1 /* Profiling of shared libraries.
2 Copyright (C) 1997-2023 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4 Based on the BSD mcount implementation.
5
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public
8 License as published by the Free Software Foundation; either
9 version 2.1 of the License, or (at your option) any later version.
10
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
15
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; if not, see
18 <https://www.gnu.org/licenses/>. */
19
20 #include <assert.h>
21 #include <errno.h>
22 #include <fcntl.h>
23 #include <inttypes.h>
24 #include <limits.h>
25 #include <stdio.h>
26 #include <stdlib.h>
27 #include <string.h>
28 #include <unistd.h>
29 #include <stdint.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 <atomic.h>
37 #include <not-cancel.h>
38
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 simultaneously 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.
49
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.
55
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.
62
63 Therefore we use a new format. This will look like this
64
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
71
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
84
85 0030+2*A ?? ?? ?? ?? <- Count data
86 ... ...
87 0030+2*A+K ?? ?? ?? ??
88
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
102
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.
105
106 `K' is the number of count entries which is computed as
107
108 textsize / HISTFRACTION
109
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
115
116 MINARCS <= textsize*(ARCDENSITY/100) <= MAXARCS
117
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. */
123
124 extern int __profile_frequency (void);
125 libc_hidden_proto (__profile_frequency)
126
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
131 {
132 uintptr_t from_pc;
133 uintptr_t self_pc;
134 /* The count field is atomically incremented in _dl_mcount, which
135 requires it to be properly aligned for its type, and for this
136 alignment to be visible to the compiler. The amount of data
137 before an array of this structure is calculated as
138 expected_size in _dl_start_profile. Everything in that
139 calculation is a multiple of 4 bytes (in the case of
140 kcountsize, because it is derived from a subtraction of
141 page-aligned values, and the corresponding calculation in
142 __monstartup also ensures it is at least a multiple of the size
143 of u_long), so all copies of this field do in fact have the
144 appropriate alignment. */
145 uint32_t count __attribute__ ((aligned (__alignof__ (uint32_t))));
146 } __attribute__ ((packed));
147
148 static struct here_cg_arc_record *data;
149
150 /* Nonzero if profiling is under way. */
151 static int running;
152
153 /* This is the number of entry which have been incorporated in the toset. */
154 static uint32_t narcs;
155 /* This is a pointer to the object representing the number of entries
156 currently in the mmaped file. At no point of time this has to be the
157 same as NARCS. If it is equal all entries from the file are in our
158 lists. */
159 static volatile uint32_t *narcsp;
160
161
162 struct here_fromstruct
163 {
164 struct here_cg_arc_record volatile *here;
165 uint16_t link;
166 };
167
168 static volatile uint16_t *tos;
169
170 static struct here_fromstruct *froms;
171 static uint32_t fromlimit;
172 static volatile uint32_t fromidx;
173
174 static uintptr_t lowpc;
175 static size_t textsize;
176 static unsigned int log_hashfraction;
177
178
179
180 /* Set up profiling data to profile object described by MAP. The output
181 file is found (or created) in OUTPUT_DIR. */
182 void
183 _dl_start_profile (void)
184 {
185 char *filename;
186 int fd;
187 struct __stat64_t64 st;
188 const ElfW(Phdr) *ph;
189 ElfW(Addr) mapstart = ~((ElfW(Addr)) 0);
190 ElfW(Addr) mapend = 0;
191 char *hist, *cp;
192 size_t idx;
193 size_t tossize;
194 size_t fromssize;
195 uintptr_t highpc;
196 uint16_t *kcount;
197 size_t kcountsize;
198 struct gmon_hdr *addr = NULL;
199 off_t expected_size;
200 /* See profil(2) where this is described. */
201 int s_scale;
202 #define SCALE_1_TO_1 0x10000L
203 const char *errstr = NULL;
204
205 /* Compute the size of the sections which contain program code. */
206 for (ph = GL(dl_profile_map)->l_phdr;
207 ph < &GL(dl_profile_map)->l_phdr[GL(dl_profile_map)->l_phnum]; ++ph)
208 if (ph->p_type == PT_LOAD && (ph->p_flags & PF_X))
209 {
210 ElfW(Addr) start = (ph->p_vaddr & ~(GLRO(dl_pagesize) - 1));
211 ElfW(Addr) end = ((ph->p_vaddr + ph->p_memsz + GLRO(dl_pagesize) - 1)
212 & ~(GLRO(dl_pagesize) - 1));
213
214 if (start < mapstart)
215 mapstart = start;
216 if (end > mapend)
217 mapend = end;
218 }
219
220 /* Now we can compute the size of the profiling data. This is done
221 with the same formulas as in `monstartup' (see gmon.c). */
222 running = 0;
223 lowpc = ROUNDDOWN (mapstart + GL(dl_profile_map)->l_addr,
224 HISTFRACTION * sizeof (HISTCOUNTER));
225 highpc = ROUNDUP (mapend + GL(dl_profile_map)->l_addr,
226 HISTFRACTION * sizeof (HISTCOUNTER));
227 textsize = highpc - lowpc;
228 kcountsize = textsize / HISTFRACTION;
229 if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
230 {
231 /* If HASHFRACTION is a power of two, mcount can use shifting
232 instead of integer division. Precompute shift amount.
233
234 This is a constant but the compiler cannot compile the
235 expression away since the __ffs implementation is not known
236 to the compiler. Help the compiler by precomputing the
237 usual cases. */
238 assert (HASHFRACTION == 2);
239
240 if (sizeof (*froms) == 8)
241 log_hashfraction = 4;
242 else if (sizeof (*froms) == 16)
243 log_hashfraction = 5;
244 else
245 log_hashfraction = __ffs (HASHFRACTION * sizeof (*froms)) - 1;
246 }
247 else
248 log_hashfraction = -1;
249 tossize = textsize / HASHFRACTION;
250 fromlimit = textsize * ARCDENSITY / 100;
251 if (fromlimit < MINARCS)
252 fromlimit = MINARCS;
253 if (fromlimit > MAXARCS)
254 fromlimit = MAXARCS;
255 fromssize = fromlimit * sizeof (struct here_fromstruct);
256
257 expected_size = (sizeof (struct gmon_hdr)
258 + 4 + sizeof (struct gmon_hist_hdr) + kcountsize
259 + 4 + 4 + fromssize * sizeof (struct here_cg_arc_record));
260
261 /* Create the gmon_hdr we expect or write. */
262 struct real_gmon_hdr
263 {
264 char cookie[4];
265 int32_t version;
266 char spare[3 * 4];
267 } gmon_hdr;
268 if (sizeof (gmon_hdr) != sizeof (struct gmon_hdr)
269 || (offsetof (struct real_gmon_hdr, cookie)
270 != offsetof (struct gmon_hdr, cookie))
271 || (offsetof (struct real_gmon_hdr, version)
272 != offsetof (struct gmon_hdr, version)))
273 abort ();
274
275 memcpy (&gmon_hdr.cookie[0], GMON_MAGIC, sizeof (gmon_hdr.cookie));
276 gmon_hdr.version = GMON_SHOBJ_VERSION;
277 memset (gmon_hdr.spare, '\0', sizeof (gmon_hdr.spare));
278
279 /* Create the hist_hdr we expect or write. */
280 struct real_gmon_hist_hdr
281 {
282 char *low_pc;
283 char *high_pc;
284 int32_t hist_size;
285 int32_t prof_rate;
286 char dimen[15];
287 char dimen_abbrev;
288 } hist_hdr;
289 if (sizeof (hist_hdr) != sizeof (struct gmon_hist_hdr)
290 || (offsetof (struct real_gmon_hist_hdr, low_pc)
291 != offsetof (struct gmon_hist_hdr, low_pc))
292 || (offsetof (struct real_gmon_hist_hdr, high_pc)
293 != offsetof (struct gmon_hist_hdr, high_pc))
294 || (offsetof (struct real_gmon_hist_hdr, hist_size)
295 != offsetof (struct gmon_hist_hdr, hist_size))
296 || (offsetof (struct real_gmon_hist_hdr, prof_rate)
297 != offsetof (struct gmon_hist_hdr, prof_rate))
298 || (offsetof (struct real_gmon_hist_hdr, dimen)
299 != offsetof (struct gmon_hist_hdr, dimen))
300 || (offsetof (struct real_gmon_hist_hdr, dimen_abbrev)
301 != offsetof (struct gmon_hist_hdr, dimen_abbrev)))
302 abort ();
303
304 hist_hdr.low_pc = (char *) mapstart;
305 hist_hdr.high_pc = (char *) mapend;
306 hist_hdr.hist_size = kcountsize / sizeof (HISTCOUNTER);
307 hist_hdr.prof_rate = __profile_frequency ();
308 if (sizeof (hist_hdr.dimen) >= sizeof ("seconds"))
309 {
310 memcpy (hist_hdr.dimen, "seconds", sizeof ("seconds"));
311 memset (hist_hdr.dimen + sizeof ("seconds"), '\0',
312 sizeof (hist_hdr.dimen) - sizeof ("seconds"));
313 }
314 else
315 strncpy (hist_hdr.dimen, "seconds", sizeof (hist_hdr.dimen));
316 hist_hdr.dimen_abbrev = 's';
317
318 /* First determine the output name. We write in the directory
319 OUTPUT_DIR and the name is composed from the shared objects
320 soname (or the file name) and the ending ".profile". */
321 filename = (char *) alloca (strlen (GLRO(dl_profile_output)) + 1
322 + strlen (GLRO(dl_profile)) + sizeof ".profile");
323 cp = __stpcpy (filename, GLRO(dl_profile_output));
324 *cp++ = '/';
325 __stpcpy (__stpcpy (cp, GLRO(dl_profile)), ".profile");
326
327 fd = __open64_nocancel (filename, O_RDWR | O_CREAT | O_NOFOLLOW
328 | O_CLOEXEC, DEFFILEMODE);
329 if (fd == -1)
330 {
331 char buf[400];
332 int errnum;
333
334 /* We cannot write the profiling data so don't do anything. */
335 errstr = "%s: cannot open file: %s\n";
336 print_error:
337 errnum = errno;
338 if (fd != -1)
339 __close_nocancel (fd);
340 _dl_error_printf (errstr, filename,
341 __strerror_r (errnum, buf, sizeof buf));
342 return;
343 }
344
345 if (__fstat64_time64 (fd, &st) < 0 || !S_ISREG (st.st_mode))
346 {
347 /* Not stat'able or not a regular file => don't use it. */
348 errstr = "%s: cannot stat file: %s\n";
349 goto print_error;
350 }
351
352 /* Test the size. If it does not match what we expect from the size
353 values in the map MAP we don't use it and warn the user. */
354 if (st.st_size == 0)
355 {
356 /* We have to create the file. */
357 char buf[GLRO(dl_pagesize)];
358
359 memset (buf, '\0', GLRO(dl_pagesize));
360
361 if (__lseek (fd, expected_size & ~(GLRO(dl_pagesize) - 1), SEEK_SET) == -1)
362 {
363 cannot_create:
364 errstr = "%s: cannot create file: %s\n";
365 goto print_error;
366 }
367
368 if (TEMP_FAILURE_RETRY
369 (__write_nocancel (fd, buf, (expected_size & (GLRO(dl_pagesize) - 1))))
370 < 0)
371 goto cannot_create;
372 }
373 else if (st.st_size != expected_size)
374 {
375 __close_nocancel (fd);
376 wrong_format:
377
378 if (addr != NULL)
379 __munmap ((void *) addr, expected_size);
380
381 _dl_error_printf ("%s: file is no correct profile data file for `%s'\n",
382 filename, GLRO(dl_profile));
383 return;
384 }
385
386 addr = (struct gmon_hdr *) __mmap (NULL, expected_size, PROT_READ|PROT_WRITE,
387 MAP_SHARED|MAP_FILE, fd, 0);
388 if (addr == (struct gmon_hdr *) MAP_FAILED)
389 {
390 errstr = "%s: cannot map file: %s\n";
391 goto print_error;
392 }
393
394 /* We don't need the file descriptor anymore. */
395 __close_nocancel (fd);
396
397 /* Pointer to data after the header. */
398 hist = (char *) (addr + 1);
399 kcount = (uint16_t *) ((char *) hist + sizeof (uint32_t)
400 + sizeof (struct gmon_hist_hdr));
401
402 /* Compute pointer to array of the arc information. */
403 narcsp = (uint32_t *) ((char *) kcount + kcountsize + sizeof (uint32_t));
404 data = (struct here_cg_arc_record *) ((char *) narcsp + sizeof (uint32_t));
405
406 if (st.st_size == 0)
407 {
408 /* Create the signature. */
409 memcpy (addr, &gmon_hdr, sizeof (struct gmon_hdr));
410
411 *(uint32_t *) hist = GMON_TAG_TIME_HIST;
412 memcpy (hist + sizeof (uint32_t), &hist_hdr,
413 sizeof (struct gmon_hist_hdr));
414
415 narcsp[-1] = GMON_TAG_CG_ARC;
416 }
417 else
418 {
419 /* Test the signature in the file. */
420 if (memcmp (addr, &gmon_hdr, sizeof (struct gmon_hdr)) != 0
421 || *(uint32_t *) hist != GMON_TAG_TIME_HIST
422 || memcmp (hist + sizeof (uint32_t), &hist_hdr,
423 sizeof (struct gmon_hist_hdr)) != 0
424 || narcsp[-1] != GMON_TAG_CG_ARC)
425 goto wrong_format;
426 }
427
428 /* Allocate memory for the froms data and the pointer to the tos records. */
429 tos = (uint16_t *) calloc (tossize + fromssize, 1);
430 if (tos == NULL)
431 {
432 __munmap ((void *) addr, expected_size);
433 _dl_fatal_printf ("Out of memory while initializing profiler\n");
434 /* NOTREACHED */
435 }
436
437 froms = (struct here_fromstruct *) ((char *) tos + tossize);
438 fromidx = 0;
439
440 /* Now we have to process all the arc count entries. BTW: it is
441 not critical whether the *NARCSP value changes meanwhile. Before
442 we enter a new entry in to toset we will check that everything is
443 available in TOS. This happens in _dl_mcount.
444
445 Loading the entries in reverse order should help to get the most
446 frequently used entries at the front of the list. */
447 for (idx = narcs = MIN (*narcsp, fromlimit); idx > 0; )
448 {
449 size_t to_index;
450 size_t newfromidx;
451 --idx;
452 to_index = (data[idx].self_pc / (HASHFRACTION * sizeof (*tos)));
453 newfromidx = fromidx++;
454 froms[newfromidx].here = &data[idx];
455 froms[newfromidx].link = tos[to_index];
456 tos[to_index] = newfromidx;
457 }
458
459 /* Setup counting data. */
460 if (kcountsize < highpc - lowpc)
461 {
462 #if 0
463 s_scale = ((double) kcountsize / (highpc - lowpc)) * SCALE_1_TO_1;
464 #else
465 size_t range = highpc - lowpc;
466 size_t quot = range / kcountsize;
467
468 if (quot >= SCALE_1_TO_1)
469 s_scale = 1;
470 else if (quot >= SCALE_1_TO_1 / 256)
471 s_scale = SCALE_1_TO_1 / quot;
472 else if (range > ULONG_MAX / 256)
473 s_scale = (SCALE_1_TO_1 * 256) / (range / (kcountsize / 256));
474 else
475 s_scale = (SCALE_1_TO_1 * 256) / ((range * 256) / kcountsize);
476 #endif
477 }
478 else
479 s_scale = SCALE_1_TO_1;
480
481 /* Start the profiler. */
482 __profil ((void *) kcount, kcountsize, lowpc, s_scale);
483
484 /* Turn on profiling. */
485 running = 1;
486 }
487
488
489 void
490 _dl_mcount (ElfW(Addr) frompc, ElfW(Addr) selfpc)
491 {
492 volatile uint16_t *topcindex;
493 size_t i, fromindex;
494 struct here_fromstruct *fromp;
495
496 if (! running)
497 return;
498
499 /* Compute relative addresses. The shared object can be loaded at
500 any address. The value of frompc could be anything. We cannot
501 restrict it in any way, just set to a fixed value (0) in case it
502 is outside the allowed range. These calls show up as calls from
503 <external> in the gprof output. */
504 frompc -= lowpc;
505 if (frompc >= textsize)
506 frompc = 0;
507 selfpc -= lowpc;
508 if (selfpc >= textsize)
509 goto done;
510
511 /* Getting here we now have to find out whether the location was
512 already used. If yes we are lucky and only have to increment a
513 counter (this also has to be atomic). If the entry is new things
514 are getting complicated... */
515
516 /* Avoid integer divide if possible. */
517 if ((HASHFRACTION & (HASHFRACTION - 1)) == 0)
518 i = selfpc >> log_hashfraction;
519 else
520 i = selfpc / (HASHFRACTION * sizeof (*tos));
521
522 topcindex = &tos[i];
523 fromindex = *topcindex;
524
525 if (fromindex == 0)
526 goto check_new_or_add;
527
528 fromp = &froms[fromindex];
529
530 /* We have to look through the chain of arcs whether there is already
531 an entry for our arc. */
532 while (fromp->here->from_pc != frompc)
533 {
534 if (fromp->link != 0)
535 do
536 fromp = &froms[fromp->link];
537 while (fromp->link != 0 && fromp->here->from_pc != frompc);
538
539 if (fromp->here->from_pc != frompc)
540 {
541 topcindex = &fromp->link;
542
543 check_new_or_add:
544 /* Our entry is not among the entries we read so far from the
545 data file. Now see whether we have to update the list. */
546 while (narcs != *narcsp && narcs < fromlimit)
547 {
548 size_t to_index;
549 size_t newfromidx;
550 to_index = (data[narcs].self_pc
551 / (HASHFRACTION * sizeof (*tos)));
552 newfromidx = atomic_fetch_add_relaxed (&fromidx, 1) + 1;
553 froms[newfromidx].here = &data[narcs];
554 froms[newfromidx].link = tos[to_index];
555 tos[to_index] = newfromidx;
556 catomic_increment (&narcs);
557 }
558
559 /* If we still have no entry stop searching and insert. */
560 if (*topcindex == 0)
561 {
562 unsigned int newarc = atomic_fetch_add_relaxed (narcsp, 1);
563
564 /* In rare cases it could happen that all entries in FROMS are
565 occupied. So we cannot count this anymore. */
566 if (newarc >= fromlimit)
567 goto done;
568
569 *topcindex = atomic_fetch_add_relaxed (&fromidx, 1) + 1;
570 fromp = &froms[*topcindex];
571
572 fromp->here = &data[newarc];
573 data[newarc].from_pc = frompc;
574 data[newarc].self_pc = selfpc;
575 data[newarc].count = 0;
576 fromp->link = 0;
577 atomic_fetch_add_relaxed (&narcs, 1);
578
579 break;
580 }
581
582 fromp = &froms[*topcindex];
583 }
584 else
585 /* Found in. */
586 break;
587 }
588
589 /* Increment the counter. */
590 atomic_fetch_add_relaxed (&fromp->here->count, 1);
591
592 done:
593 ;
594 }
595 rtld_hidden_def (_dl_mcount)