]> git.ipfire.org Git - thirdparty/glibc.git/blob - elf/dl-profile.c
Update.
[thirdparty/glibc.git] / elf / dl-profile.c
1 /* Profiling of shared libraries.
2 Copyright (C) 1997, 1998, 1999, 2000 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.
6
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.
11
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.
16
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. */
21
22 #include <errno.h>
23 #include <fcntl.h>
24 #include <inttypes.h>
25 #include <limits.h>
26 #include <stdio.h>
27 #include <stdlib.h>
28 #include <string.h>
29 #include <unistd.h>
30 #include <ldsodefs.h>
31 #include <sys/gmon.h>
32 #include <sys/gmon_out.h>
33 #include <sys/mman.h>
34 #include <sys/param.h>
35 #include <sys/stat.h>
36 #include <atomicity.h>
37
38 /* The LD_PROFILE feature has to be implemented different to the
39 normal profiling using the gmon/ functions. The problem is that an
40 arbitrary amount of processes simulataneously can be run using
41 profiling and all write the results in the same file. To provide
42 this mechanism one could implement a complicated mechanism to merge
43 the content of two profiling runs or one could extend the file
44 format to allow more than one data set. For the second solution we
45 would have the problem that the file can grow in size beyond any
46 limit and both solutions have the problem that the concurrency of
47 writing the results is a big problem.
48
49 Another much simpler method is to use mmap to map the same file in
50 all using programs and modify the data in the mmap'ed area and so
51 also automatically on the disk. Using the MAP_SHARED option of
52 mmap(2) this can be done without big problems in more than one
53 file.
54
55 This approach is very different from the normal profiling. We have
56 to use the profiling data in exactly the way they are expected to
57 be written to disk. But the normal format used by gprof is not usable
58 to do this. It is optimized for size. It writes the tags as single
59 bytes but this means that the following 32/64 bit values are
60 unaligned.
61
62 Therefore we use a new format. This will look like this
63
64 0 1 2 3 <- byte is 32 bit word
65 0000 g m o n
66 0004 *version* <- GMON_SHOBJ_VERSION
67 0008 00 00 00 00
68 000c 00 00 00 00
69 0010 00 00 00 00
70
71 0014 *tag* <- GMON_TAG_TIME_HIST
72 0018 ?? ?? ?? ??
73 ?? ?? ?? ?? <- 32/64 bit LowPC
74 0018+A ?? ?? ?? ??
75 ?? ?? ?? ?? <- 32/64 bit HighPC
76 0018+2*A *histsize*
77 001c+2*A *profrate*
78 0020+2*A s e c o
79 0024+2*A n d s \0
80 0028+2*A \0 \0 \0 \0
81 002c+2*A \0 \0 \0
82 002f+2*A s
83
84 0030+2*A ?? ?? ?? ?? <- Count data
85 ... ...
86 0030+2*A+K ?? ?? ?? ??
87
88 0030+2*A+K *tag* <- GMON_TAG_CG_ARC
89 0034+2*A+K *lastused*
90 0038+2*A+K ?? ?? ?? ??
91 ?? ?? ?? ?? <- FromPC#1
92 0038+3*A+K ?? ?? ?? ??
93 ?? ?? ?? ?? <- ToPC#1
94 0038+4*A+K ?? ?? ?? ?? <- Count#1
95 ... ... ...
96 0038+(2*(CN-1)+2)*A+(CN-1)*4+K ?? ?? ?? ??
97 ?? ?? ?? ?? <- FromPC#CGN
98 0038+(2*(CN-1)+3)*A+(CN-1)*4+K ?? ?? ?? ??
99 ?? ?? ?? ?? <- ToPC#CGN
100 0038+(2*CN+2)*A+(CN-1)*4+K ?? ?? ?? ?? <- Count#CGN
101
102 We put (for now?) no basic block information in the file since this would
103 introduce rase conditions among all the processes who want to write them.
104
105 `K' is the number of count entries which is computed as
106
107 textsize / HISTFRACTION
108
109 `CG' in the above table is the number of call graph arcs. Normally,
110 the table is sparse and the profiling code writes out only the those
111 entries which are really used in the program run. But since we must
112 not extend this table (the profiling file) we'll keep them all here.
113 So CN can be executed in advance as
114
115 MINARCS <= textsize*(ARCDENSITY/100) <= MAXARCS
116
117 Now the remaining question is: how to build the data structures we can
118 work with from this data. We need the from set and must associate the
119 froms with all the associated tos. We will do this by constructing this
120 data structures at the program start. To do this we'll simply visit all
121 entries in the call graph table and add it to the appropriate list. */
122
123 extern int __profile_frequency (void);
124
125 /* We define a special type to address the elements of the arc table.
126 This is basically the `gmon_cg_arc_record' format but it includes
127 the room for the tag and it uses real types. */
128 struct here_cg_arc_record
129 {
130 uintptr_t from_pc;
131 uintptr_t self_pc;
132 uint32_t count;
133 } __attribute__ ((packed));
134
135 static struct here_cg_arc_record *data;
136
137 /* This is the number of entry which have been incorporated in the toset. */
138 static uint32_t narcs;
139 /* This is a pointer to the object representing the number of entries
140 currently in the mmaped file. At no point of time this has to be the
141 same as NARCS. If it is equal all entries from the file are in our
142 lists. */
143 static volatile uint32_t *narcsp;
144
145 /* Description of the currently profiled object. */
146 static long int state = GMON_PROF_OFF;
147
148 static volatile uint16_t *kcount;
149 static size_t kcountsize;
150
151 struct here_fromstruct
152 {
153 struct here_cg_arc_record volatile *here;
154 uint16_t link;
155 };
156
157 static uint16_t *tos;
158 static size_t tossize;
159
160 static struct here_fromstruct *froms;
161 static size_t fromssize;
162 static size_t fromlimit;
163 static size_t fromidx;
164
165 static uintptr_t lowpc;
166 static uintptr_t highpc;
167 static size_t textsize;
168 static unsigned int hashfraction;
169 static unsigned int log_hashfraction;
170
171 /* This is the information about the mmaped memory. */
172 static struct gmon_hdr *addr;
173 static off_t expected_size;
174
175 /* See profil(2) where this is described. */
176 static int s_scale;
177 #define SCALE_1_TO_1 0x10000L
178
179
180 \f
181 /* Set up profiling data to profile object desribed by MAP. The output
182 file is found (or created) in OUTPUT_DIR. */
183 void
184 internal_function
185 _dl_start_profile (struct link_map *map, const char *output_dir)
186 {
187 char *filename;
188 int fd;
189 struct stat st;
190 const ElfW(Phdr) *ph;
191 ElfW(Addr) mapstart = ~((ElfW(Addr)) 0);
192 ElfW(Addr) mapend = 0;
193 struct gmon_hdr gmon_hdr;
194 struct gmon_hist_hdr hist_hdr;
195 char *hist, *cp;
196 size_t idx;
197
198 /* Compute the size of the sections which contain program code. */
199 for (ph = map->l_phdr; ph < &map->l_phdr[map->l_phnum]; ++ph)
200 if (ph->p_type == PT_LOAD && (ph->p_flags & PF_X))
201 {
202 ElfW(Addr) start = (ph->p_vaddr & ~(_dl_pagesize - 1));
203 ElfW(Addr) end = ((ph->p_vaddr + ph->p_memsz + _dl_pagesize - 1)
204 & ~(_dl_pagesize - 1));
205
206 if (start < mapstart)
207 mapstart = start;
208 if (end > mapend)
209 mapend = end;
210 }
211
212 /* Now we can compute the size of the profiling data. This is done
213 with the same formulars as in `monstartup' (see gmon.c). */
214 state = GMON_PROF_OFF;
215 lowpc = ROUNDDOWN (mapstart + map->l_addr,
216 HISTFRACTION * sizeof (HISTCOUNTER));
217 highpc = ROUNDUP (mapend + map->l_addr,
218 HISTFRACTION * sizeof (HISTCOUNTER));
219 textsize = highpc - lowpc;
220 kcountsize = textsize / HISTFRACTION;
221 hashfraction = HASHFRACTION;
222 if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
223 /* If HASHFRACTION is a power of two, mcount can use shifting
224 instead of integer division. Precompute shift amount. */
225 log_hashfraction = __ffs (hashfraction * sizeof (*froms)) - 1;
226 else
227 log_hashfraction = -1;
228 tossize = textsize / HASHFRACTION;
229 fromlimit = textsize * ARCDENSITY / 100;
230 if (fromlimit < MINARCS)
231 fromlimit = MINARCS;
232 if (fromlimit > MAXARCS)
233 fromlimit = MAXARCS;
234 fromssize = fromlimit * sizeof (struct here_fromstruct);
235
236 expected_size = (sizeof (struct gmon_hdr)
237 + 4 + sizeof (struct gmon_hist_hdr) + kcountsize
238 + 4 + 4 + fromssize * sizeof (struct here_cg_arc_record));
239
240 /* Create the gmon_hdr we expect or write. */
241 memset (&gmon_hdr, '\0', sizeof (struct gmon_hdr));
242 memcpy (&gmon_hdr.cookie[0], GMON_MAGIC, sizeof (gmon_hdr.cookie));
243 *(int32_t *) gmon_hdr.version = GMON_SHOBJ_VERSION;
244
245 /* Create the hist_hdr we expect or write. */
246 *(char **) hist_hdr.low_pc = (char *) mapstart;
247 *(char **) hist_hdr.high_pc = (char *) mapend;
248 *(int32_t *) hist_hdr.hist_size = kcountsize / sizeof (HISTCOUNTER);
249 *(int32_t *) hist_hdr.prof_rate = __profile_frequency ();
250 strncpy (hist_hdr.dimen, "seconds", sizeof (hist_hdr.dimen));
251 hist_hdr.dimen_abbrev = 's';
252
253 /* First determine the output name. We write in the directory
254 OUTPUT_DIR and the name is composed from the shared objects
255 soname (or the file name) and the ending ".profile". */
256 filename = (char *) alloca (strlen (output_dir) + 1 + strlen (_dl_profile)
257 + sizeof ".profile");
258 cp = __stpcpy (filename, output_dir);
259 *cp++ = '/';
260 __stpcpy (__stpcpy (cp, _dl_profile), ".profile");
261
262 fd = __open (filename, O_RDWR | O_CREAT, 0666);
263 if (fd == -1)
264 {
265 /* We cannot write the profiling data so don't do anything. */
266 char buf[400];
267 _dl_sysdep_message (filename, ": cannot open file: ",
268 __strerror_r (errno, buf, sizeof buf),
269 "\n", NULL);
270 return;
271 }
272
273 if (fstat (fd, &st) < 0 || !S_ISREG (st.st_mode))
274 {
275 /* Not stat'able or not a regular file => don't use it. */
276 char buf[400];
277 int errnum = errno;
278 __close (fd);
279 _dl_sysdep_message (filename, ": cannot stat file: ",
280 __strerror_r (errnum, buf, sizeof buf),
281 "\n", NULL);
282 return;
283 }
284
285 /* Test the size. If it does not match what we expect from the size
286 values in the map MAP we don't use it and warn the user. */
287 if (st.st_size == 0)
288 {
289 /* We have to create the file. */
290 char buf[_dl_pagesize];
291
292 memset (buf, '\0', _dl_pagesize);
293
294 if (__lseek (fd, expected_size & ~(_dl_pagesize - 1), SEEK_SET) == -1)
295 {
296 char buf[400];
297 int errnum;
298 cannot_create:
299 errnum = errno;
300 __close (fd);
301 _dl_sysdep_message (filename, ": cannot create file: ",
302 __strerror_r (errnum, buf, sizeof buf),
303 "\n", NULL);
304 return;
305 }
306
307 if (TEMP_FAILURE_RETRY (__libc_write (fd, buf, (expected_size
308 & (_dl_pagesize - 1))))
309 < 0)
310 goto cannot_create;
311 }
312 else if (st.st_size != expected_size)
313 {
314 __close (fd);
315 wrong_format:
316
317 if (addr != NULL)
318 __munmap ((void *) addr, expected_size);
319
320 _dl_sysdep_message (filename,
321 ": file is no correct profile data file for `",
322 _dl_profile, "'\n", NULL);
323 return;
324 }
325
326 addr = (struct gmon_hdr *) __mmap (NULL, expected_size, PROT_READ|PROT_WRITE,
327 MAP_SHARED|MAP_FILE, fd, 0);
328 if (addr == (struct gmon_hdr *) MAP_FAILED)
329 {
330 char buf[400];
331 int errnum = errno;
332 __close (fd);
333 _dl_sysdep_message (filename, ": cannot map file: ",
334 __strerror_r (errnum, buf, sizeof buf),
335 "\n", NULL);
336 return;
337 }
338
339 /* We don't need the file desriptor anymore. */
340 __close (fd);
341
342 /* Pointer to data after the header. */
343 hist = (char *) (addr + 1);
344 kcount = (uint16_t *) ((char *) hist + sizeof (uint32_t)
345 + sizeof (struct gmon_hist_hdr));
346
347 /* Compute pointer to array of the arc information. */
348 narcsp = (uint32_t *) ((char *) kcount + kcountsize + sizeof (uint32_t));
349 data = (struct here_cg_arc_record *) ((char *) narcsp + sizeof (uint32_t));
350
351 if (st.st_size == 0)
352 {
353 /* Create the signature. */
354 memcpy (addr, &gmon_hdr, sizeof (struct gmon_hdr));
355
356 *(uint32_t *) hist = GMON_TAG_TIME_HIST;
357 memcpy (hist + sizeof (uint32_t), &hist_hdr,
358 sizeof (struct gmon_hist_hdr));
359
360 narcsp[-1] = GMON_TAG_CG_ARC;
361 }
362 else
363 {
364 /* Test the signature in the file. */
365 if (memcmp (addr, &gmon_hdr, sizeof (struct gmon_hdr)) != 0
366 || *(uint32_t *) hist != GMON_TAG_TIME_HIST
367 || memcmp (hist + sizeof (uint32_t), &hist_hdr,
368 sizeof (struct gmon_hist_hdr)) != 0
369 || narcsp[-1] != GMON_TAG_CG_ARC)
370 goto wrong_format;
371 }
372
373 /* Allocate memory for the froms data and the pointer to the tos records. */
374 tos = (uint16_t *) calloc (tossize + fromssize, 1);
375 if (tos == NULL)
376 {
377 __munmap ((void *) addr, expected_size);
378 _dl_sysdep_fatal ("Out of memory while initializing profiler\n", NULL);
379 /* NOTREACHED */
380 }
381
382 froms = (struct here_fromstruct *) ((char *) tos + tossize);
383 fromidx = 0;
384
385 /* Now we have to process all the arc count entries. BTW: it is
386 not critical whether the *NARCSP value changes meanwhile. Before
387 we enter a new entry in to toset we will check that everything is
388 available in TOS. This happens in _dl_mcount.
389
390 Loading the entries in reverse order should help to get the most
391 frequently used entries at the front of the list. */
392 for (idx = narcs = MIN (*narcsp, fromlimit); idx > 0; )
393 {
394 size_t to_index;
395 size_t newfromidx;
396 --idx;
397 to_index = (data[idx].self_pc / (hashfraction * sizeof (*tos)));
398 newfromidx = fromidx++;
399 froms[newfromidx].here = &data[idx];
400 froms[newfromidx].link = tos[to_index];
401 tos[to_index] = newfromidx;
402 }
403
404 /* Setup counting data. */
405 if (kcountsize < highpc - lowpc)
406 {
407 #if 0
408 s_scale = ((double) kcountsize / (highpc - lowpc)) * SCALE_1_TO_1;
409 #else
410 size_t range = highpc - lowpc;
411 size_t quot = range / kcountsize;
412
413 if (quot >= SCALE_1_TO_1)
414 s_scale = 1;
415 else if (quot >= SCALE_1_TO_1 / 256)
416 s_scale = SCALE_1_TO_1 / quot;
417 else if (range > ULONG_MAX / 256)
418 s_scale = (SCALE_1_TO_1 * 256) / (range / (kcountsize / 256));
419 else
420 s_scale = (SCALE_1_TO_1 * 256) / ((range * 256) / kcountsize);
421 #endif
422 }
423 else
424 s_scale = SCALE_1_TO_1;
425
426 /* Start the profiler. */
427 __profil ((void *) kcount, kcountsize, lowpc, s_scale);
428
429 /* Turn on profiling. */
430 state = GMON_PROF_ON;
431 }
432
433
434 void
435 _dl_mcount (ElfW(Addr) frompc, ElfW(Addr) selfpc)
436 {
437 uint16_t *topcindex;
438 size_t i, fromindex;
439 struct here_fromstruct *fromp;
440
441 if (! compare_and_swap (&state, GMON_PROF_ON, GMON_PROF_BUSY))
442 return;
443
444 /* Compute relative addresses. The shared object can be loaded at
445 any address. The value of frompc could be anything. We cannot
446 restrict it in any way, just set to a fixed value (0) in case it
447 is outside the allowed range. These calls show up as calls from
448 <external> in the gprof output. */
449 frompc -= lowpc;
450 if (frompc >= textsize)
451 frompc = 0;
452 selfpc -= lowpc;
453 if (selfpc >= textsize)
454 goto done;
455
456 /* Getting here we now have to find out whether the location was
457 already used. If yes we are lucky and only have to increment a
458 counter (this also has to be atomic). If the entry is new things
459 are getting complicated... */
460
461 /* Avoid integer divide if possible. */
462 if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
463 i = selfpc >> log_hashfraction;
464 else
465 i = selfpc / (hashfraction * sizeof (*tos));
466
467 topcindex = &tos[i];
468 fromindex = *topcindex;
469
470 if (fromindex == 0)
471 goto check_new_or_add;
472
473 fromp = &froms[fromindex];
474
475 /* We have to look through the chain of arcs whether there is already
476 an entry for our arc. */
477 while (fromp->here->from_pc != frompc)
478 {
479 if (fromp->link != 0)
480 do
481 fromp = &froms[fromp->link];
482 while (fromp->link != 0 && fromp->here->from_pc != frompc);
483
484 if (fromp->here->from_pc != frompc)
485 {
486 topcindex = &fromp->link;
487
488 check_new_or_add:
489 /* Our entry is not among the entries we read so far from the
490 data file. Now see whether we have to update the list. */
491 while (narcs != *narcsp && narcs < fromlimit)
492 {
493 size_t to_index;
494 size_t newfromidx;
495 to_index = (data[narcs].self_pc
496 / (hashfraction * sizeof (*tos)));
497 newfromidx = fromidx++;
498 froms[newfromidx].here = &data[narcs];
499 froms[newfromidx].link = tos[to_index];
500 tos[to_index] = newfromidx;
501 ++narcs;
502 }
503
504 /* If we still have no entry stop searching and insert. */
505 if (*topcindex == 0)
506 {
507 size_t newarc = 1 + exchange_and_add (narcsp, 1);
508
509 /* In rare cases it could happen that all entries in FROMS are
510 occupied. So we cannot count this anymore. */
511 if (newarc >= fromlimit)
512 goto done;
513
514 fromp = &froms[*topcindex = fromidx++];
515
516 fromp->here = &data[newarc];
517 data[newarc].from_pc = frompc;
518 data[newarc].self_pc = selfpc;
519 data[newarc].count = 0;
520 fromp->link = 0;
521 ++narcs;
522
523 break;
524 }
525
526 fromp = &froms[*topcindex];
527 }
528 else
529 /* Found in. */
530 break;
531 }
532
533 /* Increment the counter. */
534 atomic_add (&fromp->here->count, 1);
535
536 done:
537 state = GMON_PROF_ON;
538 }