter display event count filter");
0551             break;
0552         case 'F':
0553             prompt_percent(&top->annotation_opts.min_pcnt,
0554                        "Enter details display event filter (percent)");
0555             break;
0556         case 'K':
0557             top->hide_kernel_symbols = !top->hide_kernel_symbols;
0558             break;
0559         case 'q':
0560         case 'Q':
0561             printf("exiting.\n");
0562             if (top->dump_symtab)
0563                 perf_session__fprintf_dsos(top->session, stderr);
0564             ret = false;
0565             break;
0566         case 's':
0567             perf_top__prompt_symbol(top, "Enter details symbol");
0568             break;
0569         case 'S':
0570             if (!top->sym_filter_entry)
0571                 break;
0572             else {
0573                 struct hist_entry *syme = top->sym_filter_entry;
0574 
0575                 top->sym_filter_entry = NULL;
0576                 __zero_source_counters(syme);
0577             }
0578             break;
0579         case 'U':
0580             top->hide_user_symbols = !top->hide_user_symbols;
0581             break;
0582         case 'z':
0583             top->zero = !top->zero;
0584             break;
0585         default:
0586             break;
0587     }
0588 
0589     return ret;
0590 }
0591 
0592 static void perf_top__sort_new_samples(void *arg)
0593 {
0594     struct perf_top *t = arg;
0595 
0596     if (t->evlist->selected != NULL)
0597         t->sym_evsel = t->evlist->selected;
0598 
0599     perf_top__resort_hists(t);
0600 
0601     if (t->lost || t->drop)
0602         pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
0603 }
0604 
0605 static void stop_top(void)
0606 {
0607     session_done = 1;
0608     done = 1;
0609 }
0610 
0611 static void *display_thread_tui(void *arg)
0612 {
0613     struct evsel *pos;
0614     struct perf_top *top = arg;
0615     const char *help = "For a higher level overview, try: perf top --sort comm,dso";
0616     struct hist_browser_timer hbt = {
0617         .timer      = perf_top__sort_new_samples,
0618         .arg        = top,
0619         .refresh    = top->delay_secs,
0620     };
0621     int ret;
0622 
0623     /* In order to read symbols from other namespaces perf to  needs to call
0624      * setns(2).  This isn't permitted if the struct_fs has multiple users.
0625      * unshare(2) the fs so that we may continue to setns into namespaces
0626      * that we're observing.
0627      */
0628     unshare(CLONE_FS);
0629 
0630     prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
0631 
0632 repeat:
0633     perf_top__sort_new_samples(top);
0634 
0635     /*
0636      * Initialize the uid_filter_str, in the future the TUI will allow
0637      * Zooming in/out UIDs. For now just use whatever the user passed
0638      * via --uid.
0639      */
0640     evlist__for_each_entry(top->evlist, pos) {
0641         struct hists *hists = evsel__hists(pos);
0642         hists->uid_filter_str = top->record_opts.target.uid_str;
0643     }
0644 
0645     ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
0646                        &top->session->header.env, !top->record_opts.overwrite,
0647                        &top->annotation_opts);
0648     if (ret == K_RELOAD) {
0649         top->zero = true;
0650         goto repeat;
0651     } else
0652         stop_top();
0653 
0654     return NULL;
0655 }
0656 
0657 static void display_sig(int sig __maybe_unused)
0658 {
0659     stop_top();
0660 }
0661 
0662 static void display_setup_sig(void)
0663 {
0664     signal(SIGSEGV, sighandler_dump_stack);
0665     signal(SIGFPE, sighandler_dump_stack);
0666     signal(SIGINT,  display_sig);
0667     signal(SIGQUIT, display_sig);
0668     signal(SIGTERM, display_sig);
0669 }
0670 
0671 static void *display_thread(void *arg)
0672 {
0673     struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
0674     struct termios save;
0675     struct perf_top *top = arg;
0676     int delay_msecs, c;
0677 
0678     /* In order to read symbols from other namespaces perf to  needs to call
0679      * setns(2).  This isn't permitted if the struct_fs has multiple users.
0680      * unshare(2) the fs so that we may continue to setns into namespaces
0681      * that we're observing.
0682      */
0683     unshare(CLONE_FS);
0684 
0685     prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
0686 
0687     display_setup_sig();
0688     pthread__unblock_sigwinch();
0689 repeat:
0690     delay_msecs = top->delay_secs * MSEC_PER_SEC;
0691     set_term_quiet_input(&save);
0692     /* trash return*/
0693     clearerr(stdin);
0694     if (poll(&stdin_poll, 1, 0) > 0)
0695         getc(stdin);
0696 
0697     while (!done) {
0698         perf_top__print_sym_table(top);
0699         /*
0700          * Either timeout expired or we got an EINTR due to SIGWINCH,
0701          * refresh screen in both cases.
0702          */
0703         switch (poll(&stdin_poll, 1, delay_msecs)) {
0704         case 0:
0705             continue;
0706         case -1:
0707             if (errno == EINTR)
0708                 continue;
0709             __fallthrough;
0710         default:
0711             c = getc(stdin);
0712             tcsetattr(0, TCSAFLUSH, &save);
0713 
0714             if (perf_top__handle_keypress(top, c))
0715                 goto repeat;
0716             stop_top();
0717         }
0718     }
0719 
0720     tcsetattr(0, TCSAFLUSH, &save);
0721     return NULL;
0722 }
0723 
0724 static int hist_iter__top_callback(struct hist_entry_iter *iter,
0725                    struct addr_location *al, bool single,
0726                    void *arg)
0727 {
0728     struct perf_top *top = arg;
0729     struct hist_entry *he = iter->he;
0730     struct evsel *evsel = iter->evsel;
0731 
0732     if (perf_hpp_list.sym && single)
0733         perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
0734 
0735     hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
0736              !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
0737              NULL);
0738     return 0;
0739 }
0740 
0741 static void perf_event__process_sample(struct perf_tool *tool,
0742                        const union perf_event *event,
0743                        struct evsel *evsel,
0744                        struct perf_sample *sample,
0745                        struct machine *machine)
0746 {
0747     struct perf_top *top = container_of(tool, struct perf_top, tool);
0748     struct addr_location al;
0749 
0750     if (!machine && perf_guest) {
0751         static struct intlist *seen;
0752 
0753         if (!seen)
0754             seen = intlist__new(NULL);
0755 
0756         if (!intlist__has_entry(seen, sample->pid)) {
0757             pr_err("Can't find guest [%d]'s kernel information\n",
0758                 sample->pid);
0759             intlist__add(seen, sample->pid);
0760         }
0761         return;
0762     }
0763 
0764     if (!machine) {
0765         pr_err("%u unprocessable samples recorded.\r",
0766                top->session->evlist->stats.nr_unprocessable_samples++);
0767         return;
0768     }
0769 
0770     if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
0771         top->exact_samples++;
0772 
0773     if (machine__resolve(machine, &al, sample) < 0)
0774         return;
0775 
0776     if (top->stitch_lbr)
0777         al.thread->lbr_stitch_enable = true;
0778 
0779     if (!machine->kptr_restrict_warned &&
0780         symbol_conf.kptr_restrict &&
0781         al.cpumode == PERF_RECORD_MISC_KERNEL) {
0782         if (!evlist__exclude_kernel(top->session->evlist)) {
0783             ui__warning(
0784 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
0785 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
0786 "Kernel%s samples will not be resolved.\n",
0787               al.map && map__has_symbols(al.map) ?
0788               " modules" : "");
0789             if (use_browser <= 0)
0790                 sleep(5);
0791         }
0792         machine->kptr_restrict_warned = true;
0793     }
0794 
0795     if (al.sym == NULL && al.map != NULL) {
0796         const char *msg = "Kernel samples will not be resolved.\n";
0797         /*
0798          * As we do lazy loading of symtabs we only will know if the
0799          * specified vmlinux file is invalid when we actually have a
0800          * hit in kernel space and then try to load it. So if we get
0801          * here and there are _no_ symbols in the DSO backing the
0802          * kernel map, bail out.
0803          *
0804          * We may never get here, for instance, if we use -K/
0805          * --hide-kernel-symbols, even if the user specifies an
0806          * invalid --vmlinux ;-)
0807          */
0808         if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
0809             __map__is_kernel(al.map) && map__has_symbols(al.map)) {
0810             if (symbol_conf.vmlinux_name) {
0811                 char serr[256];
0812                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
0813                 ui__warning("The %s file can't be used: %s\n%s",
0814                         symbol_conf.vmlinux_name, serr, msg);
0815             } else {
0816                 ui__warning("A vmlinux file was not found.\n%s",
0817                         msg);
0818             }
0819 
0820             if (use_browser <= 0)
0821                 sleep(5);
0822             top->vmlinux_warned = true;
0823         }
0824     }
0825 
0826     if (al.sym == NULL || !al.sym->idle) {
0827         struct hists *hists = evsel__hists(evsel);
0828         struct hist_entry_iter iter = {
0829             .evsel      = evsel,
0830             .sample     = sample,
0831             .add_entry_cb   = hist_iter__top_callback,
0832         };
0833 
0834         if (symbol_conf.cumulate_callchain)
0835             iter.ops = &hist_iter_cumulative;
0836         else
0837             iter.ops = &hist_iter_normal;
0838 
0839         pthread_mutex_lock(&hists->lock);
0840 
0841         if (hist_entry_iter__add(&iter, &al, top->max_stack, top) < 0)
0842             pr_err("Problem incrementing symbol period, skipping event\n");
0843 
0844         pthread_mutex_unlock(&hists->lock);
0845     }
0846 
0847     addr_location__put(&al);
0848 }
0849 
0850 static void
0851 perf_top__process_lost(struct perf_top *top, union perf_event *event,
0852                struct evsel *evsel)
0853 {
0854     top->lost += event->lost.lost;
0855     top->lost_total += event->lost.lost;
0856     evsel->evlist->stats.total_lost += event->lost.lost;
0857 }
0858 
0859 static void
0860 perf_top__process_lost_samples(struct perf_top *top,
0861                    union perf_event *event,
0862                    struct evsel *evsel)
0863 {
0864     top->lost += event->lost_samples.lost;
0865     top->lost_total += event->lost_samples.lost;
0866     evsel->evlist->stats.total_lost_samples += event->lost_samples.lost;
0867 }
0868 
0869 static u64 last_timestamp;
0870 
0871 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
0872 {
0873     struct record_opts *opts = &top->record_opts;
0874     struct evlist *evlist = top->evlist;
0875     struct mmap *md;
0876     union perf_event *event;
0877 
0878     md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
0879     if (perf_mmap__read_init(&md->core) < 0)
0880         return;
0881 
0882     while ((event = perf_mmap__read_event(&md->core)) != NULL) {
0883         int ret;
0884 
0885         ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
0886         if (ret && ret != -1)
0887             break;
0888 
0889         ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0, NULL);
0890         if (ret)
0891             break;
0892 
0893         perf_mmap__consume(&md->core);
0894 
0895         if (top->qe.rotate) {
0896             pthread_mutex_lock(&top->qe.mutex);
0897             top->qe.rotate = false;
0898             pthread_cond_signal(&top->qe.cond);
0899             pthread_mutex_unlock(&top->qe.mutex);
0900         }
0901     }
0902 
0903     perf_mmap__read_done(&md->core);
0904 }
0905 
0906 static void perf_top__mmap_read(struct perf_top *top)
0907 {
0908     bool overwrite = top->record_opts.overwrite;
0909     struct evlist *evlist = top->evlist;
0910     int i;
0911 
0912     if (overwrite)
0913         evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
0914 
0915     for (i = 0; i < top->evlist->core.nr_mmaps; i++)
0916         perf_top__mmap_read_idx(top, i);
0917 
0918     if (overwrite) {
0919         evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
0920         evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
0921     }
0922 }
0923 
0924 /*
0925  * Check per-event overwrite term.
0926  * perf top should support consistent term for all events.
0927  * - All events don't have per-event term
0928  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
0929  *   Nothing change, return 0.
0930  * - All events have same per-event term
0931  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
0932  *   Using the per-event setting to replace the opts->overwrite if
0933  *   they are different, then return 0.
0934  * - Events have different per-event term
0935  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
0936  *   Return -1
0937  * - Some of the event set per-event term, but some not.
0938  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
0939  *   Return -1
0940  */
0941 static int perf_top__overwrite_check(struct perf_top *top)
0942 {
0943     struct record_opts *opts = &top->record_opts;
0944     struct evlist *evlist = top->evlist;
0945     struct evsel_config_term *term;
0946     struct list_head *config_terms;
0947     struct evsel *evsel;
0948     int set, overwrite = -1;
0949 
0950     evlist__for_each_entry(evlist, evsel) {
0951         set = -1;
0952         config_terms = &evsel->config_terms;
0953         list_for_each_entry(term, config_terms, list) {
0954             if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
0955                 set = term->val.overwrite ? 1 : 0;
0956         }
0957 
0958         /* no term for current and previous event (likely) */
0959         if ((overwrite < 0) && (set < 0))
0960             continue;
0961 
0962         /* has term for both current and previous event, compare */
0963         if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
0964             return -1;
0965 
0966         /* no term for current event but has term for previous one */
0967         if ((overwrite >= 0) && (set < 0))
0968             return -1;
0969 
0970         /* has term for current event */
0971         if ((overwrite < 0) && (set >= 0)) {
0972             /* if it's first event, set overwrite */
0973             if (evsel == evlist__first(evlist))
0974                 overwrite = set;
0975             else
0976                 return -1;
0977         }
0978     }
0979 
0980     if ((overwrite >= 0) && (opts->overwrite != overwrite))
0981         opts->overwrite = overwrite;
0982 
0983     return 0;
0984 }
0985 
0986 static int perf_top_overwrite_fallback(struct perf_top *top,
0987                        struct evsel *evsel)
0988 {
0989     struct record_opts *opts = &top->record_opts;
0990     struct evlist *evlist = top->evlist;
0991     struct evsel *counter;
0992 
0993     if (!opts->overwrite)
0994         return 0;
0995 
0996     /* only fall back when first event fails */
0997     if (evsel != evlist__first(evlist))
0998         return 0;
0999 
1000     evlist__for_each_entry(evlist, counter)
1001         counter->core.attr.write_backward = false;
1002     opts->overwrite = false;
1003     pr_debug2("fall back to non-overwrite mode\n");
1004     return 1;
1005 }
1006 
1007 static int perf_top__start_counters(struct perf_top *top)
1008 {
1009     char msg[BUFSIZ];
1010     struct evsel *counter;
1011     struct evlist *evlist = top->evlist;
1012     struct record_opts *opts = &top->record_opts;
1013 
1014     if (perf_top__overwrite_check(top)) {
1015         ui__error("perf top only support consistent per-event "
1016               "overwrite setting for all events\n");
1017         goto out_err;
1018     }
1019 
1020     evlist__config(evlist, opts, &callchain_param);
1021 
1022     evlist__for_each_entry(evlist, counter) {
1023 try_again:
1024         if (evsel__open(counter, top->evlist->core.user_requested_cpus,
1025                      top->evlist->core.threads) < 0) {
1026 
1027             /*
1028              * Specially handle overwrite fall back.
1029              * Because perf top is the only tool which has
1030              * overwrite mode by default, support
1031              * both overwrite and non-overwrite mode, and
1032              * require consistent mode for all events.
1033              *
1034              * May move it to generic code with more tools
1035              * have similar attribute.
1036              */
1037             if (perf_missing_features.write_backward &&
1038                 perf_top_overwrite_fallback(top, counter))
1039                 goto try_again;
1040 
1041             if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
1042                 if (verbose > 0)
1043                     ui__warning("%s\n", msg);
1044                 goto try_again;
1045             }
1046 
1047             evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
1048             ui__error("%s\n", msg);
1049             goto out_err;
1050         }
1051     }
1052 
1053     if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1054         ui__error("Failed to mmap with %d (%s)\n",
1055                 errno, str_error_r(errno, msg, sizeof(msg)));
1056         goto out_err;
1057     }
1058 
1059     return 0;
1060 
1061 out_err:
1062     return -1;
1063 }
1064 
1065 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1066 {
1067     if (callchain->mode != CHAIN_NONE) {
1068         if (callchain_register_param(callchain) < 0) {
1069             ui__error("Can't register callchain params.\n");
1070             return -EINVAL;
1071         }
1072     }
1073 
1074     return 0;
1075 }
1076 
1077 static struct ordered_events *rotate_queues(struct perf_top *top)
1078 {
1079     struct ordered_events *in = top->qe.in;
1080 
1081     if (top->qe.in == &top->qe.data[1])
1082         top->qe.in = &top->qe.data[0];
1083     else
1084         top->qe.in = &top->qe.data[1];
1085 
1086     return in;
1087 }
1088 
1089 static void *process_thread(void *arg)
1090 {
1091     struct perf_top *top = arg;
1092 
1093     while (!done) {
1094         struct ordered_events *out, *in = top->qe.in;
1095 
1096         if (!in->nr_events) {
1097             usleep(100);
1098             continue;
1099         }
1100 
1101         out = rotate_queues(top);
1102 
1103         pthread_mutex_lock(&top->qe.mutex);
1104         top->qe.rotate = true;
1105         pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1106         pthread_mutex_unlock(&top->qe.mutex);
1107 
1108         if (ordered_events__flush(out, OE_FLUSH__TOP))
1109             pr_err("failed to process events\n");
1110     }
1111 
1112     return NULL;
1113 }
1114 
1115 /*
1116  * Allow only 'top->delay_secs' seconds behind samples.
1117  */
1118 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1119 {
1120     union perf_event *event = qevent->event;
1121     u64 delay_timestamp;
1122 
1123     if (event->header.type != PERF_RECORD_SAMPLE)
1124         return false;
1125 
1126     delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1127     return delay_timestamp < last_timestamp;
1128 }
1129 
1130 static int deliver_event(struct ordered_events *qe,
1131              struct ordered_event *qevent)
1132 {
1133     struct perf_top *top = qe->data;
1134     struct evlist *evlist = top->evlist;
1135     struct perf_session *session = top->session;
1136     union perf_event *event = qevent->event;
1137     struct perf_sample sample;
1138     struct evsel *evsel;
1139     struct machine *machine;
1140     int ret = -1;
1141 
1142     if (should_drop(qevent, top)) {
1143         top->drop++;
1144         top->drop_total++;
1145         return 0;
1146     }
1147 
1148     ret = evlist__parse_sample(evlist, event, &sample);
1149     if (ret) {
1150         pr_err("Can't parse sample, err = %d\n", ret);
1151         goto next_event;
1152     }
1153 
1154     evsel = evlist__id2evsel(session->evlist, sample.id);
1155     assert(evsel != NULL);
1156 
1157     if (event->header.type == PERF_RECORD_SAMPLE) {
1158         if (evswitch__discard(&top->evswitch, evsel))
1159             return 0;
1160         ++top->samples;
1161     }
1162 
1163     switch (sample.cpumode) {
1164     case PERF_RECORD_MISC_USER:
1165         ++top->us_samples;
1166         if (top->hide_user_symbols)
1167             goto next_event;
1168         machine = &session->machines.host;
1169         break;
1170     case PERF_RECORD_MISC_KERNEL:
1171         ++top->kernel_samples;
1172         if (top->hide_kernel_symbols)
1173             goto next_event;
1174         machine = &session->machines.host;
1175         break;
1176     case PERF_RECORD_MISC_GUEST_KERNEL:
1177         ++top->guest_kernel_samples;
1178         machine = perf_session__find_machine(session,
1179                              sample.pid);
1180         break;
1181     case PERF_RECORD_MISC_GUEST_USER:
1182         ++top->guest_us_samples;
1183         /*
1184          * TODO: we don't process guest user from host side
1185          * except simple counting.
1186          */
1187         goto next_event;
1188     default:
1189         if (event->header.type == PERF_RECORD_SAMPLE)
1190             goto next_event;
1191         machine = &session->machines.host;
1192         break;
1193     }
1194 
1195     if (event->header.type == PERF_RECORD_SAMPLE) {
1196         perf_event__process_sample(&top->tool, event, evsel,
1197                        &sample, machine);
1198     } else if (event->header.type == PERF_RECORD_LOST) {
1199         perf_top__process_lost(top, event, evsel);
1200     } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1201         perf_top__process_lost_samples(top, event, evsel);
1202     } else if (event->header.type < PERF_RECORD_MAX) {
1203         events_stats__inc(&session->evlist->stats, event->header.type);
1204         machine__process_event(machine, event, &sample);
1205     } else
1206         ++session->evlist->stats.nr_unknown_events;
1207 
1208     ret = 0;
1209 next_event:
1210     return ret;
1211 }
1212 
1213 static void init_process_thread(struct perf_top *top)
1214 {
1215     ordered_events__init(&top->qe.data[0], deliver_event, top);
1216     ordered_events__init(&top->qe.data[1], deliver_event, top);
1217     ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1218     ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1219     top->qe.in = &top->qe.data[0];
1220     pthread_mutex_init(&top->qe.mutex, NULL);
1221     pthread_cond_init(&top->qe.cond, NULL);
1222 }
1223 
1224 static int __cmd_top(struct perf_top *top)
1225 {
1226     struct record_opts *opts = &top->record_opts;
1227     pthread_t thread, thread_process;
1228     int ret;
1229 
1230     if (!top->annotation_opts.objdump_path) {
1231         ret = perf_env__lookup_objdump(&top->session->header.env,
1232                            &top->annotation_opts.objdump_path);
1233         if (ret)
1234             return ret;
1235     }
1236 
1237     ret = callchain_param__setup_sample_type(&callchain_param);
1238     if (ret)
1239         return ret;
1240 
1241     if (perf_session__register_idle_thread(top->session) < 0)
1242         return ret;
1243 
1244     if (top->nr_threads_synthesize > 1)
1245         perf_set_multithreaded();
1246 
1247     init_process_thread(top);
1248 
1249     if (opts->record_namespaces)
1250         top->tool.namespace_events = true;
1251     if (opts->record_cgroup) {
1252 #ifdef HAVE_FILE_HANDLE
1253         top->tool.cgroup_events = true;
1254 #else
1255         pr_err("cgroup tracking is not supported.\n");
1256         return -1;
1257 #endif
1258     }
1259 
1260     ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1261                         &top->session->machines.host,
1262                         &top->record_opts);
1263     if (ret < 0)
1264         pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1265 
1266     ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1267                          &top->session->machines.host);
1268     if (ret < 0)
1269         pr_debug("Couldn't synthesize cgroup events.\n");
1270 
1271     machine__synthesize_threads(&top->session->machines.host, &opts->target,
1272                     top->evlist->core.threads, true, false,
1273                     top->nr_threads_synthesize);
1274 
1275     if (top->nr_threads_synthesize > 1)
1276         perf_set_singlethreaded();
1277 
1278     if (perf_hpp_list.socket) {
1279         ret = perf_env__read_cpu_topology_map(&perf_env);
1280         if (ret < 0) {
1281             char errbuf[BUFSIZ];
1282             const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1283 
1284             ui__error("Could not read the CPU topology map: %s\n", err);
1285             return ret;
1286         }
1287     }
1288 
1289     ret = perf_top__start_counters(top);
1290     if (ret)
1291         return ret;
1292 
1293     top->session->evlist = top->evlist;
1294     perf_session__set_id_hdr_size(top->session);
1295 
1296     /*
1297      * When perf is starting the traced process, all the events (apart from
1298      * group members) have enable_on_exec=1 set, so don't spoil it by
1299      * prematurely enabling them.
1300      *
1301      * XXX 'top' still doesn't start workloads like record, trace, but should,
1302      * so leave the check here.
1303      */
1304         if (!target__none(&opts->target))
1305         evlist__enable(top->evlist);
1306 
1307     ret = -1;
1308     if (pthread_create(&thread_process, NULL, process_thread, top)) {
1309         ui__error("Could not create process thread.\n");
1310         return ret;
1311     }
1312 
1313     if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1314                                 display_thread), top)) {
1315         ui__error("Could not create display thread.\n");
1316         goto out_join_thread;
1317     }
1318 
1319     if (top->realtime_prio) {
1320         struct sched_param param;
1321 
1322         param.sched_priority = top->realtime_prio;
1323         if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1324             ui__error("Could not set realtime priority.\n");
1325             goto out_join;
1326         }
1327     }
1328 
1329     /* Wait for a minimal set of events before starting the snapshot */
1330     evlist__poll(top->evlist, 100);
1331 
1332     perf_top__mmap_read(top);
1333 
1334     while (!done) {
1335         u64 hits = top->samples;
1336 
1337         perf_top__mmap_read(top);
1338 
1339         if (opts->overwrite || (hits == top->samples))
1340             ret = evlist__poll(top->evlist, 100);
1341 
1342         if (resize) {
1343             perf_top__resize(top);
1344             resize = 0;
1345         }
1346     }
1347 
1348     ret = 0;
1349 out_join:
1350     pthread_join(thread, NULL);
1351 out_join_thread:
1352     pthread_cond_signal(&top->qe.cond);
1353     pthread_join(thread_process, NULL);
1354     return ret;
1355 }
1356 
1357 static int
1358 callchain_opt(const struct option *opt, const char *arg, int unset)
1359 {
1360     symbol_conf.use_callchain = true;
1361     return record_callchain_opt(opt, arg, unset);
1362 }
1363 
1364 static int
1365 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1366 {
1367     struct callchain_param *callchain = opt->value;
1368 
1369     callchain->enabled = !unset;
1370     callchain->record_mode = CALLCHAIN_FP;
1371 
1372     /*
1373      * --no-call-graph
1374      */
1375     if (unset) {
1376         symbol_conf.use_callchain = false;
1377         callchain->record_mode = CALLCHAIN_NONE;
1378         return 0;
1379     }
1380 
1381     return parse_callchain_top_opt(arg);
1382 }
1383 
1384 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1385 {
1386     if (!strcmp(var, "top.call-graph")) {
1387         var = "call-graph.record-mode";
1388         return perf_default_config(var, value, cb);
1389     }
1390     if (!strcmp(var, "top.children")) {
1391         symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1392         return 0;
1393     }
1394 
1395     return 0;
1396 }
1397 
1398 static int
1399 parse_percent_limit(const struct option *opt, const char *arg,
1400             int unset __maybe_unused)
1401 {
1402     struct perf_top *top = opt->value;
1403 
1404     top->min_percent = strtof(arg, NULL);
1405     return 0;
1406 }
1407 
1408 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1409     "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1410 
1411 int cmd_top(int argc, const char **argv)
1412 {
1413     char errbuf[BUFSIZ];
1414     struct perf_top top = {
1415         .count_filter        = 5,
1416         .delay_secs      = 2,
1417         .record_opts = {
1418             .mmap_pages = UINT_MAX,
1419             .user_freq  = UINT_MAX,
1420             .user_interval  = ULLONG_MAX,
1421             .freq       = 4000, /* 4 KHz */
1422             .target     = {
1423                 .uses_mmap   = true,
1424             },
1425             /*
1426              * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1427              * when we pause, fix that and reenable. Probably using a
1428              * separate evlist with a dummy event, i.e. a non-overwrite
1429              * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1430              * stays in overwrite mode. -acme
1431              * */
1432             .overwrite  = 0,
1433             .sample_time    = true,
1434             .sample_time_set = true,
1435         },
1436         .max_stack       = sysctl__max_stack(),
1437         .annotation_opts     = annotation__default_options,
1438         .nr_threads_synthesize = UINT_MAX,
1439     };
1440     struct record_opts *opts = &top.record_opts;
1441     struct target *target = &opts->target;
1442     const struct option options[] = {
1443     OPT_CALLBACK('e', "event", &top.evlist, "event",
1444              "event selector. use 'perf list' to list available events",
1445              parse_events_option),
1446     OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1447     OPT_STRING('p', "pid", &target->pid, "pid",
1448             "profile events on existing process id"),
1449     OPT_STRING('t', "tid", &target->tid, "tid",
1450             "profile events on existing thread id"),
1451     OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1452                 "system-wide collection from all CPUs"),
1453     OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1454             "list of cpus to monitor"),
1455     OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1456            "file", "vmlinux pathname"),
1457     OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1458             "don't load vmlinux even if found"),
1459     OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1460            "file", "kallsyms pathname"),
1461     OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1462             "hide kernel symbols"),
1463     OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1464              "number of mmap data pages", evlist__parse_mmap_pages),
1465     OPT_INTEGER('r', "realtime", &top.realtime_prio,
1466             "collect data with this RT SCHED_FIFO priority"),
1467     OPT_INTEGER('d', "delay", &top.delay_secs,
1468             "number of seconds to delay between refreshes"),
1469     OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1470                 "dump the symbol table used for profiling"),
1471     OPT_INTEGER('f', "count-filter", &top.count_filter,
1472             "only display functions with more events than this"),
1473     OPT_BOOLEAN(0, "group", &opts->group,
1474                 "put the counters into a counter group"),
1475     OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1476             "child tasks do not inherit counters"),
1477     OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1478             "symbol to annotate"),
1479     OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1480     OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1481              "profile at this frequency",
1482               record__parse_freq),
1483     OPT_INTEGER('E', "entries", &top.print_entries,
1484             "display this many functions"),
1485     OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1486             "hide user symbols"),
1487 #ifdef HAVE_SLANG_SUPPORT
1488     OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1489 #endif
1490     OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1491     OPT_INCR('v', "verbose", &verbose,
1492             "be more verbose (show counter open errors, etc)"),
1493     OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1494            "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1495            " Please refer the man page for the complete list."),
1496     OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1497            "output field(s): overhead, period, sample plus all of sort keys"),
1498     OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1499             "Show a column with the number of samples"),
1500     OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1501                NULL, "enables call-graph recording and display",
1502                &callchain_opt),
1503     OPT_CALLBACK(0, "call-graph", &callchain_param,
1504              "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1505              top_callchain_help, &parse_callchain_opt),
1506     OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1507             "Accumulate callchains of children and show total overhead as well"),
1508     OPT_INTEGER(0, "max-stack", &top.max_stack,
1509             "Set the maximum stack depth when parsing the callchain. "
1510             "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1511     OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1512            "ignore callees of these functions in call graphs",
1513            report_parse_ignore_callees_opt),
1514     OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1515             "Show a column with the sum of periods"),
1516     OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1517            "only consider symbols in these dsos"),
1518     OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1519            "only consider symbols in these comms"),
1520     OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1521            "only consider these symbols"),
1522     OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1523             "Interleave source code with assembly code (default)"),
1524     OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1525             "Display raw encoding of assembly instructions (default)"),
1526     OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1527             "Enable kernel symbol demangling"),
1528     OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1529     OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1530             "objdump binary to use for disassembly and annotations"),
1531     OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1532            "Specify disassembler style (e.g. -M intel for intel syntax)"),
1533     OPT_STRING(0, "prefix", &top.annotation_opts.prefix, "prefix",
1534             "Add prefix to source file path names in programs (with --prefix-strip)"),
1535     OPT_STRING(0, "prefix-strip", &top.annotation_opts.prefix_strip, "N",
1536             "Strip first N entries of source file path name in programs (with --prefix)"),
1537     OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1538     OPT_CALLBACK(0, "percent-limit", &top, "percent",
1539              "Don't show entries under that percent", parse_percent_limit),
1540     OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1541              "How to display percentage of filtered entries", parse_filter_percentage),
1542     OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1543            "width[,width...]",
1544            "don't try to adjust column width, use these fixed values"),
1545     OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1546             "per thread proc mmap processing timeout in ms"),
1547     OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1548              "branch any", "sample any taken branches",
1549              parse_branch_stack),
1550     OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1551              "branch filter mask", "branch stack filter modes",
1552              parse_branch_stack),
1553     OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1554             "Show raw trace event output (do not use print fmt or plugins)"),
1555     OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1556             "Show entries in a hierarchy"),
1557     OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1558             "Use a backward ring buffer, default: no"),
1559     OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1560     OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1561             "number of thread to run event synthesize"),
1562     OPT_CALLBACK('G', "cgroup", &top.evlist, "name",
1563              "monitor event in cgroup name only", parse_cgroups),
1564     OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1565             "Record namespaces events"),
1566     OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1567             "Record cgroup events"),
1568     OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1569             "Sort the output by the event at the index n in group. "
1570             "If n is invalid, sort by the first event. "
1571             "WARNING: should be used on grouped events."),
1572     OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1573             "Enable LBR callgraph stitching approach"),
1574 #ifdef HAVE_LIBPFM
1575     OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1576         "libpfm4 event selector. use 'perf list' to list available events",
1577         parse_libpfm_events_option),
1578 #endif
1579     OPTS_EVSWITCH(&top.evswitch),
1580     OPT_END()
1581     };
1582     const char * const top_usage[] = {
1583         "perf top [<options>]",
1584         NULL
1585     };
1586     int status = hists__init();
1587 
1588     if (status < 0)
1589         return status;
1590 
1591     top.annotation_opts.min_pcnt = 5;
1592     top.annotation_opts.context  = 4;
1593 
1594     top.evlist = evlist__new();
1595     if (top.evlist == NULL)
1596         return -ENOMEM;
1597 
1598     status = perf_config(perf_top_config, &top);
1599     if (status)
1600         return status;
1601     /*
1602      * Since the per arch annotation init routine may need the cpuid, read
1603      * it here, since we are not getting this from the perf.data header.
1604      */
1605     status = perf_env__read_cpuid(&perf_env);
1606     if (status) {
1607         /*
1608          * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1609          * warn the user explicitly.
1610          */
1611         eprintf(status == ENOSYS ? 1 : 0, verbose,
1612             "Couldn't read the cpuid for this machine: %s\n",
1613             str_error_r(errno, errbuf, sizeof(errbuf)));
1614     }
1615     top.evlist->env = &perf_env;
1616 
1617     argc = parse_options(argc, argv, options, top_usage, 0);
1618     if (argc)
1619         usage_with_options(top_usage, options);
1620 
1621     status = symbol__validate_sym_arguments();
1622     if (status)
1623         goto out_delete_evlist;
1624 
1625     if (annotate_check_args(&top.annotation_opts) < 0)
1626         goto out_delete_evlist;
1627 
1628     if (!top.evlist->core.nr_entries &&
1629         evlist__add_default(top.evlist) < 0) {
1630         pr_err("Not enough memory for event selector list\n");
1631         goto out_delete_evlist;
1632     }
1633 
1634     status = evswitch__init(&top.evswitch, top.evlist, stderr);
1635     if (status)
1636         goto out_delete_evlist;
1637 
1638     if (symbol_conf.report_hierarchy) {
1639         /* disable incompatible options */
1640         symbol_conf.event_group = false;
1641         symbol_conf.cumulate_callchain = false;
1642 
1643         if (field_order) {
1644             pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1645             parse_options_usage(top_usage, options, "fields", 0);
1646             parse_options_usage(NULL, options, "hierarchy", 0);
1647             goto out_delete_evlist;
1648         }
1649     }
1650 
1651     if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1652         pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1653         goto out_delete_evlist;
1654     }
1655 
1656     if (nr_cgroups > 0 && opts->record_cgroup) {
1657         pr_err("--cgroup and --all-cgroups cannot be used together\n");
1658         goto out_delete_evlist;
1659     }
1660 
1661     if (opts->branch_stack && callchain_param.enabled)
1662         symbol_conf.show_branchflag_count = true;
1663 
1664     sort__mode = SORT_MODE__TOP;
1665     /* display thread wants entries to be collapsed in a different tree */
1666     perf_hpp_list.need_collapse = 1;
1667 
1668     if (top.use_stdio)
1669         use_browser = 0;
1670 #ifdef HAVE_SLANG_SUPPORT
1671     else if (top.use_tui)
1672         use_browser = 1;
1673 #endif
1674 
1675     setup_browser(false);
1676 
1677     if (setup_sorting(top.evlist) < 0) {
1678         if (sort_order)
1679             parse_options_usage(top_usage, options, "s", 1);
1680         if (field_order)
1681             parse_options_usage(sort_order ? NULL : top_usage,
1682                         options, "fields", 0);
1683         goto out_delete_evlist;
1684     }
1685 
1686     status = target__validate(target);
1687     if (status) {
1688         target__strerror(target, status, errbuf, BUFSIZ);
1689         ui__warning("%s\n", errbuf);
1690     }
1691 
1692     status = target__parse_uid(target);
1693     if (status) {
1694         int saved_errno = errno;
1695 
1696         target__strerror(target, status, errbuf, BUFSIZ);
1697         ui__error("%s\n", errbuf);
1698 
1699         status = -saved_errno;
1700         goto out_delete_evlist;
1701     }
1702 
1703     if (target__none(target))
1704         target->system_wide = true;
1705 
1706     if (evlist__create_maps(top.evlist, target) < 0) {
1707         ui__error("Couldn't create thread/CPU maps: %s\n",
1708               errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1709         goto out_delete_evlist;
1710     }
1711 
1712     if (top.delay_secs < 1)
1713         top.delay_secs = 1;
1714 
1715     if (record_opts__config(opts)) {
1716         status = -EINVAL;
1717         goto out_delete_evlist;
1718     }
1719 
1720     top.sym_evsel = evlist__first(top.evlist);
1721 
1722     if (!callchain_param.enabled) {
1723         symbol_conf.cumulate_callchain = false;
1724         perf_hpp__cancel_cumulate();
1725     }
1726 
1727     if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1728         callchain_param.order = ORDER_CALLER;
1729 
1730     status = symbol__annotation_init();
1731     if (status < 0)
1732         goto out_delete_evlist;
1733 
1734     annotation_config__init(&top.annotation_opts);
1735 
1736     symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1737     status = symbol__init(NULL);
1738     if (status < 0)
1739         goto out_delete_evlist;
1740 
1741     sort__setup_elide(stdout);
1742 
1743     get_term_dimensions(&top.winsize);
1744     if (top.print_entries == 0) {
1745         perf_top__update_print_entries(&top);
1746         signal(SIGWINCH, winch_sig);
1747     }
1748 
1749     top.session = perf_session__new(NULL, NULL);
1750     if (IS_ERR(top.session)) {
1751         status = PTR_ERR(top.session);
1752         goto out_delete_evlist;
1753     }
1754 
1755 #ifdef HAVE_LIBBPF_SUPPORT
1756     if (!top.record_opts.no_bpf_event) {
1757         top.sb_evlist = evlist__new();
1758 
1759         if (top.sb_evlist == NULL) {
1760             pr_err("Couldn't create side band evlist.\n.");
1761             goto out_delete_evlist;
1762         }
1763 
1764         if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) {
1765             pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1766             goto out_delete_evlist;
1767         }
1768     }
1769 #endif
1770 
1771     if (evlist__start_sb_thread(top.sb_evlist, target)) {
1772         pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1773         opts->no_bpf_event = true;
1774     }
1775 
1776     status = __cmd_top(&top);
1777 
1778     if (!opts->no_bpf_event)
1779         evlist__stop_sb_thread(top.sb_evlist);
1780 
1781 out_delete_evlist:
1782     evlist__delete(top.evlist);
1783     perf_session__delete(top.session);
1784 
1785     return status;
1786 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!