Back to home page

OSCL-LXR

 
 

    


0001 // SPDX-License-Identifier: GPL-2.0
0002 #include "builtin.h"
0003 #include "perf.h"
0004 
0005 #include "util/build-id.h"
0006 #include "util/evsel.h"
0007 #include "util/evlist.h"
0008 #include "util/mmap.h"
0009 #include "util/term.h"
0010 #include "util/symbol.h"
0011 #include "util/thread.h"
0012 #include "util/header.h"
0013 #include "util/session.h"
0014 #include "util/intlist.h"
0015 #include <subcmd/pager.h>
0016 #include <subcmd/parse-options.h>
0017 #include "util/trace-event.h"
0018 #include "util/debug.h"
0019 #include "util/tool.h"
0020 #include "util/stat.h"
0021 #include "util/synthetic-events.h"
0022 #include "util/top.h"
0023 #include "util/data.h"
0024 #include "util/ordered-events.h"
0025 #include "util/kvm-stat.h"
0026 #include "ui/ui.h"
0027 #include "util/string2.h"
0028 
0029 #include <sys/prctl.h>
0030 #ifdef HAVE_TIMERFD_SUPPORT
0031 #include <sys/timerfd.h>
0032 #endif
0033 #include <sys/time.h>
0034 #include <sys/types.h>
0035 #include <sys/stat.h>
0036 #include <fcntl.h>
0037 
0038 #include <linux/err.h>
0039 #include <linux/kernel.h>
0040 #include <linux/string.h>
0041 #include <linux/time64.h>
0042 #include <linux/zalloc.h>
0043 #include <errno.h>
0044 #include <inttypes.h>
0045 #include <poll.h>
0046 #include <termios.h>
0047 #include <semaphore.h>
0048 #include <signal.h>
0049 #include <math.h>
0050 #include <perf/mmap.h>
0051 
0052 static const char *get_filename_for_perf_kvm(void)
0053 {
0054     const char *filename;
0055 
0056     if (perf_host && !perf_guest)
0057         filename = strdup("perf.data.host");
0058     else if (!perf_host && perf_guest)
0059         filename = strdup("perf.data.guest");
0060     else
0061         filename = strdup("perf.data.kvm");
0062 
0063     return filename;
0064 }
0065 
0066 #ifdef HAVE_KVM_STAT_SUPPORT
0067 
0068 void exit_event_get_key(struct evsel *evsel,
0069             struct perf_sample *sample,
0070             struct event_key *key)
0071 {
0072     key->info = 0;
0073     key->key  = evsel__intval(evsel, sample, kvm_exit_reason);
0074 }
0075 
0076 bool kvm_exit_event(struct evsel *evsel)
0077 {
0078     return !strcmp(evsel->name, kvm_exit_trace);
0079 }
0080 
0081 bool exit_event_begin(struct evsel *evsel,
0082               struct perf_sample *sample, struct event_key *key)
0083 {
0084     if (kvm_exit_event(evsel)) {
0085         exit_event_get_key(evsel, sample, key);
0086         return true;
0087     }
0088 
0089     return false;
0090 }
0091 
0092 bool kvm_entry_event(struct evsel *evsel)
0093 {
0094     return !strcmp(evsel->name, kvm_entry_trace);
0095 }
0096 
0097 bool exit_event_end(struct evsel *evsel,
0098             struct perf_sample *sample __maybe_unused,
0099             struct event_key *key __maybe_unused)
0100 {
0101     return kvm_entry_event(evsel);
0102 }
0103 
0104 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
0105                    struct exit_reasons_table *tbl,
0106                    u64 exit_code)
0107 {
0108     while (tbl->reason != NULL) {
0109         if (tbl->exit_code == exit_code)
0110             return tbl->reason;
0111         tbl++;
0112     }
0113 
0114     pr_err("unknown kvm exit code:%lld on %s\n",
0115         (unsigned long long)exit_code, kvm->exit_reasons_isa);
0116     return "UNKNOWN";
0117 }
0118 
0119 void exit_event_decode_key(struct perf_kvm_stat *kvm,
0120                struct event_key *key,
0121                char *decode)
0122 {
0123     const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
0124                           key->key);
0125 
0126     scnprintf(decode, decode_str_len, "%s", exit_reason);
0127 }
0128 
0129 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
0130 {
0131     struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
0132 
0133     for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
0134         if (!strcmp(events_ops->name, kvm->report_event)) {
0135             kvm->events_ops = events_ops->ops;
0136             return true;
0137         }
0138     }
0139 
0140     return false;
0141 }
0142 
0143 struct vcpu_event_record {
0144     int vcpu_id;
0145     u64 start_time;
0146     struct kvm_event *last_event;
0147 };
0148 
0149 
0150 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
0151 {
0152     unsigned int i;
0153 
0154     for (i = 0; i < EVENTS_CACHE_SIZE; i++)
0155         INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
0156 }
0157 
0158 #ifdef HAVE_TIMERFD_SUPPORT
0159 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
0160 {
0161     struct list_head *head;
0162     struct kvm_event *event;
0163     unsigned int i;
0164     int j;
0165 
0166     for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
0167         head = &kvm_events_cache[i];
0168         list_for_each_entry(event, head, hash_entry) {
0169             /* reset stats for event */
0170             event->total.time = 0;
0171             init_stats(&event->total.stats);
0172 
0173             for (j = 0; j < event->max_vcpu; ++j) {
0174                 event->vcpu[j].time = 0;
0175                 init_stats(&event->vcpu[j].stats);
0176             }
0177         }
0178     }
0179 }
0180 #endif
0181 
0182 static int kvm_events_hash_fn(u64 key)
0183 {
0184     return key & (EVENTS_CACHE_SIZE - 1);
0185 }
0186 
0187 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
0188 {
0189     int old_max_vcpu = event->max_vcpu;
0190     void *prev;
0191 
0192     if (vcpu_id < event->max_vcpu)
0193         return true;
0194 
0195     while (event->max_vcpu <= vcpu_id)
0196         event->max_vcpu += DEFAULT_VCPU_NUM;
0197 
0198     prev = event->vcpu;
0199     event->vcpu = realloc(event->vcpu,
0200                   event->max_vcpu * sizeof(*event->vcpu));
0201     if (!event->vcpu) {
0202         free(prev);
0203         pr_err("Not enough memory\n");
0204         return false;
0205     }
0206 
0207     memset(event->vcpu + old_max_vcpu, 0,
0208            (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
0209     return true;
0210 }
0211 
0212 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
0213 {
0214     struct kvm_event *event;
0215 
0216     event = zalloc(sizeof(*event));
0217     if (!event) {
0218         pr_err("Not enough memory\n");
0219         return NULL;
0220     }
0221 
0222     event->key = *key;
0223     init_stats(&event->total.stats);
0224     return event;
0225 }
0226 
0227 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
0228                            struct event_key *key)
0229 {
0230     struct kvm_event *event;
0231     struct list_head *head;
0232 
0233     BUG_ON(key->key == INVALID_KEY);
0234 
0235     head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
0236     list_for_each_entry(event, head, hash_entry) {
0237         if (event->key.key == key->key && event->key.info == key->info)
0238             return event;
0239     }
0240 
0241     event = kvm_alloc_init_event(key);
0242     if (!event)
0243         return NULL;
0244 
0245     list_add(&event->hash_entry, head);
0246     return event;
0247 }
0248 
0249 static bool handle_begin_event(struct perf_kvm_stat *kvm,
0250                    struct vcpu_event_record *vcpu_record,
0251                    struct event_key *key, u64 timestamp)
0252 {
0253     struct kvm_event *event = NULL;
0254 
0255     if (key->key != INVALID_KEY)
0256         event = find_create_kvm_event(kvm, key);
0257 
0258     vcpu_record->last_event = event;
0259     vcpu_record->start_time = timestamp;
0260     return true;
0261 }
0262 
0263 static void
0264 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
0265 {
0266     kvm_stats->time += time_diff;
0267     update_stats(&kvm_stats->stats, time_diff);
0268 }
0269 
0270 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
0271 {
0272     struct kvm_event_stats *kvm_stats = &event->total;
0273 
0274     if (vcpu_id != -1)
0275         kvm_stats = &event->vcpu[vcpu_id];
0276 
0277     return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
0278                 avg_stats(&kvm_stats->stats));
0279 }
0280 
0281 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
0282                  u64 time_diff)
0283 {
0284     if (vcpu_id == -1) {
0285         kvm_update_event_stats(&event->total, time_diff);
0286         return true;
0287     }
0288 
0289     if (!kvm_event_expand(event, vcpu_id))
0290         return false;
0291 
0292     kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
0293     return true;
0294 }
0295 
0296 static bool is_child_event(struct perf_kvm_stat *kvm,
0297                struct evsel *evsel,
0298                struct perf_sample *sample,
0299                struct event_key *key)
0300 {
0301     struct child_event_ops *child_ops;
0302 
0303     child_ops = kvm->events_ops->child_ops;
0304 
0305     if (!child_ops)
0306         return false;
0307 
0308     for (; child_ops->name; child_ops++) {
0309         if (!strcmp(evsel->name, child_ops->name)) {
0310             child_ops->get_key(evsel, sample, key);
0311             return true;
0312         }
0313     }
0314 
0315     return false;
0316 }
0317 
0318 static bool handle_child_event(struct perf_kvm_stat *kvm,
0319                    struct vcpu_event_record *vcpu_record,
0320                    struct event_key *key,
0321                    struct perf_sample *sample __maybe_unused)
0322 {
0323     struct kvm_event *event = NULL;
0324 
0325     if (key->key != INVALID_KEY)
0326         event = find_create_kvm_event(kvm, key);
0327 
0328     vcpu_record->last_event = event;
0329 
0330     return true;
0331 }
0332 
0333 static bool skip_event(const char *event)
0334 {
0335     const char * const *skip_events;
0336 
0337     for (skip_events = kvm_skip_events; *skip_events; skip_events++)
0338         if (!strcmp(event, *skip_events))
0339             return true;
0340 
0341     return false;
0342 }
0343 
0344 static bool handle_end_event(struct perf_kvm_stat *kvm,
0345                  struct vcpu_event_record *vcpu_record,
0346                  struct event_key *key,
0347                  struct perf_sample *sample)
0348 {
0349     struct kvm_event *event;
0350     u64 time_begin, time_diff;
0351     int vcpu;
0352 
0353     if (kvm->trace_vcpu == -1)
0354         vcpu = -1;
0355     else
0356         vcpu = vcpu_record->vcpu_id;
0357 
0358     event = vcpu_record->last_event;
0359     time_begin = vcpu_record->start_time;
0360 
0361     /* The begin event is not caught. */
0362     if (!time_begin)
0363         return true;
0364 
0365     /*
0366      * In some case, the 'begin event' only records the start timestamp,
0367      * the actual event is recognized in the 'end event' (e.g. mmio-event).
0368      */
0369 
0370     /* Both begin and end events did not get the key. */
0371     if (!event && key->key == INVALID_KEY)
0372         return true;
0373 
0374     if (!event)
0375         event = find_create_kvm_event(kvm, key);
0376 
0377     if (!event)
0378         return false;
0379 
0380     vcpu_record->last_event = NULL;
0381     vcpu_record->start_time = 0;
0382 
0383     /* seems to happen once in a while during live mode */
0384     if (sample->time < time_begin) {
0385         pr_debug("End time before begin time; skipping event.\n");
0386         return true;
0387     }
0388 
0389     time_diff = sample->time - time_begin;
0390 
0391     if (kvm->duration && time_diff > kvm->duration) {
0392         char decode[decode_str_len];
0393 
0394         kvm->events_ops->decode_key(kvm, &event->key, decode);
0395         if (!skip_event(decode)) {
0396             pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
0397                  sample->time, sample->pid, vcpu_record->vcpu_id,
0398                  decode, time_diff / NSEC_PER_USEC);
0399         }
0400     }
0401 
0402     return update_kvm_event(event, vcpu, time_diff);
0403 }
0404 
0405 static
0406 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
0407                       struct evsel *evsel,
0408                       struct perf_sample *sample)
0409 {
0410     /* Only kvm_entry records vcpu id. */
0411     if (!thread__priv(thread) && kvm_entry_event(evsel)) {
0412         struct vcpu_event_record *vcpu_record;
0413 
0414         vcpu_record = zalloc(sizeof(*vcpu_record));
0415         if (!vcpu_record) {
0416             pr_err("%s: Not enough memory\n", __func__);
0417             return NULL;
0418         }
0419 
0420         vcpu_record->vcpu_id = evsel__intval(evsel, sample, vcpu_id_str);
0421         thread__set_priv(thread, vcpu_record);
0422     }
0423 
0424     return thread__priv(thread);
0425 }
0426 
0427 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
0428                  struct thread *thread,
0429                  struct evsel *evsel,
0430                  struct perf_sample *sample)
0431 {
0432     struct vcpu_event_record *vcpu_record;
0433     struct event_key key = { .key = INVALID_KEY,
0434                  .exit_reasons = kvm->exit_reasons };
0435 
0436     vcpu_record = per_vcpu_record(thread, evsel, sample);
0437     if (!vcpu_record)
0438         return true;
0439 
0440     /* only process events for vcpus user cares about */
0441     if ((kvm->trace_vcpu != -1) &&
0442         (kvm->trace_vcpu != vcpu_record->vcpu_id))
0443         return true;
0444 
0445     if (kvm->events_ops->is_begin_event(evsel, sample, &key))
0446         return handle_begin_event(kvm, vcpu_record, &key, sample->time);
0447 
0448     if (is_child_event(kvm, evsel, sample, &key))
0449         return handle_child_event(kvm, vcpu_record, &key, sample);
0450 
0451     if (kvm->events_ops->is_end_event(evsel, sample, &key))
0452         return handle_end_event(kvm, vcpu_record, &key, sample);
0453 
0454     return true;
0455 }
0456 
0457 #define GET_EVENT_KEY(func, field)                  \
0458 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)     \
0459 {                                   \
0460     if (vcpu == -1)                         \
0461         return event->total.field;              \
0462                                     \
0463     if (vcpu >= event->max_vcpu)                    \
0464         return 0;                       \
0465                                     \
0466     return event->vcpu[vcpu].field;                 \
0467 }
0468 
0469 #define COMPARE_EVENT_KEY(func, field)                  \
0470 GET_EVENT_KEY(func, field)                      \
0471 static int compare_kvm_event_ ## func(struct kvm_event *one,        \
0472                     struct kvm_event *two, int vcpu)\
0473 {                                   \
0474     return get_event_ ##func(one, vcpu) >               \
0475                 get_event_ ##func(two, vcpu);       \
0476 }
0477 
0478 GET_EVENT_KEY(time, time);
0479 COMPARE_EVENT_KEY(count, stats.n);
0480 COMPARE_EVENT_KEY(mean, stats.mean);
0481 GET_EVENT_KEY(max, stats.max);
0482 GET_EVENT_KEY(min, stats.min);
0483 
0484 #define DEF_SORT_NAME_KEY(name, compare_key)                \
0485     { #name, compare_kvm_event_ ## compare_key }
0486 
0487 static struct kvm_event_key keys[] = {
0488     DEF_SORT_NAME_KEY(sample, count),
0489     DEF_SORT_NAME_KEY(time, mean),
0490     { NULL, NULL }
0491 };
0492 
0493 static bool select_key(struct perf_kvm_stat *kvm)
0494 {
0495     int i;
0496 
0497     for (i = 0; keys[i].name; i++) {
0498         if (!strcmp(keys[i].name, kvm->sort_key)) {
0499             kvm->compare = keys[i].key;
0500             return true;
0501         }
0502     }
0503 
0504     pr_err("Unknown compare key:%s\n", kvm->sort_key);
0505     return false;
0506 }
0507 
0508 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
0509                  key_cmp_fun bigger, int vcpu)
0510 {
0511     struct rb_node **rb = &result->rb_node;
0512     struct rb_node *parent = NULL;
0513     struct kvm_event *p;
0514 
0515     while (*rb) {
0516         p = container_of(*rb, struct kvm_event, rb);
0517         parent = *rb;
0518 
0519         if (bigger(event, p, vcpu))
0520             rb = &(*rb)->rb_left;
0521         else
0522             rb = &(*rb)->rb_right;
0523     }
0524 
0525     rb_link_node(&event->rb, parent, rb);
0526     rb_insert_color(&event->rb, result);
0527 }
0528 
0529 static void
0530 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
0531 {
0532     int vcpu = kvm->trace_vcpu;
0533 
0534     kvm->total_count += get_event_count(event, vcpu);
0535     kvm->total_time += get_event_time(event, vcpu);
0536 }
0537 
0538 static bool event_is_valid(struct kvm_event *event, int vcpu)
0539 {
0540     return !!get_event_count(event, vcpu);
0541 }
0542 
0543 static void sort_result(struct perf_kvm_stat *kvm)
0544 {
0545     unsigned int i;
0546     int vcpu = kvm->trace_vcpu;
0547     struct kvm_event *event;
0548 
0549     for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
0550         list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
0551             if (event_is_valid(event, vcpu)) {
0552                 update_total_count(kvm, event);
0553                 insert_to_result(&kvm->result, event,
0554                          kvm->compare, vcpu);
0555             }
0556         }
0557     }
0558 }
0559 
0560 /* returns left most element of result, and erase it */
0561 static struct kvm_event *pop_from_result(struct rb_root *result)
0562 {
0563     struct rb_node *node = rb_first(result);
0564 
0565     if (!node)
0566         return NULL;
0567 
0568     rb_erase(node, result);
0569     return container_of(node, struct kvm_event, rb);
0570 }
0571 
0572 static void print_vcpu_info(struct perf_kvm_stat *kvm)
0573 {
0574     int vcpu = kvm->trace_vcpu;
0575 
0576     pr_info("Analyze events for ");
0577 
0578     if (kvm->opts.target.system_wide)
0579         pr_info("all VMs, ");
0580     else if (kvm->opts.target.pid)
0581         pr_info("pid(s) %s, ", kvm->opts.target.pid);
0582     else
0583         pr_info("dazed and confused on what is monitored, ");
0584 
0585     if (vcpu == -1)
0586         pr_info("all VCPUs:\n\n");
0587     else
0588         pr_info("VCPU %d:\n\n", vcpu);
0589 }
0590 
0591 static void show_timeofday(void)
0592 {
0593     char date[64];
0594     struct timeval tv;
0595     struct tm ltime;
0596 
0597     gettimeofday(&tv, NULL);
0598     if (localtime_r(&tv.tv_sec, &ltime)) {
0599         strftime(date, sizeof(date), "%H:%M:%S", &ltime);
0600         pr_info("%s.%06ld", date, tv.tv_usec);
0601     } else
0602         pr_info("00:00:00.000000");
0603 
0604     return;
0605 }
0606 
0607 static void print_result(struct perf_kvm_stat *kvm)
0608 {
0609     char decode[decode_str_len];
0610     struct kvm_event *event;
0611     int vcpu = kvm->trace_vcpu;
0612 
0613     if (kvm->live) {
0614         puts(CONSOLE_CLEAR);
0615         show_timeofday();
0616     }
0617 
0618     pr_info("\n\n");
0619     print_vcpu_info(kvm);
0620     pr_info("%*s ", decode_str_len, kvm->events_ops->name);
0621     pr_info("%10s ", "Samples");
0622     pr_info("%9s ", "Samples%");
0623 
0624     pr_info("%9s ", "Time%");
0625     pr_info("%11s ", "Min Time");
0626     pr_info("%11s ", "Max Time");
0627     pr_info("%16s ", "Avg time");
0628     pr_info("\n\n");
0629 
0630     while ((event = pop_from_result(&kvm->result))) {
0631         u64 ecount, etime, max, min;
0632 
0633         ecount = get_event_count(event, vcpu);
0634         etime = get_event_time(event, vcpu);
0635         max = get_event_max(event, vcpu);
0636         min = get_event_min(event, vcpu);
0637 
0638         kvm->events_ops->decode_key(kvm, &event->key, decode);
0639         pr_info("%*s ", decode_str_len, decode);
0640         pr_info("%10llu ", (unsigned long long)ecount);
0641         pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
0642         pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
0643         pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
0644         pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
0645         pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
0646             kvm_event_rel_stddev(vcpu, event));
0647         pr_info("\n");
0648     }
0649 
0650     pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
0651         kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
0652 
0653     if (kvm->lost_events)
0654         pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
0655 }
0656 
0657 #ifdef HAVE_TIMERFD_SUPPORT
0658 static int process_lost_event(struct perf_tool *tool,
0659                   union perf_event *event __maybe_unused,
0660                   struct perf_sample *sample __maybe_unused,
0661                   struct machine *machine __maybe_unused)
0662 {
0663     struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
0664 
0665     kvm->lost_events++;
0666     return 0;
0667 }
0668 #endif
0669 
0670 static bool skip_sample(struct perf_kvm_stat *kvm,
0671             struct perf_sample *sample)
0672 {
0673     if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
0674         return true;
0675 
0676     return false;
0677 }
0678 
0679 static int process_sample_event(struct perf_tool *tool,
0680                 union perf_event *event,
0681                 struct perf_sample *sample,
0682                 struct evsel *evsel,
0683                 struct machine *machine)
0684 {
0685     int err = 0;
0686     struct thread *thread;
0687     struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
0688                          tool);
0689 
0690     if (skip_sample(kvm, sample))
0691         return 0;
0692 
0693     thread = machine__findnew_thread(machine, sample->pid, sample->tid);
0694     if (thread == NULL) {
0695         pr_debug("problem processing %d event, skipping it.\n",
0696             event->header.type);
0697         return -1;
0698     }
0699 
0700     if (!handle_kvm_event(kvm, thread, evsel, sample))
0701         err = -1;
0702 
0703     thread__put(thread);
0704     return err;
0705 }
0706 
0707 static int cpu_isa_config(struct perf_kvm_stat *kvm)
0708 {
0709     char buf[128], *cpuid;
0710     int err;
0711 
0712     if (kvm->live) {
0713         err = get_cpuid(buf, sizeof(buf));
0714         if (err != 0) {
0715             pr_err("Failed to look up CPU type: %s\n",
0716                    str_error_r(err, buf, sizeof(buf)));
0717             return -err;
0718         }
0719         cpuid = buf;
0720     } else
0721         cpuid = kvm->session->header.env.cpuid;
0722 
0723     if (!cpuid) {
0724         pr_err("Failed to look up CPU type\n");
0725         return -EINVAL;
0726     }
0727 
0728     err = cpu_isa_init(kvm, cpuid);
0729     if (err == -ENOTSUP)
0730         pr_err("CPU %s is not supported.\n", cpuid);
0731 
0732     return err;
0733 }
0734 
0735 static bool verify_vcpu(int vcpu)
0736 {
0737     if (vcpu != -1 && vcpu < 0) {
0738         pr_err("Invalid vcpu:%d.\n", vcpu);
0739         return false;
0740     }
0741 
0742     return true;
0743 }
0744 
0745 #ifdef HAVE_TIMERFD_SUPPORT
0746 /* keeping the max events to a modest level to keep
0747  * the processing of samples per mmap smooth.
0748  */
0749 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
0750 
0751 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
0752                    u64 *mmap_time)
0753 {
0754     struct evlist *evlist = kvm->evlist;
0755     union perf_event *event;
0756     struct mmap *md;
0757     u64 timestamp;
0758     s64 n = 0;
0759     int err;
0760 
0761     *mmap_time = ULLONG_MAX;
0762     md = &evlist->mmap[idx];
0763     err = perf_mmap__read_init(&md->core);
0764     if (err < 0)
0765         return (err == -EAGAIN) ? 0 : -1;
0766 
0767     while ((event = perf_mmap__read_event(&md->core)) != NULL) {
0768         err = evlist__parse_sample_timestamp(evlist, event, &timestamp);
0769         if (err) {
0770             perf_mmap__consume(&md->core);
0771             pr_err("Failed to parse sample\n");
0772             return -1;
0773         }
0774 
0775         err = perf_session__queue_event(kvm->session, event, timestamp, 0, NULL);
0776         /*
0777          * FIXME: Here we can't consume the event, as perf_session__queue_event will
0778          *        point to it, and it'll get possibly overwritten by the kernel.
0779          */
0780         perf_mmap__consume(&md->core);
0781 
0782         if (err) {
0783             pr_err("Failed to enqueue sample: %d\n", err);
0784             return -1;
0785         }
0786 
0787         /* save time stamp of our first sample for this mmap */
0788         if (n == 0)
0789             *mmap_time = timestamp;
0790 
0791         /* limit events per mmap handled all at once */
0792         n++;
0793         if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
0794             break;
0795     }
0796 
0797     perf_mmap__read_done(&md->core);
0798     return n;
0799 }
0800 
0801 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
0802 {
0803     int i, err, throttled = 0;
0804     s64 n, ntotal = 0;
0805     u64 flush_time = ULLONG_MAX, mmap_time;
0806 
0807     for (i = 0; i < kvm->evlist->core.nr_mmaps; i++) {
0808         n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
0809         if (n < 0)
0810             return -1;
0811 
0812         /* flush time is going to be the minimum of all the individual
0813          * mmap times. Essentially, we flush all the samples queued up
0814          * from the last pass under our minimal start time -- that leaves
0815          * a very small race for samples to come in with a lower timestamp.
0816          * The ioctl to return the perf_clock timestamp should close the
0817          * race entirely.
0818          */
0819         if (mmap_time < flush_time)
0820             flush_time = mmap_time;
0821 
0822         ntotal += n;
0823         if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
0824             throttled = 1;
0825     }
0826 
0827     /* flush queue after each round in which we processed events */
0828     if (ntotal) {
0829         struct ordered_events *oe = &kvm->session->ordered_events;
0830 
0831         oe->next_flush = flush_time;
0832         err = ordered_events__flush(oe, OE_FLUSH__ROUND);
0833         if (err) {
0834             if (kvm->lost_events)
0835                 pr_info("\nLost events: %" PRIu64 "\n\n",
0836                     kvm->lost_events);
0837             return err;
0838         }
0839     }
0840 
0841     return throttled;
0842 }
0843 
0844 static volatile int done;
0845 
0846 static void sig_handler(int sig __maybe_unused)
0847 {
0848     done = 1;
0849 }
0850 
0851 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
0852 {
0853     struct itimerspec new_value;
0854     int rc = -1;
0855 
0856     kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
0857     if (kvm->timerfd < 0) {
0858         pr_err("timerfd_create failed\n");
0859         goto out;
0860     }
0861 
0862     new_value.it_value.tv_sec = kvm->display_time;
0863     new_value.it_value.tv_nsec = 0;
0864     new_value.it_interval.tv_sec = kvm->display_time;
0865     new_value.it_interval.tv_nsec = 0;
0866 
0867     if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
0868         pr_err("timerfd_settime failed: %d\n", errno);
0869         close(kvm->timerfd);
0870         goto out;
0871     }
0872 
0873     rc = 0;
0874 out:
0875     return rc;
0876 }
0877 
0878 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
0879 {
0880     uint64_t c;
0881     int rc;
0882 
0883     rc = read(kvm->timerfd, &c, sizeof(uint64_t));
0884     if (rc < 0) {
0885         if (errno == EAGAIN)
0886             return 0;
0887 
0888         pr_err("Failed to read timer fd: %d\n", errno);
0889         return -1;
0890     }
0891 
0892     if (rc != sizeof(uint64_t)) {
0893         pr_err("Error reading timer fd - invalid size returned\n");
0894         return -1;
0895     }
0896 
0897     if (c != 1)
0898         pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
0899 
0900     /* update display */
0901     sort_result(kvm);
0902     print_result(kvm);
0903 
0904     /* reset counts */
0905     clear_events_cache_stats(kvm->kvm_events_cache);
0906     kvm->total_count = 0;
0907     kvm->total_time = 0;
0908     kvm->lost_events = 0;
0909 
0910     return 0;
0911 }
0912 
0913 static int fd_set_nonblock(int fd)
0914 {
0915     long arg = 0;
0916 
0917     arg = fcntl(fd, F_GETFL);
0918     if (arg < 0) {
0919         pr_err("Failed to get current flags for fd %d\n", fd);
0920         return -1;
0921     }
0922 
0923     if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
0924         pr_err("Failed to set non-block option on fd %d\n", fd);
0925         return -1;
0926     }
0927 
0928     return 0;
0929 }
0930 
0931 static int perf_kvm__handle_stdin(void)
0932 {
0933     int c;
0934 
0935     c = getc(stdin);
0936     if (c == 'q')
0937         return 1;
0938 
0939     return 0;
0940 }
0941 
0942 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
0943 {
0944     int nr_stdin, ret, err = -EINVAL;
0945     struct termios save;
0946 
0947     /* live flag must be set first */
0948     kvm->live = true;
0949 
0950     ret = cpu_isa_config(kvm);
0951     if (ret < 0)
0952         return ret;
0953 
0954     if (!verify_vcpu(kvm->trace_vcpu) ||
0955         !select_key(kvm) ||
0956         !register_kvm_events_ops(kvm)) {
0957         goto out;
0958     }
0959 
0960     set_term_quiet_input(&save);
0961     init_kvm_event_record(kvm);
0962 
0963     signal(SIGINT, sig_handler);
0964     signal(SIGTERM, sig_handler);
0965 
0966     /* add timer fd */
0967     if (perf_kvm__timerfd_create(kvm) < 0) {
0968         err = -1;
0969         goto out;
0970     }
0971 
0972     if (evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
0973         goto out;
0974 
0975     nr_stdin = evlist__add_pollfd(kvm->evlist, fileno(stdin));
0976     if (nr_stdin < 0)
0977         goto out;
0978 
0979     if (fd_set_nonblock(fileno(stdin)) != 0)
0980         goto out;
0981 
0982     /* everything is good - enable the events and process */
0983     evlist__enable(kvm->evlist);
0984 
0985     while (!done) {
0986         struct fdarray *fda = &kvm->evlist->core.pollfd;
0987         int rc;
0988 
0989         rc = perf_kvm__mmap_read(kvm);
0990         if (rc < 0)
0991             break;
0992 
0993         err = perf_kvm__handle_timerfd(kvm);
0994         if (err)
0995             goto out;
0996 
0997         if (fda->entries[nr_stdin].revents & POLLIN)
0998             done = perf_kvm__handle_stdin();
0999 
1000         if (!rc && !done)
1001             err = evlist__poll(kvm->evlist, 100);
1002     }
1003 
1004     evlist__disable(kvm->evlist);
1005 
1006     if (err == 0) {
1007         sort_result(kvm);
1008         print_result(kvm);
1009     }
1010 
1011 out:
1012     if (kvm->timerfd >= 0)
1013         close(kvm->timerfd);
1014 
1015     tcsetattr(0, TCSAFLUSH, &save);
1016     return err;
1017 }
1018 
1019 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1020 {
1021     int err, rc = -1;
1022     struct evsel *pos;
1023     struct evlist *evlist = kvm->evlist;
1024     char sbuf[STRERR_BUFSIZE];
1025 
1026     evlist__config(evlist, &kvm->opts, NULL);
1027 
1028     /*
1029      * Note: exclude_{guest,host} do not apply here.
1030      *       This command processes KVM tracepoints from host only
1031      */
1032     evlist__for_each_entry(evlist, pos) {
1033         struct perf_event_attr *attr = &pos->core.attr;
1034 
1035         /* make sure these *are* set */
1036         evsel__set_sample_bit(pos, TID);
1037         evsel__set_sample_bit(pos, TIME);
1038         evsel__set_sample_bit(pos, CPU);
1039         evsel__set_sample_bit(pos, RAW);
1040         /* make sure these are *not*; want as small a sample as possible */
1041         evsel__reset_sample_bit(pos, PERIOD);
1042         evsel__reset_sample_bit(pos, IP);
1043         evsel__reset_sample_bit(pos, CALLCHAIN);
1044         evsel__reset_sample_bit(pos, ADDR);
1045         evsel__reset_sample_bit(pos, READ);
1046         attr->mmap = 0;
1047         attr->comm = 0;
1048         attr->task = 0;
1049 
1050         attr->sample_period = 1;
1051 
1052         attr->watermark = 0;
1053         attr->wakeup_events = 1000;
1054 
1055         /* will enable all once we are ready */
1056         attr->disabled = 1;
1057     }
1058 
1059     err = evlist__open(evlist);
1060     if (err < 0) {
1061         printf("Couldn't create the events: %s\n",
1062                str_error_r(errno, sbuf, sizeof(sbuf)));
1063         goto out;
1064     }
1065 
1066     if (evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1067         ui__error("Failed to mmap the events: %s\n",
1068               str_error_r(errno, sbuf, sizeof(sbuf)));
1069         evlist__close(evlist);
1070         goto out;
1071     }
1072 
1073     rc = 0;
1074 
1075 out:
1076     return rc;
1077 }
1078 #endif
1079 
1080 static int read_events(struct perf_kvm_stat *kvm)
1081 {
1082     int ret;
1083 
1084     struct perf_tool eops = {
1085         .sample         = process_sample_event,
1086         .comm           = perf_event__process_comm,
1087         .namespaces     = perf_event__process_namespaces,
1088         .ordered_events     = true,
1089     };
1090     struct perf_data file = {
1091         .path  = kvm->file_name,
1092         .mode  = PERF_DATA_MODE_READ,
1093         .force = kvm->force,
1094     };
1095 
1096     kvm->tool = eops;
1097     kvm->session = perf_session__new(&file, &kvm->tool);
1098     if (IS_ERR(kvm->session)) {
1099         pr_err("Initializing perf session failed\n");
1100         return PTR_ERR(kvm->session);
1101     }
1102 
1103     symbol__init(&kvm->session->header.env);
1104 
1105     if (!perf_session__has_traces(kvm->session, "kvm record")) {
1106         ret = -EINVAL;
1107         goto out_delete;
1108     }
1109 
1110     /*
1111      * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1112      * traced in the old kernel.
1113      */
1114     ret = cpu_isa_config(kvm);
1115     if (ret < 0)
1116         goto out_delete;
1117 
1118     ret = perf_session__process_events(kvm->session);
1119 
1120 out_delete:
1121     perf_session__delete(kvm->session);
1122     return ret;
1123 }
1124 
1125 static int parse_target_str(struct perf_kvm_stat *kvm)
1126 {
1127     if (kvm->opts.target.pid) {
1128         kvm->pid_list = intlist__new(kvm->opts.target.pid);
1129         if (kvm->pid_list == NULL) {
1130             pr_err("Error parsing process id string\n");
1131             return -EINVAL;
1132         }
1133     }
1134 
1135     return 0;
1136 }
1137 
1138 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1139 {
1140     int ret = -EINVAL;
1141     int vcpu = kvm->trace_vcpu;
1142 
1143     if (parse_target_str(kvm) != 0)
1144         goto exit;
1145 
1146     if (!verify_vcpu(vcpu))
1147         goto exit;
1148 
1149     if (!select_key(kvm))
1150         goto exit;
1151 
1152     if (!register_kvm_events_ops(kvm))
1153         goto exit;
1154 
1155     init_kvm_event_record(kvm);
1156     setup_pager();
1157 
1158     ret = read_events(kvm);
1159     if (ret)
1160         goto exit;
1161 
1162     sort_result(kvm);
1163     print_result(kvm);
1164 
1165 exit:
1166     return ret;
1167 }
1168 
1169 #define STRDUP_FAIL_EXIT(s)     \
1170     ({  char *_p;       \
1171     _p = strdup(s);     \
1172         if (!_p)        \
1173             return -ENOMEM; \
1174         _p;         \
1175     })
1176 
1177 int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1178 {
1179     return 0;
1180 }
1181 
1182 static int
1183 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1184 {
1185     unsigned int rec_argc, i, j, events_tp_size;
1186     const char **rec_argv;
1187     const char * const record_args[] = {
1188         "record",
1189         "-R",
1190         "-m", "1024",
1191         "-c", "1",
1192     };
1193     const char * const kvm_stat_record_usage[] = {
1194         "perf kvm stat record [<options>]",
1195         NULL
1196     };
1197     const char * const *events_tp;
1198     int ret;
1199 
1200     events_tp_size = 0;
1201     ret = setup_kvm_events_tp(kvm);
1202     if (ret < 0) {
1203         pr_err("Unable to setup the kvm tracepoints\n");
1204         return ret;
1205     }
1206 
1207     for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1208         events_tp_size++;
1209 
1210     rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1211            2 * events_tp_size;
1212     rec_argv = calloc(rec_argc + 1, sizeof(char *));
1213 
1214     if (rec_argv == NULL)
1215         return -ENOMEM;
1216 
1217     for (i = 0; i < ARRAY_SIZE(record_args); i++)
1218         rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1219 
1220     for (j = 0; j < events_tp_size; j++) {
1221         rec_argv[i++] = "-e";
1222         rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1223     }
1224 
1225     rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1226     rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1227 
1228     for (j = 1; j < (unsigned int)argc; j++, i++)
1229         rec_argv[i] = argv[j];
1230 
1231     set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1232     set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1233     set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1234 
1235     set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1236     set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1237     set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1238     set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1239     set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1240     set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1241     set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1242     set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1243     set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1244     set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1245     set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1246     set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1247     set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1248     set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1249     set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1250 
1251     record_usage = kvm_stat_record_usage;
1252     return cmd_record(i, rec_argv);
1253 }
1254 
1255 static int
1256 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1257 {
1258     const struct option kvm_events_report_options[] = {
1259         OPT_STRING(0, "event", &kvm->report_event, "report event",
1260                "event for reporting: vmexit, "
1261                "mmio (x86 only), ioport (x86 only)"),
1262         OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1263                 "vcpu id to report"),
1264         OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1265                 "key for sorting: sample(sort by samples number)"
1266                 " time (sort by avg time)"),
1267         OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1268                "analyze events only for given process id(s)"),
1269         OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1270         OPT_END()
1271     };
1272 
1273     const char * const kvm_events_report_usage[] = {
1274         "perf kvm stat report [<options>]",
1275         NULL
1276     };
1277 
1278     if (argc) {
1279         argc = parse_options(argc, argv,
1280                      kvm_events_report_options,
1281                      kvm_events_report_usage, 0);
1282         if (argc)
1283             usage_with_options(kvm_events_report_usage,
1284                        kvm_events_report_options);
1285     }
1286 
1287     if (!kvm->opts.target.pid)
1288         kvm->opts.target.system_wide = true;
1289 
1290     return kvm_events_report_vcpu(kvm);
1291 }
1292 
1293 #ifdef HAVE_TIMERFD_SUPPORT
1294 static struct evlist *kvm_live_event_list(void)
1295 {
1296     struct evlist *evlist;
1297     char *tp, *name, *sys;
1298     int err = -1;
1299     const char * const *events_tp;
1300 
1301     evlist = evlist__new();
1302     if (evlist == NULL)
1303         return NULL;
1304 
1305     for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1306 
1307         tp = strdup(*events_tp);
1308         if (tp == NULL)
1309             goto out;
1310 
1311         /* split tracepoint into subsystem and name */
1312         sys = tp;
1313         name = strchr(tp, ':');
1314         if (name == NULL) {
1315             pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1316                    *events_tp);
1317             free(tp);
1318             goto out;
1319         }
1320         *name = '\0';
1321         name++;
1322 
1323         if (evlist__add_newtp(evlist, sys, name, NULL)) {
1324             pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1325             free(tp);
1326             goto out;
1327         }
1328 
1329         free(tp);
1330     }
1331 
1332     err = 0;
1333 
1334 out:
1335     if (err) {
1336         evlist__delete(evlist);
1337         evlist = NULL;
1338     }
1339 
1340     return evlist;
1341 }
1342 
1343 static int kvm_events_live(struct perf_kvm_stat *kvm,
1344                int argc, const char **argv)
1345 {
1346     char errbuf[BUFSIZ];
1347     int err;
1348 
1349     const struct option live_options[] = {
1350         OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1351             "record events on existing process id"),
1352         OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1353             "number of mmap data pages", evlist__parse_mmap_pages),
1354         OPT_INCR('v', "verbose", &verbose,
1355             "be more verbose (show counter open errors, etc)"),
1356         OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1357             "system-wide collection from all CPUs"),
1358         OPT_UINTEGER('d', "display", &kvm->display_time,
1359             "time in seconds between display updates"),
1360         OPT_STRING(0, "event", &kvm->report_event, "report event",
1361             "event for reporting: "
1362             "vmexit, mmio (x86 only), ioport (x86 only)"),
1363         OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1364             "vcpu id to report"),
1365         OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1366             "key for sorting: sample(sort by samples number)"
1367             " time (sort by avg time)"),
1368         OPT_U64(0, "duration", &kvm->duration,
1369             "show events other than"
1370             " HLT (x86 only) or Wait state (s390 only)"
1371             " that take longer than duration usecs"),
1372         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1373                 "per thread proc mmap processing timeout in ms"),
1374         OPT_END()
1375     };
1376     const char * const live_usage[] = {
1377         "perf kvm stat live [<options>]",
1378         NULL
1379     };
1380     struct perf_data data = {
1381         .mode = PERF_DATA_MODE_WRITE,
1382     };
1383 
1384 
1385     /* event handling */
1386     kvm->tool.sample = process_sample_event;
1387     kvm->tool.comm   = perf_event__process_comm;
1388     kvm->tool.exit   = perf_event__process_exit;
1389     kvm->tool.fork   = perf_event__process_fork;
1390     kvm->tool.lost   = process_lost_event;
1391     kvm->tool.namespaces  = perf_event__process_namespaces;
1392     kvm->tool.ordered_events = true;
1393     perf_tool__fill_defaults(&kvm->tool);
1394 
1395     /* set defaults */
1396     kvm->display_time = 1;
1397     kvm->opts.user_interval = 1;
1398     kvm->opts.mmap_pages = 512;
1399     kvm->opts.target.uses_mmap = false;
1400     kvm->opts.target.uid_str = NULL;
1401     kvm->opts.target.uid = UINT_MAX;
1402 
1403     symbol__init(NULL);
1404     disable_buildid_cache();
1405 
1406     use_browser = 0;
1407 
1408     if (argc) {
1409         argc = parse_options(argc, argv, live_options,
1410                      live_usage, 0);
1411         if (argc)
1412             usage_with_options(live_usage, live_options);
1413     }
1414 
1415     kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1416 
1417     /*
1418      * target related setups
1419      */
1420     err = target__validate(&kvm->opts.target);
1421     if (err) {
1422         target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1423         ui__warning("%s", errbuf);
1424     }
1425 
1426     if (target__none(&kvm->opts.target))
1427         kvm->opts.target.system_wide = true;
1428 
1429 
1430     /*
1431      * generate the event list
1432      */
1433     err = setup_kvm_events_tp(kvm);
1434     if (err < 0) {
1435         pr_err("Unable to setup the kvm tracepoints\n");
1436         return err;
1437     }
1438 
1439     kvm->evlist = kvm_live_event_list();
1440     if (kvm->evlist == NULL) {
1441         err = -1;
1442         goto out;
1443     }
1444 
1445     if (evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1446         usage_with_options(live_usage, live_options);
1447 
1448     /*
1449      * perf session
1450      */
1451     kvm->session = perf_session__new(&data, &kvm->tool);
1452     if (IS_ERR(kvm->session)) {
1453         err = PTR_ERR(kvm->session);
1454         goto out;
1455     }
1456     kvm->session->evlist = kvm->evlist;
1457     perf_session__set_id_hdr_size(kvm->session);
1458     ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1459     machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1460                     kvm->evlist->core.threads, true, false, 1);
1461     err = kvm_live_open_events(kvm);
1462     if (err)
1463         goto out;
1464 
1465     err = kvm_events_live_report(kvm);
1466 
1467 out:
1468     perf_session__delete(kvm->session);
1469     kvm->session = NULL;
1470     evlist__delete(kvm->evlist);
1471 
1472     return err;
1473 }
1474 #endif
1475 
1476 static void print_kvm_stat_usage(void)
1477 {
1478     printf("Usage: perf kvm stat <command>\n\n");
1479 
1480     printf("# Available commands:\n");
1481     printf("\trecord: record kvm events\n");
1482     printf("\treport: report statistical data of kvm events\n");
1483     printf("\tlive:   live reporting of statistical data of kvm events\n");
1484 
1485     printf("\nOtherwise, it is the alias of 'perf stat':\n");
1486 }
1487 
1488 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1489 {
1490     struct perf_kvm_stat kvm = {
1491         .file_name = file_name,
1492 
1493         .trace_vcpu = -1,
1494         .report_event   = "vmexit",
1495         .sort_key   = "sample",
1496 
1497     };
1498 
1499     if (argc == 1) {
1500         print_kvm_stat_usage();
1501         goto perf_stat;
1502     }
1503 
1504     if (strlen(argv[1]) > 2 && strstarts("record", argv[1]))
1505         return kvm_events_record(&kvm, argc - 1, argv + 1);
1506 
1507     if (strlen(argv[1]) > 2 && strstarts("report", argv[1]))
1508         return kvm_events_report(&kvm, argc - 1 , argv + 1);
1509 
1510 #ifdef HAVE_TIMERFD_SUPPORT
1511     if (!strncmp(argv[1], "live", 4))
1512         return kvm_events_live(&kvm, argc - 1 , argv + 1);
1513 #endif
1514 
1515 perf_stat:
1516     return cmd_stat(argc, argv);
1517 }
1518 #endif /* HAVE_KVM_STAT_SUPPORT */
1519 
1520 int __weak kvm_add_default_arch_event(int *argc __maybe_unused,
1521                     const char **argv __maybe_unused)
1522 {
1523     return 0;
1524 }
1525 
1526 static int __cmd_record(const char *file_name, int argc, const char **argv)
1527 {
1528     int rec_argc, i = 0, j, ret;
1529     const char **rec_argv;
1530 
1531     ret = kvm_add_default_arch_event(&argc, argv);
1532     if (ret)
1533         return -EINVAL;
1534 
1535     rec_argc = argc + 2;
1536     rec_argv = calloc(rec_argc + 1, sizeof(char *));
1537     rec_argv[i++] = strdup("record");
1538     rec_argv[i++] = strdup("-o");
1539     rec_argv[i++] = strdup(file_name);
1540     for (j = 1; j < argc; j++, i++)
1541         rec_argv[i] = argv[j];
1542 
1543     BUG_ON(i != rec_argc);
1544 
1545     return cmd_record(i, rec_argv);
1546 }
1547 
1548 static int __cmd_report(const char *file_name, int argc, const char **argv)
1549 {
1550     int rec_argc, i = 0, j;
1551     const char **rec_argv;
1552 
1553     rec_argc = argc + 2;
1554     rec_argv = calloc(rec_argc + 1, sizeof(char *));
1555     rec_argv[i++] = strdup("report");
1556     rec_argv[i++] = strdup("-i");
1557     rec_argv[i++] = strdup(file_name);
1558     for (j = 1; j < argc; j++, i++)
1559         rec_argv[i] = argv[j];
1560 
1561     BUG_ON(i != rec_argc);
1562 
1563     return cmd_report(i, rec_argv);
1564 }
1565 
1566 static int
1567 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1568 {
1569     int rec_argc, i = 0, j;
1570     const char **rec_argv;
1571 
1572     rec_argc = argc + 2;
1573     rec_argv = calloc(rec_argc + 1, sizeof(char *));
1574     rec_argv[i++] = strdup("buildid-list");
1575     rec_argv[i++] = strdup("-i");
1576     rec_argv[i++] = strdup(file_name);
1577     for (j = 1; j < argc; j++, i++)
1578         rec_argv[i] = argv[j];
1579 
1580     BUG_ON(i != rec_argc);
1581 
1582     return cmd_buildid_list(i, rec_argv);
1583 }
1584 
1585 int cmd_kvm(int argc, const char **argv)
1586 {
1587     const char *file_name = NULL;
1588     const struct option kvm_options[] = {
1589         OPT_STRING('i', "input", &file_name, "file",
1590                "Input file name"),
1591         OPT_STRING('o', "output", &file_name, "file",
1592                "Output file name"),
1593         OPT_BOOLEAN(0, "guest", &perf_guest,
1594                 "Collect guest os data"),
1595         OPT_BOOLEAN(0, "host", &perf_host,
1596                 "Collect host os data"),
1597         OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1598                "guest mount directory under which every guest os"
1599                " instance has a subdir"),
1600         OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1601                "file", "file saving guest os vmlinux"),
1602         OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1603                "file", "file saving guest os /proc/kallsyms"),
1604         OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1605                "file", "file saving guest os /proc/modules"),
1606         OPT_BOOLEAN(0, "guest-code", &symbol_conf.guest_code,
1607                 "Guest code can be found in hypervisor process"),
1608         OPT_INCR('v', "verbose", &verbose,
1609                 "be more verbose (show counter open errors, etc)"),
1610         OPT_END()
1611     };
1612 
1613     const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1614                         "buildid-list", "stat", NULL };
1615     const char *kvm_usage[] = { NULL, NULL };
1616 
1617     perf_host  = 0;
1618     perf_guest = 1;
1619 
1620     argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1621                     PARSE_OPT_STOP_AT_NON_OPTION);
1622     if (!argc)
1623         usage_with_options(kvm_usage, kvm_options);
1624 
1625     if (!perf_host)
1626         perf_guest = 1;
1627 
1628     if (!file_name) {
1629         file_name = get_filename_for_perf_kvm();
1630 
1631         if (!file_name) {
1632             pr_err("Failed to allocate memory for filename\n");
1633             return -ENOMEM;
1634         }
1635     }
1636 
1637     if (strlen(argv[0]) > 2 && strstarts("record", argv[0]))
1638         return __cmd_record(file_name, argc, argv);
1639     else if (strlen(argv[0]) > 2 && strstarts("report", argv[0]))
1640         return __cmd_report(file_name, argc, argv);
1641     else if (strlen(argv[0]) > 2 && strstarts("diff", argv[0]))
1642         return cmd_diff(argc, argv);
1643     else if (!strcmp(argv[0], "top"))
1644         return cmd_top(argc, argv);
1645     else if (strlen(argv[0]) > 2 && strstarts("buildid-list", argv[0]))
1646         return __cmd_buildid_list(file_name, argc, argv);
1647 #ifdef HAVE_KVM_STAT_SUPPORT
1648     else if (strlen(argv[0]) > 2 && strstarts("stat", argv[0]))
1649         return kvm_cmd_stat(file_name, argc, argv);
1650 #endif
1651     else
1652         usage_with_options(kvm_usage, kvm_options);
1653 
1654     return 0;
1655 }