Quelle  session.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
#include <errno.h>
#include <signal.h>
#include <inttypes.h>
#include <linux/err.h>
#include <linux/kernel.h>
#include <linux/zalloc.h>
#include <api/fs/fs.h>

#include <byteswap.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/mman.h>
#include <perf/cpumap.h>
#include <perf/event.h>

#include "map_symbol.h"
#include "branch.h"
#include "debug.h"
#include "env.h"
#include "evlist.h"
#include "evsel.h"
#include "memswap.h"
#include "map.h"
#include "symbol.h"
#include "session.h"
#include "tool.h"
#include "perf_regs.h"
#include "asm/bug.h"
#include "auxtrace.h"
#include "thread.h"
#include "thread-stack.h"
#include "sample-raw.h"
#include "stat.h"
#include "tsc.h"
#include "ui/progress.h"
#include "util.h"
#include "arch/common.h"
#include "units.h"
#include "annotate.h"
#include "perf.h"
#include <internal/lib.h>

static int perf_session__deliver_event(struct perf_session *session,
           union perf_event *event,
           const struct perf_tool *tool,
           u64 file_offset,
           const char *file_path);

static int perf_session__open(struct perf_session *session)
{
 struct perf_data *data = session->data;

 if (perf_session__read_header(session) < 0) {
  pr_err("incompatible file format (rerun with -v to learn more)\n");
  return -1;
 }

 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) {
  /* Auxiliary events may reference exited threads, hold onto dead ones. */
  symbol_conf.keep_exited_threads = true;
 }

 if (perf_data__is_pipe(data))
  return 0;

 if (perf_header__has_feat(&session->header, HEADER_STAT))
  return 0;

 if (!evlist__valid_sample_type(session->evlist)) {
  pr_err("non matching sample_type\n");
  return -1;
 }

 if (!evlist__valid_sample_id_all(session->evlist)) {
  pr_err("non matching sample_id_all\n");
  return -1;
 }

 if (!evlist__valid_read_format(session->evlist)) {
  pr_err("non matching read_format\n");
  return -1;
 }

 return 0;
}

void perf_session__set_id_hdr_size(struct perf_session *session)
{
 u16 id_hdr_size = evlist__id_hdr_size(session->evlist);

 machines__set_id_hdr_size(&session->machines, id_hdr_size);
}

int perf_session__create_kernel_maps(struct perf_session *session)
{
 int ret = machine__create_kernel_maps(&session->machines.host);

 if (ret >= 0)
  ret = machines__create_guest_kernel_maps(&session->machines);
 return ret;
}

static void perf_session__destroy_kernel_maps(struct perf_session *session)
{
 machines__destroy_kernel_maps(&session->machines);
}

static bool perf_session__has_comm_exec(struct perf_session *session)
{
 struct evsel *evsel;

 evlist__for_each_entry(session->evlist, evsel) {
  if (evsel->core.attr.comm_exec)
   return true;
 }

 return false;
}

static void perf_session__set_comm_exec(struct perf_session *session)
{
 bool comm_exec = perf_session__has_comm_exec(session);

 machines__set_comm_exec(&session->machines, comm_exec);
}

static int ordered_events__deliver_event(struct ordered_events *oe,
      struct ordered_event *event)
{
 struct perf_session *session = container_of(oe, struct perf_session,
          ordered_events);

 return perf_session__deliver_event(session, event->event,
        session->tool, event->file_offset,
        event->file_path);
}

struct perf_session *__perf_session__new(struct perf_data *data,
      struct perf_tool *tool,
      bool trace_event_repipe,
      struct perf_env *host_env)
{
 int ret = -ENOMEM;
 struct perf_session *session = zalloc(sizeof(*session));

 if (!session)
  goto out;

 session->trace_event_repipe = trace_event_repipe;
 session->tool   = tool;
 session->decomp_data.zstd_decomp = &session->zstd_data;
 session->active_decomp = &session->decomp_data;
 INIT_LIST_HEAD(&session->auxtrace_index);
 machines__init(&session->machines);
 ordered_events__init(&session->ordered_events,
        ordered_events__deliver_event, NULL);

 perf_env__init(&session->header.env);
 if (data) {
  ret = perf_data__open(data);
  if (ret < 0)
   goto out_delete;

  session->data = data;

  if (perf_data__is_read(data)) {
   ret = perf_session__open(session);
   if (ret < 0)
    goto out_delete;

   /*
 * set session attributes that are present in perf.data
 * but not in pipe-mode.
 */
   if (!data->is_pipe) {
    perf_session__set_id_hdr_size(session);
    perf_session__set_comm_exec(session);
   }

   evlist__init_trace_event_sample_raw(session->evlist, &session->header.env);

   /* Open the directory data. */
   if (data->is_dir) {
    ret = perf_data__open_dir(data);
    if (ret)
     goto out_delete;
   }

   if (!symbol_conf.kallsyms_name &&
       !symbol_conf.vmlinux_name)
    symbol_conf.kallsyms_name = perf_data__kallsyms_name(data);
  }
 } else  {
  assert(host_env != NULL);
  session->machines.host.env = host_env;
 }
 if (session->evlist)
  session->evlist->session = session;

 session->machines.host.single_address_space =
  perf_env__single_address_space(session->machines.host.env);

 if (!data || perf_data__is_write(data)) {
  /*
 * In O_RDONLY mode this will be performed when reading the
 * kernel MMAP event, in perf_event__process_mmap().
 */
  if (perf_session__create_kernel_maps(session) < 0)
   pr_warning("Cannot read kernel map\n");
 }

 /*
 * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is
 * processed, so evlist__sample_id_all is not meaningful here.
 */
 if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps &&
     tool->ordered_events && !evlist__sample_id_all(session->evlist)) {
  dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
  tool->ordered_events = false;
 }

 return session;

 out_delete:
 perf_session__delete(session);
 out:
 return ERR_PTR(ret);
}

static void perf_decomp__release_events(struct decomp *next)
{
 struct decomp *decomp;
 size_t mmap_len;

 do {
  decomp = next;
  if (decomp == NULL)
   break;
  next = decomp->next;
  mmap_len = decomp->mmap_len;
  munmap(decomp, mmap_len);
 } while (1);
}

void perf_session__delete(struct perf_session *session)
{
 if (session == NULL)
  return;
 auxtrace__free(session);
 auxtrace_index__free(&session->auxtrace_index);
 debuginfo_cache__delete();
 perf_session__destroy_kernel_maps(session);
 perf_decomp__release_events(session->decomp_data.decomp);
 perf_env__exit(&session->header.env);
 machines__exit(&session->machines);
 if (session->data) {
  if (perf_data__is_read(session->data))
   evlist__delete(session->evlist);
  perf_data__close(session->data);
 }
#ifdef HAVE_LIBTRACEEVENT
 trace_event__cleanup(&session->tevent);
#endif
 free(session);
}

static void swap_sample_id_all(union perf_event *event, void *data)
{
 void *end = (void *) event + event->header.size;
 int size = end - data;

 BUG_ON(size % sizeof(u64));
 mem_bswap_64(data, size);
}

static void perf_event__all64_swap(union perf_event *event,
       bool sample_id_all __maybe_unused)
{
 struct perf_event_header *hdr = &event->header;
 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
}

static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
{
 event->comm.pid = bswap_32(event->comm.pid);
 event->comm.tid = bswap_32(event->comm.tid);

 if (sample_id_all) {
  void *data = &event->comm.comm;

  data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
  swap_sample_id_all(event, data);
 }
}

static void perf_event__mmap_swap(union perf_event *event,
      bool sample_id_all)
{
 event->mmap.pid   = bswap_32(event->mmap.pid);
 event->mmap.tid   = bswap_32(event->mmap.tid);
 event->mmap.start = bswap_64(event->mmap.start);
 event->mmap.len   = bswap_64(event->mmap.len);
 event->mmap.pgoff = bswap_64(event->mmap.pgoff);

 if (sample_id_all) {
  void *data = &event->mmap.filename;

  data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
  swap_sample_id_all(event, data);
 }
}

static void perf_event__mmap2_swap(union perf_event *event,
      bool sample_id_all)
{
 event->mmap2.pid   = bswap_32(event->mmap2.pid);
 event->mmap2.tid   = bswap_32(event->mmap2.tid);
 event->mmap2.start = bswap_64(event->mmap2.start);
 event->mmap2.len   = bswap_64(event->mmap2.len);
 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);

 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID)) {
  event->mmap2.maj   = bswap_32(event->mmap2.maj);
  event->mmap2.min   = bswap_32(event->mmap2.min);
  event->mmap2.ino   = bswap_64(event->mmap2.ino);
  event->mmap2.ino_generation = bswap_64(event->mmap2.ino_generation);
 }

 if (sample_id_all) {
  void *data = &event->mmap2.filename;

  data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
  swap_sample_id_all(event, data);
 }
}
static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
{
 event->fork.pid  = bswap_32(event->fork.pid);
 event->fork.tid  = bswap_32(event->fork.tid);
 event->fork.ppid = bswap_32(event->fork.ppid);
 event->fork.ptid = bswap_32(event->fork.ptid);
 event->fork.time = bswap_64(event->fork.time);

 if (sample_id_all)
  swap_sample_id_all(event, &event->fork + 1);
}

static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
{
 event->read.pid   = bswap_32(event->read.pid);
 event->read.tid   = bswap_32(event->read.tid);
 event->read.value  = bswap_64(event->read.value);
 event->read.time_enabled = bswap_64(event->read.time_enabled);
 event->read.time_running = bswap_64(event->read.time_running);
 event->read.id   = bswap_64(event->read.id);

 if (sample_id_all)
  swap_sample_id_all(event, &event->read + 1);
}

static void perf_event__aux_swap(union perf_event *event, bool sample_id_all)
{
 event->aux.aux_offset = bswap_64(event->aux.aux_offset);
 event->aux.aux_size   = bswap_64(event->aux.aux_size);
 event->aux.flags      = bswap_64(event->aux.flags);

 if (sample_id_all)
  swap_sample_id_all(event, &event->aux + 1);
}

static void perf_event__itrace_start_swap(union perf_event *event,
       bool sample_id_all)
{
 event->itrace_start.pid  = bswap_32(event->itrace_start.pid);
 event->itrace_start.tid  = bswap_32(event->itrace_start.tid);

 if (sample_id_all)
  swap_sample_id_all(event, &event->itrace_start + 1);
}

static void perf_event__switch_swap(union perf_event *event, bool sample_id_all)
{
 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) {
  event->context_switch.next_prev_pid =
    bswap_32(event->context_switch.next_prev_pid);
  event->context_switch.next_prev_tid =
    bswap_32(event->context_switch.next_prev_tid);
 }

 if (sample_id_all)
  swap_sample_id_all(event, &event->context_switch + 1);
}

static void perf_event__text_poke_swap(union perf_event *event, bool sample_id_all)
{
 event->text_poke.addr    = bswap_64(event->text_poke.addr);
 event->text_poke.old_len = bswap_16(event->text_poke.old_len);
 event->text_poke.new_len = bswap_16(event->text_poke.new_len);

 if (sample_id_all) {
  size_t len = sizeof(event->text_poke.old_len) +
        sizeof(event->text_poke.new_len) +
        event->text_poke.old_len +
        event->text_poke.new_len;
  void *data = &event->text_poke.old_len;

  data += PERF_ALIGN(len, sizeof(u64));
  swap_sample_id_all(event, data);
 }
}

static void perf_event__throttle_swap(union perf_event *event,
          bool sample_id_all)
{
 event->throttle.time   = bswap_64(event->throttle.time);
 event->throttle.id   = bswap_64(event->throttle.id);
 event->throttle.stream_id = bswap_64(event->throttle.stream_id);

 if (sample_id_all)
  swap_sample_id_all(event, &event->throttle + 1);
}

static void perf_event__namespaces_swap(union perf_event *event,
     bool sample_id_all)
{
 u64 i;

 event->namespaces.pid  = bswap_32(event->namespaces.pid);
 event->namespaces.tid  = bswap_32(event->namespaces.tid);
 event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces);

 for (i = 0; i < event->namespaces.nr_namespaces; i++) {
  struct perf_ns_link_info *ns = &event->namespaces.link_info[i];

  ns->dev = bswap_64(ns->dev);
  ns->ino = bswap_64(ns->ino);
 }

 if (sample_id_all)
  swap_sample_id_all(event, &event->namespaces.link_info[i]);
}

static void perf_event__cgroup_swap(union perf_event *event, bool sample_id_all)
{
 event->cgroup.id = bswap_64(event->cgroup.id);

 if (sample_id_all) {
  void *data = &event->cgroup.path;

  data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
  swap_sample_id_all(event, data);
 }
}

static u8 revbyte(u8 b)
{
 int rev = (b >> 4) | ((b & 0xf) << 4);
 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 return (u8) rev;
}

/*
 * XXX this is hack in attempt to carry flags bitfield
 * through endian village. ABI says:
 *
 * Bit-fields are allocated from right to left (least to most significant)
 * on little-endian implementations and from left to right (most to least
 * significant) on big-endian implementations.
 *
 * The above seems to be byte specific, so we need to reverse each
 * byte of the bitfield. 'Internet' also says this might be implementation
 * specific and we probably need proper fix and carry perf_event_attr
 * bitfield flags in separate data file FEAT_ section. Thought this seems
 * to work for now.
 */
static void swap_bitfield(u8 *p, unsigned len)
{
 unsigned i;

 for (i = 0; i < len; i++) {
  *p = revbyte(*p);
  p++;
 }
}

/* exported for swapping attributes in file header */
void perf_event__attr_swap(struct perf_event_attr *attr)
{
 attr->type  = bswap_32(attr->type);
 attr->size  = bswap_32(attr->size);

#define bswap_safe(f, n)      \
 (attr->size > (offsetof(struct perf_event_attr, f) +  \
         sizeof(attr->f) * (n)))
#define bswap_field(f, sz)    \
do {       \
 if (bswap_safe(f, 0))   \
  attr->f = bswap_##sz(attr->f); \
} while(0)
#define bswap_field_16(f) bswap_field(f, 16)
#define bswap_field_32(f) bswap_field(f, 32)
#define bswap_field_64(f) bswap_field(f, 64)

 bswap_field_64(config);
 bswap_field_64(sample_period);
 bswap_field_64(sample_type);
 bswap_field_64(read_format);
 bswap_field_32(wakeup_events);
 bswap_field_32(bp_type);
 bswap_field_64(bp_addr);
 bswap_field_64(bp_len);
 bswap_field_64(branch_sample_type);
 bswap_field_64(sample_regs_user);
 bswap_field_32(sample_stack_user);
 bswap_field_32(aux_watermark);
 bswap_field_16(sample_max_stack);
 bswap_field_32(aux_sample_size);

 /*
 * After read_format are bitfields. Check read_format because
 * we are unable to use offsetof on bitfield.
 */
 if (bswap_safe(read_format, 1))
  swap_bitfield((u8 *) (&attr->read_format + 1),
         sizeof(u64));
#undef bswap_field_64
#undef bswap_field_32
#undef bswap_field
#undef bswap_safe
}

static void perf_event__hdr_attr_swap(union perf_event *event,
          bool sample_id_all __maybe_unused)
{
 size_t size;

 perf_event__attr_swap(&event->attr.attr);

 size = event->header.size;
 size -= perf_record_header_attr_id(event) - (void *)event;
 mem_bswap_64(perf_record_header_attr_id(event), size);
}

static void perf_event__event_update_swap(union perf_event *event,
       bool sample_id_all __maybe_unused)
{
 event->event_update.type = bswap_64(event->event_update.type);
 event->event_update.id   = bswap_64(event->event_update.id);
}

static void perf_event__event_type_swap(union perf_event *event,
     bool sample_id_all __maybe_unused)
{
 event->event_type.event_type.event_id =
  bswap_64(event->event_type.event_type.event_id);
}

static void perf_event__tracing_data_swap(union perf_event *event,
       bool sample_id_all __maybe_unused)
{
 event->tracing_data.size = bswap_32(event->tracing_data.size);
}

static void perf_event__auxtrace_info_swap(union perf_event *event,
        bool sample_id_all __maybe_unused)
{
 size_t size;

 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type);

 size = event->header.size;
 size -= (void *)&event->auxtrace_info.priv - (void *)event;
 mem_bswap_64(event->auxtrace_info.priv, size);
}

static void perf_event__auxtrace_swap(union perf_event *event,
          bool sample_id_all __maybe_unused)
{
 event->auxtrace.size      = bswap_64(event->auxtrace.size);
 event->auxtrace.offset    = bswap_64(event->auxtrace.offset);
 event->auxtrace.reference = bswap_64(event->auxtrace.reference);
 event->auxtrace.idx       = bswap_32(event->auxtrace.idx);
 event->auxtrace.tid       = bswap_32(event->auxtrace.tid);
 event->auxtrace.cpu       = bswap_32(event->auxtrace.cpu);
}

static void perf_event__auxtrace_error_swap(union perf_event *event,
         bool sample_id_all __maybe_unused)
{
 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type);
 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code);
 event->auxtrace_error.cpu  = bswap_32(event->auxtrace_error.cpu);
 event->auxtrace_error.pid  = bswap_32(event->auxtrace_error.pid);
 event->auxtrace_error.tid  = bswap_32(event->auxtrace_error.tid);
 event->auxtrace_error.fmt  = bswap_32(event->auxtrace_error.fmt);
 event->auxtrace_error.ip   = bswap_64(event->auxtrace_error.ip);
 if (event->auxtrace_error.fmt)
  event->auxtrace_error.time = bswap_64(event->auxtrace_error.time);
 if (event->auxtrace_error.fmt >= 2) {
  event->auxtrace_error.machine_pid = bswap_32(event->auxtrace_error.machine_pid);
  event->auxtrace_error.vcpu = bswap_32(event->auxtrace_error.vcpu);
 }
}

static void perf_event__thread_map_swap(union perf_event *event,
     bool sample_id_all __maybe_unused)
{
 unsigned i;

 event->thread_map.nr = bswap_64(event->thread_map.nr);

 for (i = 0; i < event->thread_map.nr; i++)
  event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid);
}

static void perf_event__cpu_map_swap(union perf_event *event,
         bool sample_id_all __maybe_unused)
{
 struct perf_record_cpu_map_data *data = &event->cpu_map.data;

 data->type = bswap_16(data->type);

 switch (data->type) {
 case PERF_CPU_MAP__CPUS:
  data->cpus_data.nr = bswap_16(data->cpus_data.nr);

  for (unsigned i = 0; i < data->cpus_data.nr; i++)
   data->cpus_data.cpu[i] = bswap_16(data->cpus_data.cpu[i]);
  break;
 case PERF_CPU_MAP__MASK:
  data->mask32_data.long_size = bswap_16(data->mask32_data.long_size);

  switch (data->mask32_data.long_size) {
  case 4:
   data->mask32_data.nr = bswap_16(data->mask32_data.nr);
   for (unsigned i = 0; i < data->mask32_data.nr; i++)
    data->mask32_data.mask[i] = bswap_32(data->mask32_data.mask[i]);
   break;
  case 8:
   data->mask64_data.nr = bswap_16(data->mask64_data.nr);
   for (unsigned i = 0; i < data->mask64_data.nr; i++)
    data->mask64_data.mask[i] = bswap_64(data->mask64_data.mask[i]);
   break;
  default:
   pr_err("cpu_map swap: unsupported long size\n");
  }
  break;
 case PERF_CPU_MAP__RANGE_CPUS:
  data->range_cpu_data.start_cpu = bswap_16(data->range_cpu_data.start_cpu);
  data->range_cpu_data.end_cpu = bswap_16(data->range_cpu_data.end_cpu);
  break;
 default:
  break;
 }
}

static void perf_event__stat_config_swap(union perf_event *event,
      bool sample_id_all __maybe_unused)
{
 u64 size;

 size  = bswap_64(event->stat_config.nr) * sizeof(event->stat_config.data[0]);
 size += 1; /* nr item itself */
 mem_bswap_64(&event->stat_config.nr, size);
}

static void perf_event__stat_swap(union perf_event *event,
      bool sample_id_all __maybe_unused)
{
 event->stat.id     = bswap_64(event->stat.id);
 event->stat.thread = bswap_32(event->stat.thread);
 event->stat.cpu    = bswap_32(event->stat.cpu);
 event->stat.val    = bswap_64(event->stat.val);
 event->stat.ena    = bswap_64(event->stat.ena);
 event->stat.run    = bswap_64(event->stat.run);
}

static void perf_event__stat_round_swap(union perf_event *event,
     bool sample_id_all __maybe_unused)
{
 event->stat_round.type = bswap_64(event->stat_round.type);
 event->stat_round.time = bswap_64(event->stat_round.time);
}

static void perf_event__time_conv_swap(union perf_event *event,
           bool sample_id_all __maybe_unused)
{
 event->time_conv.time_shift = bswap_64(event->time_conv.time_shift);
 event->time_conv.time_mult  = bswap_64(event->time_conv.time_mult);
 event->time_conv.time_zero  = bswap_64(event->time_conv.time_zero);

 if (event_contains(event->time_conv, time_cycles)) {
  event->time_conv.time_cycles = bswap_64(event->time_conv.time_cycles);
  event->time_conv.time_mask = bswap_64(event->time_conv.time_mask);
 }
}

typedef void (*perf_event__swap_op)(union perf_event *event,
        bool sample_id_all);

static perf_event__swap_op perf_event__swap_ops[] = {
 [PERF_RECORD_MMAP]    = perf_event__mmap_swap,
 [PERF_RECORD_MMAP2]    = perf_event__mmap2_swap,
 [PERF_RECORD_COMM]    = perf_event__comm_swap,
 [PERF_RECORD_FORK]    = perf_event__task_swap,
 [PERF_RECORD_EXIT]    = perf_event__task_swap,
 [PERF_RECORD_LOST]    = perf_event__all64_swap,
 [PERF_RECORD_READ]    = perf_event__read_swap,
 [PERF_RECORD_THROTTLE]    = perf_event__throttle_swap,
 [PERF_RECORD_UNTHROTTLE]   = perf_event__throttle_swap,
 [PERF_RECORD_SAMPLE]    = perf_event__all64_swap,
 [PERF_RECORD_AUX]    = perf_event__aux_swap,
 [PERF_RECORD_ITRACE_START]   = perf_event__itrace_start_swap,
 [PERF_RECORD_LOST_SAMPLES]   = perf_event__all64_swap,
 [PERF_RECORD_SWITCH]    = perf_event__switch_swap,
 [PERF_RECORD_SWITCH_CPU_WIDE]   = perf_event__switch_swap,
 [PERF_RECORD_NAMESPACES]   = perf_event__namespaces_swap,
 [PERF_RECORD_CGROUP]    = perf_event__cgroup_swap,
 [PERF_RECORD_TEXT_POKE]    = perf_event__text_poke_swap,
 [PERF_RECORD_AUX_OUTPUT_HW_ID]   = perf_event__all64_swap,
 [PERF_RECORD_HEADER_ATTR]   = perf_event__hdr_attr_swap,
 [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 [PERF_RECORD_HEADER_BUILD_ID]   = NULL,
 [PERF_RECORD_ID_INDEX]    = perf_event__all64_swap,
 [PERF_RECORD_AUXTRACE_INFO]   = perf_event__auxtrace_info_swap,
 [PERF_RECORD_AUXTRACE]    = perf_event__auxtrace_swap,
 [PERF_RECORD_AUXTRACE_ERROR]   = perf_event__auxtrace_error_swap,
 [PERF_RECORD_THREAD_MAP]   = perf_event__thread_map_swap,
 [PERF_RECORD_CPU_MAP]    = perf_event__cpu_map_swap,
 [PERF_RECORD_STAT_CONFIG]   = perf_event__stat_config_swap,
 [PERF_RECORD_STAT]    = perf_event__stat_swap,
 [PERF_RECORD_STAT_ROUND]   = perf_event__stat_round_swap,
 [PERF_RECORD_EVENT_UPDATE]   = perf_event__event_update_swap,
 [PERF_RECORD_TIME_CONV]    = perf_event__time_conv_swap,
 [PERF_RECORD_HEADER_MAX]   = NULL,
};

/*
 * When perf record finishes a pass on every buffers, it records this pseudo
 * event.
 * We record the max timestamp t found in the pass n.
 * Assuming these timestamps are monotonic across cpus, we know that if
 * a buffer still has events with timestamps below t, they will be all
 * available and then read in the pass n + 1.
 * Hence when we start to read the pass n + 2, we can safely flush every
 * events with timestamps below t.
 *
 *    ============ PASS n =================
 *       CPU 0         |   CPU 1
 *                     |
 *    cnt1 timestamps  |   cnt2 timestamps
 *          1          |         2
 *          2          |         3
 *          -          |         4  <--- max recorded
 *
 *    ============ PASS n + 1 ==============
 *       CPU 0         |   CPU 1
 *                     |
 *    cnt1 timestamps  |   cnt2 timestamps
 *          3          |         5
 *          4          |         6
 *          5          |         7 <---- max recorded
 *
 *      Flush every events below timestamp 4
 *
 *    ============ PASS n + 2 ==============
 *       CPU 0         |   CPU 1
 *                     |
 *    cnt1 timestamps  |   cnt2 timestamps
 *          6          |         8
 *          7          |         9
 *          -          |         10
 *
 *      Flush every events below timestamp 7
 *      etc...
 */
int perf_event__process_finished_round(const struct perf_tool *tool __maybe_unused,
           union perf_event *event __maybe_unused,
           struct ordered_events *oe)
{
 if (dump_trace)
  fprintf(stdout, "\n");
 return ordered_events__flush(oe, OE_FLUSH__ROUND);
}

int perf_session__queue_event(struct perf_session *s, union perf_event *event,
         u64 timestamp, u64 file_offset, const char *file_path)
{
 return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset, file_path);
}

static void callchain__lbr_callstack_printf(struct perf_sample *sample)
{
 struct ip_callchain *callchain = sample->callchain;
 struct branch_stack *lbr_stack = sample->branch_stack;
 struct branch_entry *entries = perf_sample__branch_entries(sample);
 u64 kernel_callchain_nr = callchain->nr;
 unsigned int i;

 for (i = 0; i < kernel_callchain_nr; i++) {
  if (callchain->ips[i] == PERF_CONTEXT_USER)
   break;
 }

 if ((i != kernel_callchain_nr) && lbr_stack->nr) {
  u64 total_nr;
  /*
 * LBR callstack can only get user call chain,
 * i is kernel call chain number,
 * 1 is PERF_CONTEXT_USER.
 *
 * The user call chain is stored in LBR registers.
 * LBR are pair registers. The caller is stored
 * in "from" register, while the callee is stored
 * in "to" register.
 * For example, there is a call stack
 * "A"->"B"->"C"->"D".
 * The LBR registers will be recorded like
 * "C"->"D", "B"->"C", "A"->"B".
 * So only the first "to" register and all "from"
 * registers are needed to construct the whole stack.
 */
  total_nr = i + 1 + lbr_stack->nr + 1;
  kernel_callchain_nr = i + 1;

  printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr);

  for (i = 0; i < kernel_callchain_nr; i++)
   printf("..... %2d: %016" PRIx64 "\n",
          i, callchain->ips[i]);

  printf("..... %2d: %016" PRIx64 "\n",
         (int)(kernel_callchain_nr), entries[0].to);
  for (i = 0; i < lbr_stack->nr; i++)
   printf("..... %2d: %016" PRIx64 "\n",
          (int)(i + kernel_callchain_nr + 1), entries[i].from);
 }
}

static void callchain__printf(struct evsel *evsel,
         struct perf_sample *sample)
{
 unsigned int i;
 struct ip_callchain *callchain = sample->callchain;

 if (evsel__has_branch_callstack(evsel))
  callchain__lbr_callstack_printf(sample);

 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr);

 for (i = 0; i < callchain->nr; i++)
  printf("..... %2d: %016" PRIx64 "\n",
         i, callchain->ips[i]);
}

static void branch_stack__printf(struct perf_sample *sample,
     struct evsel *evsel)
{
 struct branch_entry *entries = perf_sample__branch_entries(sample);
 bool callstack = evsel__has_branch_callstack(evsel);
 u64 *branch_stack_cntr = sample->branch_stack_cntr;
 uint64_t i;

 if (!callstack) {
  printf("%s: nr:%" PRIu64 "\n", "... branch stack", sample->branch_stack->nr);
 } else {
  /* the reason of adding 1 to nr is because after expanding
 * branch stack it generates nr + 1 callstack records. e.g.,
 *         B()->C()
 *         A()->B()
 * the final callstack should be:
 *         C()
 *         B()
 *         A()
 */
  printf("%s: nr:%" PRIu64 "\n", "... branch callstack", sample->branch_stack->nr+1);
 }

 for (i = 0; i < sample->branch_stack->nr; i++) {
  struct branch_entry *e = &entries[i];

  if (!callstack) {
   printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x %s %s\n",
    i, e->from, e->to,
    (unsigned short)e->flags.cycles,
    e->flags.mispred ? "M" : " ",
    e->flags.predicted ? "P" : " ",
    e->flags.abort ? "A" : " ",
    e->flags.in_tx ? "T" : " ",
    (unsigned)e->flags.reserved,
    get_branch_type(e),
    e->flags.spec ? branch_spec_desc(e->flags.spec) : "");
  } else {
   if (i == 0) {
    printf("..... %2"PRIu64": %016" PRIx64 "\n"
           "..... %2"PRIu64": %016" PRIx64 "\n",
      i, e->to, i+1, e->from);
   } else {
    printf("..... %2"PRIu64": %016" PRIx64 "\n", i+1, e->from);
   }
  }
 }

 if (branch_stack_cntr) {
  unsigned int br_cntr_width, br_cntr_nr;

  perf_env__find_br_cntr_info(evsel__env(evsel), &br_cntr_nr, &br_cntr_width);
  printf("... branch stack counters: nr:%" PRIu64 " (counter width: %u max counter nr:%u)\n",
   sample->branch_stack->nr, br_cntr_width, br_cntr_nr);
  for (i = 0; i < sample->branch_stack->nr; i++)
   printf("..... %2"PRIu64": %016" PRIx64 "\n", i, branch_stack_cntr[i]);
 }
}

static void regs_dump__printf(u64 mask, u64 *regs, const char *arch)
{
 unsigned rid, i = 0;

 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
  u64 val = regs[i++];

  printf(".... %-5s 0x%016" PRIx64 "\n",
         perf_reg_name(rid, arch), val);
 }
}

static const char *regs_abi[] = {
 [PERF_SAMPLE_REGS_ABI_NONE] = "none",
 [PERF_SAMPLE_REGS_ABI_32] = "32-bit",
 [PERF_SAMPLE_REGS_ABI_64] = "64-bit",
};

static inline const char *regs_dump_abi(struct regs_dump *d)
{
 if (d->abi > PERF_SAMPLE_REGS_ABI_64)
  return "unknown";

 return regs_abi[d->abi];
}

static void regs__printf(const char *type, struct regs_dump *regs, const char *arch)
{
 u64 mask = regs->mask;

 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n",
        type,
        mask,
        regs_dump_abi(regs));

 regs_dump__printf(mask, regs->regs, arch);
}

static void regs_user__printf(struct perf_sample *sample, const char *arch)
{
 struct regs_dump *user_regs;

 if (!sample->user_regs)
  return;

 user_regs = perf_sample__user_regs(sample);

 if (user_regs->regs)
  regs__printf("user", user_regs, arch);
}

static void regs_intr__printf(struct perf_sample *sample, const char *arch)
{
 struct regs_dump *intr_regs;

 if (!sample->intr_regs)
  return;

 intr_regs = perf_sample__intr_regs(sample);

 if (intr_regs->regs)
  regs__printf("intr", intr_regs, arch);
}

static void stack_user__printf(struct stack_dump *dump)
{
 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
        dump->size, dump->offset);
}

static void evlist__print_tstamp(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
{
 u64 sample_type = __evlist__combined_sample_type(evlist);

 if (event->header.type != PERF_RECORD_SAMPLE &&
     !evlist__sample_id_all(evlist)) {
  fputs("-1 -1 ", stdout);
  return;
 }

 if ((sample_type & PERF_SAMPLE_CPU))
  printf("%u ", sample->cpu);

 if (sample_type & PERF_SAMPLE_TIME)
  printf("%" PRIu64 " ", sample->time);
}

static void sample_read__printf(struct perf_sample *sample, u64 read_format)
{
 printf("... sample_read:\n");

 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
  printf("...... time enabled %016" PRIx64 "\n",
         sample->read.time_enabled);

 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
  printf("...... time running %016" PRIx64 "\n",
         sample->read.time_running);

 if (read_format & PERF_FORMAT_GROUP) {
  struct sample_read_value *value = sample->read.group.values;

  printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);

  sample_read_group__for_each(value, sample->read.group.nr, read_format) {
   printf("..... id %016" PRIx64
          ", value %016" PRIx64,
          value->id, value->value);
   if (read_format & PERF_FORMAT_LOST)
    printf(", lost %" PRIu64, value->lost);
   printf("\n");
  }
 } else {
  printf("..... id %016" PRIx64 ", value %016" PRIx64,
   sample->read.one.id, sample->read.one.value);
  if (read_format & PERF_FORMAT_LOST)
   printf(", lost %" PRIu64, sample->read.one.lost);
  printf("\n");
 }
}

static void dump_event(struct evlist *evlist, union perf_event *event,
         u64 file_offset, struct perf_sample *sample,
         const char *file_path)
{
 if (!dump_trace)
  return;

 printf("\n%#" PRIx64 "@%s [%#x]: event: %d\n",
        file_offset, file_path, event->header.size, event->header.type);

 trace_event(event);
 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw)
  evlist->trace_event_sample_raw(evlist, event, sample);

 if (sample)
  evlist__print_tstamp(evlist, event, sample);

 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
        event->header.size, perf_event__name(event->header.type));
}

char *get_page_size_name(u64 size, char *str)
{
 if (!size || !unit_number__scnprintf(str, PAGE_SIZE_NAME_LEN, size))
  snprintf(str, PAGE_SIZE_NAME_LEN, "%s", "N/A");

 return str;
}

static void dump_sample(struct evsel *evsel, union perf_event *event,
   struct perf_sample *sample, const char *arch)
{
 u64 sample_type;
 char str[PAGE_SIZE_NAME_LEN];

 if (!dump_trace)
  return;

 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
        event->header.misc, sample->pid, sample->tid, sample->ip,
        sample->period, sample->addr);

 sample_type = evsel->core.attr.sample_type;

 if (evsel__has_callchain(evsel))
  callchain__printf(evsel, sample);

 if (evsel__has_br_stack(evsel))
  branch_stack__printf(sample, evsel);

 if (sample_type & PERF_SAMPLE_REGS_USER)
  regs_user__printf(sample, arch);

 if (sample_type & PERF_SAMPLE_REGS_INTR)
  regs_intr__printf(sample, arch);

 if (sample_type & PERF_SAMPLE_STACK_USER)
  stack_user__printf(&sample->user_stack);

 if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
  printf("... weight: %" PRIu64 "", sample->weight);
   if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) {
    printf(",0x%"PRIx16"", sample->ins_lat);
    printf(",0x%"PRIx16"", sample->weight3);
   }
  printf("\n");
 }

 if (sample_type & PERF_SAMPLE_DATA_SRC)
  printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);

 if (sample_type & PERF_SAMPLE_PHYS_ADDR)
  printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr);

 if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)
  printf(" .. data page size: %s\n", get_page_size_name(sample->data_page_size, str));

 if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)
  printf(" .. code page size: %s\n", get_page_size_name(sample->code_page_size, str));

 if (sample_type & PERF_SAMPLE_TRANSACTION)
  printf("... transaction: %" PRIx64 "\n", sample->transaction);

 if (sample_type & PERF_SAMPLE_READ)
  sample_read__printf(sample, evsel->core.attr.read_format);
}

static void dump_read(struct evsel *evsel, union perf_event *event)
{
 struct perf_record_read *read_event = &event->read;
 u64 read_format;

 if (!dump_trace)
  return;

 printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid,
        evsel__name(evsel), event->read.value);

 if (!evsel)
  return;

 read_format = evsel->core.attr.read_format;

 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
  printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled);

 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
  printf("... time running : %" PRI_lu64 "\n", read_event->time_running);

 if (read_format & PERF_FORMAT_ID)
  printf("... id : %" PRI_lu64 "\n", read_event->id);

 if (read_format & PERF_FORMAT_LOST)
  printf("... lost : %" PRI_lu64 "\n", read_event->lost);
}

static struct machine *machines__find_for_cpumode(struct machines *machines,
            union perf_event *event,
            struct perf_sample *sample)
{
 if (perf_guest &&
     ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
      (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) {
  u32 pid;

  if (sample->machine_pid)
   pid = sample->machine_pid;
  else if (event->header.type == PERF_RECORD_MMAP
      || event->header.type == PERF_RECORD_MMAP2)
   pid = event->mmap.pid;
  else
   pid = sample->pid;

  /*
 * Guest code machine is created as needed and does not use
 * DEFAULT_GUEST_KERNEL_ID.
 */
  if (symbol_conf.guest_code)
   return machines__findnew(machines, pid);

  return machines__find_guest(machines, pid);
 }

 return &machines->host;
}

static int deliver_sample_value(struct evlist *evlist,
    const struct perf_tool *tool,
    union perf_event *event,
    struct perf_sample *sample,
    struct sample_read_value *v,
    struct machine *machine,
    bool per_thread)
{
 struct perf_sample_id *sid = evlist__id2sid(evlist, v->id);
 struct evsel *evsel;
 u64 *storage = NULL;

 if (sid) {
  storage = perf_sample_id__get_period_storage(sid, sample->tid, per_thread);
 }

 if (storage) {
  sample->id     = v->id;
  sample->period = v->value - *storage;
  *storage       = v->value;
 }

 if (!storage || sid->evsel == NULL) {
  ++evlist->stats.nr_unknown_id;
  return 0;
 }

 /*
 * There's no reason to deliver sample
 * for zero period, bail out.
 */
 if (!sample->period)
  return 0;

 evsel = container_of(sid->evsel, struct evsel, core);
 return tool->sample(tool, event, sample, evsel, machine);
}

static int deliver_sample_group(struct evlist *evlist,
    const struct perf_tool *tool,
    union  perf_event *event,
    struct perf_sample *sample,
    struct machine *machine,
    u64 read_format,
    bool per_thread)
{
 int ret = -EINVAL;
 struct sample_read_value *v = sample->read.group.values;

 if (tool->dont_split_sample_group)
  return deliver_sample_value(evlist, tool, event, sample, v, machine,
         per_thread);

 sample_read_group__for_each(v, sample->read.group.nr, read_format) {
  ret = deliver_sample_value(evlist, tool, event, sample, v,
        machine, per_thread);
  if (ret)
   break;
 }

 return ret;
}

static int evlist__deliver_sample(struct evlist *evlist, const struct perf_tool *tool,
      union  perf_event *event, struct perf_sample *sample,
      struct evsel *evsel, struct machine *machine)
{
 /* We know evsel != NULL. */
 u64 sample_type = evsel->core.attr.sample_type;
 u64 read_format = evsel->core.attr.read_format;
 bool per_thread = perf_evsel__attr_has_per_thread_sample_period(&evsel->core);

 /* Standard sample delivery. */
 if (!(sample_type & PERF_SAMPLE_READ))
  return tool->sample(tool, event, sample, evsel, machine);

 /* For PERF_SAMPLE_READ we have either single or group mode. */
 if (read_format & PERF_FORMAT_GROUP)
  return deliver_sample_group(evlist, tool, event, sample,
         machine, read_format, per_thread);
 else
  return deliver_sample_value(evlist, tool, event, sample,
         &sample->read.one, machine,
         per_thread);
}

static int machines__deliver_event(struct machines *machines,
       struct evlist *evlist,
       union perf_event *event,
       struct perf_sample *sample,
       const struct perf_tool *tool, u64 file_offset,
       const char *file_path)
{
 struct evsel *evsel;
 struct machine *machine;

 dump_event(evlist, event, file_offset, sample, file_path);

 evsel = evlist__id2evsel(evlist, sample->id);

 machine = machines__find_for_cpumode(machines, event, sample);

 switch (event->header.type) {
 case PERF_RECORD_SAMPLE:
  if (evsel == NULL) {
   ++evlist->stats.nr_unknown_id;
   return 0;
  }
  if (machine == NULL) {
   ++evlist->stats.nr_unprocessable_samples;
   dump_sample(evsel, event, sample, perf_env__arch(NULL));
   return 0;
  }
  dump_sample(evsel, event, sample, perf_env__arch(machine->env));
  return evlist__deliver_sample(evlist, tool, event, sample, evsel, machine);
 case PERF_RECORD_MMAP:
  return tool->mmap(tool, event, sample, machine);
 case PERF_RECORD_MMAP2:
  if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT)
   ++evlist->stats.nr_proc_map_timeout;
  return tool->mmap2(tool, event, sample, machine);
 case PERF_RECORD_COMM:
  return tool->comm(tool, event, sample, machine);
 case PERF_RECORD_NAMESPACES:
  return tool->namespaces(tool, event, sample, machine);
 case PERF_RECORD_CGROUP:
  return tool->cgroup(tool, event, sample, machine);
 case PERF_RECORD_FORK:
  return tool->fork(tool, event, sample, machine);
 case PERF_RECORD_EXIT:
  return tool->exit(tool, event, sample, machine);
 case PERF_RECORD_LOST:
  if (tool->lost == perf_event__process_lost)
   evlist->stats.total_lost += event->lost.lost;
  return tool->lost(tool, event, sample, machine);
 case PERF_RECORD_LOST_SAMPLES:
  if (event->header.misc & PERF_RECORD_MISC_LOST_SAMPLES_BPF)
   evlist->stats.total_dropped_samples += event->lost_samples.lost;
  else if (tool->lost_samples == perf_event__process_lost_samples)
   evlist->stats.total_lost_samples += event->lost_samples.lost;
  return tool->lost_samples(tool, event, sample, machine);
 case PERF_RECORD_READ:
  dump_read(evsel, event);
  return tool->read(tool, event, sample, evsel, machine);
 case PERF_RECORD_THROTTLE:
  return tool->throttle(tool, event, sample, machine);
 case PERF_RECORD_UNTHROTTLE:
  return tool->unthrottle(tool, event, sample, machine);
 case PERF_RECORD_AUX:
  if (tool->aux == perf_event__process_aux) {
   if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED)
    evlist->stats.total_aux_lost += 1;
   if (event->aux.flags & PERF_AUX_FLAG_PARTIAL)
    evlist->stats.total_aux_partial += 1;
   if (event->aux.flags & PERF_AUX_FLAG_COLLISION)
    evlist->stats.total_aux_collision += 1;
  }
  return tool->aux(tool, event, sample, machine);
 case PERF_RECORD_ITRACE_START:
  return tool->itrace_start(tool, event, sample, machine);
 case PERF_RECORD_SWITCH:
 case PERF_RECORD_SWITCH_CPU_WIDE:
  return tool->context_switch(tool, event, sample, machine);
 case PERF_RECORD_KSYMBOL:
  return tool->ksymbol(tool, event, sample, machine);
 case PERF_RECORD_BPF_EVENT:
  return tool->bpf(tool, event, sample, machine);
 case PERF_RECORD_TEXT_POKE:
  return tool->text_poke(tool, event, sample, machine);
 case PERF_RECORD_AUX_OUTPUT_HW_ID:
  return tool->aux_output_hw_id(tool, event, sample, machine);
 default:
  ++evlist->stats.nr_unknown_events;
  return -1;
 }
}

static int perf_session__deliver_event(struct perf_session *session,
           union perf_event *event,
           const struct perf_tool *tool,
           u64 file_offset,
           const char *file_path)
{
 struct perf_sample sample;
 int ret;

 perf_sample__init(&sample, /*all=*/false);
 ret = evlist__parse_sample(session->evlist, event, &sample);
 if (ret) {
  pr_err("Can't parse sample, err = %d\n", ret);
  goto out;
 }

 ret = auxtrace__process_event(session, event, &sample, tool);
 if (ret < 0)
  goto out;
 if (ret > 0) {
  ret = 0;
  goto out;
 }

 ret = machines__deliver_event(&session->machines, session->evlist,
          event, &sample, tool, file_offset, file_path);

 if (dump_trace && sample.aux_sample.size)
  auxtrace__dump_auxtrace_sample(session, &sample);
out:
 perf_sample__exit(&sample);
 return ret;
}

static s64 perf_session__process_user_event(struct perf_session *session,
         union perf_event *event,
         u64 file_offset,
         const char *file_path)
{
 struct ordered_events *oe = &session->ordered_events;
 const struct perf_tool *tool = session->tool;
 struct perf_sample sample;
 int fd = perf_data__fd(session->data);
 s64 err;

 perf_sample__init(&sample, /*all=*/true);
 if ((event->header.type != PERF_RECORD_COMPRESSED &&
      event->header.type != PERF_RECORD_COMPRESSED2) ||
     perf_tool__compressed_is_stub(tool))
  dump_event(session->evlist, event, file_offset, &sample, file_path);

 /* These events are processed right away */
 switch (event->header.type) {
 case PERF_RECORD_HEADER_ATTR:
  err = tool->attr(tool, event, &session->evlist);
  if (err == 0) {
   perf_session__set_id_hdr_size(session);
   perf_session__set_comm_exec(session);
  }
  break;
 case PERF_RECORD_EVENT_UPDATE:
  err = tool->event_update(tool, event, &session->evlist);
  break;
 case PERF_RECORD_HEADER_EVENT_TYPE:
  /*
 * Deprecated, but we need to handle it for sake
 * of old data files create in pipe mode.
 */
  err = 0;
  break;
 case PERF_RECORD_HEADER_TRACING_DATA:
  /*
 * Setup for reading amidst mmap, but only when we
 * are in 'file' mode. The 'pipe' fd is in proper
 * place already.
 */
  if (!perf_data__is_pipe(session->data))
   lseek(fd, file_offset, SEEK_SET);
  err = tool->tracing_data(session, event);
  break;
 case PERF_RECORD_HEADER_BUILD_ID:
  err = tool->build_id(session, event);
  break;
 case PERF_RECORD_FINISHED_ROUND:
  err = tool->finished_round(tool, event, oe);
  break;
 case PERF_RECORD_ID_INDEX:
  err = tool->id_index(session, event);
  break;
 case PERF_RECORD_AUXTRACE_INFO:
  err = tool->auxtrace_info(session, event);
  break;
 case PERF_RECORD_AUXTRACE:
  /*
 * Setup for reading amidst mmap, but only when we
 * are in 'file' mode.  The 'pipe' fd is in proper
 * place already.
 */
  if (!perf_data__is_pipe(session->data))
   lseek(fd, file_offset + event->header.size, SEEK_SET);
  err = tool->auxtrace(session, event);
  break;
 case PERF_RECORD_AUXTRACE_ERROR:
  perf_session__auxtrace_error_inc(session, event);
  err = tool->auxtrace_error(session, event);
  break;
 case PERF_RECORD_THREAD_MAP:
  err = tool->thread_map(session, event);
  break;
 case PERF_RECORD_CPU_MAP:
  err = tool->cpu_map(session, event);
  break;
 case PERF_RECORD_STAT_CONFIG:
  err = tool->stat_config(session, event);
  break;
 case PERF_RECORD_STAT:
  err = tool->stat(session, event);
  break;
 case PERF_RECORD_STAT_ROUND:
  err = tool->stat_round(session, event);
  break;
 case PERF_RECORD_TIME_CONV:
  session->time_conv = event->time_conv;
  err = tool->time_conv(session, event);
  break;
 case PERF_RECORD_HEADER_FEATURE:
  err = tool->feature(session, event);
  break;
 case PERF_RECORD_COMPRESSED:
 case PERF_RECORD_COMPRESSED2:
  err = tool->compressed(session, event, file_offset, file_path);
  if (err)
   dump_event(session->evlist, event, file_offset, &sample, file_path);
  break;
 case PERF_RECORD_FINISHED_INIT:
  err = tool->finished_init(session, event);
  break;
 case PERF_RECORD_BPF_METADATA:
  err = tool->bpf_metadata(session, event);
  break;
 default:
  err = -EINVAL;
  break;
 }
 perf_sample__exit(&sample);
 return err;
}

int perf_session__deliver_synth_event(struct perf_session *session,
          union perf_event *event,
          struct perf_sample *sample)
{
 struct evlist *evlist = session->evlist;
 const struct perf_tool *tool = session->tool;

 events_stats__inc(&evlist->stats, event->header.type);

 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
  return perf_session__process_user_event(session, event, 0, NULL);

 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0, NULL);
}

int perf_session__deliver_synth_attr_event(struct perf_session *session,
        const struct perf_event_attr *attr,
        u64 id)
{
 union {
  struct {
   struct perf_record_header_attr attr;
   u64 ids[1];
  } attr_id;
  union perf_event ev;
 } ev = {
  .attr_id.attr.header.type = PERF_RECORD_HEADER_ATTR,
  .attr_id.attr.header.size = sizeof(ev.attr_id),
  .attr_id.ids[0] = id,
 };

 if (attr->size != sizeof(ev.attr_id.attr.attr)) {
  pr_debug("Unexpected perf_event_attr size\n");
  return -EINVAL;
 }
 ev.attr_id.attr.attr = *attr;
 return perf_session__deliver_synth_event(session, &ev.ev, NULL);
}

static void event_swap(union perf_event *event, bool sample_id_all)
{
 perf_event__swap_op swap;

 swap = perf_event__swap_ops[event->header.type];
 if (swap)
  swap(event, sample_id_all);
}

int perf_session__peek_event(struct perf_session *session, off_t file_offset,
        void *buf, size_t buf_sz,
        union perf_event **event_ptr,
        struct perf_sample *sample)
{
 union perf_event *event;
 size_t hdr_sz, rest;
 int fd;

 if (session->one_mmap && !session->header.needs_swap) {
  event = file_offset - session->one_mmap_offset +
   session->one_mmap_addr;
  goto out_parse_sample;
 }

 if (perf_data__is_pipe(session->data))
  return -1;

 fd = perf_data__fd(session->data);
 hdr_sz = sizeof(struct perf_event_header);

 if (buf_sz < hdr_sz)
  return -1;

 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 ||
     readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz)
  return -1;

 event = (union perf_event *)buf;

 if (session->header.needs_swap)
  perf_event_header__bswap(&event->header);

 if (event->header.size < hdr_sz || event->header.size > buf_sz)
  return -1;

 buf += hdr_sz;
 rest = event->header.size - hdr_sz;

 if (readn(fd, buf, rest) != (ssize_t)rest)
  return -1;

 if (session->header.needs_swap)
  event_swap(event, evlist__sample_id_all(session->evlist));

out_parse_sample:

 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START &&
     evlist__parse_sample(session->evlist, event, sample))
  return -1;

 *event_ptr = event;

 return 0;
}

int perf_session__peek_events(struct perf_session *session, u64 offset,
         u64 size, peek_events_cb_t cb, void *data)
{
 u64 max_offset = offset + size;
 char buf[PERF_SAMPLE_MAX_SIZE];
 union perf_event *event;
 int err;

 do {
  err = perf_session__peek_event(session, offset, buf,
            PERF_SAMPLE_MAX_SIZE, &event,
            NULL);
  if (err)
   return err;

  err = cb(session, event, offset, data);
  if (err)
   return err;

  offset += event->header.size;
  if (event->header.type == PERF_RECORD_AUXTRACE)
   offset += event->auxtrace.size;

 } while (offset < max_offset);

 return err;
}

static s64 perf_session__process_event(struct perf_session *session,
           union perf_event *event, u64 file_offset,
           const char *file_path)
{
 struct evlist *evlist = session->evlist;
 const struct perf_tool *tool = session->tool;
 int ret;

 if (session->header.needs_swap)
  event_swap(event, evlist__sample_id_all(evlist));

 if (event->header.type >= PERF_RECORD_HEADER_MAX) {
  /* perf should not support unaligned event, stop here. */
  if (event->header.size % sizeof(u64))
   return -EINVAL;

  /* This perf is outdated and does not support the latest event type. */
  ui__warning("Unsupported header type %u, please consider updating perf.\n",
       event->header.type);
  /* Skip unsupported event by returning its size. */
  return event->header.size;
 }

 events_stats__inc(&evlist->stats, event->header.type);

 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
  return perf_session__process_user_event(session, event, file_offset, file_path);

 if (tool->ordered_events) {
  u64 timestamp = -1ULL;

  ret = evlist__parse_sample_timestamp(evlist, event, ×tamp);
  if (ret && ret != -1)
   return ret;

  ret = perf_session__queue_event(session, event, timestamp, file_offset, file_path);
  if (ret != -ETIME)
   return ret;
 }

 return perf_session__deliver_event(session, event, tool, file_offset, file_path);
}

void perf_event_header__bswap(struct perf_event_header *hdr)
{
 hdr->type = bswap_32(hdr->type);
 hdr->misc = bswap_16(hdr->misc);
 hdr->size = bswap_16(hdr->size);
}

struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
{
 return machine__findnew_thread(&session->machines.host, -1, pid);
}

int perf_session__register_idle_thread(struct perf_session *session)
{
 struct thread *thread = machine__idle_thread(&session->machines.host);

 /* machine__idle_thread() got the thread, so put it */
 thread__put(thread);
 return thread ? 0 : -1;
}

static void
perf_session__warn_order(const struct perf_session *session)
{
 const struct ordered_events *oe = &session->ordered_events;
 struct evsel *evsel;
 bool should_warn = true;

 evlist__for_each_entry(session->evlist, evsel) {
  if (evsel->core.attr.write_backward)
   should_warn = false;
 }

 if (!should_warn)
  return;
 if (oe->nr_unordered_events != 0)
  ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events);
}

static void perf_session__warn_about_errors(const struct perf_session *session)
{
 const struct events_stats *stats = &session->evlist->stats;

 if (session->tool->lost == perf_event__process_lost &&
     stats->nr_events[PERF_RECORD_LOST] != 0) {
  ui__warning("Processed %d events and lost %d chunks!\n\n"
       "Check IO/CPU overload!\n\n",
       stats->nr_events[0],
       stats->nr_events[PERF_RECORD_LOST]);
 }

 if (session->tool->lost_samples == perf_event__process_lost_samples) {
  double drop_rate;

  drop_rate = (double)stats->total_lost_samples /
       (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples);
  if (drop_rate > 0.05) {
   ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n",
        stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples,
        drop_rate * 100.0);
  }
 }

 if (session->tool->aux == perf_event__process_aux &&
     stats->total_aux_lost != 0) {
  ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n",
       stats->total_aux_lost,
       stats->nr_events[PERF_RECORD_AUX]);
 }

 if (session->tool->aux == perf_event__process_aux &&
     stats->total_aux_partial != 0) {
  bool vmm_exclusive = false;

  (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive",
                         &vmm_exclusive);

  ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n"
              "Are you running a KVM guest in the background?%s\n\n",
       stats->total_aux_partial,
       stats->nr_events[PERF_RECORD_AUX],
       vmm_exclusive ?
       "\nReloading kvm_intel module with vmm_exclusive=0\n"
       "will reduce the gaps to only guest's timeslices." :
       "");
 }

 if (session->tool->aux == perf_event__process_aux &&
     stats->total_aux_collision != 0) {
  ui__warning("AUX data detected collision %" PRIu64 " times out of %u!\n\n",
       stats->total_aux_collision,
       stats->nr_events[PERF_RECORD_AUX]);
 }

 if (stats->nr_unknown_events != 0) {
  ui__warning("Found %u unknown events!\n\n"
       "Is this an older tool processing a perf.data "
       "file generated by a more recent tool?\n\n"
       "If that is not the case, consider "
       "reporting to linux-kernel@vger.kernel.org.\n\n",
       stats->nr_unknown_events);
 }

 if (stats->nr_unknown_id != 0) {
  ui__warning("%u samples with id not present in the header\n",
       stats->nr_unknown_id);
 }

 if (stats->nr_invalid_chains != 0) {
  ui__warning("Found invalid callchains!\n\n"
       "%u out of %u events were discarded for this reason.\n\n"
       "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
       stats->nr_invalid_chains,
       stats->nr_events[PERF_RECORD_SAMPLE]);
 }

 if (stats->nr_unprocessable_samples != 0) {
  ui__warning("%u unprocessable samples recorded.\n"
       "Do you have a KVM guest running and not using 'perf kvm'?\n",
       stats->nr_unprocessable_samples);
 }

 perf_session__warn_order(session);

 events_stats__auxtrace_error_warn(stats);

 if (stats->nr_proc_map_timeout != 0) {
  ui__warning("%d map information files for pre-existing threads were\n"
       "not processed, if there are samples for addresses they\n"
       "will not be resolved, you may find out which are these\n"
       "threads by running with -v and redirecting the output\n"
       "to a file.\n"
       "The time limit to process proc map is too short?\n"
       "Increase it by --proc-map-timeout\n",
       stats->nr_proc_map_timeout);
 }
}

static int perf_session__flush_thread_stack(struct thread *thread,
         void *p __maybe_unused)
{
 return thread_stack__flush(thread);
}

static int perf_session__flush_thread_stacks(struct perf_session *session)
{
 return machines__for_each_thread(&session->machines,
      perf_session__flush_thread_stack,
      NULL);
}

volatile sig_atomic_t session_done;

static int __perf_session__process_decomp_events(struct perf_session *session);

static int __perf_session__process_pipe_events(struct perf_session *session)
{
 struct ordered_events *oe = &session->ordered_events;
 const struct perf_tool *tool = session->tool;
 struct ui_progress prog;
 union perf_event *event;
 uint32_t size, cur_size = 0;
 void *buf = NULL;
 s64 skip = 0;
 u64 head;
 ssize_t err;
 void *p;
 bool update_prog = false;

 /*
 * If it's from a file saving pipe data (by redirection), it would have
 * a file name other than "-".  Then we can get the total size and show
 * the progress.
 */
 if (strcmp(session->data->path, "-") && session->data->file.size) {
  ui_progress__init_size(&prog, session->data->file.size,
           "Processing events...");
  update_prog = true;
 }

 head = 0;
 cur_size = sizeof(union perf_event);

 buf = malloc(cur_size);
 if (!buf)
  return -errno;
 ordered_events__set_copy_on_queue(oe, true);
more:
 event = buf;
 err = perf_data__read(session->data, event,
         sizeof(struct perf_event_header));
 if (err <= 0) {
  if (err == 0)
   goto done;

  pr_err("failed to read event header\n");
  goto out_err;
 }

 if (session->header.needs_swap)
  perf_event_header__bswap(&event->header);

 size = event->header.size;
 if (size < sizeof(struct perf_event_header)) {
  pr_err("bad event header size\n");
  goto out_err;
 }

 if (size > cur_size) {
  void *new = realloc(buf, size);
  if (!new) {
   pr_err("failed to allocate memory to read event\n");
   goto out_err;
  }
  buf = new;
  cur_size = size;
  event = buf;
 }
 p = event;
 p += sizeof(struct perf_event_header);

 if (size - sizeof(struct perf_event_header)) {
  err = perf_data__read(session->data, p,
          size - sizeof(struct perf_event_header));
  if (err <= 0) {
   if (err == 0) {
    pr_err("unexpected end of event stream\n");
    goto done;
   }

   pr_err("failed to read event data\n");
   goto out_err;
  }
 }

 if ((skip = perf_session__process_event(session, event, head, "pipe")) < 0) {
  pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
         head, event->header.size, event->header.type);
  err = -EINVAL;
  goto out_err;
 }

 head += size;

 if (skip > 0)
  head += skip;

 err = __perf_session__process_decomp_events(session);
 if (err)
  goto out_err;

 if (update_prog)
  ui_progress__update(&prog, size);

 if (!session_done())
  goto more;
done:
 /* do the final flush for ordered samples */
 err = ordered_events__flush(oe, OE_FLUSH__FINAL);
 if (err)
  goto out_err;
 err = auxtrace__flush_events(session, tool);
 if (err)
  goto out_err;
 err = perf_session__flush_thread_stacks(session);
out_err:
 free(buf);
 if (update_prog)
  ui_progress__finish();
 if (!tool->no_warn)
  perf_session__warn_about_errors(session);
 ordered_events__free(&session->ordered_events);
 auxtrace__free_events(session);
 return err;
}

static union perf_event *
prefetch_event(char *buf, u64 head, size_t mmap_size,
        bool needs_swap, union perf_event *error)
{
 union perf_event *event;
 u16 event_size;

 /*
 * Ensure we have enough space remaining to read
 * the size of the event in the headers.
 */
 if (head + sizeof(event->header) > mmap_size)
  return NULL;

 event = (union perf_event *)(buf + head);
 if (needs_swap)
  perf_event_header__bswap(&event->header);

 event_size = event->header.size;
 if (head + event_size <= mmap_size)
  return event;

 /* We're not fetching the event so swap back again */
 if (needs_swap)
  perf_event_header__bswap(&event->header);

 /* Check if the event fits into the next mmapped buf. */
 if (event_size <= mmap_size - head % page_size) {
  /* Remap buf and fetch again. */
  return NULL;
 }

 /* Invalid input. Event size should never exceed mmap_size. */
 pr_debug("%s: head=%#" PRIx64 " event->header.size=%#x, mmap_size=%#zx:"
   " fuzzed or compressed perf.data?\n", __func__, head, event_size, mmap_size);

 return error;
}

static union perf_event *
fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap)
{
 return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL));
}

static union perf_event *
fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap)
{
 return prefetch_event(buf, head, mmap_size, needs_swap, NULL);
}

static int __perf_session__process_decomp_events(struct perf_session *session)
{
 s64 skip;
 u64 size;
 struct decomp *decomp = session->active_decomp->decomp_last;

 if (!decomp)
  return 0;

 while (decomp->head < decomp->size && !session_done()) {
  union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data,
            session->header.needs_swap);

  if (!event)
   break;

  size = event->header.size;

  if (size < sizeof(struct perf_event_header) ||
      (skip = perf_session__process_event(session, event, decomp->file_pos,
       decomp->file_path)) < 0) {
   pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
    decomp->file_pos + decomp->head, event->header.size, event->header.type);
   return -EINVAL;
  }

  if (skip)
   size += skip;

  decomp->head += size;
 }

 return 0;
}

/*
 * On 64bit we can mmap the data file in one go. No need for tiny mmap
 * slices. On 32bit we use 32MB.
 */
#if BITS_PER_LONG == 64
#define MMAP_SIZE ULLONG_MAX
#define NUM_MMAPS 1
#else
#define MMAP_SIZE (32 * 1024 * 1024ULL)
#define NUM_MMAPS 128
#endif

struct reader;

typedef s64 (*reader_cb_t)(struct perf_session *session,
      union perf_event *event,
      u64 file_offset,
      const char *file_path);

struct reader {
 int   fd;
 const char  *path;
 u64   data_size;
 u64   data_offset;
 reader_cb_t  process;
 bool   in_place_update;
 char   *mmaps[NUM_MMAPS];
 size_t   mmap_size;
 int   mmap_idx;
 char   *mmap_cur;
 u64   file_pos;
 u64   file_offset;
 u64   head;
 u64   size;
 bool   done;
 struct zstd_data   zstd_data;
 struct decomp_data decomp_data;
};

static int
reader__init(struct reader *rd, bool *one_mmap)
{
 u64 data_size = rd->data_size;
 char **mmaps = rd->mmaps;

 rd->head = rd->data_offset;
 data_size += rd->data_offset;

 rd->mmap_size = MMAP_SIZE;
 if (rd->mmap_size > data_size) {
  rd->mmap_size = data_size;
  if (one_mmap)
   *one_mmap = true;
 }

 memset(mmaps, 0, sizeof(rd->mmaps));

 if (zstd_init(&rd->zstd_data, 0))
  return -1;
 rd->decomp_data.zstd_decomp = &rd->zstd_data;

 return 0;
}

static void
reader__release_decomp(struct reader *rd)
{
 perf_decomp__release_events(rd->decomp_data.decomp);
 zstd_fini(&rd->zstd_data);
}

static int
reader__mmap(struct reader *rd, struct perf_session *session)
{
 int mmap_prot, mmap_flags;
 char *buf, **mmaps = rd->mmaps;
 u64 page_offset;

 mmap_prot  = PROT_READ;
 mmap_flags = MAP_SHARED;

 if (rd->in_place_update) {
  mmap_prot  |= PROT_WRITE;
 } else if (session->header.needs_swap) {
  mmap_prot  |= PROT_WRITE;
  mmap_flags = MAP_PRIVATE;
 }

 if (mmaps[rd->mmap_idx]) {
  munmap(mmaps[rd->mmap_idx], rd->mmap_size);
  mmaps[rd->mmap_idx] = NULL;
 }

 page_offset = page_size * (rd->head / page_size);
 rd->file_offset += page_offset;
 rd->head -= page_offset;

 buf = mmap(NULL, rd->mmap_size, mmap_prot, mmap_flags, rd->fd,
     rd->file_offset);
 if (buf == MAP_FAILED) {
  pr_err("failed to mmap file\n");
  return -errno;
 }
 mmaps[rd->mmap_idx] = rd->mmap_cur = buf;
 rd->mmap_idx = (rd->mmap_idx + 1) & (ARRAY_SIZE(rd->mmaps) - 1);
 rd->file_pos = rd->file_offset + rd->head;
 if (session->one_mmap) {
  session->one_mmap_addr = buf;
  session->one_mmap_offset = rd->file_offset;
 }

 return 0;
}

enum {
 READER_OK,
 READER_NODATA,
};

static int
reader__read_event(struct reader *rd, struct perf_session *session,
     struct ui_progress *prog)
{
 u64 size;
 int err = READER_OK;
 union perf_event *event;
 s64 skip;

 event = fetch_mmaped_event(rd->head, rd->mmap_size, rd->mmap_cur,
       session->header.needs_swap);
 if (IS_ERR(event))
  return PTR_ERR(event);

 if (!event)
  return READER_NODATA;

 size = event->header.size;

 skip = -EINVAL;

 if (size < sizeof(struct perf_event_header) ||
     (skip = rd->process(session, event, rd->file_pos, rd->path)) < 0) {
  pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n",
         rd->file_offset + rd->head, event->header.size,
         event->header.type, strerror(-skip));
  err = skip;
  goto out;
 }

 if (skip)
  size += skip;

 rd->size += size;
 rd->head += size;
 rd->file_pos += size;

 err = __perf_session__process_decomp_events(session);
 if (err)
  goto out;

 ui_progress__update(prog, size);

out:
 return err;
}

static inline bool
reader__eof(struct reader *rd)
{
 return (rd->file_pos >= rd->data_size + rd->data_offset);
}

static int
reader__process_events(struct reader *rd, struct perf_session *session,
         struct ui_progress *prog)
{
 int err;

 err = reader__init(rd, &session->one_mmap);
 if (err)
  goto out;

 session->active_decomp = &rd->decomp_data;

remap:
 err = reader__mmap(rd, session);
 if (err)
  goto out;

more:
 err = reader__read_event(rd, session, prog);
 if (err < 0)
  goto out;
 else if (err == READER_NODATA)
  goto remap;

 if (session_done())
  goto out;

 if (!reader__eof(rd))
  goto more;

out:
 session->active_decomp = &session->decomp_data;
 return err;
}

static s64 process_simple(struct perf_session *session,
     union perf_event *event,
     u64 file_offset,
     const char *file_path)
{
 return perf_session__process_event(session, event, file_offset, file_path);
}

static int __perf_session__process_events(struct perf_session *session)
{
 struct reader rd = {
  .fd  = perf_data__fd(session->data),
  .path  = session->data->file.path,
  .data_size = session->header.data_size,
  .data_offset = session->header.data_offset,
  .process = process_simple,
  .in_place_update = session->data->in_place_update,
 };
 struct ordered_events *oe = &session->ordered_events;
 const struct perf_tool *tool = session->tool;
 struct ui_progress prog;
 int err;

 if (rd.data_size == 0)
  return -1;

 ui_progress__init_size(&prog, rd.data_size, "Processing events...");

 err = reader__process_events(&rd, session, &prog);
 if (err)
  goto out_err;
 /* do the final flush for ordered samples */
 err = ordered_events__flush(oe, OE_FLUSH__FINAL);
 if (err)
  goto out_err;
 err = auxtrace__flush_events(session, tool);
 if (err)
  goto out_err;
 err = perf_session__flush_thread_stacks(session);
out_err:
 ui_progress__finish();
 if (!tool->no_warn)
  perf_session__warn_about_errors(session);
 /*
 * We may switching perf.data output, make ordered_events
 * reusable.
 */
 ordered_events__reinit(&session->ordered_events);
 auxtrace__free_events(session);
 reader__release_decomp(&rd);
 session->one_mmap = false;
 return err;
}

/*
 * Processing 2 MB of data from each reader in sequence,
 * because that's the way the ordered events sorting works
 * most efficiently.
 */
#define READER_MAX_SIZE (2 * 1024 * 1024)

/*
 * This function reads, merge and process directory data.
--> --------------------

--> maximum size reached

--> --------------------