1 /* Profiling of shared libraries.
2 Copyright (C) 1997, 1998 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4 Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
5 Based on the BSD mcount implementation.
7 The GNU C Library is free software; you can redistribute it and/or
8 modify it under the terms of the GNU Library General Public License as
9 published by the Free Software Foundation; either version 2 of the
10 License, or (at your option) any later version.
12 The GNU C Library is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 Library General Public License for more details.
17 You should have received a copy of the GNU Library General Public
18 License along with the GNU C Library; see the file COPYING.LIB. If not,
19 write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
30 #include <elf/ldsodefs.h>
32 #include <sys/gmon_out.h>
34 #include <sys/param.h>
36 #include <atomicity.h>
38 /* We have prototype anywhere. */
39 extern ssize_t __libc_write
__P ((int __fd
, __const __ptr_t __buf
,
42 /* The LD_PROFILE feature has to be implemented different to the
43 normal profiling using the gmon/ functions. The problem is that an
44 arbitrary amount of processes simulataneously can be run using
45 profiling and all write the results in the same file. To provide
46 this mechanism one could implement a complicated mechanism to merge
47 the content of two profiling runs or one could extend the file
48 format to allow more than one data set. For the second solution we
49 would have the problem that the file can grow in size beyond any
50 limit and both solutions have the problem that the concurrency of
51 writing the results is a big problem.
53 Another much simpler method is to use mmap to map the same file in
54 all using programs and modify the data in the mmap'ed area and so
55 also automatically on the disk. Using the MAP_SHARED option of
56 mmap(2) this can be done without big problems in more than one
59 This approach is very different from the normal profiling. We have
60 to use the profiling data in exactly the way they are expected to
61 be written to disk. But the normal format used by gprof is not usable
62 to do this. It is optimized for size. It writes the tags as single
63 bytes but this means that the following 32/64 bit values are
66 Therefore we use a new format. This will look like this
68 0 1 2 3 <- byte is 32 bit word
70 0004 *version* <- GMON_SHOBJ_VERSION
75 0014 *tag* <- GMON_TAG_TIME_HIST
77 ?? ?? ?? ?? <- 32/64 bit LowPC
79 ?? ?? ?? ?? <- 32/64 bit HighPC
88 0030+2*A ?? ?? ?? ?? <- Count data
90 0030+2*A+K ?? ?? ?? ??
92 0030+2*A+K *tag* <- GMON_TAG_CG_ARC
94 0038+2*A+K ?? ?? ?? ??
95 ?? ?? ?? ?? <- FromPC#1
96 0038+3*A+K ?? ?? ?? ??
98 0038+4*A+K ?? ?? ?? ?? <- Count#1
100 0038+(2*(CN-1)+2)*A+(CN-1)*4+K ?? ?? ?? ??
101 ?? ?? ?? ?? <- FromPC#CGN
102 0038+(2*(CN-1)+3)*A+(CN-1)*4+K ?? ?? ?? ??
103 ?? ?? ?? ?? <- ToPC#CGN
104 0038+(2*CN+2)*A+(CN-1)*4+K ?? ?? ?? ?? <- Count#CGN
106 We put (for now?) no basic block information in the file since this would
107 introduce rase conditions among all the processes who want to write them.
109 `K' is the number of count entries which is computed as
111 textsize / HISTFRACTION
113 `CG' in the above table is the number of call graph arcs. Normally,
114 the table is sparse and the profiling code writes out only the those
115 entries which are really used in the program run. But since we must
116 not extend this table (the profiling file) we'll keep them all here.
117 So CN can be executed in advance as
119 MINARCS <= textsize*(ARCDENSITY/100) <= MAXARCS
121 Now the remaining question is: how to build the data structures we can
122 work with from this data. We need the from set and must associate the
123 froms with all the associated tos. We will do this by constructing this
124 data structures at the program start. To do this we'll simply visit all
125 entries in the call graph table and add it to the appropriate list. */
127 extern int __profile_frequency
__P ((void));
129 /* We define a special type to address the elements of the arc table.
130 This is basically the `gmon_cg_arc_record' format but it includes
131 the room for the tag and it uses real types. */
132 struct here_cg_arc_record
137 } __attribute__ ((packed
));
139 static struct here_cg_arc_record
*data
;
141 /* This is the number of entry which have been incorporated in the toset. */
142 static uint32_t narcs
;
143 /* This is a pointer to the object representing the number of entries
144 currently in the mmaped file. At no point of time this has to be the
145 same as NARCS. If it is equal all entries from the file are in our
147 static volatile uint32_t *narcsp
;
149 /* Description of the currently profiled object. */
150 static long int state
= GMON_PROF_OFF
;
152 static volatile uint16_t *kcount
;
153 static size_t kcountsize
;
155 struct here_fromstruct
157 struct here_cg_arc_record
volatile *here
;
161 static uint16_t *tos
;
162 static size_t tossize
;
164 static struct here_fromstruct
*froms
;
165 static size_t fromssize
;
166 static size_t fromlimit
;
167 static size_t fromidx
;
169 static uintptr_t lowpc
;
170 static uintptr_t highpc
;
171 static size_t textsize
;
172 static unsigned int hashfraction
;
173 static unsigned int log_hashfraction
;
175 /* This is the information about the mmaped memory. */
176 static struct gmon_hdr
*addr
;
177 static off_t expected_size
;
179 /* See profil(2) where this is described. */
181 #define SCALE_1_TO_1 0x10000L
185 /* Set up profiling data to profile object desribed by MAP. The output
186 file is found (or created) in OUTPUT_DIR. */
189 _dl_start_profile (struct link_map
*map
, const char *output_dir
)
194 const ElfW(Phdr
) *ph
;
195 ElfW(Addr
) mapstart
= ~((ElfW(Addr
)) 0);
196 ElfW(Addr
) mapend
= 0;
197 struct gmon_hdr gmon_hdr
;
198 struct gmon_hist_hdr hist_hdr
;
202 /* Compute the size of the sections which contain program code. */
203 for (ph
= map
->l_phdr
; ph
< &map
->l_phdr
[map
->l_phnum
]; ++ph
)
204 if (ph
->p_type
== PT_LOAD
&& (ph
->p_flags
& PF_X
))
206 ElfW(Addr
) start
= (ph
->p_vaddr
& ~(_dl_pagesize
- 1));
207 ElfW(Addr
) end
= ((ph
->p_vaddr
+ ph
->p_memsz
+ _dl_pagesize
- 1)
208 & ~(_dl_pagesize
- 1));
210 if (start
< mapstart
)
216 /* Now we can compute the size of the profiling data. This is done
217 with the same formulars as in `monstartup' (see gmon.c). */
218 state
= GMON_PROF_OFF
;
219 lowpc
= ROUNDDOWN (mapstart
+ map
->l_addr
,
220 HISTFRACTION
* sizeof (HISTCOUNTER
));
221 highpc
= ROUNDUP (mapend
+ map
->l_addr
,
222 HISTFRACTION
* sizeof (HISTCOUNTER
));
223 textsize
= highpc
- lowpc
;
224 kcountsize
= textsize
/ HISTFRACTION
;
225 hashfraction
= HASHFRACTION
;
226 if ((HASHFRACTION
& (HASHFRACTION
- 1)) == 0)
227 /* If HASHFRACTION is a power of two, mcount can use shifting
228 instead of integer division. Precompute shift amount. */
229 log_hashfraction
= __ffs (hashfraction
* sizeof (*froms
)) - 1;
231 log_hashfraction
= -1;
232 tossize
= textsize
/ HASHFRACTION
;
233 fromlimit
= textsize
* ARCDENSITY
/ 100;
234 if (fromlimit
< MINARCS
)
236 if (fromlimit
> MAXARCS
)
238 fromssize
= fromlimit
* sizeof (struct here_fromstruct
);
240 expected_size
= (sizeof (struct gmon_hdr
)
241 + 4 + sizeof (struct gmon_hist_hdr
) + kcountsize
242 + 4 + 4 + fromssize
* sizeof (struct here_cg_arc_record
));
244 /* Create the gmon_hdr we expect or write. */
245 memset (&gmon_hdr
, '\0', sizeof (struct gmon_hdr
));
246 memcpy (&gmon_hdr
.cookie
[0], GMON_MAGIC
, sizeof (gmon_hdr
.cookie
));
247 *(int32_t *) gmon_hdr
.version
= GMON_SHOBJ_VERSION
;
249 /* Create the hist_hdr we expect or write. */
250 *(char **) hist_hdr
.low_pc
= (char *) mapstart
;
251 *(char **) hist_hdr
.high_pc
= (char *) mapend
;
252 *(int32_t *) hist_hdr
.hist_size
= kcountsize
/ sizeof (HISTCOUNTER
);
253 *(int32_t *) hist_hdr
.prof_rate
= __profile_frequency ();
254 strncpy (hist_hdr
.dimen
, "seconds", sizeof (hist_hdr
.dimen
));
255 hist_hdr
.dimen_abbrev
= 's';
257 /* First determine the output name. We write in the directory
258 OUTPUT_DIR and the name is composed from the shared objects
259 soname (or the file name) and the ending ".profile". */
260 filename
= (char *) alloca (strlen (output_dir
) + 1 + strlen (_dl_profile
)
261 + sizeof ".profile");
262 cp
= __stpcpy (filename
, output_dir
);
264 __stpcpy (__stpcpy (cp
, _dl_profile
), ".profile");
266 fd
= __open (filename
, O_RDWR
| O_CREAT
, 0666);
269 /* We cannot write the profiling data so don't do anything. */
271 _dl_sysdep_message (filename
, ": cannot open file: ",
272 __strerror_r (errno
, buf
, sizeof buf
),
277 if (fstat (fd
, &st
) < 0 || !S_ISREG (st
.st_mode
))
279 /* Not stat'able or not a regular file => don't use it. */
283 _dl_sysdep_message (filename
, ": cannot stat file: ",
284 __strerror_r (errnum
, buf
, sizeof buf
),
289 /* Test the size. If it does not match what we expect from the size
290 values in the map MAP we don't use it and warn the user. */
293 /* We have to create the file. */
294 char buf
[_dl_pagesize
];
296 memset (buf
, '\0', _dl_pagesize
);
298 if (__lseek (fd
, expected_size
& ~(_dl_pagesize
- 1), SEEK_SET
) == -1)
305 _dl_sysdep_message (filename
, ": cannot create file: ",
306 __strerror_r (errnum
, buf
, sizeof buf
),
311 if (TEMP_FAILURE_RETRY (__libc_write (fd
, buf
, (expected_size
312 & (_dl_pagesize
- 1))))
316 else if (st
.st_size
!= expected_size
)
322 __munmap ((void *) addr
, expected_size
);
324 _dl_sysdep_message (filename
,
325 ": file is no correct profile data file for `",
326 _dl_profile
, "'\n", NULL
);
330 addr
= (struct gmon_hdr
*) __mmap (NULL
, expected_size
, PROT_READ
|PROT_WRITE
,
331 MAP_SHARED
|MAP_FILE
, fd
, 0);
332 if (addr
== (struct gmon_hdr
*) MAP_FAILED
)
337 _dl_sysdep_message (filename
, ": cannot map file: ",
338 __strerror_r (errnum
, buf
, sizeof buf
),
343 /* We don't need the file desriptor anymore. */
346 /* Pointer to data after the header. */
347 hist
= (char *) (addr
+ 1);
348 kcount
= (uint16_t *) ((char *) hist
+ sizeof (uint32_t)
349 + sizeof (struct gmon_hist_hdr
));
351 /* Compute pointer to array of the arc information. */
352 narcsp
= (uint32_t *) ((char *) kcount
+ kcountsize
+ sizeof (uint32_t));
353 data
= (struct here_cg_arc_record
*) ((char *) narcsp
+ sizeof (uint32_t));
357 /* Create the signature. */
358 memcpy (addr
, &gmon_hdr
, sizeof (struct gmon_hdr
));
360 *(uint32_t *) hist
= GMON_TAG_TIME_HIST
;
361 memcpy (hist
+ sizeof (uint32_t), &hist_hdr
,
362 sizeof (struct gmon_hist_hdr
));
364 narcsp
[-1] = GMON_TAG_CG_ARC
;
368 /* Test the signature in the file. */
369 if (memcmp (addr
, &gmon_hdr
, sizeof (struct gmon_hdr
)) != 0
370 || *(uint32_t *) hist
!= GMON_TAG_TIME_HIST
371 || memcmp (hist
+ sizeof (uint32_t), &hist_hdr
,
372 sizeof (struct gmon_hist_hdr
)) != 0
373 || narcsp
[-1] != GMON_TAG_CG_ARC
)
377 /* Allocate memory for the froms data and the pointer to the tos records. */
378 tos
= (uint16_t *) calloc (tossize
+ fromssize
, 1);
381 __munmap ((void *) addr
, expected_size
);
382 _dl_sysdep_fatal ("Out of memory while initializing profiler\n", NULL
);
386 froms
= (struct here_fromstruct
*) ((char *) tos
+ tossize
);
389 /* Now we have to process all the arc count entries. BTW: it is
390 not critical whether the *NARCSP value changes meanwhile. Before
391 we enter a new entry in to toset we will check that everything is
392 available in TOS. This happens in _dl_mcount.
394 Loading the entries in reverse order should help to get the most
395 frequently used entries at the front of the list. */
396 for (idx
= narcs
= MIN (*narcsp
, fromlimit
); idx
> 0; )
401 to_index
= (data
[idx
].self_pc
/ (hashfraction
* sizeof (*tos
)));
402 newfromidx
= fromidx
++;
403 froms
[newfromidx
].here
= &data
[idx
];
404 froms
[newfromidx
].link
= tos
[to_index
];
405 tos
[to_index
] = newfromidx
;
408 /* Setup counting data. */
409 if (kcountsize
< highpc
- lowpc
)
412 s_scale
= ((double) kcountsize
/ (highpc
- lowpc
)) * SCALE_1_TO_1
;
414 size_t range
= highpc
- lowpc
;
415 size_t quot
= range
/ kcountsize
;
417 if (quot
>= SCALE_1_TO_1
)
419 else if (quot
>= SCALE_1_TO_1
/ 256)
420 s_scale
= SCALE_1_TO_1
/ quot
;
421 else if (range
> ULONG_MAX
/ 256)
422 s_scale
= (SCALE_1_TO_1
* 256) / (range
/ (kcountsize
/ 256));
424 s_scale
= (SCALE_1_TO_1
* 256) / ((range
* 256) / kcountsize
);
428 s_scale
= SCALE_1_TO_1
;
430 /* Start the profiler. */
431 __profil ((void *) kcount
, kcountsize
, lowpc
, s_scale
);
433 /* Turn on profiling. */
434 state
= GMON_PROF_ON
;
439 _dl_mcount (ElfW(Addr
) frompc
, ElfW(Addr
) selfpc
)
443 struct here_fromstruct
*fromp
;
445 if (! compare_and_swap (&state
, GMON_PROF_ON
, GMON_PROF_BUSY
))
448 /* Compute relative addresses. The shared object can be loaded at
449 any address. The value of frompc could be anything. We cannot
450 restrict it in any way, just set to a fixed value (0) in case it
451 is outside the allowed range. These calls show up as calls from
452 <external> in the gprof output. */
454 if (frompc
>= textsize
)
457 if (selfpc
>= textsize
)
460 /* Getting here we now have to find out whether the location was
461 already used. If yes we are lucky and only have to increment a
462 counter (this also has to be atomic). If the entry is new things
463 are getting complicated... */
465 /* Avoid integer divide if possible. */
466 if ((HASHFRACTION
& (HASHFRACTION
- 1)) == 0)
467 i
= selfpc
>> log_hashfraction
;
469 i
= selfpc
/ (hashfraction
* sizeof (*tos
));
472 fromindex
= *topcindex
;
475 goto check_new_or_add
;
477 fromp
= &froms
[fromindex
];
479 /* We have to look through the chain of arcs whether there is already
480 an entry for our arc. */
481 while (fromp
->here
->from_pc
!= frompc
)
483 if (fromp
->link
!= 0)
485 fromp
= &froms
[fromp
->link
];
486 while (fromp
->link
!= 0 && fromp
->here
->from_pc
!= frompc
);
488 if (fromp
->here
->from_pc
!= frompc
)
490 topcindex
= &fromp
->link
;
493 /* Our entry is not among the entries we read so far from the
494 data file. Now see whether we have to update the list. */
495 while (narcs
!= *narcsp
&& narcs
< fromlimit
)
499 to_index
= (data
[narcs
].self_pc
500 / (hashfraction
* sizeof (*tos
)));
501 newfromidx
= fromidx
++;
502 froms
[newfromidx
].here
= &data
[narcs
];
503 froms
[newfromidx
].link
= tos
[to_index
];
504 tos
[to_index
] = newfromidx
;
508 /* If we still have no entry stop searching and insert. */
511 size_t newarc
= 1 + exchange_and_add (narcsp
, 1);
513 /* In rare cases it could happen that all entries in FROMS are
514 occupied. So we cannot count this anymore. */
515 if (newarc
>= fromlimit
)
518 fromp
= &froms
[*topcindex
= fromidx
++];
520 fromp
->here
= &data
[newarc
];
521 data
[newarc
].from_pc
= frompc
;
522 data
[newarc
].self_pc
= selfpc
;
523 data
[newarc
].count
= 0;
530 fromp
= &froms
[*topcindex
];
537 /* Increment the counter. */
538 atomic_add (&fromp
->here
->count
, 1);
541 state
= GMON_PROF_ON
;