perf kvm: Events analysis tool
authorXiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Mon, 17 Sep 2012 08:31:15 +0000 (16:31 +0800)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Fri, 21 Sep 2012 15:51:22 +0000 (12:51 -0300)
Add 'perf kvm stat' support to analyze kvm vmexit/mmio/ioport smartly

Usage:
- kvm stat
  run a command and gather performance counter statistics, it is the alias of
  perf stat

- trace kvm events:
  perf kvm stat record, or, if other tracepoints are interesting as well, we
  can append the events like this:
  perf kvm stat record -e timer:* -a

  If many guests are running, we can track the specified guest by using -p or
  --pid, -a is used to track events generated by all guests.

- show the result:
  perf kvm stat report

The output example is following:
13005
13059

total 2 guests are running on the host

Then, track the guest whose pid is 13059:
^C[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.253 MB perf.data.guest (~11065 samples) ]

See the vmexit events:

Analyze events for all VCPUs:

             VM-EXIT    Samples  Samples%     Time%         Avg time

         APIC_ACCESS        460    70.55%     0.01%     22.44us ( +-   1.75% )
                 HLT         93    14.26%    99.98% 832077.26us ( +-  10.42% )
  EXTERNAL_INTERRUPT         64     9.82%     0.00%     35.35us ( +-  14.21% )
   PENDING_INTERRUPT         24     3.68%     0.00%      9.29us ( +-  31.39% )
           CR_ACCESS          7     1.07%     0.00%      8.12us ( +-   5.76% )
      IO_INSTRUCTION          3     0.46%     0.00%     18.00us ( +-  11.79% )
       EXCEPTION_NMI          1     0.15%     0.00%      5.83us ( +-   -nan% )

Total Samples:652, Total events handled time:77396109.80us.

See the mmio events:

Analyze events for all VCPUs:

         MMIO Access    Samples  Samples%     Time%         Avg time

        0xfee00380:W        387    84.31%    79.28%      8.29us ( +-   3.32% )
        0xfee00300:W         24     5.23%     9.96%     16.79us ( +-   1.97% )
        0xfee00300:R         24     5.23%     7.83%     13.20us ( +-   3.00% )
        0xfee00310:W         24     5.23%     2.93%      4.94us ( +-   3.84% )

Total Samples:459, Total events handled time:4044.59us.

See the ioport event:

Analyze events for all VCPUs:

      IO Port Access    Samples  Samples%     Time%         Avg time

         0xc050:POUT          3   100.00%   100.00%     13.75us ( +-  10.83% )

Total Samples:3, Total events handled time:41.26us.

And, --vcpu is used to track the specified vcpu and --key is used to sort the
result:

Analyze events for VCPU 0:

             VM-EXIT    Samples  Samples%     Time%         Avg time

                 HLT         27    13.85%    99.97% 405790.24us ( +-  12.70% )
  EXTERNAL_INTERRUPT         13     6.67%     0.00%     27.94us ( +-  22.26% )
         APIC_ACCESS        146    74.87%     0.03%     21.69us ( +-   2.91% )
      IO_INSTRUCTION          2     1.03%     0.00%     17.77us ( +-  20.56% )
           CR_ACCESS          2     1.03%     0.00%      8.55us ( +-   6.47% )
   PENDING_INTERRUPT          5     2.56%     0.00%      6.27us ( +-   3.94% )

Total Samples:195, Total events handled time:10959950.90us.

Signed-off-by: Dong Hao <haodong@linux.vnet.ibm.com>
Signed-off-by: Runzhen Wang <runzhen@linux.vnet.ibm.com>
[ Dong Hao <haodong@linux.vnet.ibm.com>
  Runzhen Wang <runzhen@linux.vnet.ibm.com>:
     - rebase it on current acme's tree
     - fix the compiling-error on i386 ]
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Acked-by: David Ahern <dsahern@gmail.com>
Cc: Avi Kivity <avi@redhat.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: kvm@vger.kernel.org
Cc: Runzhen Wang <runzhen@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/1347870675-31495-4-git-send-email-haodong@linux.vnet.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/Documentation/perf-kvm.txt
tools/perf/MANIFEST
tools/perf/builtin-kvm.c
tools/perf/util/header.c
tools/perf/util/header.h
tools/perf/util/thread.h

index dd84cb2..326f2cb 100644 (file)
@@ -12,7 +12,7 @@ SYNOPSIS
        [--guestkallsyms=<path> --guestmodules=<path> | --guestvmlinux=<path>]]
        {top|record|report|diff|buildid-list}
 'perf kvm' [--host] [--guest] [--guestkallsyms=<path> --guestmodules=<path>
-       | --guestvmlinux=<path>] {top|record|report|diff|buildid-list}
+       | --guestvmlinux=<path>] {top|record|report|diff|buildid-list|stat}
 
 DESCRIPTION
 -----------
@@ -38,6 +38,18 @@ There are a couple of variants of perf kvm:
   so that other tools can be used to fetch packages with matching symbol tables
   for use by perf report.
 
+  'perf kvm stat <command>' to run a command and gather performance counter
+  statistics.
+  Especially, perf 'kvm stat record/report' generates a statistical analysis
+  of KVM events. Currently, vmexit, mmio and ioport events are supported.
+  'perf kvm stat record <command>' records kvm events and the events between
+  start and end <command>.
+  And this command produces a file which contains tracing results of kvm
+  events.
+
+  'perf kvm stat report' reports statistical data which includes events
+  handled time, samples, and so on.
+
 OPTIONS
 -------
 -i::
@@ -68,7 +80,21 @@ OPTIONS
 --guestvmlinux=<path>::
        Guest os kernel vmlinux.
 
+STAT REPORT OPTIONS
+-------------------
+--vcpu=<value>::
+       analyze events which occures on this vcpu. (default: all vcpus)
+
+--events=<value>::
+       events to be analyzed. Possible values: vmexit, mmio, ioport.
+       (default: vmexit)
+-k::
+--key=<value>::
+       Sorting key. Possible values: sample (default, sort by samples
+       number), time (sort by average time).
+
 SEE ALSO
 --------
 linkperf:perf-top[1], linkperf:perf-record[1], linkperf:perf-report[1],
-linkperf:perf-diff[1], linkperf:perf-buildid-list[1]
+linkperf:perf-diff[1], linkperf:perf-buildid-list[1],
+linkperf:perf-stat[1]
index 0518079..80db3f4 100644 (file)
@@ -16,3 +16,6 @@ arch/*/lib/memset*.S
 include/linux/poison.h
 include/linux/magic.h
 include/linux/hw_breakpoint.h
+arch/x86/include/asm/svm.h
+arch/x86/include/asm/vmx.h
+arch/x86/include/asm/kvm_host.h
index 4d2aa2c..1878947 100644 (file)
@@ -1,6 +1,7 @@
 #include "builtin.h"
 #include "perf.h"
 
+#include "util/evsel.h"
 #include "util/util.h"
 #include "util/cache.h"
 #include "util/symbol.h"
 
 #include "util/parse-options.h"
 #include "util/trace-event.h"
-
 #include "util/debug.h"
+#include "util/debugfs.h"
+#include "util/tool.h"
+#include "util/stat.h"
 
 #include <sys/prctl.h>
 
 #include <pthread.h>
 #include <math.h>
 
-static const char              *file_name;
+#include "../../arch/x86/include/asm/svm.h"
+#include "../../arch/x86/include/asm/vmx.h"
+#include "../../arch/x86/include/asm/kvm.h"
+
+struct event_key {
+       #define INVALID_KEY     (~0ULL)
+       u64 key;
+       int info;
+};
+
+struct kvm_events_ops {
+       bool (*is_begin_event)(struct event_format *event, void *data,
+                               struct event_key *key);
+       bool (*is_end_event)(struct event_format *event, void *data,
+                               struct event_key *key);
+       void (*decode_key)(struct event_key *key, char decode[20]);
+       const char *name;
+};
+
+static void exit_event_get_key(struct event_format *event, void *data,
+                               struct event_key *key)
+{
+       key->info = 0;
+       key->key = raw_field_value(event, "exit_reason", data);
+}
+
+static bool kvm_exit_event(struct event_format *event)
+{
+       return !strcmp(event->name, "kvm_exit");
+}
+
+static bool exit_event_begin(struct event_format *event, void *data,
+                               struct event_key *key)
+{
+       if (kvm_exit_event(event)) {
+               exit_event_get_key(event, data, key);
+               return true;
+       }
+
+       return false;
+}
+
+static bool kvm_entry_event(struct event_format *event)
+{
+       return !strcmp(event->name, "kvm_entry");
+}
+
+static bool exit_event_end(struct event_format *event, void *data __maybe_unused,
+                               struct event_key *key __maybe_unused)
+{
+       return kvm_entry_event(event);
+}
+
+struct exit_reasons_table {
+       unsigned long exit_code;
+       const char *reason;
+};
+
+struct exit_reasons_table vmx_exit_reasons[] = {
+       VMX_EXIT_REASONS
+};
+
+struct exit_reasons_table svm_exit_reasons[] = {
+       SVM_EXIT_REASONS
+};
+
+static int cpu_isa;
+
+static const char *get_exit_reason(u64 exit_code)
+{
+       int table_size = ARRAY_SIZE(svm_exit_reasons);
+       struct exit_reasons_table *table = svm_exit_reasons;
+
+       if (cpu_isa == 1) {
+               table = vmx_exit_reasons;
+               table_size = ARRAY_SIZE(vmx_exit_reasons);
+       }
+
+       while (table_size--) {
+               if (table->exit_code == exit_code)
+                       return table->reason;
+               table++;
+       }
+
+       pr_err("unknown kvm exit code:%lld on %s\n",
+               (unsigned long long)exit_code, cpu_isa ? "VMX" : "SVM");
+       return "UNKNOWN";
+}
+
+static void exit_event_decode_key(struct event_key *key, char decode[20])
+{
+       const char *exit_reason = get_exit_reason(key->key);
+
+       scnprintf(decode, 20, "%s", exit_reason);
+}
+
+static struct kvm_events_ops exit_events = {
+       .is_begin_event = exit_event_begin,
+       .is_end_event = exit_event_end,
+       .decode_key = exit_event_decode_key,
+       .name = "VM-EXIT"
+};
+
+    /*
+     * For the mmio events, we treat:
+     * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
+     * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
+     */
+static void mmio_event_get_key(struct event_format *event, void *data,
+                               struct event_key *key)
+{
+       key->key = raw_field_value(event, "gpa", data);
+       key->info = raw_field_value(event, "type", data);
+}
+
+#define KVM_TRACE_MMIO_READ_UNSATISFIED 0
+#define KVM_TRACE_MMIO_READ 1
+#define KVM_TRACE_MMIO_WRITE 2
+
+static bool mmio_event_begin(struct event_format *event, void *data,
+                               struct event_key *key)
+{
+       /* MMIO read begin event in kernel. */
+       if (kvm_exit_event(event))
+               return true;
+
+       /* MMIO write begin event in kernel. */
+       if (!strcmp(event->name, "kvm_mmio") &&
+               raw_field_value(event, "type", data) == KVM_TRACE_MMIO_WRITE) {
+               mmio_event_get_key(event, data, key);
+               return true;
+       }
+
+       return false;
+}
+
+static bool mmio_event_end(struct event_format *event,  void *data,
+                       struct event_key *key)
+{
+       /* MMIO write end event in kernel. */
+       if (kvm_entry_event(event))
+               return true;
+
+       /* MMIO read end event in kernel.*/
+       if (!strcmp(event->name, "kvm_mmio") &&
+               raw_field_value(event, "type", data) == KVM_TRACE_MMIO_READ) {
+               mmio_event_get_key(event, data, key);
+               return true;
+       }
+
+       return false;
+}
+
+static void mmio_event_decode_key(struct event_key *key, char decode[20])
+{
+       scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
+                               key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
+}
+
+static struct kvm_events_ops mmio_events = {
+       .is_begin_event = mmio_event_begin,
+       .is_end_event = mmio_event_end,
+       .decode_key = mmio_event_decode_key,
+       .name = "MMIO Access"
+};
+
+ /* The time of emulation pio access is from kvm_pio to kvm_entry. */
+static void ioport_event_get_key(struct event_format *event, void *data,
+                       struct event_key *key)
+{
+       key->key = raw_field_value(event, "port", data);
+       key->info = raw_field_value(event, "rw", data);
+}
+
+static bool ioport_event_begin(struct event_format *event, void *data,
+                       struct event_key *key)
+{
+       if (!strcmp(event->name, "kvm_pio")) {
+               ioport_event_get_key(event, data, key);
+               return true;
+       }
+
+       return false;
+}
+
+static bool ioport_event_end(struct event_format *event, void *data __maybe_unused,
+                            struct event_key *key __maybe_unused)
+{
+       if (kvm_entry_event(event))
+               return true;
+
+       return false;
+}
+
+static void ioport_event_decode_key(struct event_key *key, char decode[20])
+{
+       scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
+                               key->info ? "POUT" : "PIN");
+}
+
+static struct kvm_events_ops ioport_events = {
+       .is_begin_event = ioport_event_begin,
+       .is_end_event = ioport_event_end,
+       .decode_key = ioport_event_decode_key,
+       .name = "IO Port Access"
+};
+
+static const char *report_event = "vmexit";
+struct kvm_events_ops *events_ops;
+
+static bool register_kvm_events_ops(void)
+{
+       bool ret = true;
+
+       if (!strcmp(report_event, "vmexit"))
+               events_ops = &exit_events;
+       else if (!strcmp(report_event, "mmio"))
+               events_ops = &mmio_events;
+       else if (!strcmp(report_event, "ioport"))
+               events_ops = &ioport_events;
+       else {
+               pr_err("Unknown report event:%s\n", report_event);
+               ret = false;
+       }
+
+       return ret;
+}
+
+struct kvm_event_stats {
+       u64 time;
+       struct stats stats;
+};
+
+struct kvm_event {
+       struct list_head hash_entry;
+       struct rb_node rb;
+
+       struct event_key key;
+
+       struct kvm_event_stats total;
+
+       #define DEFAULT_VCPU_NUM 8
+       int max_vcpu;
+       struct kvm_event_stats *vcpu;
+};
+
+struct vcpu_event_record {
+       int vcpu_id;
+       u64 start_time;
+       struct kvm_event *last_event;
+};
+
+#define EVENTS_BITS                    12
+#define EVENTS_CACHE_SIZE      (1UL << EVENTS_BITS)
+
+static u64 total_time;
+static u64 total_count;
+static struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
+
+static void init_kvm_event_record(void)
+{
+       int i;
+
+       for (i = 0; i < (int)EVENTS_CACHE_SIZE; i++)
+               INIT_LIST_HEAD(&kvm_events_cache[i]);
+}
+
+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;
+
+       if (vcpu_id < event->max_vcpu)
+               return true;
+
+       while (event->max_vcpu <= vcpu_id)
+               event->max_vcpu += DEFAULT_VCPU_NUM;
+
+       event->vcpu = realloc(event->vcpu,
+                             event->max_vcpu * sizeof(*event->vcpu));
+       if (!event->vcpu) {
+               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;
+       return event;
+}
+
+static struct kvm_event *find_create_kvm_event(struct event_key *key)
+{
+       struct kvm_event *event;
+       struct list_head *head;
+
+       BUG_ON(key->key == INVALID_KEY);
+
+       head = &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 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(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)
+{
+       kvm_update_event_stats(&event->total, time_diff);
+
+       if (!kvm_event_expand(event, vcpu_id))
+               return false;
+
+       kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
+       return true;
+}
+
+static bool handle_end_event(struct vcpu_event_record *vcpu_record,
+                            struct event_key *key, u64 timestamp)
+{
+       struct kvm_event *event;
+       u64 time_begin, time_diff;
+
+       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(key);
+
+       if (!event)
+               return false;
+
+       vcpu_record->last_event = NULL;
+       vcpu_record->start_time = 0;
+
+       BUG_ON(timestamp < time_begin);
+
+       time_diff = timestamp - time_begin;
+       return update_kvm_event(event, vcpu_record->vcpu_id, time_diff);
+}
+
+static struct vcpu_event_record
+*per_vcpu_record(struct thread *thread, struct event_format *event, void *data)
+{
+       /* Only kvm_entry records vcpu id. */
+       if (!thread->priv && kvm_entry_event(event)) {
+               struct vcpu_event_record *vcpu_record;
+
+               vcpu_record = zalloc(sizeof(struct vcpu_event_record));
+               if (!vcpu_record) {
+                       pr_err("Not enough memory\n");
+                       return NULL;
+               }
+
+               vcpu_record->vcpu_id = raw_field_value(event, "vcpu_id", data);
+               thread->priv = vcpu_record;
+       }
+
+       return (struct vcpu_event_record *)thread->priv;
+}
+
+static bool handle_kvm_event(struct thread *thread, struct event_format *event,
+                            void *data, u64 timestamp)
+{
+       struct vcpu_event_record *vcpu_record;
+       struct event_key key = {.key = INVALID_KEY};
+
+       vcpu_record = per_vcpu_record(thread, event, data);
+       if (!vcpu_record)
+               return true;
+
+       if (events_ops->is_begin_event(event, data, &key))
+               return handle_begin_event(vcpu_record, &key, timestamp);
+
+       if (events_ops->is_end_event(event, data, &key))
+               return handle_end_event(vcpu_record, &key, timestamp);
+
+       return true;
+}
+
+typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
+struct kvm_event_key {
+       const char *name;
+       key_cmp_fun key;
+};
+
+static int trace_vcpu = -1;
+#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);
+
+#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 const char *sort_key = "sample";
+static key_cmp_fun compare;
+
+static bool select_key(void)
+{
+       int i;
+
+       for (i = 0; keys[i].name; i++) {
+               if (!strcmp(keys[i].name, sort_key)) {
+                       compare = keys[i].key;
+                       return true;
+               }
+       }
+
+       pr_err("Unknown compare key:%s\n", sort_key);
+       return false;
+}
+
+static struct rb_root result;
+static void insert_to_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 kvm_event *event, int vcpu)
+{
+       total_count += get_event_count(event, vcpu);
+       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(int vcpu)
+{
+       unsigned int i;
+       struct kvm_event *event;
+
+       for (i = 0; i < EVENTS_CACHE_SIZE; i++)
+               list_for_each_entry(event, &kvm_events_cache[i], hash_entry)
+                       if (event_is_valid(event, vcpu)) {
+                               update_total_count(event, vcpu);
+                               insert_to_result(event, compare, vcpu);
+                       }
+}
+
+/* returns left most element of result, and erase it */
+static struct kvm_event *pop_from_result(void)
+{
+       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(int vcpu)
+{
+       pr_info("Analyze events for ");
+
+       if (vcpu == -1)
+               pr_info("all VCPUs:\n\n");
+       else
+               pr_info("VCPU %d:\n\n", vcpu);
+}
+
+static void print_result(int vcpu)
+{
+       char decode[20];
+       struct kvm_event *event;
+
+       pr_info("\n\n");
+       print_vcpu_info(vcpu);
+       pr_info("%20s ", events_ops->name);
+       pr_info("%10s ", "Samples");
+       pr_info("%9s ", "Samples%");
+
+       pr_info("%9s ", "Time%");
+       pr_info("%16s ", "Avg time");
+       pr_info("\n\n");
+
+       while ((event = pop_from_result())) {
+               u64 ecount, etime;
+
+               ecount = get_event_count(event, vcpu);
+               etime = get_event_time(event, vcpu);
+
+               events_ops->decode_key(&event->key, decode);
+               pr_info("%20s ", decode);
+               pr_info("%10llu ", (unsigned long long)ecount);
+               pr_info("%8.2f%% ", (double)ecount / total_count * 100);
+               pr_info("%8.2f%% ", (double)etime / total_time * 100);
+               pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
+                       kvm_event_rel_stddev(vcpu, event));
+               pr_info("\n");
+       }
+
+       pr_info("\nTotal Samples:%lld, Total events handled time:%.2fus.\n\n",
+               (unsigned long long)total_count, total_time / 1e3);
+}
+
+static int process_sample_event(struct perf_tool *tool __maybe_unused,
+                               union perf_event *event,
+                               struct perf_sample *sample,
+                               struct perf_evsel *evsel,
+                               struct machine *machine)
+{
+       struct thread *thread = machine__findnew_thread(machine, sample->tid);
+
+       if (thread == NULL) {
+               pr_debug("problem processing %d event, skipping it.\n",
+                       event->header.type);
+               return -1;
+       }
+
+       if (!handle_kvm_event(thread, evsel->tp_format, sample->raw_data,
+                        sample->time))
+               return -1;
+
+       return 0;
+}
+
+static struct perf_tool eops = {
+       .sample                 = process_sample_event,
+       .comm                   = perf_event__process_comm,
+       .ordered_samples        = true,
+};
+
+static int get_cpu_isa(struct perf_session *session)
+{
+       char *cpuid;
+       int isa;
+
+       cpuid = perf_header__read_feature(session, HEADER_CPUID);
+
+       if (!cpuid) {
+               pr_err("read HEADER_CPUID failed.\n");
+               return -ENOTSUP;
+       }
+
+       if (strstr(cpuid, "Intel"))
+               isa = 1;
+       else if (strstr(cpuid, "AMD"))
+               isa = 0;
+       else {
+               pr_err("CPU %s is not supported.\n", cpuid);
+               isa = -ENOTSUP;
+       }
+
+       free(cpuid);
+       return isa;
+}
+
+static const char *file_name;
+
+static int read_events(void)
+{
+       struct perf_session *kvm_session;
+       int ret;
+
+       kvm_session = perf_session__new(file_name, O_RDONLY, 0, false, &eops);
+       if (!kvm_session) {
+               pr_err("Initializing perf session failed\n");
+               return -EINVAL;
+       }
+
+       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 = get_cpu_isa(kvm_session);
+
+       if (ret < 0)
+               return ret;
+
+       cpu_isa = ret;
+
+       return perf_session__process_events(kvm_session, &eops);
+}
+
+static bool verify_vcpu(int vcpu)
+{
+       if (vcpu != -1 && vcpu < 0) {
+               pr_err("Invalid vcpu:%d.\n", vcpu);
+               return false;
+       }
+
+       return true;
+}
+
+static int kvm_events_report_vcpu(int vcpu)
+{
+       int ret = -EINVAL;
+
+       if (!verify_vcpu(vcpu))
+               goto exit;
+
+       if (!select_key())
+               goto exit;
+
+       if (!register_kvm_events_ops())
+               goto exit;
+
+       init_kvm_event_record();
+       setup_pager();
+
+       ret = read_events();
+       if (ret)
+               goto exit;
+
+       sort_result(vcpu);
+       print_result(vcpu);
+exit:
+       return ret;
+}
+
+static const char * const record_args[] = {
+       "record",
+       "-R",
+       "-f",
+       "-m", "1024",
+       "-c", "1",
+       "-e", "kvm:kvm_entry",
+       "-e", "kvm:kvm_exit",
+       "-e", "kvm:kvm_mmio",
+       "-e", "kvm:kvm_pio",
+};
+
+#define STRDUP_FAIL_EXIT(s)            \
+       ({      char *_p;               \
+       _p = strdup(s);         \
+               if (!_p)                \
+                       return -ENOMEM; \
+               _p;                     \
+       })
+
+static int kvm_events_record(int argc, const char **argv)
+{
+       unsigned int rec_argc, i, j;
+       const char **rec_argv;
+
+       rec_argc = ARRAY_SIZE(record_args) + argc + 2;
+       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]);
+
+       rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
+       rec_argv[i++] = STRDUP_FAIL_EXIT(file_name);
+
+       for (j = 1; j < (unsigned int)argc; j++, i++)
+               rec_argv[i] = argv[j];
+
+       return cmd_record(i, rec_argv, NULL);
+}
+
+static const char * const kvm_events_report_usage[] = {
+       "perf kvm stat report [<options>]",
+       NULL
+};
+
+static const struct option kvm_events_report_options[] = {
+       OPT_STRING(0, "event", &report_event, "report event",
+                   "event for reporting: vmexit, mmio, ioport"),
+       OPT_INTEGER(0, "vcpu", &trace_vcpu,
+                   "vcpu id to report"),
+       OPT_STRING('k', "key", &sort_key, "sort-key",
+                   "key for sorting: sample(sort by samples number)"
+                   " time (sort by avg time)"),
+       OPT_END()
+};
+
+static int kvm_events_report(int argc, const char **argv)
+{
+       symbol__init();
+
+       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);
+       }
+
+       return kvm_events_report_vcpu(trace_vcpu);
+}
+
+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("\nOtherwise, it is the alias of 'perf stat':\n");
+}
+
+static int kvm_cmd_stat(int argc, const char **argv)
+{
+       if (argc == 1) {
+               print_kvm_stat_usage();
+               goto perf_stat;
+       }
+
+       if (!strncmp(argv[1], "rec", 3))
+               return kvm_events_record(argc - 1, argv + 1);
+
+       if (!strncmp(argv[1], "rep", 3))
+               return kvm_events_report(argc - 1 , argv + 1);
+
+perf_stat:
+       return cmd_stat(argc, argv, NULL);
+}
+
 static char                    name_buffer[256];
 
 static const char * const kvm_usage[] = {
-       "perf kvm [<options>] {top|record|report|diff|buildid-list}",
+       "perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
        NULL
 };
 
@@ -135,6 +967,8 @@ int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
                return cmd_top(argc, argv, NULL);
        else if (!strncmp(argv[0], "buildid-list", 12))
                return __cmd_buildid_list(argc, argv);
+       else if (!strncmp(argv[0], "stat", 4))
+               return kvm_cmd_stat(argc, argv);
        else
                usage_with_options(kvm_usage, kvm_options);
 
index acbf633..ad72b28 100644 (file)
@@ -1673,6 +1673,11 @@ static int process_build_id(struct perf_file_section *section,
        return 0;
 }
 
+static char *read_cpuid(struct perf_header *ph, int fd)
+{
+       return do_read_string(fd, ph);
+}
+
 static struct perf_evsel *
 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
 {
@@ -1726,6 +1731,7 @@ process_event_desc(struct perf_file_section *section __maybe_unused,
 struct feature_ops {
        int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
        void (*print)(struct perf_header *h, int fd, FILE *fp);
+       char *(*read)(struct perf_header *h, int fd);
        int (*process)(struct perf_file_section *section,
                       struct perf_header *h, int feat, int fd, void *data);
        const char *name;
@@ -1740,6 +1746,9 @@ struct feature_ops {
 #define FEAT_OPF(n, func) \
        [n] = { .name = #n, .write = write_##func, .print = print_##func, \
                .full_only = true }
+#define FEAT_OPA_R(n, func) \
+       [n] = { .name = #n, .write = write_##func, .print = print_##func, \
+               .read  = read_##func }
 
 /* feature_ops not implemented: */
 #define print_tracing_data     NULL
@@ -1754,7 +1763,7 @@ static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
        FEAT_OPA(HEADER_ARCH,           arch),
        FEAT_OPA(HEADER_NRCPUS,         nrcpus),
        FEAT_OPA(HEADER_CPUDESC,        cpudesc),
-       FEAT_OPA(HEADER_CPUID,          cpuid),
+       FEAT_OPA_R(HEADER_CPUID,        cpuid),
        FEAT_OPA(HEADER_TOTAL_MEM,      total_mem),
        FEAT_OPP(HEADER_EVENT_DESC,     event_desc),
        FEAT_OPA(HEADER_CMDLINE,        cmdline),
@@ -1809,6 +1818,54 @@ int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
        return 0;
 }
 
+struct header_read_data {
+       int feat;
+       char *result;
+};
+
+static int perf_file_section__read_feature(struct perf_file_section *section,
+                               struct perf_header *ph,
+                               int feat, int fd, void *data)
+{
+       struct header_read_data *hd = data;
+
+       if (feat != hd->feat)
+               return 0;
+
+       if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
+               pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
+                               "%d, continuing...\n", section->offset, feat);
+       return 0;
+       }
+
+       if (feat >= HEADER_LAST_FEATURE) {
+               pr_warning("unknown feature %d\n", feat);
+               return 0;
+       }
+
+       if (!feat_ops[feat].read) {
+               pr_warning("read is not supported for feature %d\n", feat);
+               return 0;
+       }
+
+       hd->result = feat_ops[feat].read(ph, fd);
+       return 0;
+}
+
+char *perf_header__read_feature(struct perf_session *session, int feat)
+{
+       struct perf_header *header = &session->header;
+       struct header_read_data hd;
+       int fd = session->fd;
+
+       hd.feat = feat;
+       hd.result = NULL;
+
+       perf_header__process_sections(header, fd, &hd,
+                               perf_file_section__read_feature);
+       return hd.result;
+}
+
 static int do_write_feat(int fd, struct perf_header *h, int type,
                         struct perf_file_section **p,
                         struct perf_evlist *evlist)
index 209dad4..58de08b 100644 (file)
@@ -94,6 +94,7 @@ int perf_header__process_sections(struct perf_header *header, int fd,
                                  int feat, int fd, void *data));
 
 int perf_header__fprintf_info(struct perf_session *s, FILE *fp, bool full);
+char *perf_header__read_feature(struct perf_session *session, int feat);
 
 int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
                          const char *name, bool is_kallsyms, bool is_vdso);
index 70c2c13..f66610b 100644 (file)
@@ -16,6 +16,8 @@ struct thread {
        bool                    comm_set;
        char                    *comm;
        int                     comm_len;
+
+       void                    *priv;
 };
 
 struct machine;