]> 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 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 <link.h>
26 #include <stdio.h>
27 #include <stdlib.h>
28 #include <string.h>
29 #include <unistd.h>
30 #include <sys/gmon.h>
31 #include <sys/gmon_out.h>
32 #include <sys/mman.h>
33 #include <sys/param.h>
34 #include <sys/stat.h>
35 #include <atomicity.h>
36
37 /* The LD_PROFILE feature has to be implemented different to the
38 normal profiling using the gmon/ functions. The problem is that an
39 arbitrary amount of processes simulataneously can be run using
40 profiling and all write the results in the same file. To provide
41 this mechanism one could implement a complicated mechanism to merge
42 the content of two profiling runs or one could extend the file
43 format to allow more than one data set. For the second solution we
44 would have the problem that the file can grow in size beyond any
45 limit and both solutions have the problem that the concurrency of
46 writing the results is a big problem.
47
48 Another much simpler method is to use mmap to map the same file in
49 all using programs and modify the data in the mmap'ed area and so
50 also automatically on the disk. Using the MAP_SHARED option of
51 mmap(2) this can be done without big problems in more than one
52 file.
53
54 This approach is very different from the normal profiling. We have
55 to use the profiling data in exactly the way they are expected to
56 be written to disk. But the normal format used by gprof is not usable
57 to do this. It is optimized for size. It writes the tags as single
58 bytes but this means that the following 32/64 bit values are
59 unaligned.
60
61 Therefore we use a new format. This will look like this
62
63 0 1 2 3 <- byte is 32 bit word
64 0000 g m o n
65 0004 *version* <- GMON_SHOBJ_VERSION
66 0008 00 00 00 00
67 000c 00 00 00 00
68 0010 00 00 00 00
69
70 0014 *tag* <- GMON_TAG_TIME_HIST
71 0018 ?? ?? ?? ??
72 ?? ?? ?? ?? <- 32/64 bit LowPC
73 0018+A ?? ?? ?? ??
74 ?? ?? ?? ?? <- 32/64 bit HighPC
75 0018+2*A *histsize*
76 001c+2*A *profrate*
77 0020+2*A s e c o
78 0024+2*A n d s \0
79 0028+2*A \0 \0 \0 \0
80 002c+2*A \0 \0 \0
81 002f+2*A s
82
83 0030+2*A ?? ?? ?? ?? <- Count data
84 ... ...
85 0030+2*A+K ?? ?? ?? ??
86
87 0030+2*A+K *tag* <- GMON_TAG_CG_ARC
88 0034+2*A+K *lastused*
89 0038+2*A+K ?? ?? ?? ??
90 ?? ?? ?? ?? <- FromPC#1
91 0038+3*A+K ?? ?? ?? ??
92 ?? ?? ?? ?? <- ToPC#1
93 0038+4*A+K ?? ?? ?? ?? <- Count#1
94 ... ... ...
95 0038+(2*(CN-1)+2)*A+(CN-1)*4+K ?? ?? ?? ??
96 ?? ?? ?? ?? <- FromPC#CGN
97 0038+(2*(CN-1)+3)*A+(CN-1)*4+K ?? ?? ?? ??
98 ?? ?? ?? ?? <- ToPC#CGN
99 0038+(2*CN+2)*A+(CN-1)*4+K ?? ?? ?? ?? <- Count#CGN
100
101 We put (for now?) no basic block information in the file since this would
102 introduce rase conditions among all the processes who want to write them.
103
104 `K' is the number of count entries which is computed as
105
106 textsize / HISTFRACTION
107
108 `CG' in the above table is the number of call graph arcs. Normally,
109 the table is sparse and the profiling code writes out only the those
110 entries which are really used in the program run. But since we must
111 not extend this table (the profiling file) we'll keep them all here.
112 So CN can be executed in advance as
113
114 MINARCS <= textsize*(ARCDENSITY/100) <= MAXARCS
115
116 Now the remaining question is: how to build the data structures we can
117 work with from this data. We need the from set and must associate the
118 froms with all the associated tos. We will do this by constructing this
119 data structures at the program start. To do this we'll simply visit all
120 entries in the call graph table and add it to the appropriate list. */
121
122 extern char *_strerror_internal __P ((int, char *buf, size_t));
123
124 extern int __profile_frequency __P ((void));
125
126 /* We define a special type to address the elements of the arc table.
127 This is basically the `gmon_cg_arc_record' format but it includes
128 the room for the tag and it uses real types. */
129 struct here_cg_arc_record
130 {
131 uintptr_t from_pc;
132 uintptr_t self_pc;
133 uint32_t count;
134 } __attribute__ ((packed));
135
136 static struct here_cg_arc_record *data;
137
138 /* This is the number of entry which have been incorporated in the toset. */
139 static uint32_t narcs;
140 /* This is a pointer to the object representing the number of entries
141 currently in the mmaped file. At no point of time this has to be the
142 same as NARCS. If it is equal all entries from the file are in our
143 lists. */
144 static volatile uint32_t *narcsp;
145
146 /* Description of the currently profiled object. */
147 static long int state = GMON_PROF_OFF;
148
149 static volatile uint16_t *kcount;
150 static size_t kcountsize;
151
152 struct here_fromstruct
153 {
154 struct here_cg_arc_record volatile *here;
155 uint16_t link;
156 };
157
158 static uint16_t *tos;
159 static size_t tossize;
160
161 static struct here_fromstruct *froms;
162 static size_t fromssize;
163 static size_t fromlimit;
164 static size_t fromidx;
165
166 static uintptr_t lowpc;
167 static uintptr_t highpc;
168 static size_t textsize;
169 static unsigned int hashfraction;
170 static unsigned int log_hashfraction;
171
172 /* This is the information about the mmaped memory. */
173 static struct gmon_hdr *addr;
174 static off_t expected_size;
175
176 /* See profil(2) where this is described. */
177 static int s_scale;
178 #define SCALE_1_TO_1 0x10000L
179
180
181 \f
182 /* Set up profiling data to profile object desribed by MAP. The output
183 file is found (or created) in OUTPUT_DIR. */
184 void
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 = __builtin_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_internal (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_internal (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_internal (errnum, buf, sizeof buf),
303 "\n", NULL);
304 return;
305 }
306
307 if (TEMP_FAILURE_RETRY (__write (fd, buf, (expected_size
308 & (_dl_pagesize - 1)))) < 0)
309 goto cannot_create;
310 }
311 else if (st.st_size != expected_size)
312 {
313 __close (fd);
314 wrong_format:
315
316 if (addr != NULL)
317 __munmap ((void *) addr, expected_size);
318
319 _dl_sysdep_message (filename,
320 ": file is no correct profile data file for `",
321 _dl_profile, "'\n", NULL);
322 return;
323 }
324
325 addr = (struct gmon_hdr *) __mmap (NULL, expected_size, PROT_READ|PROT_WRITE,
326 MAP_SHARED|MAP_FILE, fd, 0);
327 if (addr == (struct gmon_hdr *) MAP_FAILED)
328 {
329 char buf[400];
330 int errnum = errno;
331 __close (fd);
332 _dl_sysdep_message (filename, ": cannot map file: ",
333 _strerror_internal (errnum, buf, sizeof buf),
334 "\n", NULL);
335 return;
336 }
337
338 /* We don't need the file desriptor anymore. */
339 __close (fd);
340
341 /* Pointer to data after the header. */
342 hist = (char *) (addr + 1);
343 kcount = (uint16_t *) ((char *) hist + sizeof (uint32_t)
344 + sizeof (struct gmon_hist_hdr));
345
346 /* Compute pointer to array of the arc information. */
347 narcsp = (uint32_t *) ((char *) kcount + kcountsize + sizeof (uint32_t));
348 data = (struct here_cg_arc_record *) ((char *) narcsp + sizeof (uint32_t));
349
350 if (st.st_size == 0)
351 {
352 /* Create the signature. */
353 memcpy (addr, &gmon_hdr, sizeof (struct gmon_hdr));
354
355 *(uint32_t *) hist = GMON_TAG_TIME_HIST;
356 memcpy (hist + sizeof (uint32_t), &hist_hdr,
357 sizeof (struct gmon_hist_hdr));
358
359 narcsp[-1] = GMON_TAG_CG_ARC;
360 }
361 else
362 {
363 /* Test the signature in the file. */
364 if (memcmp (addr, &gmon_hdr, sizeof (struct gmon_hdr)) != 0
365 || *(uint32_t *) hist != GMON_TAG_TIME_HIST
366 || memcmp (hist + sizeof (uint32_t), &hist_hdr,
367 sizeof (struct gmon_hist_hdr)) != 0
368 || narcsp[-1] != GMON_TAG_CG_ARC)
369 goto wrong_format;
370 }
371
372 /* Allocate memory for the froms data and the pointer to the tos records. */
373 tos = (uint16_t *) calloc (tossize + fromssize, 1);
374 if (tos == NULL)
375 {
376 __munmap ((void *) addr, expected_size);
377 _dl_sysdep_fatal ("Out of memory while initializing profiler\n", NULL);
378 /* NOTREACHED */
379 }
380
381 froms = (struct here_fromstruct *) ((char *) tos + tossize);
382 fromidx = 0;
383
384 /* Now we have to process all the arc count entries. BTW: it is
385 not critical whether the *NARCSP value changes meanwhile. Before
386 we enter a new entry in to toset we will check that everything is
387 available in TOS. This happens in _dl_mcount.
388
389 Loading the entries in reverse order should help to get the most
390 frequently used entries at the front of the list. */
391 for (idx = narcs = MIN (*narcsp, fromlimit); idx > 0; )
392 {
393 size_t to_index;
394 size_t newfromidx;
395 --idx;
396 to_index = (data[idx].self_pc / (hashfraction * sizeof (*tos)));
397 newfromidx = fromidx++;
398 froms[newfromidx].here = &data[idx];
399 froms[newfromidx].link = tos[to_index];
400 tos[to_index] = newfromidx;
401 }
402
403 /* Setup counting data. */
404 if (kcountsize < highpc - lowpc)
405 {
406 /* XXX I've not yet verified that the second expression is really
407 well suited but something like this is in any case necessary
408 for targets without hardware FP support. --drepper */
409 #if 0
410 s_scale = ((double) kcountsize / (highpc - lowpc)) * SCALE_1_TO_1;
411 #else
412 s_scale = (kcountsize * SCALE_1_TO_1) / (highpc - lowpc);
413 #endif
414 }
415 else
416 s_scale = SCALE_1_TO_1;
417
418 /* Start the profiler. */
419 __profil ((void *) kcount, kcountsize, lowpc, s_scale);
420
421 /* Turn on profiling. */
422 state = GMON_PROF_ON;
423 }
424
425
426 void
427 _dl_mcount (ElfW(Addr) frompc, ElfW(Addr) selfpc)
428 {
429 uint16_t *topcindex;
430 size_t i, fromindex;
431 struct here_fromstruct *fromp;
432
433 if (! compare_and_swap (&state, GMON_PROF_ON, GMON_PROF_BUSY))
434 return;
435
436 /* Compute relative addresses. The shared object can be loaded at
437 any address. The value of frompc could be anything. We cannot
438 restrict it in any way, just set to a fixed value (0) in case it
439 is outside the allowed range. These calls show up as calls from
440 <external> in the gprof output. */
441 frompc -= lowpc;
442 if (frompc >= textsize)
443 frompc = 0;
444 selfpc -= lowpc;
445 if (selfpc >= textsize)
446 goto done;
447
448 /* Getting here we now have to find out whether the location was
449 already used. If yes we are lucky and only have to increment a
450 counter (this also has to be atomic). If the entry is new things
451 are getting complicated... */
452
453 /* Avoid integer divide if possible. */
454 if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
455 i = selfpc >> log_hashfraction;
456 else
457 i = selfpc / (hashfraction * sizeof (*tos));
458
459 topcindex = &tos[i];
460 fromindex = *topcindex;
461
462 if (fromindex == 0)
463 goto check_new_or_add;
464
465 fromp = &froms[fromindex];
466
467 /* We have to look through the chain of arcs whether there is already
468 an entry for our arc. */
469 while (fromp->here->from_pc == frompc)
470 {
471 if (fromp->link != 0)
472 do
473 fromp = &froms[fromp->link];
474 while (fromp->link != 0 && fromp->here->from_pc != frompc);
475
476 if (fromp->here->from_pc != frompc)
477 {
478 topcindex = &fromp->link;
479
480 check_new_or_add:
481 /* Our entry is not among the entries we read so far from the
482 data file. Now see whether we have to update the list. */
483 while (narcs != *narcsp && narcs < fromlimit)
484 {
485 size_t to_index;
486 size_t newfromidx;
487 to_index = (data[narcs].self_pc
488 / (hashfraction * sizeof (*tos)));
489 newfromidx = fromidx++;
490 froms[newfromidx].here = &data[narcs];
491 froms[newfromidx].link = tos[to_index];
492 tos[to_index] = newfromidx;
493 ++narcs;
494 }
495
496 /* If we still have no entry stop searching and insert. */
497 if (*topcindex == 0)
498 {
499 size_t newarc = 1 + exchange_and_add (narcsp, 1);
500
501 /* In rare cases it could happen that all entries in FROMS are
502 occupied. So we cannot count this anymore. */
503 if (newarc >= fromlimit)
504 goto done;
505
506 fromp = &froms[*topcindex = fromidx++];
507
508 fromp->here = &data[newarc];
509 data[newarc].from_pc = frompc;
510 data[newarc].self_pc = selfpc;
511 data[newarc].count = 0;
512 fromp->link = 0;
513
514 narcs++;
515
516 break;
517 }
518
519 fromp = &froms[*topcindex];
520 }
521 else
522 /* Found in. */
523 break;
524 }
525
526 /* Increment the counter. */
527 atomic_add (&fromp->here->count, 1);
528
529 done:
530 state = GMON_PROF_ON;
531 }