xref: /titanic_41/usr/src/lib/libdtrace/common/dt_link.c (revision 56b2bdd1f04d465cfe4a95b88ae5cba5884154e4)
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 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #define	ELF_TARGET_ALL
30 #include <elf.h>
31 
32 #include <sys/types.h>
33 #include <sys/sysmacros.h>
34 
35 #include <unistd.h>
36 #include <strings.h>
37 #include <alloca.h>
38 #include <limits.h>
39 #include <stddef.h>
40 #include <stdlib.h>
41 #include <stdio.h>
42 #include <fcntl.h>
43 #include <errno.h>
44 #include <wait.h>
45 #include <assert.h>
46 #include <sys/ipc.h>
47 
48 #include <dt_impl.h>
49 #include <dt_provider.h>
50 #include <dt_program.h>
51 #include <dt_string.h>
52 
53 #define	ESHDR_NULL	0
54 #define	ESHDR_SHSTRTAB	1
55 #define	ESHDR_DOF	2
56 #define	ESHDR_STRTAB	3
57 #define	ESHDR_SYMTAB	4
58 #define	ESHDR_REL	5
59 #define	ESHDR_NUM	6
60 
61 #define	PWRITE_SCN(index, data) \
62 	(lseek64(fd, (off64_t)elf_file.shdr[(index)].sh_offset, SEEK_SET) != \
63 	(off64_t)elf_file.shdr[(index)].sh_offset || \
64 	dt_write(dtp, fd, (data), elf_file.shdr[(index)].sh_size) != \
65 	elf_file.shdr[(index)].sh_size)
66 
67 static const char DTRACE_SHSTRTAB32[] = "\0"
68 ".shstrtab\0"		/* 1 */
69 ".SUNW_dof\0"		/* 11 */
70 ".strtab\0"		/* 21 */
71 ".symtab\0"		/* 29 */
72 #ifdef __sparc
73 ".rela.SUNW_dof";	/* 37 */
74 #else
75 ".rel.SUNW_dof";	/* 37 */
76 #endif
77 
78 static const char DTRACE_SHSTRTAB64[] = "\0"
79 ".shstrtab\0"		/* 1 */
80 ".SUNW_dof\0"		/* 11 */
81 ".strtab\0"		/* 21 */
82 ".symtab\0"		/* 29 */
83 ".rela.SUNW_dof";	/* 37 */
84 
85 static const char DOFSTR[] = "__SUNW_dof";
86 static const char DOFLAZYSTR[] = "___SUNW_dof";
87 
88 typedef struct dt_link_pair {
89 	struct dt_link_pair *dlp_next;	/* next pair in linked list */
90 	void *dlp_str;			/* buffer for string table */
91 	void *dlp_sym;			/* buffer for symbol table */
92 } dt_link_pair_t;
93 
94 typedef struct dof_elf32 {
95 	uint32_t de_nrel;		/* relocation count */
96 #ifdef __sparc
97 	Elf32_Rela *de_rel;		/* array of relocations for sparc */
98 #else
99 	Elf32_Rel *de_rel;		/* array of relocations for x86 */
100 #endif
101 	uint32_t de_nsym;		/* symbol count */
102 	Elf32_Sym *de_sym;		/* array of symbols */
103 	uint32_t de_strlen;		/* size of of string table */
104 	char *de_strtab;		/* string table */
105 	uint32_t de_global;		/* index of the first global symbol */
106 } dof_elf32_t;
107 
108 static int
109 prepare_elf32(dtrace_hdl_t *dtp, const dof_hdr_t *dof, dof_elf32_t *dep)
110 {
111 	dof_sec_t *dofs, *s;
112 	dof_relohdr_t *dofrh;
113 	dof_relodesc_t *dofr;
114 	char *strtab;
115 	int i, j, nrel;
116 	size_t strtabsz = 1;
117 	uint32_t count = 0;
118 	size_t base;
119 	Elf32_Sym *sym;
120 #ifdef __sparc
121 	Elf32_Rela *rel;
122 #else
123 	Elf32_Rel *rel;
124 #endif
125 
126 	/*LINTED*/
127 	dofs = (dof_sec_t *)((char *)dof + dof->dofh_secoff);
128 
129 	/*
130 	 * First compute the size of the string table and the number of
131 	 * relocations present in the DOF.
132 	 */
133 	for (i = 0; i < dof->dofh_secnum; i++) {
134 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
135 			continue;
136 
137 		/*LINTED*/
138 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
139 
140 		s = &dofs[dofrh->dofr_strtab];
141 		strtab = (char *)dof + s->dofs_offset;
142 		assert(strtab[0] == '\0');
143 		strtabsz += s->dofs_size - 1;
144 
145 		s = &dofs[dofrh->dofr_relsec];
146 		/*LINTED*/
147 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
148 		count += s->dofs_size / s->dofs_entsize;
149 	}
150 
151 	dep->de_strlen = strtabsz;
152 	dep->de_nrel = count;
153 	dep->de_nsym = count + 1; /* the first symbol is always null */
154 
155 	if (dtp->dt_lazyload) {
156 		dep->de_strlen += sizeof (DOFLAZYSTR);
157 		dep->de_nsym++;
158 	} else {
159 		dep->de_strlen += sizeof (DOFSTR);
160 		dep->de_nsym++;
161 	}
162 
163 	if ((dep->de_rel = calloc(dep->de_nrel,
164 	    sizeof (dep->de_rel[0]))) == NULL) {
165 		return (dt_set_errno(dtp, EDT_NOMEM));
166 	}
167 
168 	if ((dep->de_sym = calloc(dep->de_nsym, sizeof (Elf32_Sym))) == NULL) {
169 		free(dep->de_rel);
170 		return (dt_set_errno(dtp, EDT_NOMEM));
171 	}
172 
173 	if ((dep->de_strtab = calloc(dep->de_strlen, 1)) == NULL) {
174 		free(dep->de_rel);
175 		free(dep->de_sym);
176 		return (dt_set_errno(dtp, EDT_NOMEM));
177 	}
178 
179 	count = 0;
180 	strtabsz = 1;
181 	dep->de_strtab[0] = '\0';
182 	rel = dep->de_rel;
183 	sym = dep->de_sym;
184 	dep->de_global = 1;
185 
186 	/*
187 	 * The first symbol table entry must be zeroed and is always ignored.
188 	 */
189 	bzero(sym, sizeof (Elf32_Sym));
190 	sym++;
191 
192 	/*
193 	 * Take a second pass through the DOF sections filling in the
194 	 * memory we allocated.
195 	 */
196 	for (i = 0; i < dof->dofh_secnum; i++) {
197 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
198 			continue;
199 
200 		/*LINTED*/
201 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
202 
203 		s = &dofs[dofrh->dofr_strtab];
204 		strtab = (char *)dof + s->dofs_offset;
205 		bcopy(strtab + 1, dep->de_strtab + strtabsz, s->dofs_size);
206 		base = strtabsz;
207 		strtabsz += s->dofs_size - 1;
208 
209 		s = &dofs[dofrh->dofr_relsec];
210 		/*LINTED*/
211 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
212 		nrel = s->dofs_size / s->dofs_entsize;
213 
214 		s = &dofs[dofrh->dofr_tgtsec];
215 
216 		for (j = 0; j < nrel; j++) {
217 #if defined(__i386) || defined(__amd64)
218 			rel->r_offset = s->dofs_offset +
219 			    dofr[j].dofr_offset;
220 			rel->r_info = ELF32_R_INFO(count + dep->de_global,
221 			    R_386_32);
222 #elif defined(__sparc)
223 			/*
224 			 * Add 4 bytes to hit the low half of this 64-bit
225 			 * big-endian address.
226 			 */
227 			rel->r_offset = s->dofs_offset +
228 			    dofr[j].dofr_offset + 4;
229 			rel->r_info = ELF32_R_INFO(count + dep->de_global,
230 			    R_SPARC_32);
231 #else
232 #error unknown ISA
233 #endif
234 
235 			sym->st_name = base + dofr[j].dofr_name - 1;
236 			sym->st_value = 0;
237 			sym->st_size = 0;
238 			sym->st_info = ELF32_ST_INFO(STB_GLOBAL, STT_FUNC);
239 			sym->st_other = 0;
240 			sym->st_shndx = SHN_UNDEF;
241 
242 			rel++;
243 			sym++;
244 			count++;
245 		}
246 	}
247 
248 	/*
249 	 * Add a symbol for the DOF itself. We use a different symbol for
250 	 * lazily and actively loaded DOF to make them easy to distinguish.
251 	 */
252 	sym->st_name = strtabsz;
253 	sym->st_value = 0;
254 	sym->st_size = dof->dofh_filesz;
255 	sym->st_info = ELF32_ST_INFO(STB_GLOBAL, STT_OBJECT);
256 	sym->st_other = 0;
257 	sym->st_shndx = ESHDR_DOF;
258 	sym++;
259 
260 	if (dtp->dt_lazyload) {
261 		bcopy(DOFLAZYSTR, dep->de_strtab + strtabsz,
262 		    sizeof (DOFLAZYSTR));
263 		strtabsz += sizeof (DOFLAZYSTR);
264 	} else {
265 		bcopy(DOFSTR, dep->de_strtab + strtabsz, sizeof (DOFSTR));
266 		strtabsz += sizeof (DOFSTR);
267 	}
268 
269 	assert(count == dep->de_nrel);
270 	assert(strtabsz == dep->de_strlen);
271 
272 	return (0);
273 }
274 
275 
276 typedef struct dof_elf64 {
277 	uint32_t de_nrel;
278 	Elf64_Rela *de_rel;
279 	uint32_t de_nsym;
280 	Elf64_Sym *de_sym;
281 
282 	uint32_t de_strlen;
283 	char *de_strtab;
284 
285 	uint32_t de_global;
286 } dof_elf64_t;
287 
288 static int
289 prepare_elf64(dtrace_hdl_t *dtp, const dof_hdr_t *dof, dof_elf64_t *dep)
290 {
291 	dof_sec_t *dofs, *s;
292 	dof_relohdr_t *dofrh;
293 	dof_relodesc_t *dofr;
294 	char *strtab;
295 	int i, j, nrel;
296 	size_t strtabsz = 1;
297 	uint32_t count = 0;
298 	size_t base;
299 	Elf64_Sym *sym;
300 	Elf64_Rela *rel;
301 
302 	/*LINTED*/
303 	dofs = (dof_sec_t *)((char *)dof + dof->dofh_secoff);
304 
305 	/*
306 	 * First compute the size of the string table and the number of
307 	 * relocations present in the DOF.
308 	 */
309 	for (i = 0; i < dof->dofh_secnum; i++) {
310 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
311 			continue;
312 
313 		/*LINTED*/
314 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
315 
316 		s = &dofs[dofrh->dofr_strtab];
317 		strtab = (char *)dof + s->dofs_offset;
318 		assert(strtab[0] == '\0');
319 		strtabsz += s->dofs_size - 1;
320 
321 		s = &dofs[dofrh->dofr_relsec];
322 		/*LINTED*/
323 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
324 		count += s->dofs_size / s->dofs_entsize;
325 	}
326 
327 	dep->de_strlen = strtabsz;
328 	dep->de_nrel = count;
329 	dep->de_nsym = count + 1; /* the first symbol is always null */
330 
331 	if (dtp->dt_lazyload) {
332 		dep->de_strlen += sizeof (DOFLAZYSTR);
333 		dep->de_nsym++;
334 	} else {
335 		dep->de_strlen += sizeof (DOFSTR);
336 		dep->de_nsym++;
337 	}
338 
339 	if ((dep->de_rel = calloc(dep->de_nrel,
340 	    sizeof (dep->de_rel[0]))) == NULL) {
341 		return (dt_set_errno(dtp, EDT_NOMEM));
342 	}
343 
344 	if ((dep->de_sym = calloc(dep->de_nsym, sizeof (Elf64_Sym))) == NULL) {
345 		free(dep->de_rel);
346 		return (dt_set_errno(dtp, EDT_NOMEM));
347 	}
348 
349 	if ((dep->de_strtab = calloc(dep->de_strlen, 1)) == NULL) {
350 		free(dep->de_rel);
351 		free(dep->de_sym);
352 		return (dt_set_errno(dtp, EDT_NOMEM));
353 	}
354 
355 	count = 0;
356 	strtabsz = 1;
357 	dep->de_strtab[0] = '\0';
358 	rel = dep->de_rel;
359 	sym = dep->de_sym;
360 	dep->de_global = 1;
361 
362 	/*
363 	 * The first symbol table entry must be zeroed and is always ignored.
364 	 */
365 	bzero(sym, sizeof (Elf64_Sym));
366 	sym++;
367 
368 	/*
369 	 * Take a second pass through the DOF sections filling in the
370 	 * memory we allocated.
371 	 */
372 	for (i = 0; i < dof->dofh_secnum; i++) {
373 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
374 			continue;
375 
376 		/*LINTED*/
377 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
378 
379 		s = &dofs[dofrh->dofr_strtab];
380 		strtab = (char *)dof + s->dofs_offset;
381 		bcopy(strtab + 1, dep->de_strtab + strtabsz, s->dofs_size);
382 		base = strtabsz;
383 		strtabsz += s->dofs_size - 1;
384 
385 		s = &dofs[dofrh->dofr_relsec];
386 		/*LINTED*/
387 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
388 		nrel = s->dofs_size / s->dofs_entsize;
389 
390 		s = &dofs[dofrh->dofr_tgtsec];
391 
392 		for (j = 0; j < nrel; j++) {
393 #if defined(__i386) || defined(__amd64)
394 			rel->r_offset = s->dofs_offset +
395 			    dofr[j].dofr_offset;
396 			rel->r_info = ELF64_R_INFO(count + dep->de_global,
397 			    R_AMD64_64);
398 #elif defined(__sparc)
399 			rel->r_offset = s->dofs_offset +
400 			    dofr[j].dofr_offset;
401 			rel->r_info = ELF64_R_INFO(count + dep->de_global,
402 			    R_SPARC_64);
403 #else
404 #error unknown ISA
405 #endif
406 
407 			sym->st_name = base + dofr[j].dofr_name - 1;
408 			sym->st_value = 0;
409 			sym->st_size = 0;
410 			sym->st_info = GELF_ST_INFO(STB_GLOBAL, STT_FUNC);
411 			sym->st_other = 0;
412 			sym->st_shndx = SHN_UNDEF;
413 
414 			rel++;
415 			sym++;
416 			count++;
417 		}
418 	}
419 
420 	/*
421 	 * Add a symbol for the DOF itself. We use a different symbol for
422 	 * lazily and actively loaded DOF to make them easy to distinguish.
423 	 */
424 	sym->st_name = strtabsz;
425 	sym->st_value = 0;
426 	sym->st_size = dof->dofh_filesz;
427 	sym->st_info = GELF_ST_INFO(STB_GLOBAL, STT_OBJECT);
428 	sym->st_other = 0;
429 	sym->st_shndx = ESHDR_DOF;
430 	sym++;
431 
432 	if (dtp->dt_lazyload) {
433 		bcopy(DOFLAZYSTR, dep->de_strtab + strtabsz,
434 		    sizeof (DOFLAZYSTR));
435 		strtabsz += sizeof (DOFLAZYSTR);
436 	} else {
437 		bcopy(DOFSTR, dep->de_strtab + strtabsz, sizeof (DOFSTR));
438 		strtabsz += sizeof (DOFSTR);
439 	}
440 
441 	assert(count == dep->de_nrel);
442 	assert(strtabsz == dep->de_strlen);
443 
444 	return (0);
445 }
446 
447 /*
448  * Write out an ELF32 file prologue consisting of a header, section headers,
449  * and a section header string table.  The DOF data will follow this prologue
450  * and complete the contents of the given ELF file.
451  */
452 static int
453 dump_elf32(dtrace_hdl_t *dtp, const dof_hdr_t *dof, int fd)
454 {
455 	struct {
456 		Elf32_Ehdr ehdr;
457 		Elf32_Shdr shdr[ESHDR_NUM];
458 	} elf_file;
459 
460 	Elf32_Shdr *shp;
461 	Elf32_Off off;
462 	dof_elf32_t de;
463 	int ret = 0;
464 	uint_t nshdr;
465 
466 	if (prepare_elf32(dtp, dof, &de) != 0)
467 		return (-1); /* errno is set for us */
468 
469 	/*
470 	 * If there are no relocations, we only need enough sections for
471 	 * the shstrtab and the DOF.
472 	 */
473 	nshdr = de.de_nrel == 0 ? ESHDR_SYMTAB + 1 : ESHDR_NUM;
474 
475 	bzero(&elf_file, sizeof (elf_file));
476 
477 	elf_file.ehdr.e_ident[EI_MAG0] = ELFMAG0;
478 	elf_file.ehdr.e_ident[EI_MAG1] = ELFMAG1;
479 	elf_file.ehdr.e_ident[EI_MAG2] = ELFMAG2;
480 	elf_file.ehdr.e_ident[EI_MAG3] = ELFMAG3;
481 	elf_file.ehdr.e_ident[EI_VERSION] = EV_CURRENT;
482 	elf_file.ehdr.e_ident[EI_CLASS] = ELFCLASS32;
483 #if defined(_BIG_ENDIAN)
484 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2MSB;
485 #elif defined(_LITTLE_ENDIAN)
486 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2LSB;
487 #endif
488 	elf_file.ehdr.e_type = ET_REL;
489 #if defined(__sparc)
490 	elf_file.ehdr.e_machine = EM_SPARC;
491 #elif defined(__i386) || defined(__amd64)
492 	elf_file.ehdr.e_machine = EM_386;
493 #endif
494 	elf_file.ehdr.e_version = EV_CURRENT;
495 	elf_file.ehdr.e_shoff = sizeof (Elf32_Ehdr);
496 	elf_file.ehdr.e_ehsize = sizeof (Elf32_Ehdr);
497 	elf_file.ehdr.e_phentsize = sizeof (Elf32_Phdr);
498 	elf_file.ehdr.e_shentsize = sizeof (Elf32_Shdr);
499 	elf_file.ehdr.e_shnum = nshdr;
500 	elf_file.ehdr.e_shstrndx = ESHDR_SHSTRTAB;
501 	off = sizeof (elf_file) + nshdr * sizeof (Elf32_Shdr);
502 
503 	shp = &elf_file.shdr[ESHDR_SHSTRTAB];
504 	shp->sh_name = 1; /* DTRACE_SHSTRTAB32[1] = ".shstrtab" */
505 	shp->sh_type = SHT_STRTAB;
506 	shp->sh_offset = off;
507 	shp->sh_size = sizeof (DTRACE_SHSTRTAB32);
508 	shp->sh_addralign = sizeof (char);
509 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
510 
511 	shp = &elf_file.shdr[ESHDR_DOF];
512 	shp->sh_name = 11; /* DTRACE_SHSTRTAB32[11] = ".SUNW_dof" */
513 	shp->sh_flags = SHF_ALLOC;
514 	shp->sh_type = SHT_SUNW_dof;
515 	shp->sh_offset = off;
516 	shp->sh_size = dof->dofh_filesz;
517 	shp->sh_addralign = 8;
518 	off = shp->sh_offset + shp->sh_size;
519 
520 	shp = &elf_file.shdr[ESHDR_STRTAB];
521 	shp->sh_name = 21; /* DTRACE_SHSTRTAB32[21] = ".strtab" */
522 	shp->sh_flags = SHF_ALLOC;
523 	shp->sh_type = SHT_STRTAB;
524 	shp->sh_offset = off;
525 	shp->sh_size = de.de_strlen;
526 	shp->sh_addralign = sizeof (char);
527 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 4);
528 
529 	shp = &elf_file.shdr[ESHDR_SYMTAB];
530 	shp->sh_name = 29; /* DTRACE_SHSTRTAB32[29] = ".symtab" */
531 	shp->sh_flags = SHF_ALLOC;
532 	shp->sh_type = SHT_SYMTAB;
533 	shp->sh_entsize = sizeof (Elf32_Sym);
534 	shp->sh_link = ESHDR_STRTAB;
535 	shp->sh_offset = off;
536 	shp->sh_info = de.de_global;
537 	shp->sh_size = de.de_nsym * sizeof (Elf32_Sym);
538 	shp->sh_addralign = 4;
539 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 4);
540 
541 	if (de.de_nrel == 0) {
542 		if (dt_write(dtp, fd, &elf_file,
543 		    sizeof (elf_file)) != sizeof (elf_file) ||
544 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB32) ||
545 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
546 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
547 		    PWRITE_SCN(ESHDR_DOF, dof)) {
548 			ret = dt_set_errno(dtp, errno);
549 		}
550 	} else {
551 		shp = &elf_file.shdr[ESHDR_REL];
552 		shp->sh_name = 37; /* DTRACE_SHSTRTAB32[37] = ".rel.SUNW_dof" */
553 		shp->sh_flags = SHF_ALLOC;
554 #ifdef __sparc
555 		shp->sh_type = SHT_RELA;
556 #else
557 		shp->sh_type = SHT_REL;
558 #endif
559 		shp->sh_entsize = sizeof (de.de_rel[0]);
560 		shp->sh_link = ESHDR_SYMTAB;
561 		shp->sh_info = ESHDR_DOF;
562 		shp->sh_offset = off;
563 		shp->sh_size = de.de_nrel * sizeof (de.de_rel[0]);
564 		shp->sh_addralign = 4;
565 
566 		if (dt_write(dtp, fd, &elf_file,
567 		    sizeof (elf_file)) != sizeof (elf_file) ||
568 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB32) ||
569 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
570 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
571 		    PWRITE_SCN(ESHDR_REL, de.de_rel) ||
572 		    PWRITE_SCN(ESHDR_DOF, dof)) {
573 			ret = dt_set_errno(dtp, errno);
574 		}
575 	}
576 
577 	free(de.de_strtab);
578 	free(de.de_sym);
579 	free(de.de_rel);
580 
581 	return (ret);
582 }
583 
584 /*
585  * Write out an ELF64 file prologue consisting of a header, section headers,
586  * and a section header string table.  The DOF data will follow this prologue
587  * and complete the contents of the given ELF file.
588  */
589 static int
590 dump_elf64(dtrace_hdl_t *dtp, const dof_hdr_t *dof, int fd)
591 {
592 	struct {
593 		Elf64_Ehdr ehdr;
594 		Elf64_Shdr shdr[ESHDR_NUM];
595 	} elf_file;
596 
597 	Elf64_Shdr *shp;
598 	Elf64_Off off;
599 	dof_elf64_t de;
600 	int ret = 0;
601 	uint_t nshdr;
602 
603 	if (prepare_elf64(dtp, dof, &de) != 0)
604 		return (-1); /* errno is set for us */
605 
606 	/*
607 	 * If there are no relocations, we only need enough sections for
608 	 * the shstrtab and the DOF.
609 	 */
610 	nshdr = de.de_nrel == 0 ? ESHDR_SYMTAB + 1 : ESHDR_NUM;
611 
612 	bzero(&elf_file, sizeof (elf_file));
613 
614 	elf_file.ehdr.e_ident[EI_MAG0] = ELFMAG0;
615 	elf_file.ehdr.e_ident[EI_MAG1] = ELFMAG1;
616 	elf_file.ehdr.e_ident[EI_MAG2] = ELFMAG2;
617 	elf_file.ehdr.e_ident[EI_MAG3] = ELFMAG3;
618 	elf_file.ehdr.e_ident[EI_VERSION] = EV_CURRENT;
619 	elf_file.ehdr.e_ident[EI_CLASS] = ELFCLASS64;
620 #if defined(_BIG_ENDIAN)
621 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2MSB;
622 #elif defined(_LITTLE_ENDIAN)
623 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2LSB;
624 #endif
625 	elf_file.ehdr.e_type = ET_REL;
626 #if defined(__sparc)
627 	elf_file.ehdr.e_machine = EM_SPARCV9;
628 #elif defined(__i386) || defined(__amd64)
629 	elf_file.ehdr.e_machine = EM_AMD64;
630 #endif
631 	elf_file.ehdr.e_version = EV_CURRENT;
632 	elf_file.ehdr.e_shoff = sizeof (Elf64_Ehdr);
633 	elf_file.ehdr.e_ehsize = sizeof (Elf64_Ehdr);
634 	elf_file.ehdr.e_phentsize = sizeof (Elf64_Phdr);
635 	elf_file.ehdr.e_shentsize = sizeof (Elf64_Shdr);
636 	elf_file.ehdr.e_shnum = nshdr;
637 	elf_file.ehdr.e_shstrndx = ESHDR_SHSTRTAB;
638 	off = sizeof (elf_file) + nshdr * sizeof (Elf64_Shdr);
639 
640 	shp = &elf_file.shdr[ESHDR_SHSTRTAB];
641 	shp->sh_name = 1; /* DTRACE_SHSTRTAB64[1] = ".shstrtab" */
642 	shp->sh_type = SHT_STRTAB;
643 	shp->sh_offset = off;
644 	shp->sh_size = sizeof (DTRACE_SHSTRTAB64);
645 	shp->sh_addralign = sizeof (char);
646 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
647 
648 	shp = &elf_file.shdr[ESHDR_DOF];
649 	shp->sh_name = 11; /* DTRACE_SHSTRTAB64[11] = ".SUNW_dof" */
650 	shp->sh_flags = SHF_ALLOC;
651 	shp->sh_type = SHT_SUNW_dof;
652 	shp->sh_offset = off;
653 	shp->sh_size = dof->dofh_filesz;
654 	shp->sh_addralign = 8;
655 	off = shp->sh_offset + shp->sh_size;
656 
657 	shp = &elf_file.shdr[ESHDR_STRTAB];
658 	shp->sh_name = 21; /* DTRACE_SHSTRTAB64[21] = ".strtab" */
659 	shp->sh_flags = SHF_ALLOC;
660 	shp->sh_type = SHT_STRTAB;
661 	shp->sh_offset = off;
662 	shp->sh_size = de.de_strlen;
663 	shp->sh_addralign = sizeof (char);
664 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
665 
666 	shp = &elf_file.shdr[ESHDR_SYMTAB];
667 	shp->sh_name = 29; /* DTRACE_SHSTRTAB64[29] = ".symtab" */
668 	shp->sh_flags = SHF_ALLOC;
669 	shp->sh_type = SHT_SYMTAB;
670 	shp->sh_entsize = sizeof (Elf64_Sym);
671 	shp->sh_link = ESHDR_STRTAB;
672 	shp->sh_offset = off;
673 	shp->sh_info = de.de_global;
674 	shp->sh_size = de.de_nsym * sizeof (Elf64_Sym);
675 	shp->sh_addralign = 8;
676 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
677 
678 	if (de.de_nrel == 0) {
679 		if (dt_write(dtp, fd, &elf_file,
680 		    sizeof (elf_file)) != sizeof (elf_file) ||
681 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB64) ||
682 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
683 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
684 		    PWRITE_SCN(ESHDR_DOF, dof)) {
685 			ret = dt_set_errno(dtp, errno);
686 		}
687 	} else {
688 		shp = &elf_file.shdr[ESHDR_REL];
689 		shp->sh_name = 37; /* DTRACE_SHSTRTAB64[37] = ".rel.SUNW_dof" */
690 		shp->sh_flags = SHF_ALLOC;
691 		shp->sh_type = SHT_RELA;
692 		shp->sh_entsize = sizeof (de.de_rel[0]);
693 		shp->sh_link = ESHDR_SYMTAB;
694 		shp->sh_info = ESHDR_DOF;
695 		shp->sh_offset = off;
696 		shp->sh_size = de.de_nrel * sizeof (de.de_rel[0]);
697 		shp->sh_addralign = 8;
698 
699 		if (dt_write(dtp, fd, &elf_file,
700 		    sizeof (elf_file)) != sizeof (elf_file) ||
701 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB64) ||
702 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
703 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
704 		    PWRITE_SCN(ESHDR_REL, de.de_rel) ||
705 		    PWRITE_SCN(ESHDR_DOF, dof)) {
706 			ret = dt_set_errno(dtp, errno);
707 		}
708 	}
709 
710 	free(de.de_strtab);
711 	free(de.de_sym);
712 	free(de.de_rel);
713 
714 	return (ret);
715 }
716 
717 static int
718 dt_symtab_lookup(Elf_Data *data_sym, int nsym, uintptr_t addr, uint_t shn,
719     GElf_Sym *sym)
720 {
721 	int i, ret = -1;
722 	GElf_Sym s;
723 
724 	for (i = 0; i < nsym && gelf_getsym(data_sym, i, sym) != NULL; i++) {
725 		if (GELF_ST_TYPE(sym->st_info) == STT_FUNC &&
726 		    shn == sym->st_shndx &&
727 		    sym->st_value <= addr &&
728 		    addr < sym->st_value + sym->st_size) {
729 			if (GELF_ST_BIND(sym->st_info) == STB_GLOBAL)
730 				return (0);
731 
732 			ret = 0;
733 			s = *sym;
734 		}
735 	}
736 
737 	if (ret == 0)
738 		*sym = s;
739 	return (ret);
740 }
741 
742 #if defined(__sparc)
743 
744 #define	DT_OP_RET		0x81c7e008
745 #define	DT_OP_NOP		0x01000000
746 #define	DT_OP_CALL		0x40000000
747 #define	DT_OP_CLR_O0		0x90102000
748 
749 #define	DT_IS_MOV_O7(inst)	(((inst) & 0xffffe000) == 0x9e100000)
750 #define	DT_IS_RESTORE(inst)	(((inst) & 0xc1f80000) == 0x81e80000)
751 #define	DT_IS_RETL(inst)	(((inst) & 0xfff83fff) == 0x81c02008)
752 
753 #define	DT_RS2(inst)		((inst) & 0x1f)
754 #define	DT_MAKE_RETL(reg)	(0x81c02008 | ((reg) << 14))
755 
756 /*ARGSUSED*/
757 static int
758 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
759     uint32_t *off)
760 {
761 	uint32_t *ip;
762 
763 	if ((rela->r_offset & (sizeof (uint32_t) - 1)) != 0)
764 		return (-1);
765 
766 	/*LINTED*/
767 	ip = (uint32_t *)(p + rela->r_offset);
768 
769 	/*
770 	 * We only know about some specific relocation types.
771 	 */
772 	if (GELF_R_TYPE(rela->r_info) != R_SPARC_WDISP30 &&
773 	    GELF_R_TYPE(rela->r_info) != R_SPARC_WPLT30)
774 		return (-1);
775 
776 	/*
777 	 * We may have already processed this object file in an earlier linker
778 	 * invocation. Check to see if the present instruction sequence matches
779 	 * the one we would install below.
780 	 */
781 	if (isenabled) {
782 		if (ip[0] == DT_OP_NOP) {
783 			(*off) += sizeof (ip[0]);
784 			return (0);
785 		}
786 	} else {
787 		if (DT_IS_RESTORE(ip[1])) {
788 			if (ip[0] == DT_OP_RET) {
789 				(*off) += sizeof (ip[0]);
790 				return (0);
791 			}
792 		} else if (DT_IS_MOV_O7(ip[1])) {
793 			if (DT_IS_RETL(ip[0]))
794 				return (0);
795 		} else {
796 			if (ip[0] == DT_OP_NOP) {
797 				(*off) += sizeof (ip[0]);
798 				return (0);
799 			}
800 		}
801 	}
802 
803 	/*
804 	 * We only expect call instructions with a displacement of 0.
805 	 */
806 	if (ip[0] != DT_OP_CALL) {
807 		dt_dprintf("found %x instead of a call instruction at %llx\n",
808 		    ip[0], (u_longlong_t)rela->r_offset);
809 		return (-1);
810 	}
811 
812 	if (isenabled) {
813 		/*
814 		 * It would necessarily indicate incorrect usage if an is-
815 		 * enabled probe were tail-called so flag that as an error.
816 		 * It's also potentially (very) tricky to handle gracefully,
817 		 * but could be done if this were a desired use scenario.
818 		 */
819 		if (DT_IS_RESTORE(ip[1]) || DT_IS_MOV_O7(ip[1])) {
820 			dt_dprintf("tail call to is-enabled probe at %llx\n",
821 			    (u_longlong_t)rela->r_offset);
822 			return (-1);
823 		}
824 
825 
826 		/*
827 		 * On SPARC, we take advantage of the fact that the first
828 		 * argument shares the same register as for the return value.
829 		 * The macro handles the work of zeroing that register so we
830 		 * don't need to do anything special here. We instrument the
831 		 * instruction in the delay slot as we'll need to modify the
832 		 * return register after that instruction has been emulated.
833 		 */
834 		ip[0] = DT_OP_NOP;
835 		(*off) += sizeof (ip[0]);
836 	} else {
837 		/*
838 		 * If the call is followed by a restore, it's a tail call so
839 		 * change the call to a ret. If the call if followed by a mov
840 		 * of a register into %o7, it's a tail call in leaf context
841 		 * so change the call to a retl-like instruction that returns
842 		 * to that register value + 8 (rather than the typical %o7 +
843 		 * 8); the delay slot instruction is left, but should have no
844 		 * effect. Otherwise we change the call to be a nop. We
845 		 * identify the subsequent instruction as the probe point in
846 		 * all but the leaf tail-call case to ensure that arguments to
847 		 * the probe are complete and consistent. An astute, though
848 		 * largely hypothetical, observer would note that there is the
849 		 * possibility of a false-positive probe firing if the function
850 		 * contained a branch to the instruction in the delay slot of
851 		 * the call. Fixing this would require significant in-kernel
852 		 * modifications, and isn't worth doing until we see it in the
853 		 * wild.
854 		 */
855 		if (DT_IS_RESTORE(ip[1])) {
856 			ip[0] = DT_OP_RET;
857 			(*off) += sizeof (ip[0]);
858 		} else if (DT_IS_MOV_O7(ip[1])) {
859 			ip[0] = DT_MAKE_RETL(DT_RS2(ip[1]));
860 		} else {
861 			ip[0] = DT_OP_NOP;
862 			(*off) += sizeof (ip[0]);
863 		}
864 	}
865 
866 	return (0);
867 }
868 
869 #elif defined(__i386) || defined(__amd64)
870 
871 #define	DT_OP_NOP		0x90
872 #define	DT_OP_RET		0xc3
873 #define	DT_OP_CALL		0xe8
874 #define	DT_OP_JMP32		0xe9
875 #define	DT_OP_REX_RAX		0x48
876 #define	DT_OP_XOR_EAX_0		0x33
877 #define	DT_OP_XOR_EAX_1		0xc0
878 
879 static int
880 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
881     uint32_t *off)
882 {
883 	uint8_t *ip = (uint8_t *)(p + rela->r_offset - 1);
884 	uint8_t ret;
885 
886 	/*
887 	 * On x86, the first byte of the instruction is the call opcode and
888 	 * the next four bytes are the 32-bit address; the relocation is for
889 	 * the address operand. We back up the offset to the first byte of
890 	 * the instruction. For is-enabled probes, we later advance the offset
891 	 * so that it hits the first nop in the instruction sequence.
892 	 */
893 	(*off) -= 1;
894 
895 	/*
896 	 * We only know about some specific relocation types. Luckily
897 	 * these types have the same values on both 32-bit and 64-bit
898 	 * x86 architectures.
899 	 */
900 	if (GELF_R_TYPE(rela->r_info) != R_386_PC32 &&
901 	    GELF_R_TYPE(rela->r_info) != R_386_PLT32)
902 		return (-1);
903 
904 	/*
905 	 * We may have already processed this object file in an earlier linker
906 	 * invocation. Check to see if the present instruction sequence matches
907 	 * the one we would install. For is-enabled probes, we advance the
908 	 * offset to the first nop instruction in the sequence to match the
909 	 * text modification code below.
910 	 */
911 	if (!isenabled) {
912 		if ((ip[0] == DT_OP_NOP || ip[0] == DT_OP_RET) &&
913 		    ip[1] == DT_OP_NOP && ip[2] == DT_OP_NOP &&
914 		    ip[3] == DT_OP_NOP && ip[4] == DT_OP_NOP)
915 			return (0);
916 	} else if (dtp->dt_oflags & DTRACE_O_LP64) {
917 		if (ip[0] == DT_OP_REX_RAX &&
918 		    ip[1] == DT_OP_XOR_EAX_0 && ip[2] == DT_OP_XOR_EAX_1 &&
919 		    (ip[3] == DT_OP_NOP || ip[3] == DT_OP_RET) &&
920 		    ip[4] == DT_OP_NOP) {
921 			(*off) += 3;
922 			return (0);
923 		}
924 	} else {
925 		if (ip[0] == DT_OP_XOR_EAX_0 && ip[1] == DT_OP_XOR_EAX_1 &&
926 		    (ip[2] == DT_OP_NOP || ip[2] == DT_OP_RET) &&
927 		    ip[3] == DT_OP_NOP && ip[4] == DT_OP_NOP) {
928 			(*off) += 2;
929 			return (0);
930 		}
931 	}
932 
933 	/*
934 	 * We expect either a call instrution with a 32-bit displacement or a
935 	 * jmp instruction with a 32-bit displacement acting as a tail-call.
936 	 */
937 	if (ip[0] != DT_OP_CALL && ip[0] != DT_OP_JMP32) {
938 		dt_dprintf("found %x instead of a call or jmp instruction at "
939 		    "%llx\n", ip[0], (u_longlong_t)rela->r_offset);
940 		return (-1);
941 	}
942 
943 	ret = (ip[0] == DT_OP_JMP32) ? DT_OP_RET : DT_OP_NOP;
944 
945 	/*
946 	 * Establish the instruction sequence -- all nops for probes, and an
947 	 * instruction to clear the return value register (%eax/%rax) followed
948 	 * by nops for is-enabled probes. For is-enabled probes, we advance
949 	 * the offset to the first nop. This isn't stricly necessary but makes
950 	 * for more readable disassembly when the probe is enabled.
951 	 */
952 	if (!isenabled) {
953 		ip[0] = ret;
954 		ip[1] = DT_OP_NOP;
955 		ip[2] = DT_OP_NOP;
956 		ip[3] = DT_OP_NOP;
957 		ip[4] = DT_OP_NOP;
958 	} else if (dtp->dt_oflags & DTRACE_O_LP64) {
959 		ip[0] = DT_OP_REX_RAX;
960 		ip[1] = DT_OP_XOR_EAX_0;
961 		ip[2] = DT_OP_XOR_EAX_1;
962 		ip[3] = ret;
963 		ip[4] = DT_OP_NOP;
964 		(*off) += 3;
965 	} else {
966 		ip[0] = DT_OP_XOR_EAX_0;
967 		ip[1] = DT_OP_XOR_EAX_1;
968 		ip[2] = ret;
969 		ip[3] = DT_OP_NOP;
970 		ip[4] = DT_OP_NOP;
971 		(*off) += 2;
972 	}
973 
974 	return (0);
975 }
976 
977 #else
978 #error unknown ISA
979 #endif
980 
981 /*PRINTFLIKE5*/
982 static int
983 dt_link_error(dtrace_hdl_t *dtp, Elf *elf, int fd, dt_link_pair_t *bufs,
984     const char *format, ...)
985 {
986 	va_list ap;
987 	dt_link_pair_t *pair;
988 
989 	va_start(ap, format);
990 	dt_set_errmsg(dtp, NULL, NULL, NULL, 0, format, ap);
991 	va_end(ap);
992 
993 	if (elf != NULL)
994 		(void) elf_end(elf);
995 
996 	if (fd >= 0)
997 		(void) close(fd);
998 
999 	while ((pair = bufs) != NULL) {
1000 		bufs = pair->dlp_next;
1001 		dt_free(dtp, pair->dlp_str);
1002 		dt_free(dtp, pair->dlp_sym);
1003 		dt_free(dtp, pair);
1004 	}
1005 
1006 	return (dt_set_errno(dtp, EDT_COMPILER));
1007 }
1008 
1009 static int
1010 process_obj(dtrace_hdl_t *dtp, const char *obj, int *eprobesp)
1011 {
1012 	static const char dt_prefix[] = "__dtrace";
1013 	static const char dt_enabled[] = "enabled";
1014 	static const char dt_symprefix[] = "$dtrace";
1015 	static const char dt_symfmt[] = "%s%d.%s";
1016 	int fd, i, ndx, eprobe, mod = 0;
1017 	Elf *elf = NULL;
1018 	GElf_Ehdr ehdr;
1019 	Elf_Scn *scn_rel, *scn_sym, *scn_str, *scn_tgt;
1020 	Elf_Data *data_rel, *data_sym, *data_str, *data_tgt;
1021 	GElf_Shdr shdr_rel, shdr_sym, shdr_str, shdr_tgt;
1022 	GElf_Sym rsym, fsym, dsym;
1023 	GElf_Rela rela;
1024 	char *s, *p, *r;
1025 	char pname[DTRACE_PROVNAMELEN];
1026 	dt_provider_t *pvp;
1027 	dt_probe_t *prp;
1028 	uint32_t off, eclass, emachine1, emachine2;
1029 	size_t symsize, nsym, isym, istr, len;
1030 	key_t objkey;
1031 	dt_link_pair_t *pair, *bufs = NULL;
1032 	dt_strtab_t *strtab;
1033 
1034 	if ((fd = open64(obj, O_RDWR)) == -1) {
1035 		return (dt_link_error(dtp, elf, fd, bufs,
1036 		    "failed to open %s: %s", obj, strerror(errno)));
1037 	}
1038 
1039 	if ((elf = elf_begin(fd, ELF_C_RDWR, NULL)) == NULL) {
1040 		return (dt_link_error(dtp, elf, fd, bufs,
1041 		    "failed to process %s: %s", obj, elf_errmsg(elf_errno())));
1042 	}
1043 
1044 	switch (elf_kind(elf)) {
1045 	case ELF_K_ELF:
1046 		break;
1047 	case ELF_K_AR:
1048 		return (dt_link_error(dtp, elf, fd, bufs, "archives are not "
1049 		    "permitted; use the contents of the archive instead: %s",
1050 		    obj));
1051 	default:
1052 		return (dt_link_error(dtp, elf, fd, bufs,
1053 		    "invalid file type: %s", obj));
1054 	}
1055 
1056 	if (gelf_getehdr(elf, &ehdr) == NULL) {
1057 		return (dt_link_error(dtp, elf, fd, bufs, "corrupt file: %s",
1058 		    obj));
1059 	}
1060 
1061 	if (dtp->dt_oflags & DTRACE_O_LP64) {
1062 		eclass = ELFCLASS64;
1063 #if defined(__sparc)
1064 		emachine1 = emachine2 = EM_SPARCV9;
1065 #elif defined(__i386) || defined(__amd64)
1066 		emachine1 = emachine2 = EM_AMD64;
1067 #endif
1068 		symsize = sizeof (Elf64_Sym);
1069 	} else {
1070 		eclass = ELFCLASS32;
1071 #if defined(__sparc)
1072 		emachine1 = EM_SPARC;
1073 		emachine2 = EM_SPARC32PLUS;
1074 #elif defined(__i386) || defined(__amd64)
1075 		emachine1 = emachine2 = EM_386;
1076 #endif
1077 		symsize = sizeof (Elf32_Sym);
1078 	}
1079 
1080 	if (ehdr.e_ident[EI_CLASS] != eclass) {
1081 		return (dt_link_error(dtp, elf, fd, bufs,
1082 		    "incorrect ELF class for object file: %s", obj));
1083 	}
1084 
1085 	if (ehdr.e_machine != emachine1 && ehdr.e_machine != emachine2) {
1086 		return (dt_link_error(dtp, elf, fd, bufs,
1087 		    "incorrect ELF machine type for object file: %s", obj));
1088 	}
1089 
1090 	/*
1091 	 * We use this token as a relatively unique handle for this file on the
1092 	 * system in order to disambiguate potential conflicts between files of
1093 	 * the same name which contain identially named local symbols.
1094 	 */
1095 	if ((objkey = ftok(obj, 0)) == (key_t)-1) {
1096 		return (dt_link_error(dtp, elf, fd, bufs,
1097 		    "failed to generate unique key for object file: %s", obj));
1098 	}
1099 
1100 	scn_rel = NULL;
1101 	while ((scn_rel = elf_nextscn(elf, scn_rel)) != NULL) {
1102 		if (gelf_getshdr(scn_rel, &shdr_rel) == NULL)
1103 			goto err;
1104 
1105 		/*
1106 		 * Skip any non-relocation sections.
1107 		 */
1108 		if (shdr_rel.sh_type != SHT_RELA && shdr_rel.sh_type != SHT_REL)
1109 			continue;
1110 
1111 		if ((data_rel = elf_getdata(scn_rel, NULL)) == NULL)
1112 			goto err;
1113 
1114 		/*
1115 		 * Grab the section, section header and section data for the
1116 		 * symbol table that this relocation section references.
1117 		 */
1118 		if ((scn_sym = elf_getscn(elf, shdr_rel.sh_link)) == NULL ||
1119 		    gelf_getshdr(scn_sym, &shdr_sym) == NULL ||
1120 		    (data_sym = elf_getdata(scn_sym, NULL)) == NULL)
1121 			goto err;
1122 
1123 		/*
1124 		 * Ditto for that symbol table's string table.
1125 		 */
1126 		if ((scn_str = elf_getscn(elf, shdr_sym.sh_link)) == NULL ||
1127 		    gelf_getshdr(scn_str, &shdr_str) == NULL ||
1128 		    (data_str = elf_getdata(scn_str, NULL)) == NULL)
1129 			goto err;
1130 
1131 		/*
1132 		 * Grab the section, section header and section data for the
1133 		 * target section for the relocations. For the relocations
1134 		 * we're looking for -- this will typically be the text of the
1135 		 * object file.
1136 		 */
1137 		if ((scn_tgt = elf_getscn(elf, shdr_rel.sh_info)) == NULL ||
1138 		    gelf_getshdr(scn_tgt, &shdr_tgt) == NULL ||
1139 		    (data_tgt = elf_getdata(scn_tgt, NULL)) == NULL)
1140 			goto err;
1141 
1142 		/*
1143 		 * We're looking for relocations to symbols matching this form:
1144 		 *
1145 		 *   __dtrace[enabled]_<prov>___<probe>
1146 		 *
1147 		 * For the generated object, we need to record the location
1148 		 * identified by the relocation, and create a new relocation
1149 		 * in the generated object that will be resolved at link time
1150 		 * to the location of the function in which the probe is
1151 		 * embedded. In the target object, we change the matched symbol
1152 		 * so that it will be ignored at link time, and we modify the
1153 		 * target (text) section to replace the call instruction with
1154 		 * one or more nops.
1155 		 *
1156 		 * If the function containing the probe is locally scoped
1157 		 * (static), we create an alias used by the relocation in the
1158 		 * generated object. The alias, a new symbol, will be global
1159 		 * (so that the relocation from the generated object can be
1160 		 * resolved), and hidden (so that it is converted to a local
1161 		 * symbol at link time). Such aliases have this form:
1162 		 *
1163 		 *   $dtrace<key>.<function>
1164 		 *
1165 		 * We take a first pass through all the relocations to
1166 		 * populate our string table and count the number of extra
1167 		 * symbols we'll require.
1168 		 */
1169 		strtab = dt_strtab_create(1);
1170 		nsym = 0;
1171 		isym = data_sym->d_size / symsize;
1172 		istr = data_str->d_size;
1173 
1174 		for (i = 0; i < shdr_rel.sh_size / shdr_rel.sh_entsize; i++) {
1175 
1176 			if (shdr_rel.sh_type == SHT_RELA) {
1177 				if (gelf_getrela(data_rel, i, &rela) == NULL)
1178 					continue;
1179 			} else {
1180 				GElf_Rel rel;
1181 				if (gelf_getrel(data_rel, i, &rel) == NULL)
1182 					continue;
1183 				rela.r_offset = rel.r_offset;
1184 				rela.r_info = rel.r_info;
1185 				rela.r_addend = 0;
1186 			}
1187 
1188 			if (gelf_getsym(data_sym, GELF_R_SYM(rela.r_info),
1189 			    &rsym) == NULL) {
1190 				dt_strtab_destroy(strtab);
1191 				goto err;
1192 			}
1193 
1194 			s = (char *)data_str->d_buf + rsym.st_name;
1195 
1196 			if (strncmp(s, dt_prefix, sizeof (dt_prefix) - 1) != 0)
1197 				continue;
1198 
1199 			if (dt_symtab_lookup(data_sym, isym, rela.r_offset,
1200 			    shdr_rel.sh_info, &fsym) != 0) {
1201 				dt_strtab_destroy(strtab);
1202 				goto err;
1203 			}
1204 
1205 			if (GELF_ST_BIND(fsym.st_info) != STB_LOCAL)
1206 				continue;
1207 
1208 			if (fsym.st_name > data_str->d_size) {
1209 				dt_strtab_destroy(strtab);
1210 				goto err;
1211 			}
1212 
1213 			s = (char *)data_str->d_buf + fsym.st_name;
1214 
1215 			/*
1216 			 * If this symbol isn't of type function, we've really
1217 			 * driven off the rails or the object file is corrupt.
1218 			 */
1219 			if (GELF_ST_TYPE(fsym.st_info) != STT_FUNC) {
1220 				dt_strtab_destroy(strtab);
1221 				return (dt_link_error(dtp, elf, fd, bufs,
1222 				    "expected %s to be of type function", s));
1223 			}
1224 
1225 			len = snprintf(NULL, 0, dt_symfmt, dt_symprefix,
1226 			    objkey, s) + 1;
1227 			if ((p = dt_alloc(dtp, len)) == NULL) {
1228 				dt_strtab_destroy(strtab);
1229 				goto err;
1230 			}
1231 			(void) snprintf(p, len, dt_symfmt, dt_symprefix,
1232 			    objkey, s);
1233 
1234 			if (dt_strtab_index(strtab, p) == -1) {
1235 				nsym++;
1236 				(void) dt_strtab_insert(strtab, p);
1237 			}
1238 
1239 			dt_free(dtp, p);
1240 		}
1241 
1242 		/*
1243 		 * If needed, allocate the additional space for the symbol
1244 		 * table and string table copying the old data into the new
1245 		 * buffers, and marking the buffers as dirty. We inject those
1246 		 * newly allocated buffers into the libelf data structures, but
1247 		 * are still responsible for freeing them once we're done with
1248 		 * the elf handle.
1249 		 */
1250 		if (nsym > 0) {
1251 			/*
1252 			 * The first byte of the string table is reserved for
1253 			 * the \0 entry.
1254 			 */
1255 			len = dt_strtab_size(strtab) - 1;
1256 
1257 			assert(len > 0);
1258 			assert(dt_strtab_index(strtab, "") == 0);
1259 
1260 			dt_strtab_destroy(strtab);
1261 
1262 			if ((pair = dt_alloc(dtp, sizeof (*pair))) == NULL)
1263 				goto err;
1264 
1265 			if ((pair->dlp_str = dt_alloc(dtp, data_str->d_size +
1266 			    len)) == NULL) {
1267 				dt_free(dtp, pair);
1268 				goto err;
1269 			}
1270 
1271 			if ((pair->dlp_sym = dt_alloc(dtp, data_sym->d_size +
1272 			    nsym * symsize)) == NULL) {
1273 				dt_free(dtp, pair->dlp_str);
1274 				dt_free(dtp, pair);
1275 				goto err;
1276 			}
1277 
1278 			pair->dlp_next = bufs;
1279 			bufs = pair;
1280 
1281 			bcopy(data_str->d_buf, pair->dlp_str, data_str->d_size);
1282 			data_str->d_buf = pair->dlp_str;
1283 			data_str->d_size += len;
1284 			(void) elf_flagdata(data_str, ELF_C_SET, ELF_F_DIRTY);
1285 
1286 			shdr_str.sh_size += len;
1287 			(void) gelf_update_shdr(scn_str, &shdr_str);
1288 
1289 			bcopy(data_sym->d_buf, pair->dlp_sym, data_sym->d_size);
1290 			data_sym->d_buf = pair->dlp_sym;
1291 			data_sym->d_size += nsym * symsize;
1292 			(void) elf_flagdata(data_sym, ELF_C_SET, ELF_F_DIRTY);
1293 
1294 			shdr_sym.sh_size += nsym * symsize;
1295 			(void) gelf_update_shdr(scn_sym, &shdr_sym);
1296 
1297 			nsym += isym;
1298 		} else {
1299 			dt_strtab_destroy(strtab);
1300 		}
1301 
1302 		/*
1303 		 * Now that the tables have been allocated, perform the
1304 		 * modifications described above.
1305 		 */
1306 		for (i = 0; i < shdr_rel.sh_size / shdr_rel.sh_entsize; i++) {
1307 
1308 			if (shdr_rel.sh_type == SHT_RELA) {
1309 				if (gelf_getrela(data_rel, i, &rela) == NULL)
1310 					continue;
1311 			} else {
1312 				GElf_Rel rel;
1313 				if (gelf_getrel(data_rel, i, &rel) == NULL)
1314 					continue;
1315 				rela.r_offset = rel.r_offset;
1316 				rela.r_info = rel.r_info;
1317 				rela.r_addend = 0;
1318 			}
1319 
1320 			ndx = GELF_R_SYM(rela.r_info);
1321 
1322 			if (gelf_getsym(data_sym, ndx, &rsym) == NULL ||
1323 			    rsym.st_name > data_str->d_size)
1324 				goto err;
1325 
1326 			s = (char *)data_str->d_buf + rsym.st_name;
1327 
1328 			if (strncmp(s, dt_prefix, sizeof (dt_prefix) - 1) != 0)
1329 				continue;
1330 
1331 			s += sizeof (dt_prefix) - 1;
1332 
1333 			/*
1334 			 * Check to see if this is an 'is-enabled' check as
1335 			 * opposed to a normal probe.
1336 			 */
1337 			if (strncmp(s, dt_enabled,
1338 			    sizeof (dt_enabled) - 1) == 0) {
1339 				s += sizeof (dt_enabled) - 1;
1340 				eprobe = 1;
1341 				*eprobesp = 1;
1342 				dt_dprintf("is-enabled probe\n");
1343 			} else {
1344 				eprobe = 0;
1345 				dt_dprintf("normal probe\n");
1346 			}
1347 
1348 			if (*s++ != '_')
1349 				goto err;
1350 
1351 			if ((p = strstr(s, "___")) == NULL ||
1352 			    p - s >= sizeof (pname))
1353 				goto err;
1354 
1355 			bcopy(s, pname, p - s);
1356 			pname[p - s] = '\0';
1357 
1358 			p = strhyphenate(p + 3); /* strlen("___") */
1359 
1360 			if (dt_symtab_lookup(data_sym, isym, rela.r_offset,
1361 			    shdr_rel.sh_info, &fsym) != 0)
1362 				goto err;
1363 
1364 			if (fsym.st_name > data_str->d_size)
1365 				goto err;
1366 
1367 			assert(GELF_ST_TYPE(fsym.st_info) == STT_FUNC);
1368 
1369 			/*
1370 			 * If a NULL relocation name is passed to
1371 			 * dt_probe_define(), the function name is used for the
1372 			 * relocation. The relocation needs to use a mangled
1373 			 * name if the symbol is locally scoped; the function
1374 			 * name may need to change if we've found the global
1375 			 * alias for the locally scoped symbol (we prefer
1376 			 * global symbols to locals in dt_symtab_lookup()).
1377 			 */
1378 			s = (char *)data_str->d_buf + fsym.st_name;
1379 			r = NULL;
1380 
1381 			if (GELF_ST_BIND(fsym.st_info) == STB_LOCAL) {
1382 				dsym = fsym;
1383 				dsym.st_name = istr;
1384 				dsym.st_info = GELF_ST_INFO(STB_GLOBAL,
1385 				    STT_FUNC);
1386 				dsym.st_other =
1387 				    ELF64_ST_VISIBILITY(STV_ELIMINATE);
1388 				(void) gelf_update_sym(data_sym, isym, &dsym);
1389 
1390 				r = (char *)data_str->d_buf + istr;
1391 				istr += 1 + sprintf(r, dt_symfmt,
1392 				    dt_symprefix, objkey, s);
1393 				isym++;
1394 				assert(isym <= nsym);
1395 
1396 			} else if (strncmp(s, dt_symprefix,
1397 			    strlen(dt_symprefix)) == 0) {
1398 				r = s;
1399 				if ((s = strchr(s, '.')) == NULL)
1400 					goto err;
1401 				s++;
1402 			}
1403 
1404 			if ((pvp = dt_provider_lookup(dtp, pname)) == NULL) {
1405 				return (dt_link_error(dtp, elf, fd, bufs,
1406 				    "no such provider %s", pname));
1407 			}
1408 
1409 			if ((prp = dt_probe_lookup(pvp, p)) == NULL) {
1410 				return (dt_link_error(dtp, elf, fd, bufs,
1411 				    "no such probe %s", p));
1412 			}
1413 
1414 			assert(fsym.st_value <= rela.r_offset);
1415 
1416 			off = rela.r_offset - fsym.st_value;
1417 			if (dt_modtext(dtp, data_tgt->d_buf, eprobe,
1418 			    &rela, &off) != 0) {
1419 				goto err;
1420 			}
1421 
1422 			if (dt_probe_define(pvp, prp, s, r, off, eprobe) != 0) {
1423 				return (dt_link_error(dtp, elf, fd, bufs,
1424 				    "failed to allocate space for probe"));
1425 			}
1426 
1427 			mod = 1;
1428 			(void) elf_flagdata(data_tgt, ELF_C_SET, ELF_F_DIRTY);
1429 
1430 			/*
1431 			 * This symbol may already have been marked to
1432 			 * be ignored by another relocation referencing
1433 			 * the same symbol or if this object file has
1434 			 * already been processed by an earlier link
1435 			 * invocation.
1436 			 */
1437 			if (rsym.st_shndx != SHN_SUNW_IGNORE) {
1438 				rsym.st_shndx = SHN_SUNW_IGNORE;
1439 				(void) gelf_update_sym(data_sym, ndx, &rsym);
1440 			}
1441 		}
1442 	}
1443 
1444 	if (mod && elf_update(elf, ELF_C_WRITE) == -1)
1445 		goto err;
1446 
1447 	(void) elf_end(elf);
1448 	(void) close(fd);
1449 
1450 	while ((pair = bufs) != NULL) {
1451 		bufs = pair->dlp_next;
1452 		dt_free(dtp, pair->dlp_str);
1453 		dt_free(dtp, pair->dlp_sym);
1454 		dt_free(dtp, pair);
1455 	}
1456 
1457 	return (0);
1458 
1459 err:
1460 	return (dt_link_error(dtp, elf, fd, bufs,
1461 	    "an error was encountered while processing %s", obj));
1462 }
1463 
1464 int
1465 dtrace_program_link(dtrace_hdl_t *dtp, dtrace_prog_t *pgp, uint_t dflags,
1466     const char *file, int objc, char *const objv[])
1467 {
1468 	char drti[PATH_MAX];
1469 	dof_hdr_t *dof;
1470 	int fd, status, i, cur;
1471 	char *cmd, tmp;
1472 	size_t len;
1473 	int eprobes = 0, ret = 0;
1474 
1475 	/*
1476 	 * A NULL program indicates a special use in which we just link
1477 	 * together a bunch of object files specified in objv and then
1478 	 * unlink(2) those object files.
1479 	 */
1480 	if (pgp == NULL) {
1481 		const char *fmt = "%s -o %s -r";
1482 
1483 		len = snprintf(&tmp, 1, fmt, dtp->dt_ld_path, file) + 1;
1484 
1485 		for (i = 0; i < objc; i++)
1486 			len += strlen(objv[i]) + 1;
1487 
1488 		cmd = alloca(len);
1489 
1490 		cur = snprintf(cmd, len, fmt, dtp->dt_ld_path, file);
1491 
1492 		for (i = 0; i < objc; i++)
1493 			cur += snprintf(cmd + cur, len - cur, " %s", objv[i]);
1494 
1495 		if ((status = system(cmd)) == -1) {
1496 			return (dt_link_error(dtp, NULL, -1, NULL,
1497 			    "failed to run %s: %s", dtp->dt_ld_path,
1498 			    strerror(errno)));
1499 		}
1500 
1501 		if (WIFSIGNALED(status)) {
1502 			return (dt_link_error(dtp, NULL, -1, NULL,
1503 			    "failed to link %s: %s failed due to signal %d",
1504 			    file, dtp->dt_ld_path, WTERMSIG(status)));
1505 		}
1506 
1507 		if (WEXITSTATUS(status) != 0) {
1508 			return (dt_link_error(dtp, NULL, -1, NULL,
1509 			    "failed to link %s: %s exited with status %d\n",
1510 			    file, dtp->dt_ld_path, WEXITSTATUS(status)));
1511 		}
1512 
1513 		for (i = 0; i < objc; i++) {
1514 			if (strcmp(objv[i], file) != 0)
1515 				(void) unlink(objv[i]);
1516 		}
1517 
1518 		return (0);
1519 	}
1520 
1521 	for (i = 0; i < objc; i++) {
1522 		if (process_obj(dtp, objv[i], &eprobes) != 0)
1523 			return (-1); /* errno is set for us */
1524 	}
1525 
1526 	/*
1527 	 * If there are is-enabled probes then we need to force use of DOF
1528 	 * version 2.
1529 	 */
1530 	if (eprobes && pgp->dp_dofversion < DOF_VERSION_2)
1531 		pgp->dp_dofversion = DOF_VERSION_2;
1532 
1533 	if ((dof = dtrace_dof_create(dtp, pgp, dflags)) == NULL)
1534 		return (-1); /* errno is set for us */
1535 
1536 	/*
1537 	 * Create a temporary file and then unlink it if we're going to
1538 	 * combine it with drti.o later.  We can still refer to it in child
1539 	 * processes as /dev/fd/<fd>.
1540 	 */
1541 	if ((fd = open64(file, O_RDWR | O_CREAT | O_TRUNC, 0666)) == -1) {
1542 		return (dt_link_error(dtp, NULL, -1, NULL,
1543 		    "failed to open %s: %s", file, strerror(errno)));
1544 	}
1545 
1546 	/*
1547 	 * If -xlinktype=DOF has been selected, just write out the DOF.
1548 	 * Otherwise proceed to the default of generating and linking ELF.
1549 	 */
1550 	switch (dtp->dt_linktype) {
1551 	case DT_LTYP_DOF:
1552 		if (dt_write(dtp, fd, dof, dof->dofh_filesz) < dof->dofh_filesz)
1553 			ret = errno;
1554 
1555 		if (close(fd) != 0 && ret == 0)
1556 			ret = errno;
1557 
1558 		if (ret != 0) {
1559 			return (dt_link_error(dtp, NULL, -1, NULL,
1560 			    "failed to write %s: %s", file, strerror(ret)));
1561 		}
1562 
1563 		return (0);
1564 
1565 	case DT_LTYP_ELF:
1566 		break; /* fall through to the rest of dtrace_program_link() */
1567 
1568 	default:
1569 		return (dt_link_error(dtp, NULL, -1, NULL,
1570 		    "invalid link type %u\n", dtp->dt_linktype));
1571 	}
1572 
1573 
1574 	if (!dtp->dt_lazyload)
1575 		(void) unlink(file);
1576 
1577 	if (dtp->dt_oflags & DTRACE_O_LP64)
1578 		status = dump_elf64(dtp, dof, fd);
1579 	else
1580 		status = dump_elf32(dtp, dof, fd);
1581 
1582 	if (status != 0 || lseek(fd, 0, SEEK_SET) != 0) {
1583 		return (dt_link_error(dtp, NULL, -1, NULL,
1584 		    "failed to write %s: %s", file, strerror(errno)));
1585 	}
1586 
1587 	if (!dtp->dt_lazyload) {
1588 		const char *fmt = "%s -o %s -r -Blocal -Breduce /dev/fd/%d %s";
1589 
1590 		if (dtp->dt_oflags & DTRACE_O_LP64) {
1591 			(void) snprintf(drti, sizeof (drti),
1592 			    "%s/64/drti.o", _dtrace_libdir);
1593 		} else {
1594 			(void) snprintf(drti, sizeof (drti),
1595 			    "%s/drti.o", _dtrace_libdir);
1596 		}
1597 
1598 		len = snprintf(&tmp, 1, fmt, dtp->dt_ld_path, file, fd,
1599 		    drti) + 1;
1600 
1601 		cmd = alloca(len);
1602 
1603 		(void) snprintf(cmd, len, fmt, dtp->dt_ld_path, file, fd, drti);
1604 
1605 		if ((status = system(cmd)) == -1) {
1606 			ret = dt_link_error(dtp, NULL, -1, NULL,
1607 			    "failed to run %s: %s", dtp->dt_ld_path,
1608 			    strerror(errno));
1609 			goto done;
1610 		}
1611 
1612 		(void) close(fd); /* release temporary file */
1613 
1614 		if (WIFSIGNALED(status)) {
1615 			ret = dt_link_error(dtp, NULL, -1, NULL,
1616 			    "failed to link %s: %s failed due to signal %d",
1617 			    file, dtp->dt_ld_path, WTERMSIG(status));
1618 			goto done;
1619 		}
1620 
1621 		if (WEXITSTATUS(status) != 0) {
1622 			ret = dt_link_error(dtp, NULL, -1, NULL,
1623 			    "failed to link %s: %s exited with status %d\n",
1624 			    file, dtp->dt_ld_path, WEXITSTATUS(status));
1625 			goto done;
1626 		}
1627 	} else {
1628 		(void) close(fd);
1629 	}
1630 
1631 done:
1632 	dtrace_dof_destroy(dtp, dof);
1633 	return (ret);
1634 }
1635