dl-profile.c

Newlib 嵌入式 C库 标准实现代码

      char buf[400];
      _dl_error_printf ("%s: cannot open file: %s\n", filename,
			__strerror_r (errno, buf, sizeof buf));
      return;
    }

  if (fstat64 (fd, &st) < 0 || !S_ISREG (st.st_mode))
    {
      /* Not stat'able or not a regular file => don't use it.  */
      char buf[400];
      int errnum = errno;
      __close (fd);
      _dl_error_printf ("%s: cannot stat file: %s\n", filename,
			__strerror_r (errnum, buf, sizeof buf));
      return;
    }

  /* Test the size.  If it does not match what we expect from the size
     values in the map MAP we don't use it and warn the user.  */
  if (st.st_size == 0)
    {
      /* We have to create the file.  */
      char buf[_dl_pagesize];

      memset (buf, '\0', _dl_pagesize);

      if (__lseek (fd, expected_size & ~(_dl_pagesize - 1), SEEK_SET) == -1)
	{
	  char buf[400];
	  int errnum;
	cannot_create:
	  errnum = errno;
	  __close (fd);
	  _dl_error_printf ("%s: cannot create file: %s\n", filename,
			    __strerror_r (errnum, buf, sizeof buf));
	  return;
	}

      if (TEMP_FAILURE_RETRY (__libc_write (fd, buf, (expected_size
						      & (_dl_pagesize - 1))))
	  < 0)
	goto cannot_create;
    }
  else if (st.st_size != expected_size)
    {
      __close (fd);
    wrong_format:

      if (addr != NULL)
	__munmap ((void *) addr, expected_size);

      _dl_error_printf ("%s: file is no correct profile data file for `%s'\n",
			filename, _dl_profile);
      return;
    }

  addr = (struct gmon_hdr *) __mmap (NULL, expected_size, PROT_READ|PROT_WRITE,
				     MAP_SHARED|MAP_FILE, fd, 0);
  if (addr == (struct gmon_hdr *) MAP_FAILED)
    {
      char buf[400];
      int errnum = errno;
      __close (fd);
      _dl_error_printf ("%s: cannot map file: %s\n", filename,
			__strerror_r (errnum, buf, sizeof buf));
      return;
    }

  /* We don't need the file desriptor anymore.  */
  __close (fd);

  /* Pointer to data after the header.  */
  hist = (char *) (addr + 1);
  kcount = (uint16_t *) ((char *) hist + sizeof (uint32_t)
			 + sizeof (struct gmon_hist_hdr));

  /* Compute pointer to array of the arc information.  */
  narcsp = (uint32_t *) ((char *) kcount + kcountsize + sizeof (uint32_t));
  data = (struct here_cg_arc_record *) ((char *) narcsp + sizeof (uint32_t));

  if (st.st_size == 0)
    {
      /* Create the signature.  */
      memcpy (addr, &gmon_hdr, sizeof (struct gmon_hdr));

      *(uint32_t *) hist = GMON_TAG_TIME_HIST;
      memcpy (hist + sizeof (uint32_t), &hist_hdr,
	      sizeof (struct gmon_hist_hdr));

      narcsp[-1] = GMON_TAG_CG_ARC;
    }
  else
    {
      /* Test the signature in the file.  */
      if (memcmp (addr, &gmon_hdr, sizeof (struct gmon_hdr)) != 0
	  || *(uint32_t *) hist != GMON_TAG_TIME_HIST
	  || memcmp (hist + sizeof (uint32_t), &hist_hdr,
		     sizeof (struct gmon_hist_hdr)) != 0
	  || narcsp[-1] != GMON_TAG_CG_ARC)
	goto wrong_format;
    }

  /* Allocate memory for the froms data and the pointer to the tos records.  */
  tos = (uint16_t *) calloc (tossize + fromssize, 1);
  if (tos == NULL)
    {
      __munmap ((void *) addr, expected_size);
      _dl_fatal_printf ("Out of memory while initializing profiler\n");
      /* NOTREACHED */
    }

  froms = (struct here_fromstruct *) ((char *) tos + tossize);
  fromidx = 0;

  /* Now we have to process all the arc count entries.  BTW: it is
     not critical whether the *NARCSP value changes meanwhile.  Before
     we enter a new entry in to toset we will check that everything is
     available in TOS.  This happens in _dl_mcount.

     Loading the entries in reverse order should help to get the most
     frequently used entries at the front of the list.  */
  for (idx = narcs = MIN (*narcsp, fromlimit); idx > 0; )
    {
      size_t to_index;
      size_t newfromidx;
      --idx;
      to_index = (data[idx].self_pc / (hashfraction * sizeof (*tos)));
      newfromidx = fromidx++;
      froms[newfromidx].here = &data[idx];
      froms[newfromidx].link = tos[to_index];
      tos[to_index] = newfromidx;
    }

  /* Setup counting data.  */
  if (kcountsize < highpc - lowpc)
    {
#if 0
      s_scale = ((double) kcountsize / (highpc - lowpc)) * SCALE_1_TO_1;
#else
      size_t range = highpc - lowpc;
      size_t quot = range / kcountsize;

      if (quot >= SCALE_1_TO_1)
	s_scale = 1;
      else if (quot >= SCALE_1_TO_1 / 256)
	s_scale = SCALE_1_TO_1 / quot;
      else if (range > ULONG_MAX / 256)
	s_scale = (SCALE_1_TO_1 * 256) / (range / (kcountsize / 256));
      else
	s_scale = (SCALE_1_TO_1 * 256) / ((range * 256) / kcountsize);
#endif
    }
  else
    s_scale = SCALE_1_TO_1;

  /* Start the profiler.  */
  profil ((void *) kcount, kcountsize, lowpc, s_scale);

  /* Turn on profiling.  */
  running = 1;
}


void
_dl_mcount (ElfW(Addr) frompc, ElfW(Addr) selfpc)
{
  volatile uint16_t *topcindex;
  size_t i, fromindex;
  struct here_fromstruct *fromp;

  if (! running)
    return;

  /* Compute relative addresses.  The shared object can be loaded at
     any address.  The value of frompc could be anything.  We cannot
     restrict it in any way, just set to a fixed value (0) in case it
     is outside the allowed range.  These calls show up as calls from
     <external> in the gprof output.  */
  frompc -= lowpc;
  if (frompc >= textsize)
    frompc = 0;
  selfpc -= lowpc;
  if (selfpc >= textsize)
    goto done;

  /* Getting here we now have to find out whether the location was
     already used.  If yes we are lucky and only have to increment a
     counter (this also has to be atomic).  If the entry is new things
     are getting complicated...  */

  /* Avoid integer divide if possible.  */
  if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
    i = selfpc >> log_hashfraction;
  else
    i = selfpc / (hashfraction * sizeof (*tos));

  topcindex = &tos[i];
  fromindex = *topcindex;

  if (fromindex == 0)
    goto check_new_or_add;

  fromp = &froms[fromindex];

  /* We have to look through the chain of arcs whether there is already
     an entry for our arc.  */
  while (fromp->here->from_pc != frompc)
    {
      if (fromp->link != 0)
	do
	  fromp = &froms[fromp->link];
	while (fromp->link != 0 && fromp->here->from_pc != frompc);

      if (fromp->here->from_pc != frompc)
	{
	  topcindex = &fromp->link;

	check_new_or_add:
	  /* Our entry is not among the entries we read so far from the
	     data file.  Now see whether we have to update the list.  */
	  while (narcs != *narcsp && narcs < fromlimit)
	    {
	      size_t to_index;
	      size_t newfromidx;
	      to_index = (data[narcs].self_pc
			  / (hashfraction * sizeof (*tos)));
	      newfromidx = exchange_and_add (&fromidx, 1) + 1;
	      froms[newfromidx].here = &data[narcs];
	      froms[newfromidx].link = tos[to_index];
	      tos[to_index] = newfromidx;
	      atomic_add (&narcs, 1);
	    }

	  /* If we still have no entry stop searching and insert.  */
	  if (*topcindex == 0)
	    {
	      uint32_t newarc = exchange_and_add (narcsp, 1);

	      /* In rare cases it could happen that all entries in FROMS are
		 occupied.  So we cannot count this anymore.  */
	      if (newarc >= fromlimit)
		goto done;

	      *topcindex = exchange_and_add (&fromidx, 1) + 1;
	      fromp = &froms[*topcindex];

	      fromp->here = &data[newarc];
	      data[newarc].from_pc = frompc;
	      data[newarc].self_pc = selfpc;
	      data[newarc].count = 0;
	      fromp->link = 0;
	      atomic_add (&narcs, 1);

	      break;
	    }

	  fromp = &froms[*topcindex];
	}
      else
	/* Found in.  */
	break;
    }

  /* Increment the counter.  */
  atomic_add (&fromp->here->count, 1);

 done:
  ;
}