Define old CPU_* macros in temers of __CPU_*_S macros. Define CPU_*_S macros.
[glibc/pb-stable.git] / elf / dl-profile.c
blob47033f32eff6a4ed55cc65dd5c285095767ba95d
1 /* Profiling of shared libraries.
2 Copyright (C) 1997-2002, 2003, 2004, 2006 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 <atomic.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;
151 struct here_fromstruct
153 struct here_cg_arc_record volatile *here;
154 uint16_t link;
157 static volatile uint16_t *tos;
159 static struct here_fromstruct *froms;
160 static uint32_t fromlimit;
161 static volatile uint32_t fromidx;
163 static uintptr_t lowpc;
164 static size_t textsize;
165 static unsigned int log_hashfraction;
169 /* Set up profiling data to profile object desribed by MAP. The output
170 file is found (or created) in OUTPUT_DIR. */
171 void
172 internal_function
173 _dl_start_profile (void)
175 char *filename;
176 int fd;
177 struct stat64 st;
178 const ElfW(Phdr) *ph;
179 ElfW(Addr) mapstart = ~((ElfW(Addr)) 0);
180 ElfW(Addr) mapend = 0;
181 struct gmon_hdr gmon_hdr;
182 struct gmon_hist_hdr hist_hdr;
183 char *hist, *cp;
184 size_t idx;
185 size_t tossize;
186 size_t fromssize;
187 uintptr_t highpc;
188 uint16_t *kcount;
189 size_t kcountsize;
190 struct gmon_hdr *addr = NULL;
191 off_t expected_size;
192 /* See profil(2) where this is described. */
193 int s_scale;
194 #define SCALE_1_TO_1 0x10000L
195 const char *errstr = NULL;
197 /* Compute the size of the sections which contain program code. */
198 for (ph = GL(dl_profile_map)->l_phdr;
199 ph < &GL(dl_profile_map)->l_phdr[GL(dl_profile_map)->l_phnum]; ++ph)
200 if (ph->p_type == PT_LOAD && (ph->p_flags & PF_X))
202 ElfW(Addr) start = (ph->p_vaddr & ~(GLRO(dl_pagesize) - 1));
203 ElfW(Addr) end = ((ph->p_vaddr + ph->p_memsz + GLRO(dl_pagesize) - 1)
204 & ~(GLRO(dl_pagesize) - 1));
206 if (start < mapstart)
207 mapstart = start;
208 if (end > mapend)
209 mapend = end;
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 running = 0;
215 lowpc = ROUNDDOWN (mapstart + GL(dl_profile_map)->l_addr,
216 HISTFRACTION * sizeof (HISTCOUNTER));
217 highpc = ROUNDUP (mapend + GL(dl_profile_map)->l_addr,
218 HISTFRACTION * sizeof (HISTCOUNTER));
219 textsize = highpc - lowpc;
220 kcountsize = textsize / HISTFRACTION;
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 (GLRO(dl_profile_output)) + 1
277 + strlen (GLRO(dl_profile)) + sizeof ".profile");
278 cp = __stpcpy (filename, GLRO(dl_profile_output));
279 *cp++ = '/';
280 __stpcpy (__stpcpy (cp, GLRO(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 char buf[400];
291 int errnum;
293 /* We cannot write the profiling data so don't do anything. */
294 errstr = "%s: cannot open file: %s\n";
295 print_error:
296 errnum = errno;
297 if (fd != -1)
298 __close (fd);
299 _dl_error_printf (errstr, filename,
300 __strerror_r (errnum, buf, sizeof buf));
301 return;
304 if (__fxstat64 (_STAT_VER, fd, &st) < 0 || !S_ISREG (st.st_mode))
306 /* Not stat'able or not a regular file => don't use it. */
307 errstr = "%s: cannot stat file: %s\n";
308 goto print_error;
311 /* Test the size. If it does not match what we expect from the size
312 values in the map MAP we don't use it and warn the user. */
313 if (st.st_size == 0)
315 /* We have to create the file. */
316 char buf[GLRO(dl_pagesize)];
318 memset (buf, '\0', GLRO(dl_pagesize));
320 if (__lseek (fd, expected_size & ~(GLRO(dl_pagesize) - 1), SEEK_SET) == -1)
322 cannot_create:
323 errstr = "%s: cannot create file: %s\n";
324 goto print_error;
327 if (TEMP_FAILURE_RETRY (__libc_write (fd, buf, (expected_size
328 & (GLRO(dl_pagesize)
329 - 1))))
330 < 0)
331 goto cannot_create;
333 else if (st.st_size != expected_size)
335 __close (fd);
336 wrong_format:
338 if (addr != NULL)
339 __munmap ((void *) addr, expected_size);
341 _dl_error_printf ("%s: file is no correct profile data file for `%s'\n",
342 filename, GLRO(dl_profile));
343 return;
346 addr = (struct gmon_hdr *) __mmap (NULL, expected_size, PROT_READ|PROT_WRITE,
347 MAP_SHARED|MAP_FILE, fd, 0);
348 if (addr == (struct gmon_hdr *) MAP_FAILED)
350 errstr = "%s: cannot map file: %s\n";
351 goto print_error;
354 /* We don't need the file descriptor anymore. */
355 __close (fd);
357 /* Pointer to data after the header. */
358 hist = (char *) (addr + 1);
359 kcount = (uint16_t *) ((char *) hist + sizeof (uint32_t)
360 + sizeof (struct gmon_hist_hdr));
362 /* Compute pointer to array of the arc information. */
363 narcsp = (uint32_t *) ((char *) kcount + kcountsize + sizeof (uint32_t));
364 data = (struct here_cg_arc_record *) ((char *) narcsp + sizeof (uint32_t));
366 if (st.st_size == 0)
368 /* Create the signature. */
369 memcpy (addr, &gmon_hdr, sizeof (struct gmon_hdr));
371 *(uint32_t *) hist = GMON_TAG_TIME_HIST;
372 memcpy (hist + sizeof (uint32_t), &hist_hdr,
373 sizeof (struct gmon_hist_hdr));
375 narcsp[-1] = GMON_TAG_CG_ARC;
377 else
379 /* Test the signature in the file. */
380 if (memcmp (addr, &gmon_hdr, sizeof (struct gmon_hdr)) != 0
381 || *(uint32_t *) hist != GMON_TAG_TIME_HIST
382 || memcmp (hist + sizeof (uint32_t), &hist_hdr,
383 sizeof (struct gmon_hist_hdr)) != 0
384 || narcsp[-1] != GMON_TAG_CG_ARC)
385 goto wrong_format;
388 /* Allocate memory for the froms data and the pointer to the tos records. */
389 tos = (uint16_t *) calloc (tossize + fromssize, 1);
390 if (tos == NULL)
392 __munmap ((void *) addr, expected_size);
393 _dl_fatal_printf ("Out of memory while initializing profiler\n");
394 /* NOTREACHED */
397 froms = (struct here_fromstruct *) ((char *) tos + tossize);
398 fromidx = 0;
400 /* Now we have to process all the arc count entries. BTW: it is
401 not critical whether the *NARCSP value changes meanwhile. Before
402 we enter a new entry in to toset we will check that everything is
403 available in TOS. This happens in _dl_mcount.
405 Loading the entries in reverse order should help to get the most
406 frequently used entries at the front of the list. */
407 for (idx = narcs = MIN (*narcsp, fromlimit); idx > 0; )
409 size_t to_index;
410 size_t newfromidx;
411 --idx;
412 to_index = (data[idx].self_pc / (HASHFRACTION * sizeof (*tos)));
413 newfromidx = fromidx++;
414 froms[newfromidx].here = &data[idx];
415 froms[newfromidx].link = tos[to_index];
416 tos[to_index] = newfromidx;
419 /* Setup counting data. */
420 if (kcountsize < highpc - lowpc)
422 #if 0
423 s_scale = ((double) kcountsize / (highpc - lowpc)) * SCALE_1_TO_1;
424 #else
425 size_t range = highpc - lowpc;
426 size_t quot = range / kcountsize;
428 if (quot >= SCALE_1_TO_1)
429 s_scale = 1;
430 else if (quot >= SCALE_1_TO_1 / 256)
431 s_scale = SCALE_1_TO_1 / quot;
432 else if (range > ULONG_MAX / 256)
433 s_scale = (SCALE_1_TO_1 * 256) / (range / (kcountsize / 256));
434 else
435 s_scale = (SCALE_1_TO_1 * 256) / ((range * 256) / kcountsize);
436 #endif
438 else
439 s_scale = SCALE_1_TO_1;
441 /* Start the profiler. */
442 __profil ((void *) kcount, kcountsize, lowpc, s_scale);
444 /* Turn on profiling. */
445 running = 1;
449 void
450 _dl_mcount (ElfW(Addr) frompc, ElfW(Addr) selfpc)
452 volatile uint16_t *topcindex;
453 size_t i, fromindex;
454 struct here_fromstruct *fromp;
456 if (! running)
457 return;
459 /* Compute relative addresses. The shared object can be loaded at
460 any address. The value of frompc could be anything. We cannot
461 restrict it in any way, just set to a fixed value (0) in case it
462 is outside the allowed range. These calls show up as calls from
463 <external> in the gprof output. */
464 frompc -= lowpc;
465 if (frompc >= textsize)
466 frompc = 0;
467 selfpc -= lowpc;
468 if (selfpc >= textsize)
469 goto done;
471 /* Getting here we now have to find out whether the location was
472 already used. If yes we are lucky and only have to increment a
473 counter (this also has to be atomic). If the entry is new things
474 are getting complicated... */
476 /* Avoid integer divide if possible. */
477 if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
478 i = selfpc >> log_hashfraction;
479 else
480 i = selfpc / (HASHFRACTION * sizeof (*tos));
482 topcindex = &tos[i];
483 fromindex = *topcindex;
485 if (fromindex == 0)
486 goto check_new_or_add;
488 fromp = &froms[fromindex];
490 /* We have to look through the chain of arcs whether there is already
491 an entry for our arc. */
492 while (fromp->here->from_pc != frompc)
494 if (fromp->link != 0)
496 fromp = &froms[fromp->link];
497 while (fromp->link != 0 && fromp->here->from_pc != frompc);
499 if (fromp->here->from_pc != frompc)
501 topcindex = &fromp->link;
503 check_new_or_add:
504 /* Our entry is not among the entries we read so far from the
505 data file. Now see whether we have to update the list. */
506 while (narcs != *narcsp && narcs < fromlimit)
508 size_t to_index;
509 size_t newfromidx;
510 to_index = (data[narcs].self_pc
511 / (HASHFRACTION * sizeof (*tos)));
512 newfromidx = catomic_exchange_and_add (&fromidx, 1) + 1;
513 froms[newfromidx].here = &data[narcs];
514 froms[newfromidx].link = tos[to_index];
515 tos[to_index] = newfromidx;
516 catomic_increment (&narcs);
519 /* If we still have no entry stop searching and insert. */
520 if (*topcindex == 0)
522 uint_fast32_t newarc = catomic_exchange_and_add (narcsp, 1);
524 /* In rare cases it could happen that all entries in FROMS are
525 occupied. So we cannot count this anymore. */
526 if (newarc >= fromlimit)
527 goto done;
529 *topcindex = catomic_exchange_and_add (&fromidx, 1) + 1;
530 fromp = &froms[*topcindex];
532 fromp->here = &data[newarc];
533 data[newarc].from_pc = frompc;
534 data[newarc].self_pc = selfpc;
535 data[newarc].count = 0;
536 fromp->link = 0;
537 catomic_increment (&narcs);
539 break;
542 fromp = &froms[*topcindex];
544 else
545 /* Found in. */
546 break;
549 /* Increment the counter. */
550 catomic_increment (&fromp->here->count);
552 done:
555 INTDEF(_dl_mcount)