dl-profile.c

Newlib 嵌入式 C库 标准实现代码

/* Profiling of shared libraries.
   Copyright (C) 1997, 1998, 1999, 2000, 2001 Free Software Foundation, Inc.
   This file is part of the GNU C Library.
   Contributed by Ulrich Drepper <[e-mail]>, 1997.
   Based on the BSD mcount implementation.

   The GNU C Library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Lesser General Public
   License as published by the Free Software Foundation; either
   version 2.1 of the License, or (at your option) any later version.

   The GNU C Library is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   Lesser General Public License for more details.

   You should have received a copy of the GNU Lesser General Public
   License along with the GNU C Library; if not, write to the Free
   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
   02111-1307 USA.  */

#include <errno.h>
#include <fcntl.h>
#include <limits.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <ldsodefs.h>
#include <sys/gmon.h>
#include <sys/gmon_out.h>
#include <sys/mman.h>
#include <sys/param.h>
#include <sys/stat.h>
#include <atomicity.h>
#include <config.h>

/* The LD_PROFILE feature has to be implemented different to the
   normal profiling using the gmon/ functions.  The problem is that an
   arbitrary amount of processes simulataneously can be run using
   profiling and all write the results in the same file.  To provide
   this mechanism one could implement a complicated mechanism to merge
   the content of two profiling runs or one could extend the file
   format to allow more than one data set.  For the second solution we
   would have the problem that the file can grow in size beyond any
   limit and both solutions have the problem that the concurrency of
   writing the results is a big problem.

   Another much simpler method is to use mmap to map the same file in
   all using programs and modify the data in the mmap'ed area and so
   also automatically on the disk.  Using the MAP_SHARED option of
   mmap(2) this can be done without big problems in more than one
   file.

   This approach is very different from the normal profiling.  We have
   to use the profiling data in exactly the way they are expected to
   be written to disk.  But the normal format used by gprof is not usable
   to do this.  It is optimized for size.  It writes the tags as single
   bytes but this means that the following 32/64 bit values are
   unaligned.

   Therefore we use a new format.  This will look like this

					0  1  2  3	<- byte is 32 bit word
	0000				g  m  o  n
	0004				*version*	<- GMON_SHOBJ_VERSION
	0008				00 00 00 00
	000c				00 00 00 00
	0010				00 00 00 00

	0014				*tag*		<- GMON_TAG_TIME_HIST
	0018				?? ?? ?? ??
					?? ?? ?? ??	<- 32/64 bit LowPC
	0018+A				?? ?? ?? ??
					?? ?? ?? ??	<- 32/64 bit HighPC
	0018+2*A			*histsize*
	001c+2*A			*profrate*
	0020+2*A			s  e  c  o
	0024+2*A			n  d  s  \0
	0028+2*A			\0 \0 \0 \0
	002c+2*A			\0 \0 \0
	002f+2*A			s

	0030+2*A			?? ?? ?? ??	<- Count data
	...				...
	0030+2*A+K			?? ?? ?? ??

	0030+2*A+K			*tag*		<- GMON_TAG_CG_ARC
	0034+2*A+K			*lastused*
	0038+2*A+K			?? ?? ?? ??
					?? ?? ?? ??	<- FromPC#1
	0038+3*A+K			?? ?? ?? ??
					?? ?? ?? ??	<- ToPC#1
	0038+4*A+K			?? ?? ?? ??	<- Count#1
	...				...		   ...
	0038+(2*(CN-1)+2)*A+(CN-1)*4+K	?? ?? ?? ??
					?? ?? ?? ??	<- FromPC#CGN
	0038+(2*(CN-1)+3)*A+(CN-1)*4+K	?? ?? ?? ??
					?? ?? ?? ??	<- ToPC#CGN
	0038+(2*CN+2)*A+(CN-1)*4+K	?? ?? ?? ??	<- Count#CGN

   We put (for now?) no basic block information in the file since this would
   introduce rase conditions among all the processes who want to write them.

   `K' is the number of count entries which is computed as

 		textsize / HISTFRACTION

   `CG' in the above table is the number of call graph arcs.  Normally,
   the table is sparse and the profiling code writes out only the those
   entries which are really used in the program run.  But since we must
   not extend this table (the profiling file) we'll keep them all here.
   So CN can be executed in advance as

		MINARCS <= textsize*(ARCDENSITY/100) <= MAXARCS

   Now the remaining question is: how to build the data structures we can
   work with from this data.  We need the from set and must associate the
   froms with all the associated tos.  We will do this by constructing this
   data structures at the program start.  To do this we'll simply visit all
   entries in the call graph table and add it to the appropriate list.  */

extern int __profile_frequency (void);

/* We define a special type to address the elements of the arc table.
   This is basically the `gmon_cg_arc_record' format but it includes
   the room for the tag and it uses real types.  */
struct here_cg_arc_record
  {
    uintptr_t from_pc;
    uintptr_t self_pc;
    uint32_t count;
  } __attribute__ ((packed));

static struct here_cg_arc_record *data;

/* Nonzero if profiling is under way.  */
static int running;

/* This is the number of entry which have been incorporated in the toset.  */
static uint32_t narcs;
/* This is a pointer to the object representing the number of entries
   currently in the mmaped file.  At no point of time this has to be the
   same as NARCS.  If it is equal all entries from the file are in our
   lists.  */
static volatile uint32_t *narcsp;

static volatile uint16_t *kcount;
static size_t kcountsize;

struct here_fromstruct
  {
    struct here_cg_arc_record volatile *here;
    uint16_t link;
  };

static volatile uint16_t *tos;

static struct here_fromstruct *froms;
static uint32_t fromlimit;
static volatile uint32_t fromidx;

static uintptr_t lowpc;
static size_t textsize;
static unsigned int hashfraction;
static unsigned int log_hashfraction;



/* Set up profiling data to profile object desribed by MAP.  The output
   file is found (or created) in OUTPUT_DIR.  */
void
internal_function
_dl_start_profile (struct link_map *map, const char *output_dir)
{
  char *filename;
  int fd;
  struct stat64 st;
  const ElfW(Phdr) *ph;
  ElfW(Addr) mapstart = ~((ElfW(Addr)) 0);
  ElfW(Addr) mapend = 0;
  struct gmon_hdr gmon_hdr;
  struct gmon_hist_hdr hist_hdr;
  char *hist, *cp, *tmp;
  size_t idx;
  size_t tossize;
  size_t fromssize;
  uintptr_t highpc;
  struct gmon_hdr *addr = NULL;
  off_t expected_size;
  /* See profil(2) where this is described.  */
  int s_scale;
#define SCALE_1_TO_1	0x10000L

  /* Compute the size of the sections which contain program code.  */
  for (ph = map->l_phdr; ph < &map->l_phdr[map->l_phnum]; ++ph)
    if (ph->p_type == PT_LOAD && (ph->p_flags & PF_X))
      {
	ElfW(Addr) start = (ph->p_vaddr & ~(_dl_pagesize - 1));
	ElfW(Addr) end = ((ph->p_vaddr + ph->p_memsz + _dl_pagesize - 1)
			  & ~(_dl_pagesize - 1));

	if (start < mapstart)
	  mapstart = start;
	if (end > mapend)
	  mapend = end;
      }

  /* Now we can compute the size of the profiling data.  This is done
     with the same formulars as in `monstartup' (see gmon.c).  */
  running = 0;
  lowpc = ROUNDDOWN (mapstart + map->l_addr,
		     HISTFRACTION * sizeof (HISTCOUNTER));
  highpc = ROUNDUP (mapend + map->l_addr,
		    HISTFRACTION * sizeof (HISTCOUNTER));
  textsize = highpc - lowpc;
  kcountsize = textsize / HISTFRACTION;
  hashfraction = HASHFRACTION;
  if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
    /* If HASHFRACTION is a power of two, mcount can use shifting
       instead of integer division.  Precompute shift amount.  */
    log_hashfraction = ffs (hashfraction * sizeof (*froms)) - 1;
  else
    log_hashfraction = -1;
  tossize = textsize / HASHFRACTION;
  fromlimit = textsize * ARCDENSITY / 100;
  if (fromlimit < MINARCS)
    fromlimit = MINARCS;
  if (fromlimit > MAXARCS)
    fromlimit = MAXARCS;
  fromssize = fromlimit * sizeof (struct here_fromstruct);

  expected_size = (sizeof (struct gmon_hdr)
		   + 4 + sizeof (struct gmon_hist_hdr) + kcountsize
		   + 4 + 4 + fromssize * sizeof (struct here_cg_arc_record));

  /* Create the gmon_hdr we expect or write.  */
  memset (&gmon_hdr, '\0', sizeof (struct gmon_hdr));
  memcpy (&gmon_hdr.cookie[0], GMON_MAGIC, sizeof (gmon_hdr.cookie));
  *(int32_t *) gmon_hdr.version = GMON_SHOBJ_VERSION;

  /* Create the hist_hdr we expect or write.  */
  *(char **) hist_hdr.low_pc = (char *) mapstart;
  *(char **) hist_hdr.high_pc = (char *) mapend;
  *(int32_t *) hist_hdr.hist_size = kcountsize / sizeof (HISTCOUNTER);
  *(int32_t *) hist_hdr.prof_rate = __profile_frequency ();
  strncpy (hist_hdr.dimen, "seconds", sizeof (hist_hdr.dimen));
  hist_hdr.dimen_abbrev = 's';

  /* First determine the output name.  We write in the directory
     OUTPUT_DIR and the name is composed from the shared objects
     soname (or the file name) and the ending ".profile".  */
  filename = (char *) alloca (strlen (output_dir) + 1 + strlen (_dl_profile)
			      + sizeof ".profile");
  cp = strcpy (filename, output_dir);
  cp += strlen (output_dir);
  *cp++ = '/';
  tmp = strcpy (cp, _dl_profile);
  tmp += strlen (_dl_profile);
  strcpy (tmp, ".profile");

#ifdef O_NOFOLLOW
# define EXTRA_FLAGS | O_NOFOLLOW
#else
# define EXTRA_FLAGS
#endif
  fd = __open (filename, O_RDWR | O_CREAT EXTRA_FLAGS);
  if (fd == -1)
    {
      /* We cannot write the profiling data so don't do anything.  */