Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / tools / perf / builtin-kvm.c
diff --git a/kernel/tools/perf/builtin-kvm.c b/kernel/tools/perf/builtin-kvm.c
new file mode 100644 (file)
index 0000000..1f9338f
--- /dev/null
@@ -0,0 +1,1581 @@
+#include "builtin.h"
+#include "perf.h"
+
+#include "util/evsel.h"
+#include "util/evlist.h"
+#include "util/util.h"
+#include "util/cache.h"
+#include "util/symbol.h"
+#include "util/thread.h"
+#include "util/header.h"
+#include "util/session.h"
+#include "util/intlist.h"
+#include "util/parse-options.h"
+#include "util/trace-event.h"
+#include "util/debug.h"
+#include <api/fs/debugfs.h>
+#include "util/tool.h"
+#include "util/stat.h"
+#include "util/top.h"
+#include "util/data.h"
+#include "util/ordered-events.h"
+
+#include <sys/prctl.h>
+#ifdef HAVE_TIMERFD_SUPPORT
+#include <sys/timerfd.h>
+#endif
+
+#include <termios.h>
+#include <semaphore.h>
+#include <pthread.h>
+#include <math.h>
+
+#ifdef HAVE_KVM_STAT_SUPPORT
+#include <asm/kvm_perf.h>
+#include "util/kvm-stat.h"
+
+void exit_event_get_key(struct perf_evsel *evsel,
+                       struct perf_sample *sample,
+                       struct event_key *key)
+{
+       key->info = 0;
+       key->key = perf_evsel__intval(evsel, sample, KVM_EXIT_REASON);
+}
+
+bool kvm_exit_event(struct perf_evsel *evsel)
+{
+       return !strcmp(evsel->name, KVM_EXIT_TRACE);
+}
+
+bool exit_event_begin(struct perf_evsel *evsel,
+                     struct perf_sample *sample, struct event_key *key)
+{
+       if (kvm_exit_event(evsel)) {
+               exit_event_get_key(evsel, sample, key);
+               return true;
+       }
+
+       return false;
+}
+
+bool kvm_entry_event(struct perf_evsel *evsel)
+{
+       return !strcmp(evsel->name, KVM_ENTRY_TRACE);
+}
+
+bool exit_event_end(struct perf_evsel *evsel,
+                   struct perf_sample *sample __maybe_unused,
+                   struct event_key *key __maybe_unused)
+{
+       return kvm_entry_event(evsel);
+}
+
+static const char *get_exit_reason(struct perf_kvm_stat *kvm,
+                                  struct exit_reasons_table *tbl,
+                                  u64 exit_code)
+{
+       while (tbl->reason != NULL) {
+               if (tbl->exit_code == exit_code)
+                       return tbl->reason;
+               tbl++;
+       }
+
+       pr_err("unknown kvm exit code:%lld on %s\n",
+               (unsigned long long)exit_code, kvm->exit_reasons_isa);
+       return "UNKNOWN";
+}
+
+void exit_event_decode_key(struct perf_kvm_stat *kvm,
+                          struct event_key *key,
+                          char *decode)
+{
+       const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
+                                                 key->key);
+
+       scnprintf(decode, DECODE_STR_LEN, "%s", exit_reason);
+}
+
+static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
+{
+       struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
+
+       for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
+               if (!strcmp(events_ops->name, kvm->report_event)) {
+                       kvm->events_ops = events_ops->ops;
+                       return true;
+               }
+       }
+
+       return false;
+}
+
+struct vcpu_event_record {
+       int vcpu_id;
+       u64 start_time;
+       struct kvm_event *last_event;
+};
+
+
+static void init_kvm_event_record(struct perf_kvm_stat *kvm)
+{
+       unsigned int i;
+
+       for (i = 0; i < EVENTS_CACHE_SIZE; i++)
+               INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
+}
+
+#ifdef HAVE_TIMERFD_SUPPORT
+static void clear_events_cache_stats(struct list_head *kvm_events_cache)
+{
+       struct list_head *head;
+       struct kvm_event *event;
+       unsigned int i;
+       int j;
+
+       for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
+               head = &kvm_events_cache[i];
+               list_for_each_entry(event, head, hash_entry) {
+                       /* reset stats for event */
+                       event->total.time = 0;
+                       init_stats(&event->total.stats);
+
+                       for (j = 0; j < event->max_vcpu; ++j) {
+                               event->vcpu[j].time = 0;
+                               init_stats(&event->vcpu[j].stats);
+                       }
+               }
+       }
+}
+#endif
+
+static int kvm_events_hash_fn(u64 key)
+{
+       return key & (EVENTS_CACHE_SIZE - 1);
+}
+
+static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
+{
+       int old_max_vcpu = event->max_vcpu;
+       void *prev;
+
+       if (vcpu_id < event->max_vcpu)
+               return true;
+
+       while (event->max_vcpu <= vcpu_id)
+               event->max_vcpu += DEFAULT_VCPU_NUM;
+
+       prev = event->vcpu;
+       event->vcpu = realloc(event->vcpu,
+                             event->max_vcpu * sizeof(*event->vcpu));
+       if (!event->vcpu) {
+               free(prev);
+               pr_err("Not enough memory\n");
+               return false;
+       }
+
+       memset(event->vcpu + old_max_vcpu, 0,
+              (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
+       return true;
+}
+
+static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
+{
+       struct kvm_event *event;
+
+       event = zalloc(sizeof(*event));
+       if (!event) {
+               pr_err("Not enough memory\n");
+               return NULL;
+       }
+
+       event->key = *key;
+       init_stats(&event->total.stats);
+       return event;
+}
+
+static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
+                                              struct event_key *key)
+{
+       struct kvm_event *event;
+       struct list_head *head;
+
+       BUG_ON(key->key == INVALID_KEY);
+
+       head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
+       list_for_each_entry(event, head, hash_entry) {
+               if (event->key.key == key->key && event->key.info == key->info)
+                       return event;
+       }
+
+       event = kvm_alloc_init_event(key);
+       if (!event)
+               return NULL;
+
+       list_add(&event->hash_entry, head);
+       return event;
+}
+
+static bool handle_begin_event(struct perf_kvm_stat *kvm,
+                              struct vcpu_event_record *vcpu_record,
+                              struct event_key *key, u64 timestamp)
+{
+       struct kvm_event *event = NULL;
+
+       if (key->key != INVALID_KEY)
+               event = find_create_kvm_event(kvm, key);
+
+       vcpu_record->last_event = event;
+       vcpu_record->start_time = timestamp;
+       return true;
+}
+
+static void
+kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
+{
+       kvm_stats->time += time_diff;
+       update_stats(&kvm_stats->stats, time_diff);
+}
+
+static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
+{
+       struct kvm_event_stats *kvm_stats = &event->total;
+
+       if (vcpu_id != -1)
+               kvm_stats = &event->vcpu[vcpu_id];
+
+       return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
+                               avg_stats(&kvm_stats->stats));
+}
+
+static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
+                            u64 time_diff)
+{
+       if (vcpu_id == -1) {
+               kvm_update_event_stats(&event->total, time_diff);
+               return true;
+       }
+
+       if (!kvm_event_expand(event, vcpu_id))
+               return false;
+
+       kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
+       return true;
+}
+
+static bool is_child_event(struct perf_kvm_stat *kvm,
+                          struct perf_evsel *evsel,
+                          struct perf_sample *sample,
+                          struct event_key *key)
+{
+       struct child_event_ops *child_ops;
+
+       child_ops = kvm->events_ops->child_ops;
+
+       if (!child_ops)
+               return false;
+
+       for (; child_ops->name; child_ops++) {
+               if (!strcmp(evsel->name, child_ops->name)) {
+                       child_ops->get_key(evsel, sample, key);
+                       return true;
+               }
+       }
+
+       return false;
+}
+
+static bool handle_child_event(struct perf_kvm_stat *kvm,
+                              struct vcpu_event_record *vcpu_record,
+                              struct event_key *key,
+                              struct perf_sample *sample __maybe_unused)
+{
+       struct kvm_event *event = NULL;
+
+       if (key->key != INVALID_KEY)
+               event = find_create_kvm_event(kvm, key);
+
+       vcpu_record->last_event = event;
+
+       return true;
+}
+
+static bool skip_event(const char *event)
+{
+       const char * const *skip_events;
+
+       for (skip_events = kvm_skip_events; *skip_events; skip_events++)
+               if (!strcmp(event, *skip_events))
+                       return true;
+
+       return false;
+}
+
+static bool handle_end_event(struct perf_kvm_stat *kvm,
+                            struct vcpu_event_record *vcpu_record,
+                            struct event_key *key,
+                            struct perf_sample *sample)
+{
+       struct kvm_event *event;
+       u64 time_begin, time_diff;
+       int vcpu;
+
+       if (kvm->trace_vcpu == -1)
+               vcpu = -1;
+       else
+               vcpu = vcpu_record->vcpu_id;
+
+       event = vcpu_record->last_event;
+       time_begin = vcpu_record->start_time;
+
+       /* The begin event is not caught. */
+       if (!time_begin)
+               return true;
+
+       /*
+        * In some case, the 'begin event' only records the start timestamp,
+        * the actual event is recognized in the 'end event' (e.g. mmio-event).
+        */
+
+       /* Both begin and end events did not get the key. */
+       if (!event && key->key == INVALID_KEY)
+               return true;
+
+       if (!event)
+               event = find_create_kvm_event(kvm, key);
+
+       if (!event)
+               return false;
+
+       vcpu_record->last_event = NULL;
+       vcpu_record->start_time = 0;
+
+       /* seems to happen once in a while during live mode */
+       if (sample->time < time_begin) {
+               pr_debug("End time before begin time; skipping event.\n");
+               return true;
+       }
+
+       time_diff = sample->time - time_begin;
+
+       if (kvm->duration && time_diff > kvm->duration) {
+               char decode[DECODE_STR_LEN];
+
+               kvm->events_ops->decode_key(kvm, &event->key, decode);
+               if (!skip_event(decode)) {
+                       pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
+                                sample->time, sample->pid, vcpu_record->vcpu_id,
+                                decode, time_diff/1000);
+               }
+       }
+
+       return update_kvm_event(event, vcpu, time_diff);
+}
+
+static
+struct vcpu_event_record *per_vcpu_record(struct thread *thread,
+                                         struct perf_evsel *evsel,
+                                         struct perf_sample *sample)
+{
+       /* Only kvm_entry records vcpu id. */
+       if (!thread__priv(thread) && kvm_entry_event(evsel)) {
+               struct vcpu_event_record *vcpu_record;
+
+               vcpu_record = zalloc(sizeof(*vcpu_record));
+               if (!vcpu_record) {
+                       pr_err("%s: Not enough memory\n", __func__);
+                       return NULL;
+               }
+
+               vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, VCPU_ID);
+               thread__set_priv(thread, vcpu_record);
+       }
+
+       return thread__priv(thread);
+}
+
+static bool handle_kvm_event(struct perf_kvm_stat *kvm,
+                            struct thread *thread,
+                            struct perf_evsel *evsel,
+                            struct perf_sample *sample)
+{
+       struct vcpu_event_record *vcpu_record;
+       struct event_key key = { .key = INVALID_KEY,
+                                .exit_reasons = kvm->exit_reasons };
+
+       vcpu_record = per_vcpu_record(thread, evsel, sample);
+       if (!vcpu_record)
+               return true;
+
+       /* only process events for vcpus user cares about */
+       if ((kvm->trace_vcpu != -1) &&
+           (kvm->trace_vcpu != vcpu_record->vcpu_id))
+               return true;
+
+       if (kvm->events_ops->is_begin_event(evsel, sample, &key))
+               return handle_begin_event(kvm, vcpu_record, &key, sample->time);
+
+       if (is_child_event(kvm, evsel, sample, &key))
+               return handle_child_event(kvm, vcpu_record, &key, sample);
+
+       if (kvm->events_ops->is_end_event(evsel, sample, &key))
+               return handle_end_event(kvm, vcpu_record, &key, sample);
+
+       return true;
+}
+
+#define GET_EVENT_KEY(func, field)                                     \
+static u64 get_event_ ##func(struct kvm_event *event, int vcpu)                \
+{                                                                      \
+       if (vcpu == -1)                                                 \
+               return event->total.field;                              \
+                                                                       \
+       if (vcpu >= event->max_vcpu)                                    \
+               return 0;                                               \
+                                                                       \
+       return event->vcpu[vcpu].field;                                 \
+}
+
+#define COMPARE_EVENT_KEY(func, field)                                 \
+GET_EVENT_KEY(func, field)                                             \
+static int compare_kvm_event_ ## func(struct kvm_event *one,           \
+                                       struct kvm_event *two, int vcpu)\
+{                                                                      \
+       return get_event_ ##func(one, vcpu) >                           \
+                               get_event_ ##func(two, vcpu);           \
+}
+
+GET_EVENT_KEY(time, time);
+COMPARE_EVENT_KEY(count, stats.n);
+COMPARE_EVENT_KEY(mean, stats.mean);
+GET_EVENT_KEY(max, stats.max);
+GET_EVENT_KEY(min, stats.min);
+
+#define DEF_SORT_NAME_KEY(name, compare_key)                           \
+       { #name, compare_kvm_event_ ## compare_key }
+
+static struct kvm_event_key keys[] = {
+       DEF_SORT_NAME_KEY(sample, count),
+       DEF_SORT_NAME_KEY(time, mean),
+       { NULL, NULL }
+};
+
+static bool select_key(struct perf_kvm_stat *kvm)
+{
+       int i;
+
+       for (i = 0; keys[i].name; i++) {
+               if (!strcmp(keys[i].name, kvm->sort_key)) {
+                       kvm->compare = keys[i].key;
+                       return true;
+               }
+       }
+
+       pr_err("Unknown compare key:%s\n", kvm->sort_key);
+       return false;
+}
+
+static void insert_to_result(struct rb_root *result, struct kvm_event *event,
+                            key_cmp_fun bigger, int vcpu)
+{
+       struct rb_node **rb = &result->rb_node;
+       struct rb_node *parent = NULL;
+       struct kvm_event *p;
+
+       while (*rb) {
+               p = container_of(*rb, struct kvm_event, rb);
+               parent = *rb;
+
+               if (bigger(event, p, vcpu))
+                       rb = &(*rb)->rb_left;
+               else
+                       rb = &(*rb)->rb_right;
+       }
+
+       rb_link_node(&event->rb, parent, rb);
+       rb_insert_color(&event->rb, result);
+}
+
+static void
+update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
+{
+       int vcpu = kvm->trace_vcpu;
+
+       kvm->total_count += get_event_count(event, vcpu);
+       kvm->total_time += get_event_time(event, vcpu);
+}
+
+static bool event_is_valid(struct kvm_event *event, int vcpu)
+{
+       return !!get_event_count(event, vcpu);
+}
+
+static void sort_result(struct perf_kvm_stat *kvm)
+{
+       unsigned int i;
+       int vcpu = kvm->trace_vcpu;
+       struct kvm_event *event;
+
+       for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
+               list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
+                       if (event_is_valid(event, vcpu)) {
+                               update_total_count(kvm, event);
+                               insert_to_result(&kvm->result, event,
+                                                kvm->compare, vcpu);
+                       }
+               }
+       }
+}
+
+/* returns left most element of result, and erase it */
+static struct kvm_event *pop_from_result(struct rb_root *result)
+{
+       struct rb_node *node = rb_first(result);
+
+       if (!node)
+               return NULL;
+
+       rb_erase(node, result);
+       return container_of(node, struct kvm_event, rb);
+}
+
+static void print_vcpu_info(struct perf_kvm_stat *kvm)
+{
+       int vcpu = kvm->trace_vcpu;
+
+       pr_info("Analyze events for ");
+
+       if (kvm->opts.target.system_wide)
+               pr_info("all VMs, ");
+       else if (kvm->opts.target.pid)
+               pr_info("pid(s) %s, ", kvm->opts.target.pid);
+       else
+               pr_info("dazed and confused on what is monitored, ");
+
+       if (vcpu == -1)
+               pr_info("all VCPUs:\n\n");
+       else
+               pr_info("VCPU %d:\n\n", vcpu);
+}
+
+static void show_timeofday(void)
+{
+       char date[64];
+       struct timeval tv;
+       struct tm ltime;
+
+       gettimeofday(&tv, NULL);
+       if (localtime_r(&tv.tv_sec, &ltime)) {
+               strftime(date, sizeof(date), "%H:%M:%S", &ltime);
+               pr_info("%s.%06ld", date, tv.tv_usec);
+       } else
+               pr_info("00:00:00.000000");
+
+       return;
+}
+
+static void print_result(struct perf_kvm_stat *kvm)
+{
+       char decode[DECODE_STR_LEN];
+       struct kvm_event *event;
+       int vcpu = kvm->trace_vcpu;
+
+       if (kvm->live) {
+               puts(CONSOLE_CLEAR);
+               show_timeofday();
+       }
+
+       pr_info("\n\n");
+       print_vcpu_info(kvm);
+       pr_info("%*s ", DECODE_STR_LEN, kvm->events_ops->name);
+       pr_info("%10s ", "Samples");
+       pr_info("%9s ", "Samples%");
+
+       pr_info("%9s ", "Time%");
+       pr_info("%11s ", "Min Time");
+       pr_info("%11s ", "Max Time");
+       pr_info("%16s ", "Avg time");
+       pr_info("\n\n");
+
+       while ((event = pop_from_result(&kvm->result))) {
+               u64 ecount, etime, max, min;
+
+               ecount = get_event_count(event, vcpu);
+               etime = get_event_time(event, vcpu);
+               max = get_event_max(event, vcpu);
+               min = get_event_min(event, vcpu);
+
+               kvm->events_ops->decode_key(kvm, &event->key, decode);
+               pr_info("%*s ", DECODE_STR_LEN, decode);
+               pr_info("%10llu ", (unsigned long long)ecount);
+               pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
+               pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
+               pr_info("%9.2fus ", (double)min / 1e3);
+               pr_info("%9.2fus ", (double)max / 1e3);
+               pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
+                       kvm_event_rel_stddev(vcpu, event));
+               pr_info("\n");
+       }
+
+       pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
+               kvm->total_count, kvm->total_time / 1e3);
+
+       if (kvm->lost_events)
+               pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
+}
+
+#ifdef HAVE_TIMERFD_SUPPORT
+static int process_lost_event(struct perf_tool *tool,
+                             union perf_event *event __maybe_unused,
+                             struct perf_sample *sample __maybe_unused,
+                             struct machine *machine __maybe_unused)
+{
+       struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
+
+       kvm->lost_events++;
+       return 0;
+}
+#endif
+
+static bool skip_sample(struct perf_kvm_stat *kvm,
+                       struct perf_sample *sample)
+{
+       if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
+               return true;
+
+       return false;
+}
+
+static int process_sample_event(struct perf_tool *tool,
+                               union perf_event *event,
+                               struct perf_sample *sample,
+                               struct perf_evsel *evsel,
+                               struct machine *machine)
+{
+       struct thread *thread;
+       struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
+                                                tool);
+
+       if (skip_sample(kvm, sample))
+               return 0;
+
+       thread = machine__findnew_thread(machine, sample->pid, sample->tid);
+       if (thread == NULL) {
+               pr_debug("problem processing %d event, skipping it.\n",
+                       event->header.type);
+               return -1;
+       }
+
+       if (!handle_kvm_event(kvm, thread, evsel, sample))
+               return -1;
+
+       return 0;
+}
+
+static int cpu_isa_config(struct perf_kvm_stat *kvm)
+{
+       char buf[64], *cpuid;
+       int err;
+
+       if (kvm->live) {
+               err = get_cpuid(buf, sizeof(buf));
+               if (err != 0) {
+                       pr_err("Failed to look up CPU type\n");
+                       return err;
+               }
+               cpuid = buf;
+       } else
+               cpuid = kvm->session->header.env.cpuid;
+
+       if (!cpuid) {
+               pr_err("Failed to look up CPU type\n");
+               return -EINVAL;
+       }
+
+       err = cpu_isa_init(kvm, cpuid);
+       if (err == -ENOTSUP)
+               pr_err("CPU %s is not supported.\n", cpuid);
+
+       return err;
+}
+
+static bool verify_vcpu(int vcpu)
+{
+       if (vcpu != -1 && vcpu < 0) {
+               pr_err("Invalid vcpu:%d.\n", vcpu);
+               return false;
+       }
+
+       return true;
+}
+
+#ifdef HAVE_TIMERFD_SUPPORT
+/* keeping the max events to a modest level to keep
+ * the processing of samples per mmap smooth.
+ */
+#define PERF_KVM__MAX_EVENTS_PER_MMAP  25
+
+static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
+                                  u64 *mmap_time)
+{
+       union perf_event *event;
+       struct perf_sample sample;
+       s64 n = 0;
+       int err;
+
+       *mmap_time = ULLONG_MAX;
+       while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
+               err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
+               if (err) {
+                       perf_evlist__mmap_consume(kvm->evlist, idx);
+                       pr_err("Failed to parse sample\n");
+                       return -1;
+               }
+
+               err = perf_session__queue_event(kvm->session, event, &sample, 0);
+               /*
+                * FIXME: Here we can't consume the event, as perf_session__queue_event will
+                *        point to it, and it'll get possibly overwritten by the kernel.
+                */
+               perf_evlist__mmap_consume(kvm->evlist, idx);
+
+               if (err) {
+                       pr_err("Failed to enqueue sample: %d\n", err);
+                       return -1;
+               }
+
+               /* save time stamp of our first sample for this mmap */
+               if (n == 0)
+                       *mmap_time = sample.time;
+
+               /* limit events per mmap handled all at once */
+               n++;
+               if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
+                       break;
+       }
+
+       return n;
+}
+
+static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
+{
+       int i, err, throttled = 0;
+       s64 n, ntotal = 0;
+       u64 flush_time = ULLONG_MAX, mmap_time;
+
+       for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
+               n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
+               if (n < 0)
+                       return -1;
+
+               /* flush time is going to be the minimum of all the individual
+                * mmap times. Essentially, we flush all the samples queued up
+                * from the last pass under our minimal start time -- that leaves
+                * a very small race for samples to come in with a lower timestamp.
+                * The ioctl to return the perf_clock timestamp should close the
+                * race entirely.
+                */
+               if (mmap_time < flush_time)
+                       flush_time = mmap_time;
+
+               ntotal += n;
+               if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
+                       throttled = 1;
+       }
+
+       /* flush queue after each round in which we processed events */
+       if (ntotal) {
+               struct ordered_events *oe = &kvm->session->ordered_events;
+
+               oe->next_flush = flush_time;
+               err = ordered_events__flush(oe, OE_FLUSH__ROUND);
+               if (err) {
+                       if (kvm->lost_events)
+                               pr_info("\nLost events: %" PRIu64 "\n\n",
+                                       kvm->lost_events);
+                       return err;
+               }
+       }
+
+       return throttled;
+}
+
+static volatile int done;
+
+static void sig_handler(int sig __maybe_unused)
+{
+       done = 1;
+}
+
+static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
+{
+       struct itimerspec new_value;
+       int rc = -1;
+
+       kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
+       if (kvm->timerfd < 0) {
+               pr_err("timerfd_create failed\n");
+               goto out;
+       }
+
+       new_value.it_value.tv_sec = kvm->display_time;
+       new_value.it_value.tv_nsec = 0;
+       new_value.it_interval.tv_sec = kvm->display_time;
+       new_value.it_interval.tv_nsec = 0;
+
+       if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
+               pr_err("timerfd_settime failed: %d\n", errno);
+               close(kvm->timerfd);
+               goto out;
+       }
+
+       rc = 0;
+out:
+       return rc;
+}
+
+static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
+{
+       uint64_t c;
+       int rc;
+
+       rc = read(kvm->timerfd, &c, sizeof(uint64_t));
+       if (rc < 0) {
+               if (errno == EAGAIN)
+                       return 0;
+
+               pr_err("Failed to read timer fd: %d\n", errno);
+               return -1;
+       }
+
+       if (rc != sizeof(uint64_t)) {
+               pr_err("Error reading timer fd - invalid size returned\n");
+               return -1;
+       }
+
+       if (c != 1)
+               pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
+
+       /* update display */
+       sort_result(kvm);
+       print_result(kvm);
+
+       /* reset counts */
+       clear_events_cache_stats(kvm->kvm_events_cache);
+       kvm->total_count = 0;
+       kvm->total_time = 0;
+       kvm->lost_events = 0;
+
+       return 0;
+}
+
+static int fd_set_nonblock(int fd)
+{
+       long arg = 0;
+
+       arg = fcntl(fd, F_GETFL);
+       if (arg < 0) {
+               pr_err("Failed to get current flags for fd %d\n", fd);
+               return -1;
+       }
+
+       if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
+               pr_err("Failed to set non-block option on fd %d\n", fd);
+               return -1;
+       }
+
+       return 0;
+}
+
+static int perf_kvm__handle_stdin(void)
+{
+       int c;
+
+       c = getc(stdin);
+       if (c == 'q')
+               return 1;
+
+       return 0;
+}
+
+static int kvm_events_live_report(struct perf_kvm_stat *kvm)
+{
+       int nr_stdin, ret, err = -EINVAL;
+       struct termios save;
+
+       /* live flag must be set first */
+       kvm->live = true;
+
+       ret = cpu_isa_config(kvm);
+       if (ret < 0)
+               return ret;
+
+       if (!verify_vcpu(kvm->trace_vcpu) ||
+           !select_key(kvm) ||
+           !register_kvm_events_ops(kvm)) {
+               goto out;
+       }
+
+       set_term_quiet_input(&save);
+       init_kvm_event_record(kvm);
+
+       signal(SIGINT, sig_handler);
+       signal(SIGTERM, sig_handler);
+
+       /* add timer fd */
+       if (perf_kvm__timerfd_create(kvm) < 0) {
+               err = -1;
+               goto out;
+       }
+
+       if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
+               goto out;
+
+       nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
+       if (nr_stdin < 0)
+               goto out;
+
+       if (fd_set_nonblock(fileno(stdin)) != 0)
+               goto out;
+
+       /* everything is good - enable the events and process */
+       perf_evlist__enable(kvm->evlist);
+
+       while (!done) {
+               struct fdarray *fda = &kvm->evlist->pollfd;
+               int rc;
+
+               rc = perf_kvm__mmap_read(kvm);
+               if (rc < 0)
+                       break;
+
+               err = perf_kvm__handle_timerfd(kvm);
+               if (err)
+                       goto out;
+
+               if (fda->entries[nr_stdin].revents & POLLIN)
+                       done = perf_kvm__handle_stdin();
+
+               if (!rc && !done)
+                       err = fdarray__poll(fda, 100);
+       }
+
+       perf_evlist__disable(kvm->evlist);
+
+       if (err == 0) {
+               sort_result(kvm);
+               print_result(kvm);
+       }
+
+out:
+       if (kvm->timerfd >= 0)
+               close(kvm->timerfd);
+
+       tcsetattr(0, TCSAFLUSH, &save);
+       return err;
+}
+
+static int kvm_live_open_events(struct perf_kvm_stat *kvm)
+{
+       int err, rc = -1;
+       struct perf_evsel *pos;
+       struct perf_evlist *evlist = kvm->evlist;
+       char sbuf[STRERR_BUFSIZE];
+
+       perf_evlist__config(evlist, &kvm->opts);
+
+       /*
+        * Note: exclude_{guest,host} do not apply here.
+        *       This command processes KVM tracepoints from host only
+        */
+       evlist__for_each(evlist, pos) {
+               struct perf_event_attr *attr = &pos->attr;
+
+               /* make sure these *are* set */
+               perf_evsel__set_sample_bit(pos, TID);
+               perf_evsel__set_sample_bit(pos, TIME);
+               perf_evsel__set_sample_bit(pos, CPU);
+               perf_evsel__set_sample_bit(pos, RAW);
+               /* make sure these are *not*; want as small a sample as possible */
+               perf_evsel__reset_sample_bit(pos, PERIOD);
+               perf_evsel__reset_sample_bit(pos, IP);
+               perf_evsel__reset_sample_bit(pos, CALLCHAIN);
+               perf_evsel__reset_sample_bit(pos, ADDR);
+               perf_evsel__reset_sample_bit(pos, READ);
+               attr->mmap = 0;
+               attr->comm = 0;
+               attr->task = 0;
+
+               attr->sample_period = 1;
+
+               attr->watermark = 0;
+               attr->wakeup_events = 1000;
+
+               /* will enable all once we are ready */
+               attr->disabled = 1;
+       }
+
+       err = perf_evlist__open(evlist);
+       if (err < 0) {
+               printf("Couldn't create the events: %s\n",
+                      strerror_r(errno, sbuf, sizeof(sbuf)));
+               goto out;
+       }
+
+       if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
+               ui__error("Failed to mmap the events: %s\n",
+                         strerror_r(errno, sbuf, sizeof(sbuf)));
+               perf_evlist__close(evlist);
+               goto out;
+       }
+
+       rc = 0;
+
+out:
+       return rc;
+}
+#endif
+
+static int read_events(struct perf_kvm_stat *kvm)
+{
+       int ret;
+
+       struct perf_tool eops = {
+               .sample                 = process_sample_event,
+               .comm                   = perf_event__process_comm,
+               .ordered_events         = true,
+       };
+       struct perf_data_file file = {
+               .path = kvm->file_name,
+               .mode = PERF_DATA_MODE_READ,
+               .force = kvm->force,
+       };
+
+       kvm->tool = eops;
+       kvm->session = perf_session__new(&file, false, &kvm->tool);
+       if (!kvm->session) {
+               pr_err("Initializing perf session failed\n");
+               return -1;
+       }
+
+       symbol__init(&kvm->session->header.env);
+
+       if (!perf_session__has_traces(kvm->session, "kvm record"))
+               return -EINVAL;
+
+       /*
+        * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
+        * traced in the old kernel.
+        */
+       ret = cpu_isa_config(kvm);
+       if (ret < 0)
+               return ret;
+
+       return perf_session__process_events(kvm->session);
+}
+
+static int parse_target_str(struct perf_kvm_stat *kvm)
+{
+       if (kvm->opts.target.pid) {
+               kvm->pid_list = intlist__new(kvm->opts.target.pid);
+               if (kvm->pid_list == NULL) {
+                       pr_err("Error parsing process id string\n");
+                       return -EINVAL;
+               }
+       }
+
+       return 0;
+}
+
+static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
+{
+       int ret = -EINVAL;
+       int vcpu = kvm->trace_vcpu;
+
+       if (parse_target_str(kvm) != 0)
+               goto exit;
+
+       if (!verify_vcpu(vcpu))
+               goto exit;
+
+       if (!select_key(kvm))
+               goto exit;
+
+       if (!register_kvm_events_ops(kvm))
+               goto exit;
+
+       init_kvm_event_record(kvm);
+       setup_pager();
+
+       ret = read_events(kvm);
+       if (ret)
+               goto exit;
+
+       sort_result(kvm);
+       print_result(kvm);
+
+exit:
+       return ret;
+}
+
+#define STRDUP_FAIL_EXIT(s)            \
+       ({      char *_p;               \
+       _p = strdup(s);         \
+               if (!_p)                \
+                       return -ENOMEM; \
+               _p;                     \
+       })
+
+static int
+kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
+{
+       unsigned int rec_argc, i, j, events_tp_size;
+       const char **rec_argv;
+       const char * const record_args[] = {
+               "record",
+               "-R",
+               "-m", "1024",
+               "-c", "1",
+       };
+       const char * const kvm_stat_record_usage[] = {
+               "perf kvm stat record [<options>]",
+               NULL
+       };
+       const char * const *events_tp;
+       events_tp_size = 0;
+
+       for (events_tp = kvm_events_tp; *events_tp; events_tp++)
+               events_tp_size++;
+
+       rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
+                  2 * events_tp_size;
+       rec_argv = calloc(rec_argc + 1, sizeof(char *));
+
+       if (rec_argv == NULL)
+               return -ENOMEM;
+
+       for (i = 0; i < ARRAY_SIZE(record_args); i++)
+               rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
+
+       for (j = 0; j < events_tp_size; j++) {
+               rec_argv[i++] = "-e";
+               rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
+       }
+
+       rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
+       rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
+
+       for (j = 1; j < (unsigned int)argc; j++, i++)
+               rec_argv[i] = argv[j];
+
+       set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
+       set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
+       set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
+
+       set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
+       set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
+
+       record_usage = kvm_stat_record_usage;
+       return cmd_record(i, rec_argv, NULL);
+}
+
+static int
+kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
+{
+       const struct option kvm_events_report_options[] = {
+               OPT_STRING(0, "event", &kvm->report_event, "report event",
+                          "event for reporting: vmexit, "
+                          "mmio (x86 only), ioport (x86 only)"),
+               OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
+                           "vcpu id to report"),
+               OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
+                           "key for sorting: sample(sort by samples number)"
+                           " time (sort by avg time)"),
+               OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
+                          "analyze events only for given process id(s)"),
+               OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
+               OPT_END()
+       };
+
+       const char * const kvm_events_report_usage[] = {
+               "perf kvm stat report [<options>]",
+               NULL
+       };
+
+       if (argc) {
+               argc = parse_options(argc, argv,
+                                    kvm_events_report_options,
+                                    kvm_events_report_usage, 0);
+               if (argc)
+                       usage_with_options(kvm_events_report_usage,
+                                          kvm_events_report_options);
+       }
+
+       if (!kvm->opts.target.pid)
+               kvm->opts.target.system_wide = true;
+
+       return kvm_events_report_vcpu(kvm);
+}
+
+#ifdef HAVE_TIMERFD_SUPPORT
+static struct perf_evlist *kvm_live_event_list(void)
+{
+       struct perf_evlist *evlist;
+       char *tp, *name, *sys;
+       int err = -1;
+       const char * const *events_tp;
+
+       evlist = perf_evlist__new();
+       if (evlist == NULL)
+               return NULL;
+
+       for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
+
+               tp = strdup(*events_tp);
+               if (tp == NULL)
+                       goto out;
+
+               /* split tracepoint into subsystem and name */
+               sys = tp;
+               name = strchr(tp, ':');
+               if (name == NULL) {
+                       pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
+                              *events_tp);
+                       free(tp);
+                       goto out;
+               }
+               *name = '\0';
+               name++;
+
+               if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
+                       pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
+                       free(tp);
+                       goto out;
+               }
+
+               free(tp);
+       }
+
+       err = 0;
+
+out:
+       if (err) {
+               perf_evlist__delete(evlist);
+               evlist = NULL;
+       }
+
+       return evlist;
+}
+
+static int kvm_events_live(struct perf_kvm_stat *kvm,
+                          int argc, const char **argv)
+{
+       char errbuf[BUFSIZ];
+       int err;
+
+       const struct option live_options[] = {
+               OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
+                       "record events on existing process id"),
+               OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
+                       "number of mmap data pages",
+                       perf_evlist__parse_mmap_pages),
+               OPT_INCR('v', "verbose", &verbose,
+                       "be more verbose (show counter open errors, etc)"),
+               OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
+                       "system-wide collection from all CPUs"),
+               OPT_UINTEGER('d', "display", &kvm->display_time,
+                       "time in seconds between display updates"),
+               OPT_STRING(0, "event", &kvm->report_event, "report event",
+                       "event for reporting: "
+                       "vmexit, mmio (x86 only), ioport (x86 only)"),
+               OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
+                       "vcpu id to report"),
+               OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
+                       "key for sorting: sample(sort by samples number)"
+                       " time (sort by avg time)"),
+               OPT_U64(0, "duration", &kvm->duration,
+                       "show events other than"
+                       " HLT (x86 only) or Wait state (s390 only)"
+                       " that take longer than duration usecs"),
+               OPT_END()
+       };
+       const char * const live_usage[] = {
+               "perf kvm stat live [<options>]",
+               NULL
+       };
+       struct perf_data_file file = {
+               .mode = PERF_DATA_MODE_WRITE,
+       };
+
+
+       /* event handling */
+       kvm->tool.sample = process_sample_event;
+       kvm->tool.comm   = perf_event__process_comm;
+       kvm->tool.exit   = perf_event__process_exit;
+       kvm->tool.fork   = perf_event__process_fork;
+       kvm->tool.lost   = process_lost_event;
+       kvm->tool.ordered_events = true;
+       perf_tool__fill_defaults(&kvm->tool);
+
+       /* set defaults */
+       kvm->display_time = 1;
+       kvm->opts.user_interval = 1;
+       kvm->opts.mmap_pages = 512;
+       kvm->opts.target.uses_mmap = false;
+       kvm->opts.target.uid_str = NULL;
+       kvm->opts.target.uid = UINT_MAX;
+
+       symbol__init(NULL);
+       disable_buildid_cache();
+
+       use_browser = 0;
+       setup_browser(false);
+
+       if (argc) {
+               argc = parse_options(argc, argv, live_options,
+                                    live_usage, 0);
+               if (argc)
+                       usage_with_options(live_usage, live_options);
+       }
+
+       kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
+
+       /*
+        * target related setups
+        */
+       err = target__validate(&kvm->opts.target);
+       if (err) {
+               target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
+               ui__warning("%s", errbuf);
+       }
+
+       if (target__none(&kvm->opts.target))
+               kvm->opts.target.system_wide = true;
+
+
+       /*
+        * generate the event list
+        */
+       kvm->evlist = kvm_live_event_list();
+       if (kvm->evlist == NULL) {
+               err = -1;
+               goto out;
+       }
+
+       symbol_conf.nr_events = kvm->evlist->nr_entries;
+
+       if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
+               usage_with_options(live_usage, live_options);
+
+       /*
+        * perf session
+        */
+       kvm->session = perf_session__new(&file, false, &kvm->tool);
+       if (kvm->session == NULL) {
+               err = -1;
+               goto out;
+       }
+       kvm->session->evlist = kvm->evlist;
+       perf_session__set_id_hdr_size(kvm->session);
+       ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
+       machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
+                                   kvm->evlist->threads, false);
+       err = kvm_live_open_events(kvm);
+       if (err)
+               goto out;
+
+       err = kvm_events_live_report(kvm);
+
+out:
+       exit_browser(0);
+
+       if (kvm->session)
+               perf_session__delete(kvm->session);
+       kvm->session = NULL;
+       if (kvm->evlist)
+               perf_evlist__delete(kvm->evlist);
+
+       return err;
+}
+#endif
+
+static void print_kvm_stat_usage(void)
+{
+       printf("Usage: perf kvm stat <command>\n\n");
+
+       printf("# Available commands:\n");
+       printf("\trecord: record kvm events\n");
+       printf("\treport: report statistical data of kvm events\n");
+       printf("\tlive:   live reporting of statistical data of kvm events\n");
+
+       printf("\nOtherwise, it is the alias of 'perf stat':\n");
+}
+
+static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
+{
+       struct perf_kvm_stat kvm = {
+               .file_name = file_name,
+
+               .trace_vcpu     = -1,
+               .report_event   = "vmexit",
+               .sort_key       = "sample",
+
+       };
+
+       if (argc == 1) {
+               print_kvm_stat_usage();
+               goto perf_stat;
+       }
+
+       if (!strncmp(argv[1], "rec", 3))
+               return kvm_events_record(&kvm, argc - 1, argv + 1);
+
+       if (!strncmp(argv[1], "rep", 3))
+               return kvm_events_report(&kvm, argc - 1 , argv + 1);
+
+#ifdef HAVE_TIMERFD_SUPPORT
+       if (!strncmp(argv[1], "live", 4))
+               return kvm_events_live(&kvm, argc - 1 , argv + 1);
+#endif
+
+perf_stat:
+       return cmd_stat(argc, argv, NULL);
+}
+#endif /* HAVE_KVM_STAT_SUPPORT */
+
+static int __cmd_record(const char *file_name, int argc, const char **argv)
+{
+       int rec_argc, i = 0, j;
+       const char **rec_argv;
+
+       rec_argc = argc + 2;
+       rec_argv = calloc(rec_argc + 1, sizeof(char *));
+       rec_argv[i++] = strdup("record");
+       rec_argv[i++] = strdup("-o");
+       rec_argv[i++] = strdup(file_name);
+       for (j = 1; j < argc; j++, i++)
+               rec_argv[i] = argv[j];
+
+       BUG_ON(i != rec_argc);
+
+       return cmd_record(i, rec_argv, NULL);
+}
+
+static int __cmd_report(const char *file_name, int argc, const char **argv)
+{
+       int rec_argc, i = 0, j;
+       const char **rec_argv;
+
+       rec_argc = argc + 2;
+       rec_argv = calloc(rec_argc + 1, sizeof(char *));
+       rec_argv[i++] = strdup("report");
+       rec_argv[i++] = strdup("-i");
+       rec_argv[i++] = strdup(file_name);
+       for (j = 1; j < argc; j++, i++)
+               rec_argv[i] = argv[j];
+
+       BUG_ON(i != rec_argc);
+
+       return cmd_report(i, rec_argv, NULL);
+}
+
+static int
+__cmd_buildid_list(const char *file_name, int argc, const char **argv)
+{
+       int rec_argc, i = 0, j;
+       const char **rec_argv;
+
+       rec_argc = argc + 2;
+       rec_argv = calloc(rec_argc + 1, sizeof(char *));
+       rec_argv[i++] = strdup("buildid-list");
+       rec_argv[i++] = strdup("-i");
+       rec_argv[i++] = strdup(file_name);
+       for (j = 1; j < argc; j++, i++)
+               rec_argv[i] = argv[j];
+
+       BUG_ON(i != rec_argc);
+
+       return cmd_buildid_list(i, rec_argv, NULL);
+}
+
+int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
+{
+       const char *file_name = NULL;
+       const struct option kvm_options[] = {
+               OPT_STRING('i', "input", &file_name, "file",
+                          "Input file name"),
+               OPT_STRING('o', "output", &file_name, "file",
+                          "Output file name"),
+               OPT_BOOLEAN(0, "guest", &perf_guest,
+                           "Collect guest os data"),
+               OPT_BOOLEAN(0, "host", &perf_host,
+                           "Collect host os data"),
+               OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
+                          "guest mount directory under which every guest os"
+                          " instance has a subdir"),
+               OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
+                          "file", "file saving guest os vmlinux"),
+               OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
+                          "file", "file saving guest os /proc/kallsyms"),
+               OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
+                          "file", "file saving guest os /proc/modules"),
+               OPT_INCR('v', "verbose", &verbose,
+                           "be more verbose (show counter open errors, etc)"),
+               OPT_END()
+       };
+
+       const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
+                                               "buildid-list", "stat", NULL };
+       const char *kvm_usage[] = { NULL, NULL };
+
+       perf_host  = 0;
+       perf_guest = 1;
+
+       argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
+                                       PARSE_OPT_STOP_AT_NON_OPTION);
+       if (!argc)
+               usage_with_options(kvm_usage, kvm_options);
+
+       if (!perf_host)
+               perf_guest = 1;
+
+       if (!file_name) {
+               file_name = get_filename_for_perf_kvm();
+
+               if (!file_name) {
+                       pr_err("Failed to allocate memory for filename\n");
+                       return -ENOMEM;
+               }
+       }
+
+       if (!strncmp(argv[0], "rec", 3))
+               return __cmd_record(file_name, argc, argv);
+       else if (!strncmp(argv[0], "rep", 3))
+               return __cmd_report(file_name, argc, argv);
+       else if (!strncmp(argv[0], "diff", 4))
+               return cmd_diff(argc, argv, NULL);
+       else if (!strncmp(argv[0], "top", 3))
+               return cmd_top(argc, argv, NULL);
+       else if (!strncmp(argv[0], "buildid-list", 12))
+               return __cmd_buildid_list(file_name, argc, argv);
+#ifdef HAVE_KVM_STAT_SUPPORT
+       else if (!strncmp(argv[0], "stat", 4))
+               return kvm_cmd_stat(file_name, argc, argv);
+#endif
+       else
+               usage_with_options(kvm_usage, kvm_options);
+
+       return 0;
+}