(elf_get_dynamic_info): Also relocate DT_HASH entry if necessary.
[glibc.git] / elf / dl-profile.c
blob3a18881d263b07a83d158fe5b200dd9f82cdbce2
1 /* Profiling of shared libraries.
2 Copyright (C) 1997,1998,1999,2000,2001,2002 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 Lesser General Public
9 License as published by the Free Software Foundation; either
10 version 2.1 of the 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 Lesser General Public License for more details.
17 You should have received a copy of the GNU Lesser General Public
18 License along with the GNU C Library; if not, write to the Free
19 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
20 02111-1307 USA. */
22 #include <assert.h>
23 #include <errno.h>
24 #include <fcntl.h>
25 #include <inttypes.h>
26 #include <limits.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <string.h>
30 #include <unistd.h>
31 #include <ldsodefs.h>
32 #include <sys/gmon.h>
33 #include <sys/gmon_out.h>
34 #include <sys/mman.h>
35 #include <sys/param.h>
36 #include <sys/stat.h>
37 #include <atomicity.h>
39 /* The LD_PROFILE feature has to be implemented different to the
40 normal profiling using the gmon/ functions. The problem is that an
41 arbitrary amount of processes simulataneously can be run using
42 profiling and all write the results in the same file. To provide
43 this mechanism one could implement a complicated mechanism to merge
44 the content of two profiling runs or one could extend the file
45 format to allow more than one data set. For the second solution we
46 would have the problem that the file can grow in size beyond any
47 limit and both solutions have the problem that the concurrency of
48 writing the results is a big problem.
50 Another much simpler method is to use mmap to map the same file in
51 all using programs and modify the data in the mmap'ed area and so
52 also automatically on the disk. Using the MAP_SHARED option of
53 mmap(2) this can be done without big problems in more than one
54 file.
56 This approach is very different from the normal profiling. We have
57 to use the profiling data in exactly the way they are expected to
58 be written to disk. But the normal format used by gprof is not usable
59 to do this. It is optimized for size. It writes the tags as single
60 bytes but this means that the following 32/64 bit values are
61 unaligned.
63 Therefore we use a new format. This will look like this
65 0 1 2 3 <- byte is 32 bit word
66 0000 g m o n
67 0004 *version* <- GMON_SHOBJ_VERSION
68 0008 00 00 00 00
69 000c 00 00 00 00
70 0010 00 00 00 00
72 0014 *tag* <- GMON_TAG_TIME_HIST
73 0018 ?? ?? ?? ??
74 ?? ?? ?? ?? <- 32/64 bit LowPC
75 0018+A ?? ?? ?? ??
76 ?? ?? ?? ?? <- 32/64 bit HighPC
77 0018+2*A *histsize*
78 001c+2*A *profrate*
79 0020+2*A s e c o
80 0024+2*A n d s \0
81 0028+2*A \0 \0 \0 \0
82 002c+2*A \0 \0 \0
83 002f+2*A s
85 0030+2*A ?? ?? ?? ?? <- Count data
86 ... ...
87 0030+2*A+K ?? ?? ?? ??
89 0030+2*A+K *tag* <- GMON_TAG_CG_ARC
90 0034+2*A+K *lastused*
91 0038+2*A+K ?? ?? ?? ??
92 ?? ?? ?? ?? <- FromPC#1
93 0038+3*A+K ?? ?? ?? ??
94 ?? ?? ?? ?? <- ToPC#1
95 0038+4*A+K ?? ?? ?? ?? <- Count#1
96 ... ... ...
97 0038+(2*(CN-1)+2)*A+(CN-1)*4+K ?? ?? ?? ??
98 ?? ?? ?? ?? <- FromPC#CGN
99 0038+(2*(CN-1)+3)*A+(CN-1)*4+K ?? ?? ?? ??
100 ?? ?? ?? ?? <- ToPC#CGN
101 0038+(2*CN+2)*A+(CN-1)*4+K ?? ?? ?? ?? <- Count#CGN
103 We put (for now?) no basic block information in the file since this would
104 introduce rase conditions among all the processes who want to write them.
106 `K' is the number of count entries which is computed as
108 textsize / HISTFRACTION
110 `CG' in the above table is the number of call graph arcs. Normally,
111 the table is sparse and the profiling code writes out only the those
112 entries which are really used in the program run. But since we must
113 not extend this table (the profiling file) we'll keep them all here.
114 So CN can be executed in advance as
116 MINARCS <= textsize*(ARCDENSITY/100) <= MAXARCS
118 Now the remaining question is: how to build the data structures we can
119 work with from this data. We need the from set and must associate the
120 froms with all the associated tos. We will do this by constructing this
121 data structures at the program start. To do this we'll simply visit all
122 entries in the call graph table and add it to the appropriate list. */
124 extern int __profile_frequency (void);
125 libc_hidden_proto (__profile_frequency)
127 /* We define a special type to address the elements of the arc table.
128 This is basically the `gmon_cg_arc_record' format but it includes
129 the room for the tag and it uses real types. */
130 struct here_cg_arc_record
132 uintptr_t from_pc;
133 uintptr_t self_pc;
134 uint32_t count;
135 } __attribute__ ((packed));
137 static struct here_cg_arc_record *data;
139 /* Nonzero if profiling is under way. */
140 static int running;
142 /* This is the number of entry which have been incorporated in the toset. */
143 static uint32_t narcs;
144 /* This is a pointer to the object representing the number of entries
145 currently in the mmaped file. At no point of time this has to be the
146 same as NARCS. If it is equal all entries from the file are in our
147 lists. */
148 static volatile uint32_t *narcsp;
150 static volatile uint16_t *kcount;
151 static size_t kcountsize;
153 struct here_fromstruct
155 struct here_cg_arc_record volatile *here;
156 uint16_t link;
159 static volatile uint16_t *tos;
161 static struct here_fromstruct *froms;
162 static uint32_t fromlimit;
163 static volatile uint32_t fromidx;
165 static uintptr_t lowpc;
166 static size_t textsize;
167 static unsigned int hashfraction;
168 static unsigned int log_hashfraction;
172 /* Set up profiling data to profile object desribed by MAP. The output
173 file is found (or created) in OUTPUT_DIR. */
174 void
175 internal_function
176 _dl_start_profile (struct link_map *map, const char *output_dir)
178 char *filename;
179 int fd;
180 struct stat64 st;
181 const ElfW(Phdr) *ph;
182 ElfW(Addr) mapstart = ~((ElfW(Addr)) 0);
183 ElfW(Addr) mapend = 0;
184 struct gmon_hdr gmon_hdr;
185 struct gmon_hist_hdr hist_hdr;
186 char *hist, *cp;
187 size_t idx;
188 size_t tossize;
189 size_t fromssize;
190 uintptr_t highpc;
191 struct gmon_hdr *addr = NULL;
192 off_t expected_size;
193 /* See profil(2) where this is described. */
194 int s_scale;
195 #define SCALE_1_TO_1 0x10000L
197 /* Compute the size of the sections which contain program code. */
198 for (ph = map->l_phdr; ph < &map->l_phdr[map->l_phnum]; ++ph)
199 if (ph->p_type == PT_LOAD && (ph->p_flags & PF_X))
201 ElfW(Addr) start = (ph->p_vaddr & ~(GL(dl_pagesize) - 1));
202 ElfW(Addr) end = ((ph->p_vaddr + ph->p_memsz + GL(dl_pagesize) - 1)
203 & ~(GL(dl_pagesize) - 1));
205 if (start < mapstart)
206 mapstart = start;
207 if (end > mapend)
208 mapend = end;
211 /* Now we can compute the size of the profiling data. This is done
212 with the same formulars as in `monstartup' (see gmon.c). */
213 running = 0;
214 lowpc = ROUNDDOWN (mapstart + map->l_addr,
215 HISTFRACTION * sizeof (HISTCOUNTER));
216 highpc = ROUNDUP (mapend + map->l_addr,
217 HISTFRACTION * sizeof (HISTCOUNTER));
218 textsize = highpc - lowpc;
219 kcountsize = textsize / HISTFRACTION;
220 hashfraction = HASHFRACTION;
221 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.
226 This is a constant but the compiler cannot compile the
227 expression away since the __ffs implementation is not known
228 to the compiler. Help the compiler by precomputing the
229 usual cases. */
230 assert (hashfraction == 2);
232 if (sizeof (*froms) == 8)
233 log_hashfraction = 4;
234 else if (sizeof (*froms) == 16)
235 log_hashfraction = 5;
236 else
237 log_hashfraction = __ffs (hashfraction * sizeof (*froms)) - 1;
239 else
240 log_hashfraction = -1;
241 tossize = textsize / HASHFRACTION;
242 fromlimit = textsize * ARCDENSITY / 100;
243 if (fromlimit < MINARCS)
244 fromlimit = MINARCS;
245 if (fromlimit > MAXARCS)
246 fromlimit = MAXARCS;
247 fromssize = fromlimit * sizeof (struct here_fromstruct);
249 expected_size = (sizeof (struct gmon_hdr)
250 + 4 + sizeof (struct gmon_hist_hdr) + kcountsize
251 + 4 + 4 + fromssize * sizeof (struct here_cg_arc_record));
253 /* Create the gmon_hdr we expect or write. */
254 memset (&gmon_hdr, '\0', sizeof (struct gmon_hdr));
255 memcpy (&gmon_hdr.cookie[0], GMON_MAGIC, sizeof (gmon_hdr.cookie));
256 *(int32_t *) gmon_hdr.version = GMON_SHOBJ_VERSION;
258 /* Create the hist_hdr we expect or write. */
259 *(char **) hist_hdr.low_pc = (char *) mapstart;
260 *(char **) hist_hdr.high_pc = (char *) mapend;
261 *(int32_t *) hist_hdr.hist_size = kcountsize / sizeof (HISTCOUNTER);
262 *(int32_t *) hist_hdr.prof_rate = __profile_frequency ();
263 if (sizeof (hist_hdr.dimen) >= sizeof ("seconds"))
265 memcpy (hist_hdr.dimen, "seconds", sizeof ("seconds"));
266 memset (hist_hdr.dimen + sizeof ("seconds"), '\0',
267 sizeof (hist_hdr.dimen) - sizeof ("seconds"));
269 else
270 strncpy (hist_hdr.dimen, "seconds", sizeof (hist_hdr.dimen));
271 hist_hdr.dimen_abbrev = 's';
273 /* First determine the output name. We write in the directory
274 OUTPUT_DIR and the name is composed from the shared objects
275 soname (or the file name) and the ending ".profile". */
276 filename = (char *) alloca (strlen (output_dir) + 1 + strlen (GL(dl_profile))
277 + sizeof ".profile");
278 cp = __stpcpy (filename, output_dir);
279 *cp++ = '/';
280 __stpcpy (__stpcpy (cp, GL(dl_profile)), ".profile");
282 #ifdef O_NOFOLLOW
283 # define EXTRA_FLAGS | O_NOFOLLOW
284 #else
285 # define EXTRA_FLAGS
286 #endif
287 fd = __open (filename, O_RDWR | O_CREAT EXTRA_FLAGS, DEFFILEMODE);
288 if (fd == -1)
290 /* We cannot write the profiling data so don't do anything. */
291 char buf[400];
292 _dl_error_printf ("%s: cannot open file: %s\n", filename,
293 __strerror_r (errno, buf, sizeof buf));
294 return;
297 if (__fxstat64 (_STAT_VER, fd, &st) < 0 || !S_ISREG (st.st_mode))
299 /* Not stat'able or not a regular file => don't use it. */
300 char buf[400];
301 int errnum = errno;
302 __close (fd);
303 _dl_error_printf ("%s: cannot stat file: %s\n", filename,
304 __strerror_r (errnum, buf, sizeof buf));
305 return;
308 /* Test the size. If it does not match what we expect from the size
309 values in the map MAP we don't use it and warn the user. */
310 if (st.st_size == 0)
312 /* We have to create the file. */
313 char buf[GL(dl_pagesize)];
315 memset (buf, '\0', GL(dl_pagesize));
317 if (__lseek (fd, expected_size & ~(GL(dl_pagesize) - 1), SEEK_SET) == -1)
319 char buf[400];
320 int errnum;
321 cannot_create:
322 errnum = errno;
323 __close (fd);
324 _dl_error_printf ("%s: cannot create file: %s\n", filename,
325 __strerror_r (errnum, buf, sizeof buf));
326 return;
329 if (TEMP_FAILURE_RETRY (__libc_write (fd, buf, (expected_size
330 & (GL(dl_pagesize)
331 - 1))))
332 < 0)
333 goto cannot_create;
335 else if (st.st_size != expected_size)
337 __close (fd);
338 wrong_format:
340 if (addr != NULL)
341 __munmap ((void *) addr, expected_size);
343 _dl_error_printf ("%s: file is no correct profile data file for `%s'\n",
344 filename, GL(dl_profile));
345 return;
348 addr = (struct gmon_hdr *) __mmap (NULL, expected_size, PROT_READ|PROT_WRITE,
349 MAP_SHARED|MAP_FILE, fd, 0);
350 if (addr == (struct gmon_hdr *) MAP_FAILED)
352 char buf[400];
353 int errnum = errno;
354 __close (fd);
355 _dl_error_printf ("%s: cannot map file: %s\n", filename,
356 __strerror_r (errnum, buf, sizeof buf));
357 return;
360 /* We don't need the file desriptor anymore. */
361 __close (fd);
363 /* Pointer to data after the header. */
364 hist = (char *) (addr + 1);
365 kcount = (uint16_t *) ((char *) hist + sizeof (uint32_t)
366 + sizeof (struct gmon_hist_hdr));
368 /* Compute pointer to array of the arc information. */
369 narcsp = (uint32_t *) ((char *) kcount + kcountsize + sizeof (uint32_t));
370 data = (struct here_cg_arc_record *) ((char *) narcsp + sizeof (uint32_t));
372 if (st.st_size == 0)
374 /* Create the signature. */
375 memcpy (addr, &gmon_hdr, sizeof (struct gmon_hdr));
377 *(uint32_t *) hist = GMON_TAG_TIME_HIST;
378 memcpy (hist + sizeof (uint32_t), &hist_hdr,
379 sizeof (struct gmon_hist_hdr));
381 narcsp[-1] = GMON_TAG_CG_ARC;
383 else
385 /* Test the signature in the file. */
386 if (memcmp (addr, &gmon_hdr, sizeof (struct gmon_hdr)) != 0
387 || *(uint32_t *) hist != GMON_TAG_TIME_HIST
388 || memcmp (hist + sizeof (uint32_t), &hist_hdr,
389 sizeof (struct gmon_hist_hdr)) != 0
390 || narcsp[-1] != GMON_TAG_CG_ARC)
391 goto wrong_format;
394 /* Allocate memory for the froms data and the pointer to the tos records. */
395 tos = (uint16_t *) calloc (tossize + fromssize, 1);
396 if (tos == NULL)
398 __munmap ((void *) addr, expected_size);
399 _dl_fatal_printf ("Out of memory while initializing profiler\n");
400 /* NOTREACHED */
403 froms = (struct here_fromstruct *) ((char *) tos + tossize);
404 fromidx = 0;
406 /* Now we have to process all the arc count entries. BTW: it is
407 not critical whether the *NARCSP value changes meanwhile. Before
408 we enter a new entry in to toset we will check that everything is
409 available in TOS. This happens in _dl_mcount.
411 Loading the entries in reverse order should help to get the most
412 frequently used entries at the front of the list. */
413 for (idx = narcs = MIN (*narcsp, fromlimit); idx > 0; )
415 size_t to_index;
416 size_t newfromidx;
417 --idx;
418 to_index = (data[idx].self_pc / (hashfraction * sizeof (*tos)));
419 newfromidx = fromidx++;
420 froms[newfromidx].here = &data[idx];
421 froms[newfromidx].link = tos[to_index];
422 tos[to_index] = newfromidx;
425 /* Setup counting data. */
426 if (kcountsize < highpc - lowpc)
428 #if 0
429 s_scale = ((double) kcountsize / (highpc - lowpc)) * SCALE_1_TO_1;
430 #else
431 size_t range = highpc - lowpc;
432 size_t quot = range / kcountsize;
434 if (quot >= SCALE_1_TO_1)
435 s_scale = 1;
436 else if (quot >= SCALE_1_TO_1 / 256)
437 s_scale = SCALE_1_TO_1 / quot;
438 else if (range > ULONG_MAX / 256)
439 s_scale = (SCALE_1_TO_1 * 256) / (range / (kcountsize / 256));
440 else
441 s_scale = (SCALE_1_TO_1 * 256) / ((range * 256) / kcountsize);
442 #endif
444 else
445 s_scale = SCALE_1_TO_1;
447 /* Start the profiler. */
448 __profil ((void *) kcount, kcountsize, lowpc, s_scale);
450 /* Turn on profiling. */
451 running = 1;
453 INTDEF (_dl_start_profile)
456 void
457 _dl_mcount (ElfW(Addr) frompc, ElfW(Addr) selfpc)
459 volatile uint16_t *topcindex;
460 size_t i, fromindex;
461 struct here_fromstruct *fromp;
463 if (! running)
464 return;
466 /* Compute relative addresses. The shared object can be loaded at
467 any address. The value of frompc could be anything. We cannot
468 restrict it in any way, just set to a fixed value (0) in case it
469 is outside the allowed range. These calls show up as calls from
470 <external> in the gprof output. */
471 frompc -= lowpc;
472 if (frompc >= textsize)
473 frompc = 0;
474 selfpc -= lowpc;
475 if (selfpc >= textsize)
476 goto done;
478 /* Getting here we now have to find out whether the location was
479 already used. If yes we are lucky and only have to increment a
480 counter (this also has to be atomic). If the entry is new things
481 are getting complicated... */
483 /* Avoid integer divide if possible. */
484 if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
485 i = selfpc >> log_hashfraction;
486 else
487 i = selfpc / (hashfraction * sizeof (*tos));
489 topcindex = &tos[i];
490 fromindex = *topcindex;
492 if (fromindex == 0)
493 goto check_new_or_add;
495 fromp = &froms[fromindex];
497 /* We have to look through the chain of arcs whether there is already
498 an entry for our arc. */
499 while (fromp->here->from_pc != frompc)
501 if (fromp->link != 0)
503 fromp = &froms[fromp->link];
504 while (fromp->link != 0 && fromp->here->from_pc != frompc);
506 if (fromp->here->from_pc != frompc)
508 topcindex = &fromp->link;
510 check_new_or_add:
511 /* Our entry is not among the entries we read so far from the
512 data file. Now see whether we have to update the list. */
513 while (narcs != *narcsp && narcs < fromlimit)
515 size_t to_index;
516 size_t newfromidx;
517 to_index = (data[narcs].self_pc
518 / (hashfraction * sizeof (*tos)));
519 newfromidx = exchange_and_add (&fromidx, 1) + 1;
520 froms[newfromidx].here = &data[narcs];
521 froms[newfromidx].link = tos[to_index];
522 tos[to_index] = newfromidx;
523 atomic_add (&narcs, 1);
526 /* If we still have no entry stop searching and insert. */
527 if (*topcindex == 0)
529 uint_fast32_t newarc = exchange_and_add (narcsp, 1);
531 /* In rare cases it could happen that all entries in FROMS are
532 occupied. So we cannot count this anymore. */
533 if (newarc >= fromlimit)
534 goto done;
536 *topcindex = exchange_and_add (&fromidx, 1) + 1;
537 fromp = &froms[*topcindex];
539 fromp->here = &data[newarc];
540 data[newarc].from_pc = frompc;
541 data[newarc].self_pc = selfpc;
542 data[newarc].count = 0;
543 fromp->link = 0;
544 atomic_add (&narcs, 1);
546 break;
549 fromp = &froms[*topcindex];
551 else
552 /* Found in. */
553 break;
556 /* Increment the counter. */
557 atomic_add (&fromp->here->count, 1);
559 done:
562 INTDEF(_dl_mcount)