1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
24 */
25
26 /*
27 * Routines to provide profiling of shared libraries required by the called
28 * executable.
29 */
30 #include <stdio.h>
31 #include <fcntl.h>
32 #include <sys/mman.h>
33 #include <unistd.h>
34 #include <stdlib.h>
35 #include <string.h>
36 #include <sys/types.h>
37 #include <sys/stat.h>
38 #include <synch.h>
39 #include <signal.h>
40 #include <synch.h>
41 #include <link.h>
42 #include <libintl.h>
43 #include <sys/param.h>
44 #include <procfs.h>
45 #include "msg.h"
46 #include "sgs.h"
47 #include "profile.h"
48 #include "_rtld.h"
49
50
51 static char Profile[MAXPATHLEN]; /* Profile buffer pathname */
52 static char *pname = 0; /* name of object to profile */
53 static L_hdr *Hptr; /* profile buffer header pointer */
54 static L_cgarc *Cptr; /* profile buffer call graph pointer */
55 static caddr_t Hpc, Lpc; /* Range of addresses being monitored */
56 static size_t Fsize; /* Size of mapped in profile buffer */
57 uintptr_t profcookie = 0;
58
59 /*
60 * When handling mutex's locally we need to mask signals. The signal
61 * mask is for everything except SIGWAITING.
62 */
63 static const sigset_t iset = { ~0U, ~0U, ~0U, ~0U };
64
65 static lwp_mutex_t sharedmutex = SHAREDMUTEX;
66
67 static int
prof_mutex_init(lwp_mutex_t * mp)68 prof_mutex_init(lwp_mutex_t *mp)
69 {
70 (void) memcpy(mp, &sharedmutex, sizeof (lwp_mutex_t));
71 return (0);
72 }
73
74 static int
prof_mutex_lock(lwp_mutex_t * mp,sigset_t * oset)75 prof_mutex_lock(lwp_mutex_t *mp, sigset_t *oset)
76 {
77 if (oset)
78 (void) sigprocmask(SIG_BLOCK, &iset, oset);
79 (void) _lwp_mutex_lock(mp);
80 return (0);
81 }
82
83 static int
prof_mutex_unlock(mutex_t * mp,sigset_t * oset)84 prof_mutex_unlock(mutex_t *mp, sigset_t *oset)
85 {
86 (void) _lwp_mutex_unlock(mp);
87 if (oset)
88 (void) sigprocmask(SIG_SETMASK, oset, NULL);
89 return (0);
90 }
91
92 const char *
_ldprof_msg(Msg mid)93 _ldprof_msg(Msg mid)
94 {
95 return (dgettext(MSG_ORIG(MSG_SUNW_OST_SGS), MSG_ORIG(mid)));
96 }
97
98 /*
99 * Determine whether a set (of arbitrary size) is in use - used to analyze proc
100 * status information.
101 */
102 static int
setisinuse(uint32_t * sp,uint_t n)103 setisinuse(uint32_t *sp, uint_t n)
104 {
105 while (n--)
106 if (*sp++)
107 return (1);
108 return (0);
109 }
110
111 #define prisinuse(sp) \
112 setisinuse((uint32_t *)(sp), \
113 (uint_t)(sizeof (*(sp)) / sizeof (uint32_t)))
114
115 uint_t
la_version(uint_t version)116 la_version(uint_t version)
117 {
118 int fd;
119 ssize_t num;
120 pstatus_t status;
121
122 if (version < LAV_CURRENT) {
123 (void) fprintf(stderr, MSG_INTL(MSG_GEN_AUDITVERSION),
124 LAV_CURRENT, version);
125 return (LAV_CURRENT);
126 }
127
128 /*
129 * To reduce the potential for deadlock conditions that can arise from
130 * being monitored (say by truss(1)) while setting a lock in the profile
131 * buffer, determine if someone is monitoring us. If so silently
132 * disable profiling.
133 */
134 if ((fd = open(MSG_ORIG(MSG_FMT_PROCSELF), O_RDONLY)) < 0)
135 return (LAV_CURRENT);
136
137 num = read(fd, &status, sizeof (status));
138 (void) close(fd);
139
140 if ((num != sizeof (status)) ||
141 prisinuse(&status.pr_sigtrace) || prisinuse(&status.pr_flttrace) ||
142 prisinuse(&status.pr_sysentry) || prisinuse(&status.pr_sysexit)) {
143 return (LAV_CURRENT);
144 }
145
146 /*
147 * We're presently not being monitored (although there's no control of
148 * someone attaching to us later), so retrieve the profile target name.
149 */
150 if (dlinfo((void *)NULL, RTLD_DI_PROFILENAME, &pname) == -1)
151 (void) fprintf(stderr, MSG_INTL(MSG_GEN_PROFNOTSET));
152
153 return (LAV_CURRENT);
154 }
155
156
157 int
profile_open(const char * fname,Link_map * lmp)158 profile_open(const char *fname, Link_map *lmp)
159 {
160 size_t hsize; /* struct hdr size */
161 size_t psize; /* profile histogram size */
162 size_t csize; /* call graph array size */
163 size_t msize; /* size of memory being profiled */
164 int i, fd, fixed = 0;
165 caddr_t lpc;
166 caddr_t hpc;
167 caddr_t addr;
168 struct stat status;
169 int new_buffer = 0;
170 sigset_t mask;
171 int err;
172 Ehdr * ehdr; /* ELF header for file */
173 Phdr * phdr; /* program headers for file */
174 Dyn * dynp = 0; /* Dynamic section */
175 Word nsym = 0; /* no. of symtab ntries */
176
177 if (*Profile == '\0') {
178 const char *dir, *suf;
179 char *tmp;
180
181 /*
182 * From the basename of the specified filename generate the
183 * appropriate profile buffer name. The profile file is created
184 * if it does not already exist.
185 */
186 if (((tmp = strrchr(fname, '/')) != 0) && (*(++tmp)))
187 fname = tmp;
188
189 #if defined(_ELF64)
190 suf = MSG_ORIG(MSG_SUF_PROFILE_64);
191 #else
192 suf = MSG_ORIG(MSG_SUF_PROFILE);
193 #endif
194 if (dlinfo((void *)NULL, RTLD_DI_PROFILEOUT, &dir) == -1)
195 dir = MSG_ORIG(MSG_PTH_VARTMP);
196
197 (void) snprintf(Profile, MAXPATHLEN, MSG_ORIG(MSG_FMT_PROFILE),
198 dir, fname, suf);
199 }
200
201 if ((fd = open(Profile, (O_RDWR | O_CREAT), 0666)) == -1) {
202 err = errno;
203 (void) fprintf(stderr, MSG_INTL(MSG_SYS_OPEN), Profile,
204 strerror(err));
205 return (0);
206 }
207
208 /*
209 * Now we determine the valid pc range for this object. The lpc is easy
210 * (lmp->l_addr), to determine the hpc we must examine the Phdrs.
211 */
212 lpc = hpc = (caddr_t)lmp->l_addr;
213 /* LINTED */
214 ehdr = (Ehdr *)lpc;
215 if (ehdr->e_phnum == 0) {
216 (void) close(fd);
217 return (0);
218 }
219 if (ehdr->e_type == ET_EXEC)
220 fixed = 1;
221 /* LINTED */
222 phdr = (Phdr *)(ehdr->e_phoff + lpc);
223 for (i = 0; i < ehdr->e_phnum; i++, phdr++) {
224 caddr_t _hpc;
225
226 if (phdr->p_type == PT_DYNAMIC) {
227 dynp = (Dyn *)phdr->p_vaddr;
228 if (fixed == 0) {
229 dynp = (Dyn *)((unsigned long)dynp +
230 (unsigned long)lpc);
231 }
232 continue;
233 }
234
235 if (phdr->p_type != PT_LOAD)
236 continue;
237
238 _hpc = (caddr_t)(phdr->p_vaddr + phdr->p_memsz);
239 if (fixed == 0) {
240 _hpc = (caddr_t)((unsigned long)_hpc +
241 (unsigned long)lpc);
242 }
243 if (_hpc > hpc)
244 hpc = _hpc;
245 }
246 if (lpc == hpc) {
247 (void) close(fd);
248 return (0);
249 }
250
251 /*
252 * In order to determine the number of symbols in the object scan the
253 * dynamic section until we find the DT_HASH entry (hash[1] == symcnt).
254 */
255 if (dynp) {
256 for (; dynp->d_tag != DT_NULL; dynp++) {
257 unsigned int *hashp;
258
259 if (dynp->d_tag != DT_HASH)
260 continue;
261
262 hashp = (unsigned int *)dynp->d_un.d_ptr;
263 if (fixed == 0) {
264 hashp = (unsigned int *)((unsigned long)hashp +
265 (unsigned long)lpc);
266 }
267 nsym = hashp[1];
268 break;
269 }
270 }
271
272 /*
273 * Determine the (minimum) size of the buffer to allocate
274 */
275 Lpc = lpc = (caddr_t)PRF_ROUNDWN((long)lpc, sizeof (long));
276 Hpc = hpc = (caddr_t)PRF_ROUNDUP((long)hpc, sizeof (long));
277
278 hsize = sizeof (L_hdr);
279 msize = (size_t)(hpc - lpc);
280 psize = (size_t)PRF_ROUNDUP((msize / PRF_BARSIZE), sizeof (long));
281 csize = (nsym + 1) * PRF_CGINIT * sizeof (L_cgarc);
282 Fsize = (hsize + psize + csize);
283
284 /*
285 * If the file size is zero (ie. we just created it), truncate it
286 * to the minimum size.
287 */
288 (void) fstat(fd, &status);
289 if (status.st_size == 0) {
290 if (ftruncate(fd, Fsize) == -1) {
291 err = errno;
292 (void) fprintf(stderr, MSG_INTL(MSG_SYS_FTRUNC),
293 Profile, strerror(err));
294 (void) close(fd);
295 return (0);
296 }
297 new_buffer++;
298 } else
299 Fsize = status.st_size;
300
301 /*
302 * Map the file in.
303 */
304 if ((addr = (caddr_t)mmap(0, Fsize, (PROT_READ | PROT_WRITE),
305 MAP_SHARED, fd, 0)) == (char *)-1) {
306 err = errno;
307 (void) fprintf(stderr, MSG_INTL(MSG_SYS_MMAP), Profile,
308 strerror(err));
309 (void) close(fd);
310 return (0);
311 }
312 (void) close(fd);
313
314 /*
315 * Initialize the remaining elements of the header. All pc addresses
316 * that are recorded are relative to zero thus allowing the recorded
317 * entries to be correlated with the symbols in the original file,
318 * and to compensate for any differences in where the file is mapped.
319 * If the high pc address has been initialized from a previous run,
320 * and the new entry is different from the original then a new library
321 * must have been installed. In this case bale out.
322 */
323 /* LINTED */
324 Hptr = (L_hdr *)addr;
325
326 if (new_buffer)
327 (void) prof_mutex_init((lwp_mutex_t *)&Hptr->hd_mutex);
328
329 (void) prof_mutex_lock((mutex_t *)&Hptr->hd_mutex, &mask);
330 if (Hptr->hd_hpc) {
331 if (Hptr->hd_hpc != (caddr_t)(hpc - lpc)) {
332 (void) fprintf(stderr, MSG_INTL(MSG_GEN_PROFSZCHG),
333 Profile);
334 (void) prof_mutex_unlock((mutex_t *)&Hptr->
335 hd_mutex, &mask);
336 (void) munmap((caddr_t)Hptr, Fsize);
337 return (0);
338 }
339 } else {
340 /*
341 * Initialize the header information as we must have just
342 * created the output file.
343 */
344 Hptr->hd_magic = (unsigned int)PRF_MAGIC;
345 #if defined(_ELF64)
346 Hptr->hd_version = (unsigned int)PRF_VERSION_64;
347 #else
348 Hptr->hd_version = (unsigned int)PRF_VERSION;
349 #endif
350 Hptr->hd_hpc = (caddr_t)(hpc - lpc);
351 /* LINTED */
352 Hptr->hd_psize = (unsigned int)psize;
353 /* LINTED */
354 Hptr->hd_fsize = (unsigned int)Fsize;
355 Hptr->hd_ncndx = nsym;
356 Hptr->hd_lcndx = (nsym + 1) * PRF_CGINIT;
357 }
358
359 (void) prof_mutex_unlock((mutex_t *)&Hptr->hd_mutex, &mask);
360 /* LINTED */
361 Cptr = (L_cgarc *)(addr + hsize + psize);
362
363 /*
364 * Turn on profiling
365 */
366 /* LINTED */
367 profil((unsigned short *)(addr + hsize),
368 psize, (unsigned long)lpc, (unsigned int) PRF_SCALE);
369
370 return (1);
371 }
372
373
374 uint_t
375 /* ARGSUSED1 */
la_objopen(Link_map * lmp,Lmid_t lmid,uintptr_t * cookie)376 la_objopen(Link_map *lmp, Lmid_t lmid, uintptr_t *cookie)
377 {
378 char *objname;
379
380 /*
381 * This would only occur if the getenv() in la_version() failed.
382 * at this point there is nothing for us to do.
383 */
384 if (pname == 0)
385 return (0);
386
387 /*
388 * Just grab the 'basename' of the object current object for
389 * comparing against the 'profiled object name'
390 */
391 if (((objname = strrchr(lmp->l_name, '/')) == 0) ||
392 (*(++objname) == 0))
393 objname = lmp->l_name;
394
395 /*
396 * Is this the object we are going to profile. If not
397 * just set the 'BINDFROM' flag for this object.
398 */
399 if ((strcmp(pname, objname) != 0) &&
400 (strcmp(pname, lmp->l_name) != 0))
401 return (LA_FLG_BINDFROM);
402
403 /*
404 * Don't even try to profile an object that does not have
405 * auditing enabled on it's link-map. This catches 'ld.so.1'.
406 */
407 if (LIST(LINKMAP_TO_RTMAP(lmp))->lm_tflags & LML_TFLG_NOAUDIT)
408 return (LA_FLG_BINDFROM);
409
410 if (profile_open(pname, lmp) == 0)
411 return (0);
412
413 profcookie = *cookie;
414
415 return (LA_FLG_BINDFROM | LA_FLG_BINDTO);
416 }
417
418
419
420 uint_t
la_objclose(uintptr_t * cookie)421 la_objclose(uintptr_t *cookie)
422 {
423 if (*cookie != profcookie)
424 return (0);
425
426 profcookie = 0;
427 /*
428 * Turn profil() off.
429 */
430 profil(0, 0, 0, 0);
431 (void) munmap((caddr_t)Hptr, Fsize);
432 return (0);
433 }
434
435
436 static int
remap_profile(int fd)437 remap_profile(int fd)
438 {
439 caddr_t addr;
440 size_t l_fsize;
441
442 l_fsize = Hptr->hd_fsize;
443
444 if ((addr = (caddr_t)mmap(0, l_fsize, (PROT_READ | PROT_WRITE),
445 MAP_SHARED, fd, 0)) == (char *)-1) {
446 int err = errno;
447
448 (void) fprintf(stderr, MSG_INTL(MSG_SYS_MMAP), Profile,
449 strerror(err));
450 return (0);
451 }
452 (void) munmap((caddr_t)Hptr, Fsize);
453
454 Fsize = l_fsize;
455 /* LINTED */
456 Hptr = (L_hdr*) addr;
457 /* LINTED */
458 Cptr = (L_cgarc *)(addr + sizeof (L_hdr) + Hptr->hd_psize);
459 return (1);
460 }
461
462
463 /*
464 * Update a call graph arc entry. This routine can be called three ways;
465 * o On initialization from one of the bndr() functions.
466 * In this case the `to' address is known, and may be used to
467 * initialize the call graph entry if this function has not
468 * been entered before.
469 * o On initial relocation (ie. LD_BIND_NOW). In this case the `to'
470 * address is known but the `from' isn't. The call graph entry
471 * is initialized to hold this dummy `to' address, but will be
472 * re-initialized later when a function is first called.
473 * o From an initialized plt entry. When profiling, the plt entries
474 * are filled in with the calling functions symbol index and
475 * the plt_cg_elf interface function. This interface function
476 * calls here to determine the `to' functions address, and in so
477 * doing increments the call count.
478 */
479 uintptr_t
plt_cg_interp(uint_t ndx,caddr_t from,caddr_t to)480 plt_cg_interp(uint_t ndx, caddr_t from, caddr_t to)
481 {
482 L_cgarc * cptr, cbucket;
483 sigset_t mask;
484
485 /*
486 * If the from address is outside of the address range being profiled,
487 * simply assign it to the `outside' address.
488 */
489 if (from != PRF_UNKNOWN) {
490 if ((from > Hpc) || (from < Lpc))
491 from = PRF_OUTADDR;
492 else
493 from = (caddr_t)(from - Lpc);
494 }
495
496 (void) prof_mutex_lock((mutex_t *)&Hptr->hd_mutex, &mask);
497 /*
498 * Has the buffer grown since last we looked at it (another processes
499 * could have grown it...).
500 */
501 /* LINTED */
502 if (Hptr->hd_fsize != (unsigned int)Fsize) {
503 int fd;
504 fd = open(Profile, O_RDWR, 0);
505 if (remap_profile(fd) == 0) {
506 (void) prof_mutex_unlock((mutex_t *)&Hptr->hd_mutex,
507 &mask);
508 exit(1);
509 }
510 (void) close(fd);
511 }
512
513 cptr = &Cptr[ndx];
514
515 if (cptr->cg_to == 0) {
516 /*
517 * If this is the first time this function has been called we
518 * got here from one of the binders or an initial relocation
519 * (ie. LD_BIND_NOW). In this case the `to' address is
520 * provided. Initialize this functions call graph entry with
521 * the functions address (retained as a relative offset).
522 * If we know where the function call originated from
523 * initialize the count field.
524 */
525 cptr->cg_to = (caddr_t)(to - Lpc);
526 cptr->cg_from = from;
527 if (from != PRF_UNKNOWN)
528 cptr->cg_count = 1;
529 } else {
530 /*
531 * If a function has been called from a previous run, but we
532 * don't know where we came from (ie. LD_BIND_NOW), then later
533 * calls through the plt will be able to obtain the required
534 * functions address, thus there is no need to proceed further.
535 */
536 if (from != PRF_UNKNOWN) {
537 /*
538 * If the from addresses match simply bump the count.
539 * If not scan the link list to find a match for this
540 * `from' address. If one doesn't exit create a new
541 * entry and link it in.
542 */
543 while ((cptr->cg_from != from) &&
544 (cptr->cg_from != PRF_UNKNOWN)) {
545 if (cptr->cg_next != 0)
546 cptr = &Cptr[cptr->cg_next];
547 else {
548 to = cptr->cg_to;
549 cptr->cg_next = Hptr->hd_ncndx++;
550 cptr = &Cptr[cptr->cg_next];
551 /*
552 * If we've run out of file, extend it.
553 */
554 if (Hptr->hd_ncndx == Hptr->hd_lcndx) {
555 caddr_t addr;
556 int fd;
557
558 /* LINTED */
559 Hptr->hd_fsize += (unsigned int)
560 PRF_CGNUMB *
561 sizeof (L_cgarc);
562 fd = open(Profile, O_RDWR, 0);
563 if (ftruncate(fd,
564 Hptr->hd_fsize) == -1) {
565 int err = errno;
566
567 (void) fprintf(stderr,
568 MSG_INTL(
569 MSG_SYS_FTRUNC),
570 Profile,
571 strerror(err));
572 (void) close(fd);
573 cptr = &cbucket;
574 }
575 /*
576 * Since the buffer will be
577 * remapped, we need to be
578 * prepared to adjust cptr.
579 */
580 addr = (caddr_t)((Addr)cptr -
581 (Addr)Cptr);
582 if (remap_profile(fd) == 0) {
583 /* CSTYLED */
584 (void) prof_mutex_unlock(
585 (mutex_t *)&Hptr->
586 hd_mutex, &mask);
587 exit(1);
588 }
589 cptr = (L_cgarc *)((Addr)addr +
590 (Addr)Cptr);
591 (void) close(fd);
592 Hptr->hd_lcndx += PRF_CGNUMB;
593 }
594 cptr->cg_from = from;
595 cptr->cg_to = to;
596 }
597 }
598 /*
599 * If we're updating an entry from an unknown call
600 * address initialize this element, otherwise
601 * increment the call count.
602 */
603 if (cptr->cg_from == PRF_UNKNOWN) {
604 cptr->cg_from = from;
605 cptr->cg_count = 1;
606 } else
607 cptr->cg_count++;
608 }
609 }
610 /*
611 * Return the real address of the function.
612 */
613 (void) prof_mutex_unlock((mutex_t *)&Hptr->hd_mutex, &mask);
614
615 return ((uintptr_t)((Addr)cptr->cg_to + (Addr)Lpc));
616 }
617
618 /* ARGSUSED2 */
619 #if defined(__sparcv9)
620 uintptr_t
la_sparcv9_pltenter(Elf64_Sym * symp,uint_t symndx,uintptr_t * refcookie,uintptr_t * defcookie,La_sparcv9_regs * regset,uint_t * sbflags,const char * sym_name)621 la_sparcv9_pltenter(Elf64_Sym *symp, uint_t symndx, uintptr_t *refcookie,
622 uintptr_t *defcookie, La_sparcv9_regs *regset, uint_t *sbflags,
623 const char *sym_name)
624 #elif defined(__sparc)
625 uintptr_t
626 la_sparcv8_pltenter(Elf32_Sym *symp, uint_t symndx, uintptr_t *refcookie,
627 uintptr_t *defcookie, La_sparcv8_regs *regset, uint_t *sbflags)
628 #elif defined(__amd64)
629 uintptr_t
630 la_amd64_pltenter(Elf64_Sym *symp, uint_t symndx, uintptr_t *refcookie,
631 uintptr_t *defcookie, La_amd64_regs *regset, uint_t *sbflags,
632 const char *sym_name)
633 #elif defined(__i386)
634 uintptr_t
635 la_i86_pltenter(Elf32_Sym *symp, uint_t symndx, uintptr_t *refcookie,
636 uintptr_t *defcookie, La_i86_regs *regset, uint_t *sbflags)
637 #else
638 #error unexpected architecture!
639 #endif
640 {
641 caddr_t from;
642
643 /*
644 * profiling has been disabled.
645 */
646 if (profcookie == 0)
647 return (symp->st_value);
648 #if defined(__sparc)
649 /*
650 * The callers return address is currently stored in O7 (which
651 * will become I7 when the window shift occurs).
652 */
653 from = (caddr_t)regset->lr_rego7;
654 #elif defined(__amd64)
655 /*
656 * The callers return address is on the top of the stack for amd64
657 */
658 from = *(caddr_t *)(regset->lr_rsp);
659 #elif defined(__i386)
660 /*
661 * The callers return address is on the top of the stack for i386
662 */
663 from = *(caddr_t *)(regset->lr_esp);
664 #else
665 #error unexpected architecture!
666 #endif
667 return (plt_cg_interp(symndx, (caddr_t)from, (caddr_t)symp->st_value));
668 }
669