xref: /netbsd-src/external/cddl/osnet/dist/lib/libdtrace/common/dt_link.c (revision 9df6d2eefa8d7ebc0f702b139753132fc1699e9f)
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 #ifdef __NetBSD__
31 #ifdef __x86_64__
32 #include <i386/elf_machdep.h>
33 #undef ELF32_MACHDEP_ID_CASES
34 #undef ELF64_MACHDEP_ID_CASES
35 #undef ELF64_MACHDEP_ENDIANNESS
36 #undef KERN_ELFSIZE
37 #undef ARCH_ELFSIZE
38 #undef R_TYPE
39 #endif
40 #endif
41 #include <elf.h>
42 
43 #include <sys/types.h>
44 #ifdef illumos
45 #include <sys/sysmacros.h>
46 #else
47 #define	P2ROUNDUP(x, align)		(-(-(x) & -(align)))
48 #endif
49 
50 #include <unistd.h>
51 #include <strings.h>
52 #ifdef illumos
53 #include <alloca.h>
54 #endif
55 #include <limits.h>
56 #include <stddef.h>
57 #include <stdlib.h>
58 #include <stdio.h>
59 #include <fcntl.h>
60 #include <errno.h>
61 #ifdef illumos
62 #include <wait.h>
63 #else
64 #include <sys/wait.h>
65 #include <libelf.h>
66 #include <gelf.h>
67 #include <sys/mman.h>
68 #endif
69 #include <assert.h>
70 #include <sys/ipc.h>
71 
72 #include <dt_impl.h>
73 #include <dt_provider.h>
74 #include <dt_program.h>
75 #include <dt_string.h>
76 
77 #define	ESHDR_NULL	0
78 #define	ESHDR_SHSTRTAB	1
79 #define	ESHDR_DOF	2
80 #define	ESHDR_STRTAB	3
81 #define	ESHDR_SYMTAB	4
82 #define	ESHDR_REL	5
83 #define	ESHDR_NUM	6
84 
85 #define	PWRITE_SCN(index, data) \
86 	(lseek64(fd, (off64_t)elf_file.shdr[(index)].sh_offset, SEEK_SET) != \
87 	(off64_t)elf_file.shdr[(index)].sh_offset || \
88 	dt_write(dtp, fd, (data), elf_file.shdr[(index)].sh_size) != \
89 	elf_file.shdr[(index)].sh_size)
90 
91 static const char DTRACE_SHSTRTAB32[] = "\0"
92 ".shstrtab\0"		/* 1 */
93 ".SUNW_dof\0"		/* 11 */
94 ".strtab\0"		/* 21 */
95 ".symtab\0"		/* 29 */
96 #ifdef __sparc
97 ".rela.SUNW_dof";	/* 37 */
98 #else
99 ".rel.SUNW_dof";	/* 37 */
100 #endif
101 
102 static const char DTRACE_SHSTRTAB64[] = "\0"
103 ".shstrtab\0"		/* 1 */
104 ".SUNW_dof\0"		/* 11 */
105 ".strtab\0"		/* 21 */
106 ".symtab\0"		/* 29 */
107 ".rela.SUNW_dof";	/* 37 */
108 
109 static const char DOFSTR[] = "__SUNW_dof";
110 static const char DOFLAZYSTR[] = "___SUNW_dof";
111 
112 typedef struct dt_link_pair {
113 	struct dt_link_pair *dlp_next;	/* next pair in linked list */
114 	void *dlp_str;			/* buffer for string table */
115 	void *dlp_sym;			/* buffer for symbol table */
116 } dt_link_pair_t;
117 
118 typedef struct dof_elf32 {
119 	uint32_t de_nrel;		/* relocation count */
120 #ifdef __sparc
121 	Elf32_Rela *de_rel;		/* array of relocations for sparc */
122 #else
123 	Elf32_Rel *de_rel;		/* array of relocations for x86 */
124 #endif
125 	uint32_t de_nsym;		/* symbol count */
126 	Elf32_Sym *de_sym;		/* array of symbols */
127 	uint32_t de_strlen;		/* size of of string table */
128 	char *de_strtab;		/* string table */
129 	uint32_t de_global;		/* index of the first global symbol */
130 } dof_elf32_t;
131 
132 static int
prepare_elf32(dtrace_hdl_t * dtp,const dof_hdr_t * dof,dof_elf32_t * dep)133 prepare_elf32(dtrace_hdl_t *dtp, const dof_hdr_t *dof, dof_elf32_t *dep)
134 {
135 	dof_sec_t *dofs, *s;
136 	dof_relohdr_t *dofrh;
137 	dof_relodesc_t *dofr;
138 	char *strtab;
139 	int i, j, nrel;
140 	size_t strtabsz = 1;
141 	uint32_t count = 0;
142 	size_t base;
143 	Elf32_Sym *sym;
144 #ifdef __sparc
145 	Elf32_Rela *rel;
146 #else
147 	Elf32_Rel *rel;
148 #endif
149 
150 	/*LINTED*/
151 	dofs = (dof_sec_t *)((char *)dof + dof->dofh_secoff);
152 
153 	/*
154 	 * First compute the size of the string table and the number of
155 	 * relocations present in the DOF.
156 	 */
157 	for (i = 0; i < dof->dofh_secnum; i++) {
158 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
159 			continue;
160 
161 		/*LINTED*/
162 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
163 
164 		s = &dofs[dofrh->dofr_strtab];
165 		strtab = (char *)dof + s->dofs_offset;
166 		assert(strtab[0] == '\0');
167 		strtabsz += s->dofs_size - 1;
168 
169 		s = &dofs[dofrh->dofr_relsec];
170 		/*LINTED*/
171 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
172 		count += s->dofs_size / s->dofs_entsize;
173 	}
174 
175 	dep->de_strlen = strtabsz;
176 	dep->de_nrel = count;
177 	dep->de_nsym = count + 1; /* the first symbol is always null */
178 
179 	if (dtp->dt_lazyload) {
180 		dep->de_strlen += sizeof (DOFLAZYSTR);
181 		dep->de_nsym++;
182 	} else {
183 		dep->de_strlen += sizeof (DOFSTR);
184 		dep->de_nsym++;
185 	}
186 
187 	if ((dep->de_rel = calloc(dep->de_nrel,
188 	    sizeof (dep->de_rel[0]))) == NULL) {
189 		return (dt_set_errno(dtp, EDT_NOMEM));
190 	}
191 
192 	if ((dep->de_sym = calloc(dep->de_nsym, sizeof (Elf32_Sym))) == NULL) {
193 		free(dep->de_rel);
194 		return (dt_set_errno(dtp, EDT_NOMEM));
195 	}
196 
197 	if ((dep->de_strtab = calloc(dep->de_strlen, 1)) == NULL) {
198 		free(dep->de_rel);
199 		free(dep->de_sym);
200 		return (dt_set_errno(dtp, EDT_NOMEM));
201 	}
202 
203 	count = 0;
204 	strtabsz = 1;
205 	dep->de_strtab[0] = '\0';
206 	rel = dep->de_rel;
207 	sym = dep->de_sym;
208 	dep->de_global = 1;
209 
210 	/*
211 	 * The first symbol table entry must be zeroed and is always ignored.
212 	 */
213 	bzero(sym, sizeof (Elf32_Sym));
214 	sym++;
215 
216 	/*
217 	 * Take a second pass through the DOF sections filling in the
218 	 * memory we allocated.
219 	 */
220 	for (i = 0; i < dof->dofh_secnum; i++) {
221 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
222 			continue;
223 
224 		/*LINTED*/
225 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
226 
227 		s = &dofs[dofrh->dofr_strtab];
228 		strtab = (char *)dof + s->dofs_offset;
229 		bcopy(strtab + 1, dep->de_strtab + strtabsz, s->dofs_size);
230 		base = strtabsz;
231 		strtabsz += s->dofs_size - 1;
232 
233 		s = &dofs[dofrh->dofr_relsec];
234 		/*LINTED*/
235 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
236 		nrel = s->dofs_size / s->dofs_entsize;
237 
238 		s = &dofs[dofrh->dofr_tgtsec];
239 
240 		for (j = 0; j < nrel; j++) {
241 #if defined(__aarch64__)
242 /* XXX */
243 printf("%s:%s(%d): DOODAD\n",__FUNCTION__,__FILE__,__LINE__);
244 #elif defined(__arm__)
245 /* XXX */
246 printf("%s:%s(%d): DOODAD\n",__FUNCTION__,__FILE__,__LINE__);
247 #elif defined(__i386) || defined(__amd64)
248 			rel->r_offset = s->dofs_offset +
249 			    dofr[j].dofr_offset;
250 			rel->r_info = ELF32_R_INFO(count + dep->de_global,
251 			    R_386_PC32);
252 #elif defined(__mips__)
253 /* XXX */
254 printf("%s:%s(%d): DOODAD\n",__FUNCTION__,__FILE__,__LINE__);
255 #elif defined(__powerpc__)
256 			/*
257 			 * Add 4 bytes to hit the low half of this 64-bit
258 			 * big-endian address.
259 			 */
260 			rel->r_offset = s->dofs_offset +
261 			    dofr[j].dofr_offset + 4;
262 			rel->r_info = ELF32_R_INFO(count + dep->de_global,
263 			    R_PPC_REL32);
264 #elif defined(__sparc)
265 			/*
266 			 * Add 4 bytes to hit the low half of this 64-bit
267 			 * big-endian address.
268 			 */
269 			rel->r_offset = s->dofs_offset +
270 			    dofr[j].dofr_offset + 4;
271 			rel->r_info = ELF32_R_INFO(count + dep->de_global,
272 			    R_SPARC_32);
273 #elif defined(__riscv__)
274 /* XXX */
275 printf("%s:%s(%d): DOODAD\n",__FUNCTION__,__FILE__,__LINE__);
276 #else
277 #error unknown ISA
278 #endif
279 
280 			sym->st_name = base + dofr[j].dofr_name - 1;
281 			sym->st_value = 0;
282 			sym->st_size = 0;
283 			sym->st_info = ELF32_ST_INFO(STB_GLOBAL, STT_FUNC);
284 			sym->st_other = 0;
285 			sym->st_shndx = SHN_UNDEF;
286 
287 			rel++;
288 			sym++;
289 			count++;
290 		}
291 	}
292 
293 	/*
294 	 * Add a symbol for the DOF itself. We use a different symbol for
295 	 * lazily and actively loaded DOF to make them easy to distinguish.
296 	 */
297 	sym->st_name = strtabsz;
298 	sym->st_value = 0;
299 	sym->st_size = dof->dofh_filesz;
300 	sym->st_info = ELF32_ST_INFO(STB_GLOBAL, STT_OBJECT);
301 	sym->st_other = ELF32_ST_VISIBILITY(STV_HIDDEN);
302 	sym->st_shndx = ESHDR_DOF;
303 	sym++;
304 
305 	if (dtp->dt_lazyload) {
306 		bcopy(DOFLAZYSTR, dep->de_strtab + strtabsz,
307 		    sizeof (DOFLAZYSTR));
308 		strtabsz += sizeof (DOFLAZYSTR);
309 	} else {
310 		bcopy(DOFSTR, dep->de_strtab + strtabsz, sizeof (DOFSTR));
311 		strtabsz += sizeof (DOFSTR);
312 	}
313 
314 	assert(count == dep->de_nrel);
315 	assert(strtabsz == dep->de_strlen);
316 
317 	return (0);
318 }
319 
320 
321 typedef struct dof_elf64 {
322 	uint32_t de_nrel;
323 	Elf64_Rela *de_rel;
324 	uint32_t de_nsym;
325 	Elf64_Sym *de_sym;
326 
327 	uint32_t de_strlen;
328 	char *de_strtab;
329 
330 	uint32_t de_global;
331 } dof_elf64_t;
332 
333 static int
prepare_elf64(dtrace_hdl_t * dtp,const dof_hdr_t * dof,dof_elf64_t * dep)334 prepare_elf64(dtrace_hdl_t *dtp, const dof_hdr_t *dof, dof_elf64_t *dep)
335 {
336 	dof_sec_t *dofs, *s;
337 	dof_relohdr_t *dofrh;
338 	dof_relodesc_t *dofr;
339 	char *strtab;
340 	int i, j, nrel;
341 	size_t strtabsz = 1;
342 #ifdef illumos
343 	uint32_t count = 0;
344 #else
345 	uint64_t count = 0;
346 #endif
347 	size_t base;
348 	Elf64_Sym *sym;
349 	Elf64_Rela *rel;
350 
351 	/*LINTED*/
352 	dofs = (dof_sec_t *)((char *)dof + dof->dofh_secoff);
353 
354 	/*
355 	 * First compute the size of the string table and the number of
356 	 * relocations present in the DOF.
357 	 */
358 	for (i = 0; i < dof->dofh_secnum; i++) {
359 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
360 			continue;
361 
362 		/*LINTED*/
363 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
364 
365 		s = &dofs[dofrh->dofr_strtab];
366 		strtab = (char *)dof + s->dofs_offset;
367 		assert(strtab[0] == '\0');
368 		strtabsz += s->dofs_size - 1;
369 
370 		s = &dofs[dofrh->dofr_relsec];
371 		/*LINTED*/
372 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
373 		count += s->dofs_size / s->dofs_entsize;
374 	}
375 
376 	dep->de_strlen = strtabsz;
377 	dep->de_nrel = count;
378 	dep->de_nsym = count + 1; /* the first symbol is always null */
379 
380 	if (dtp->dt_lazyload) {
381 		dep->de_strlen += sizeof (DOFLAZYSTR);
382 		dep->de_nsym++;
383 	} else {
384 		dep->de_strlen += sizeof (DOFSTR);
385 		dep->de_nsym++;
386 	}
387 
388 	if ((dep->de_rel = calloc(dep->de_nrel,
389 	    sizeof (dep->de_rel[0]))) == NULL) {
390 		return (dt_set_errno(dtp, EDT_NOMEM));
391 	}
392 
393 	if ((dep->de_sym = calloc(dep->de_nsym, sizeof (Elf64_Sym))) == NULL) {
394 		free(dep->de_rel);
395 		return (dt_set_errno(dtp, EDT_NOMEM));
396 	}
397 
398 	if ((dep->de_strtab = calloc(dep->de_strlen, 1)) == NULL) {
399 		free(dep->de_rel);
400 		free(dep->de_sym);
401 		return (dt_set_errno(dtp, EDT_NOMEM));
402 	}
403 
404 	count = 0;
405 	strtabsz = 1;
406 	dep->de_strtab[0] = '\0';
407 	rel = dep->de_rel;
408 	sym = dep->de_sym;
409 	dep->de_global = 1;
410 
411 	/*
412 	 * The first symbol table entry must be zeroed and is always ignored.
413 	 */
414 	bzero(sym, sizeof (Elf64_Sym));
415 	sym++;
416 
417 	/*
418 	 * Take a second pass through the DOF sections filling in the
419 	 * memory we allocated.
420 	 */
421 	for (i = 0; i < dof->dofh_secnum; i++) {
422 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
423 			continue;
424 
425 		/*LINTED*/
426 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
427 
428 		s = &dofs[dofrh->dofr_strtab];
429 		strtab = (char *)dof + s->dofs_offset;
430 		bcopy(strtab + 1, dep->de_strtab + strtabsz, s->dofs_size);
431 		base = strtabsz;
432 		strtabsz += s->dofs_size - 1;
433 
434 		s = &dofs[dofrh->dofr_relsec];
435 		/*LINTED*/
436 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
437 		nrel = s->dofs_size / s->dofs_entsize;
438 
439 		s = &dofs[dofrh->dofr_tgtsec];
440 
441 		for (j = 0; j < nrel; j++) {
442 #if defined(__aarch64__)
443 /* XXX */
444 #elif defined(__arm__)
445 /* XXX */
446 #elif defined(__mips__)
447 /* XXX */
448 #elif defined(__powerpc__)
449 			rel->r_offset = s->dofs_offset +
450 			    dofr[j].dofr_offset;
451 			rel->r_info = ELF64_R_INFO(count + dep->de_global,
452 			    R_PPC64_REL64);
453 #elif defined(__riscv__)
454 /* XXX */
455 #elif defined(__i386) || defined(__amd64)
456 #ifndef R_X86_64_PC64
457 #define R_X86_64_PC64           24
458 #endif
459 			rel->r_offset = s->dofs_offset +
460 			    dofr[j].dofr_offset;
461 			rel->r_info = ELF64_R_INFO(count + dep->de_global,
462 			    R_X86_64_PC64);
463 #else
464 #error unknown ISA
465 #endif
466 
467 			sym->st_name = base + dofr[j].dofr_name - 1;
468 			sym->st_value = 0;
469 			sym->st_size = 0;
470 			sym->st_info = GELF_ST_INFO(STB_GLOBAL, STT_FUNC);
471 			sym->st_other = 0;
472 			sym->st_shndx = SHN_UNDEF;
473 
474 			rel++;
475 			sym++;
476 			count++;
477 		}
478 	}
479 
480 	/*
481 	 * Add a symbol for the DOF itself. We use a different symbol for
482 	 * lazily and actively loaded DOF to make them easy to distinguish.
483 	 */
484 	sym->st_name = strtabsz;
485 	sym->st_value = 0;
486 	sym->st_size = dof->dofh_filesz;
487 	sym->st_info = GELF_ST_INFO(STB_GLOBAL, STT_OBJECT);
488 	sym->st_other = ELF64_ST_VISIBILITY(STV_HIDDEN);
489 	sym->st_shndx = ESHDR_DOF;
490 	sym++;
491 
492 	if (dtp->dt_lazyload) {
493 		bcopy(DOFLAZYSTR, dep->de_strtab + strtabsz,
494 		    sizeof (DOFLAZYSTR));
495 		strtabsz += sizeof (DOFLAZYSTR);
496 	} else {
497 		bcopy(DOFSTR, dep->de_strtab + strtabsz, sizeof (DOFSTR));
498 		strtabsz += sizeof (DOFSTR);
499 	}
500 
501 	assert(count == dep->de_nrel);
502 	assert(strtabsz == dep->de_strlen);
503 
504 	return (0);
505 }
506 
507 /*
508  * Write out an ELF32 file prologue consisting of a header, section headers,
509  * and a section header string table.  The DOF data will follow this prologue
510  * and complete the contents of the given ELF file.
511  */
512 static int
dump_elf32(dtrace_hdl_t * dtp,const dof_hdr_t * dof,int fd)513 dump_elf32(dtrace_hdl_t *dtp, const dof_hdr_t *dof, int fd)
514 {
515 	struct {
516 		Elf32_Ehdr ehdr;
517 		Elf32_Shdr shdr[ESHDR_NUM];
518 	} elf_file;
519 
520 	Elf32_Shdr *shp;
521 	Elf32_Off off;
522 	dof_elf32_t de;
523 	int ret = 0;
524 	uint_t nshdr;
525 
526 	memset(&de, 0, sizeof(de));	// XXX: gcc
527 	if (prepare_elf32(dtp, dof, &de) != 0)
528 		return (-1); /* errno is set for us */
529 
530 	/*
531 	 * If there are no relocations, we only need enough sections for
532 	 * the shstrtab and the DOF.
533 	 */
534 	nshdr = de.de_nrel == 0 ? ESHDR_SYMTAB + 1 : ESHDR_NUM;
535 
536 	bzero(&elf_file, sizeof (elf_file));
537 
538 	elf_file.ehdr.e_ident[EI_MAG0] = ELFMAG0;
539 	elf_file.ehdr.e_ident[EI_MAG1] = ELFMAG1;
540 	elf_file.ehdr.e_ident[EI_MAG2] = ELFMAG2;
541 	elf_file.ehdr.e_ident[EI_MAG3] = ELFMAG3;
542 	elf_file.ehdr.e_ident[EI_VERSION] = EV_CURRENT;
543 	elf_file.ehdr.e_ident[EI_CLASS] = ELFCLASS32;
544 #if BYTE_ORDER == _BIG_ENDIAN
545 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2MSB;
546 #else
547 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2LSB;
548 #endif
549 #if defined(__FreeBSD__) || defined(__NetBSD__)
550 	elf_file.ehdr.e_ident[EI_OSABI] = ELFOSABI_FREEBSD;
551 #endif
552 	elf_file.ehdr.e_type = ET_REL;
553 #if defined(__aarch64__)
554 	elf_file.ehdr.e_machine = EM_AARCH64;
555 #elif defined(__arm__)
556 	elf_file.ehdr.e_machine = EM_ARM;
557 #elif defined(__mips__)
558 	elf_file.ehdr.e_machine = EM_MIPS;
559 #elif defined(__powerpc__)
560 	elf_file.ehdr.e_machine = EM_PPC;
561 #elif defined(__sparc)
562 	elf_file.ehdr.e_machine = EM_SPARC;
563 #elif defined(__i386) || defined(__amd64)
564 	elf_file.ehdr.e_machine = EM_386;
565 #endif
566 	elf_file.ehdr.e_version = EV_CURRENT;
567 	elf_file.ehdr.e_shoff = sizeof (Elf32_Ehdr);
568 	elf_file.ehdr.e_ehsize = sizeof (Elf32_Ehdr);
569 	elf_file.ehdr.e_phentsize = sizeof (Elf32_Phdr);
570 	elf_file.ehdr.e_shentsize = sizeof (Elf32_Shdr);
571 	elf_file.ehdr.e_shnum = nshdr;
572 	elf_file.ehdr.e_shstrndx = ESHDR_SHSTRTAB;
573 	off = sizeof (elf_file) + nshdr * sizeof (Elf32_Shdr);
574 
575 	shp = &elf_file.shdr[ESHDR_SHSTRTAB];
576 	shp->sh_name = 1; /* DTRACE_SHSTRTAB32[1] = ".shstrtab" */
577 	shp->sh_type = SHT_STRTAB;
578 	shp->sh_offset = off;
579 	shp->sh_size = sizeof (DTRACE_SHSTRTAB32);
580 	shp->sh_addralign = sizeof (char);
581 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
582 
583 	shp = &elf_file.shdr[ESHDR_DOF];
584 	shp->sh_name = 11; /* DTRACE_SHSTRTAB32[11] = ".SUNW_dof" */
585 	shp->sh_flags = SHF_ALLOC;
586 	shp->sh_type = SHT_SUNW_dof;
587 	shp->sh_offset = off;
588 	shp->sh_size = dof->dofh_filesz;
589 	shp->sh_addralign = 8;
590 	off = shp->sh_offset + shp->sh_size;
591 
592 	shp = &elf_file.shdr[ESHDR_STRTAB];
593 	shp->sh_name = 21; /* DTRACE_SHSTRTAB32[21] = ".strtab" */
594 	shp->sh_flags = SHF_ALLOC;
595 	shp->sh_type = SHT_STRTAB;
596 	shp->sh_offset = off;
597 	shp->sh_size = de.de_strlen;
598 	shp->sh_addralign = sizeof (char);
599 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 4);
600 
601 	shp = &elf_file.shdr[ESHDR_SYMTAB];
602 	shp->sh_name = 29; /* DTRACE_SHSTRTAB32[29] = ".symtab" */
603 	shp->sh_flags = SHF_ALLOC;
604 	shp->sh_type = SHT_SYMTAB;
605 	shp->sh_entsize = sizeof (Elf32_Sym);
606 	shp->sh_link = ESHDR_STRTAB;
607 	shp->sh_offset = off;
608 	shp->sh_info = de.de_global;
609 	shp->sh_size = de.de_nsym * sizeof (Elf32_Sym);
610 	shp->sh_addralign = 4;
611 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 4);
612 
613 	if (de.de_nrel == 0) {
614 		if (dt_write(dtp, fd, &elf_file,
615 		    sizeof (elf_file)) != sizeof (elf_file) ||
616 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB32) ||
617 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
618 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
619 		    PWRITE_SCN(ESHDR_DOF, dof)) {
620 			ret = dt_set_errno(dtp, errno);
621 		}
622 	} else {
623 		shp = &elf_file.shdr[ESHDR_REL];
624 		shp->sh_name = 37; /* DTRACE_SHSTRTAB32[37] = ".rel.SUNW_dof" */
625 		shp->sh_flags = SHF_ALLOC;
626 #ifdef __sparc
627 		shp->sh_type = SHT_RELA;
628 #else
629 		shp->sh_type = SHT_REL;
630 #endif
631 		shp->sh_entsize = sizeof (de.de_rel[0]);
632 		shp->sh_link = ESHDR_SYMTAB;
633 		shp->sh_info = ESHDR_DOF;
634 		shp->sh_offset = off;
635 		shp->sh_size = de.de_nrel * sizeof (de.de_rel[0]);
636 		shp->sh_addralign = 4;
637 
638 		if (dt_write(dtp, fd, &elf_file,
639 		    sizeof (elf_file)) != sizeof (elf_file) ||
640 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB32) ||
641 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
642 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
643 		    PWRITE_SCN(ESHDR_REL, de.de_rel) ||
644 		    PWRITE_SCN(ESHDR_DOF, dof)) {
645 			ret = dt_set_errno(dtp, errno);
646 		}
647 	}
648 
649 	free(de.de_strtab);
650 	free(de.de_sym);
651 	free(de.de_rel);
652 
653 	return (ret);
654 }
655 
656 /*
657  * Write out an ELF64 file prologue consisting of a header, section headers,
658  * and a section header string table.  The DOF data will follow this prologue
659  * and complete the contents of the given ELF file.
660  */
661 static int
dump_elf64(dtrace_hdl_t * dtp,const dof_hdr_t * dof,int fd)662 dump_elf64(dtrace_hdl_t *dtp, const dof_hdr_t *dof, int fd)
663 {
664 	struct {
665 		Elf64_Ehdr ehdr;
666 		Elf64_Shdr shdr[ESHDR_NUM];
667 	} elf_file;
668 
669 	Elf64_Shdr *shp;
670 	Elf64_Off off;
671 	dof_elf64_t de;
672 	int ret = 0;
673 	uint_t nshdr;
674 
675 	memset(&de, 0, sizeof(de));	// XXX: gcc
676 	if (prepare_elf64(dtp, dof, &de) != 0)
677 		return (-1); /* errno is set for us */
678 
679 	/*
680 	 * If there are no relocations, we only need enough sections for
681 	 * the shstrtab and the DOF.
682 	 */
683 	nshdr = de.de_nrel == 0 ? ESHDR_SYMTAB + 1 : ESHDR_NUM;
684 
685 	bzero(&elf_file, sizeof (elf_file));
686 
687 	elf_file.ehdr.e_ident[EI_MAG0] = ELFMAG0;
688 	elf_file.ehdr.e_ident[EI_MAG1] = ELFMAG1;
689 	elf_file.ehdr.e_ident[EI_MAG2] = ELFMAG2;
690 	elf_file.ehdr.e_ident[EI_MAG3] = ELFMAG3;
691 	elf_file.ehdr.e_ident[EI_VERSION] = EV_CURRENT;
692 	elf_file.ehdr.e_ident[EI_CLASS] = ELFCLASS64;
693 #if BYTE_ORDER == _BIG_ENDIAN
694 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2MSB;
695 #else
696 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2LSB;
697 #endif
698 #if defined(__FreeBSD__) || defined(__NetBSD__)
699 	elf_file.ehdr.e_ident[EI_OSABI] = ELFOSABI_FREEBSD;
700 #endif
701 	elf_file.ehdr.e_type = ET_REL;
702 #if defined(__arm__)
703 	elf_file.ehdr.e_machine = EM_ARM;
704 #elif defined(__mips__)
705 	elf_file.ehdr.e_machine = EM_MIPS;
706 #elif defined(__powerpc64__)
707 	elf_file.ehdr.e_machine = EM_PPC64;
708 #elif defined(__sparc)
709 	elf_file.ehdr.e_machine = EM_SPARCV9;
710 #elif defined(__i386) || defined(__amd64)
711 	elf_file.ehdr.e_machine = EM_AMD64;
712 #endif
713 	elf_file.ehdr.e_version = EV_CURRENT;
714 	elf_file.ehdr.e_shoff = sizeof (Elf64_Ehdr);
715 	elf_file.ehdr.e_ehsize = sizeof (Elf64_Ehdr);
716 	elf_file.ehdr.e_phentsize = sizeof (Elf64_Phdr);
717 	elf_file.ehdr.e_shentsize = sizeof (Elf64_Shdr);
718 	elf_file.ehdr.e_shnum = nshdr;
719 	elf_file.ehdr.e_shstrndx = ESHDR_SHSTRTAB;
720 	off = sizeof (elf_file) + nshdr * sizeof (Elf64_Shdr);
721 
722 	shp = &elf_file.shdr[ESHDR_SHSTRTAB];
723 	shp->sh_name = 1; /* DTRACE_SHSTRTAB64[1] = ".shstrtab" */
724 	shp->sh_type = SHT_STRTAB;
725 	shp->sh_offset = off;
726 	shp->sh_size = sizeof (DTRACE_SHSTRTAB64);
727 	shp->sh_addralign = sizeof (char);
728 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
729 
730 	shp = &elf_file.shdr[ESHDR_DOF];
731 	shp->sh_name = 11; /* DTRACE_SHSTRTAB64[11] = ".SUNW_dof" */
732 	shp->sh_flags = SHF_ALLOC;
733 	shp->sh_type = SHT_SUNW_dof;
734 	shp->sh_offset = off;
735 	shp->sh_size = dof->dofh_filesz;
736 	shp->sh_addralign = 8;
737 	off = shp->sh_offset + shp->sh_size;
738 
739 	shp = &elf_file.shdr[ESHDR_STRTAB];
740 	shp->sh_name = 21; /* DTRACE_SHSTRTAB64[21] = ".strtab" */
741 	shp->sh_flags = SHF_ALLOC;
742 	shp->sh_type = SHT_STRTAB;
743 	shp->sh_offset = off;
744 	shp->sh_size = de.de_strlen;
745 	shp->sh_addralign = sizeof (char);
746 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
747 
748 	shp = &elf_file.shdr[ESHDR_SYMTAB];
749 	shp->sh_name = 29; /* DTRACE_SHSTRTAB64[29] = ".symtab" */
750 	shp->sh_flags = SHF_ALLOC;
751 	shp->sh_type = SHT_SYMTAB;
752 	shp->sh_entsize = sizeof (Elf64_Sym);
753 	shp->sh_link = ESHDR_STRTAB;
754 	shp->sh_offset = off;
755 	shp->sh_info = de.de_global;
756 	shp->sh_size = de.de_nsym * sizeof (Elf64_Sym);
757 	shp->sh_addralign = 8;
758 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
759 
760 	if (de.de_nrel == 0) {
761 		if (dt_write(dtp, fd, &elf_file,
762 		    sizeof (elf_file)) != sizeof (elf_file) ||
763 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB64) ||
764 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
765 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
766 		    PWRITE_SCN(ESHDR_DOF, dof)) {
767 			ret = dt_set_errno(dtp, errno);
768 		}
769 	} else {
770 		shp = &elf_file.shdr[ESHDR_REL];
771 		shp->sh_name = 37; /* DTRACE_SHSTRTAB64[37] = ".rel.SUNW_dof" */
772 		shp->sh_flags = SHF_ALLOC;
773 		shp->sh_type = SHT_RELA;
774 		shp->sh_entsize = sizeof (de.de_rel[0]);
775 		shp->sh_link = ESHDR_SYMTAB;
776 		shp->sh_info = ESHDR_DOF;
777 		shp->sh_offset = off;
778 		shp->sh_size = de.de_nrel * sizeof (de.de_rel[0]);
779 		shp->sh_addralign = 8;
780 
781 		if (dt_write(dtp, fd, &elf_file,
782 		    sizeof (elf_file)) != sizeof (elf_file) ||
783 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB64) ||
784 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
785 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
786 		    PWRITE_SCN(ESHDR_REL, de.de_rel) ||
787 		    PWRITE_SCN(ESHDR_DOF, dof)) {
788 			ret = dt_set_errno(dtp, errno);
789 		}
790 	}
791 
792 	free(de.de_strtab);
793 	free(de.de_sym);
794 	free(de.de_rel);
795 
796 	return (ret);
797 }
798 
799 static int
dt_symtab_lookup(Elf_Data * data_sym,int start,int end,uintptr_t addr,uint_t shn,GElf_Sym * sym,int uses_funcdesc,Elf * elf)800 dt_symtab_lookup(Elf_Data *data_sym, int start, int end, uintptr_t addr,
801     uint_t shn, GElf_Sym *sym, int uses_funcdesc, Elf *elf)
802 {
803 	Elf64_Addr symval;
804 	Elf_Scn *opd_scn;
805 	Elf_Data *opd_desc;
806 	int i;
807 
808 	for (i = start; i < end && gelf_getsym(data_sym, i, sym) != NULL; i++) {
809 		if (GELF_ST_TYPE(sym->st_info) == STT_FUNC) {
810 			symval = sym->st_value;
811 			if (uses_funcdesc) {
812 				opd_scn = elf_getscn(elf, sym->st_shndx);
813 				opd_desc = elf_rawdata(opd_scn, NULL);
814 				symval =
815 				    *(uint64_t*)((char *)opd_desc->d_buf + symval);
816 			}
817 			if ((uses_funcdesc || shn == sym->st_shndx) &&
818 			    symval <= addr && addr < symval + sym->st_size)
819 				return (0);
820 		}
821 	}
822 
823 	return (-1);
824 }
825 
826 #if defined(__aarch64__)
827 /* XXX */
828 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)829 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
830     uint32_t *off)
831 {
832 printf("%s:%s(%d): DOODAD\n",__FUNCTION__,__FILE__,__LINE__);
833 	return (0);
834 }
835 #elif defined(__arm__)
836 /* XXX */
837 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)838 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
839     uint32_t *off)
840 {
841 printf("%s:%s(%d): DOODAD\n",__FUNCTION__,__FILE__,__LINE__);
842 	return (0);
843 }
844 #elif defined(__mips__)
845 /* XXX */
846 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)847 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
848     uint32_t *off)
849 {
850 printf("%s:%s(%d): DOODAD\n",__FUNCTION__,__FILE__,__LINE__);
851 	return (0);
852 }
853 #elif defined(__powerpc__)
854 /* The sentinel is 'xor r3,r3,r3'. */
855 #define DT_OP_XOR_R3	0x7c631a78
856 
857 #define DT_OP_NOP		0x60000000
858 #define DT_OP_BLR		0x4e800020
859 
860 /* This captures all forms of branching to address. */
861 #define DT_IS_BRANCH(inst)	((inst & 0xfc000000) == 0x48000000)
862 #define DT_IS_BL(inst)	(DT_IS_BRANCH(inst) && (inst & 0x01))
863 
864 /* XXX */
865 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)866 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
867     uint32_t *off)
868 {
869 	uint32_t *ip;
870 
871 	if ((rela->r_offset & (sizeof (uint32_t) - 1)) != 0)
872 		return (-1);
873 
874 	/*LINTED*/
875 	ip = (uint32_t *)(p + rela->r_offset);
876 
877 	/*
878 	 * We only know about some specific relocation types.
879 	 */
880 	if (GELF_R_TYPE(rela->r_info) != R_PPC_REL24 &&
881 	    GELF_R_TYPE(rela->r_info) != R_PPC_PLTREL24)
882 		return (-1);
883 
884 	/*
885 	 * We may have already processed this object file in an earlier linker
886 	 * invocation. Check to see if the present instruction sequence matches
887 	 * the one we would install below.
888 	 */
889 	if (isenabled) {
890 		if (ip[0] == DT_OP_XOR_R3) {
891 			(*off) += sizeof (ip[0]);
892 			return (0);
893 		}
894 	} else {
895 		if (ip[0] == DT_OP_NOP) {
896 			(*off) += sizeof (ip[0]);
897 			return (0);
898 		}
899 	}
900 
901 	/*
902 	 * We only expect branch to address instructions.
903 	 */
904 	if (!DT_IS_BRANCH(ip[0])) {
905 		dt_dprintf("found %x instead of a branch instruction at %llx\n",
906 		    ip[0], (u_longlong_t)rela->r_offset);
907 		return (-1);
908 	}
909 
910 	if (isenabled) {
911 		/*
912 		 * It would necessarily indicate incorrect usage if an is-
913 		 * enabled probe were tail-called so flag that as an error.
914 		 * It's also potentially (very) tricky to handle gracefully,
915 		 * but could be done if this were a desired use scenario.
916 		 */
917 		if (!DT_IS_BL(ip[0])) {
918 			dt_dprintf("tail call to is-enabled probe at %llx\n",
919 			    (u_longlong_t)rela->r_offset);
920 			return (-1);
921 		}
922 
923 		ip[0] = DT_OP_XOR_R3;
924 		(*off) += sizeof (ip[0]);
925 	} else {
926 		if (DT_IS_BL(ip[0]))
927 			ip[0] = DT_OP_NOP;
928 		else
929 			ip[0] = DT_OP_BLR;
930 	}
931 
932 	return (0);
933 }
934 #elif defined(__riscv__)
935 /* XXX */
936 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)937 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
938     uint32_t *off)
939 {
940 printf("%s:%s(%d): DOODAD\n",__FUNCTION__,__FILE__,__LINE__);
941 	return (0);
942 }
943 #elif defined(__sparc)
944 
945 #define	DT_OP_RET		0x81c7e008
946 #define	DT_OP_NOP		0x01000000
947 #define	DT_OP_CALL		0x40000000
948 #define	DT_OP_CLR_O0		0x90102000
949 
950 #define	DT_IS_MOV_O7(inst)	(((inst) & 0xffffe000) == 0x9e100000)
951 #define	DT_IS_RESTORE(inst)	(((inst) & 0xc1f80000) == 0x81e80000)
952 #define	DT_IS_RETL(inst)	(((inst) & 0xfff83fff) == 0x81c02008)
953 
954 #define	DT_RS2(inst)		((inst) & 0x1f)
955 #define	DT_MAKE_RETL(reg)	(0x81c02008 | ((reg) << 14))
956 
957 /*ARGSUSED*/
958 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)959 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
960     uint32_t *off)
961 {
962 	uint32_t *ip;
963 
964 	if ((rela->r_offset & (sizeof (uint32_t) - 1)) != 0)
965 		return (-1);
966 
967 	/*LINTED*/
968 	ip = (uint32_t *)(p + rela->r_offset);
969 
970 	/*
971 	 * We only know about some specific relocation types.
972 	 */
973 	if (GELF_R_TYPE(rela->r_info) != R_SPARC_WDISP30 &&
974 	    GELF_R_TYPE(rela->r_info) != R_SPARC_WPLT30)
975 		return (-1);
976 
977 	/*
978 	 * We may have already processed this object file in an earlier linker
979 	 * invocation. Check to see if the present instruction sequence matches
980 	 * the one we would install below.
981 	 */
982 	if (isenabled) {
983 		if (ip[0] == DT_OP_NOP) {
984 			(*off) += sizeof (ip[0]);
985 			return (0);
986 		}
987 	} else {
988 		if (DT_IS_RESTORE(ip[1])) {
989 			if (ip[0] == DT_OP_RET) {
990 				(*off) += sizeof (ip[0]);
991 				return (0);
992 			}
993 		} else if (DT_IS_MOV_O7(ip[1])) {
994 			if (DT_IS_RETL(ip[0]))
995 				return (0);
996 		} else {
997 			if (ip[0] == DT_OP_NOP) {
998 				(*off) += sizeof (ip[0]);
999 				return (0);
1000 			}
1001 		}
1002 	}
1003 
1004 	/*
1005 	 * We only expect call instructions with a displacement of 0.
1006 	 */
1007 	if (ip[0] != DT_OP_CALL) {
1008 		dt_dprintf("found %x instead of a call instruction at %llx\n",
1009 		    ip[0], (u_longlong_t)rela->r_offset);
1010 		return (-1);
1011 	}
1012 
1013 	if (isenabled) {
1014 		/*
1015 		 * It would necessarily indicate incorrect usage if an is-
1016 		 * enabled probe were tail-called so flag that as an error.
1017 		 * It's also potentially (very) tricky to handle gracefully,
1018 		 * but could be done if this were a desired use scenario.
1019 		 */
1020 		if (DT_IS_RESTORE(ip[1]) || DT_IS_MOV_O7(ip[1])) {
1021 			dt_dprintf("tail call to is-enabled probe at %llx\n",
1022 			    (u_longlong_t)rela->r_offset);
1023 			return (-1);
1024 		}
1025 
1026 
1027 		/*
1028 		 * On SPARC, we take advantage of the fact that the first
1029 		 * argument shares the same register as for the return value.
1030 		 * The macro handles the work of zeroing that register so we
1031 		 * don't need to do anything special here. We instrument the
1032 		 * instruction in the delay slot as we'll need to modify the
1033 		 * return register after that instruction has been emulated.
1034 		 */
1035 		ip[0] = DT_OP_NOP;
1036 		(*off) += sizeof (ip[0]);
1037 	} else {
1038 		/*
1039 		 * If the call is followed by a restore, it's a tail call so
1040 		 * change the call to a ret. If the call if followed by a mov
1041 		 * of a register into %o7, it's a tail call in leaf context
1042 		 * so change the call to a retl-like instruction that returns
1043 		 * to that register value + 8 (rather than the typical %o7 +
1044 		 * 8); the delay slot instruction is left, but should have no
1045 		 * effect. Otherwise we change the call to be a nop. We
1046 		 * identify the subsequent instruction as the probe point in
1047 		 * all but the leaf tail-call case to ensure that arguments to
1048 		 * the probe are complete and consistent. An astute, though
1049 		 * largely hypothetical, observer would note that there is the
1050 		 * possibility of a false-positive probe firing if the function
1051 		 * contained a branch to the instruction in the delay slot of
1052 		 * the call. Fixing this would require significant in-kernel
1053 		 * modifications, and isn't worth doing until we see it in the
1054 		 * wild.
1055 		 */
1056 		if (DT_IS_RESTORE(ip[1])) {
1057 			ip[0] = DT_OP_RET;
1058 			(*off) += sizeof (ip[0]);
1059 		} else if (DT_IS_MOV_O7(ip[1])) {
1060 			ip[0] = DT_MAKE_RETL(DT_RS2(ip[1]));
1061 		} else {
1062 			ip[0] = DT_OP_NOP;
1063 			(*off) += sizeof (ip[0]);
1064 		}
1065 	}
1066 
1067 	return (0);
1068 }
1069 
1070 #elif defined(__i386) || defined(__amd64)
1071 
1072 #define	DT_OP_NOP		0x90
1073 #define	DT_OP_RET		0xc3
1074 #define	DT_OP_CALL		0xe8
1075 #define	DT_OP_JMP32		0xe9
1076 #define	DT_OP_REX_RAX		0x48
1077 #define	DT_OP_XOR_EAX_0		0x33
1078 #define	DT_OP_XOR_EAX_1		0xc0
1079 
1080 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)1081 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
1082     uint32_t *off)
1083 {
1084 	uint8_t *ip = (uint8_t *)(p + rela->r_offset - 1);
1085 	uint8_t ret;
1086 
1087 	/*
1088 	 * On x86, the first byte of the instruction is the call opcode and
1089 	 * the next four bytes are the 32-bit address; the relocation is for
1090 	 * the address operand. We back up the offset to the first byte of
1091 	 * the instruction. For is-enabled probes, we later advance the offset
1092 	 * so that it hits the first nop in the instruction sequence.
1093 	 */
1094 	(*off) -= 1;
1095 
1096 	/*
1097 	 * We only know about some specific relocation types. Luckily
1098 	 * these types have the same values on both 32-bit and 64-bit
1099 	 * x86 architectures.
1100 	 */
1101 	if (GELF_R_TYPE(rela->r_info) != R_386_PC32 &&
1102 	    GELF_R_TYPE(rela->r_info) != R_386_PLT32)
1103 		return (-1);
1104 
1105 	/*
1106 	 * We may have already processed this object file in an earlier linker
1107 	 * invocation. Check to see if the present instruction sequence matches
1108 	 * the one we would install. For is-enabled probes, we advance the
1109 	 * offset to the first nop instruction in the sequence to match the
1110 	 * text modification code below.
1111 	 */
1112 	if (!isenabled) {
1113 		if ((ip[0] == DT_OP_NOP || ip[0] == DT_OP_RET) &&
1114 		    ip[1] == DT_OP_NOP && ip[2] == DT_OP_NOP &&
1115 		    ip[3] == DT_OP_NOP && ip[4] == DT_OP_NOP)
1116 			return (0);
1117 	} else if (dtp->dt_oflags & DTRACE_O_LP64) {
1118 		if (ip[0] == DT_OP_REX_RAX &&
1119 		    ip[1] == DT_OP_XOR_EAX_0 && ip[2] == DT_OP_XOR_EAX_1 &&
1120 		    (ip[3] == DT_OP_NOP || ip[3] == DT_OP_RET) &&
1121 		    ip[4] == DT_OP_NOP) {
1122 			(*off) += 3;
1123 			return (0);
1124 		}
1125 	} else {
1126 		if (ip[0] == DT_OP_XOR_EAX_0 && ip[1] == DT_OP_XOR_EAX_1 &&
1127 		    (ip[2] == DT_OP_NOP || ip[2] == DT_OP_RET) &&
1128 		    ip[3] == DT_OP_NOP && ip[4] == DT_OP_NOP) {
1129 			(*off) += 2;
1130 			return (0);
1131 		}
1132 	}
1133 
1134 	/*
1135 	 * We expect either a call instrution with a 32-bit displacement or a
1136 	 * jmp instruction with a 32-bit displacement acting as a tail-call.
1137 	 */
1138 	if (ip[0] != DT_OP_CALL && ip[0] != DT_OP_JMP32) {
1139 		dt_dprintf("found %x instead of a call or jmp instruction at "
1140 		    "%llx\n", ip[0], (u_longlong_t)rela->r_offset);
1141 		return (-1);
1142 	}
1143 
1144 	ret = (ip[0] == DT_OP_JMP32) ? DT_OP_RET : DT_OP_NOP;
1145 
1146 	/*
1147 	 * Establish the instruction sequence -- all nops for probes, and an
1148 	 * instruction to clear the return value register (%eax/%rax) followed
1149 	 * by nops for is-enabled probes. For is-enabled probes, we advance
1150 	 * the offset to the first nop. This isn't stricly necessary but makes
1151 	 * for more readable disassembly when the probe is enabled.
1152 	 */
1153 	if (!isenabled) {
1154 		ip[0] = ret;
1155 		ip[1] = DT_OP_NOP;
1156 		ip[2] = DT_OP_NOP;
1157 		ip[3] = DT_OP_NOP;
1158 		ip[4] = DT_OP_NOP;
1159 	} else if (dtp->dt_oflags & DTRACE_O_LP64) {
1160 		ip[0] = DT_OP_REX_RAX;
1161 		ip[1] = DT_OP_XOR_EAX_0;
1162 		ip[2] = DT_OP_XOR_EAX_1;
1163 		ip[3] = ret;
1164 		ip[4] = DT_OP_NOP;
1165 		(*off) += 3;
1166 	} else {
1167 		ip[0] = DT_OP_XOR_EAX_0;
1168 		ip[1] = DT_OP_XOR_EAX_1;
1169 		ip[2] = ret;
1170 		ip[3] = DT_OP_NOP;
1171 		ip[4] = DT_OP_NOP;
1172 		(*off) += 2;
1173 	}
1174 
1175 	return (0);
1176 }
1177 
1178 #else
1179 #error unknown ISA
1180 #endif
1181 
1182 /*PRINTFLIKE5*/
1183 static int
dt_link_error(dtrace_hdl_t * dtp,Elf * elf,int fd,dt_link_pair_t * bufs,const char * format,...)1184 dt_link_error(dtrace_hdl_t *dtp, Elf *elf, int fd, dt_link_pair_t *bufs,
1185     const char *format, ...)
1186 {
1187 	va_list ap;
1188 	dt_link_pair_t *pair;
1189 
1190 	va_start(ap, format);
1191 	dt_set_errmsg(dtp, NULL, NULL, NULL, 0, format, ap);
1192 	va_end(ap);
1193 
1194 	if (elf != NULL)
1195 		(void) elf_end(elf);
1196 
1197 	if (fd >= 0)
1198 		(void) close(fd);
1199 
1200 	while ((pair = bufs) != NULL) {
1201 		bufs = pair->dlp_next;
1202 		dt_free(dtp, pair->dlp_str);
1203 		dt_free(dtp, pair->dlp_sym);
1204 		dt_free(dtp, pair);
1205 	}
1206 
1207 	return (dt_set_errno(dtp, EDT_COMPILER));
1208 }
1209 
1210 static int
process_obj(dtrace_hdl_t * dtp,const char * obj,int * eprobesp)1211 process_obj(dtrace_hdl_t *dtp, const char *obj, int *eprobesp)
1212 {
1213 	static const char dt_prefix[] = "__dtrace";
1214 	static const char dt_enabled[] = "enabled";
1215 	static const char dt_symprefix[] = "$dtrace";
1216 	static const char dt_symfmt[] = "%s%ld.%s";
1217 	static const char dt_weaksymfmt[] = "%s.%s";
1218 	char probename[DTRACE_NAMELEN];
1219 	int fd, i, ndx, eprobe, mod = 0;
1220 	Elf *elf = NULL;
1221 	GElf_Ehdr ehdr;
1222 	Elf_Scn *scn_rel, *scn_sym, *scn_str, *scn_tgt;
1223 	Elf_Data *data_rel, *data_sym, *data_str, *data_tgt;
1224 	GElf_Shdr shdr_rel, shdr_sym, shdr_str, shdr_tgt;
1225 	GElf_Sym rsym, fsym, dsym;
1226 	GElf_Rela rela;
1227 	char *s, *p, *r;
1228 	char pname[DTRACE_PROVNAMELEN];
1229 	dt_provider_t *pvp;
1230 	dt_probe_t *prp;
1231 	uint32_t off, eclass, emachine1, emachine2;
1232 	size_t symsize, osym, nsym, isym, istr, len;
1233 	key_t objkey;
1234 	dt_link_pair_t *pair, *bufs = NULL;
1235 	dt_strtab_t *strtab;
1236 	void *tmp;
1237 
1238 	if ((fd = open64(obj, O_RDWR)) == -1) {
1239 		return (dt_link_error(dtp, elf, fd, bufs,
1240 		    "failed to open %s: %s", obj, strerror(errno)));
1241 	}
1242 
1243 	if ((elf = elf_begin(fd, ELF_C_RDWR, NULL)) == NULL) {
1244 		return (dt_link_error(dtp, elf, fd, bufs,
1245 		    "failed to process %s: %s", obj, elf_errmsg(elf_errno())));
1246 	}
1247 
1248 	switch (elf_kind(elf)) {
1249 	case ELF_K_ELF:
1250 		break;
1251 	case ELF_K_AR:
1252 		return (dt_link_error(dtp, elf, fd, bufs, "archives are not "
1253 		    "permitted; use the contents of the archive instead: %s",
1254 		    obj));
1255 	default:
1256 		return (dt_link_error(dtp, elf, fd, bufs,
1257 		    "invalid file type: %s", obj));
1258 	}
1259 
1260 	if (gelf_getehdr(elf, &ehdr) == NULL) {
1261 		return (dt_link_error(dtp, elf, fd, bufs, "corrupt file: %s",
1262 		    obj));
1263 	}
1264 
1265 	if (dtp->dt_oflags & DTRACE_O_LP64) {
1266 		eclass = ELFCLASS64;
1267 #if defined(__aarch64__)
1268 		emachine1 = emachine2 = EM_AARCH64;
1269 #elif defined(__mips__)
1270 		emachine1 = emachine2 = EM_MIPS;
1271 #elif defined(__powerpc__)
1272 		emachine1 = emachine2 = EM_PPC64;
1273 #elif defined(__sparc)
1274 		emachine1 = emachine2 = EM_SPARCV9;
1275 #elif defined(__i386) || defined(__amd64)
1276 		emachine1 = emachine2 = EM_AMD64;
1277 #endif
1278 		symsize = sizeof (Elf64_Sym);
1279 	} else {
1280 		eclass = ELFCLASS32;
1281 #if defined(__arm__) || defined(__aarch64__)
1282 		emachine1 = emachine2 = EM_ARM;
1283 #elif defined(__mips__)
1284 		emachine1 = emachine2 = EM_MIPS;
1285 #elif defined(__powerpc__)
1286 		emachine1 = emachine2 = EM_PPC;
1287 #elif defined(__sparc)
1288 		emachine1 = EM_SPARC;
1289 		emachine2 = EM_SPARC32PLUS;
1290 #elif defined(__i386) || defined(__amd64)
1291 		emachine1 = emachine2 = EM_386;
1292 #endif
1293 		symsize = sizeof (Elf32_Sym);
1294 	}
1295 
1296 	if (ehdr.e_ident[EI_CLASS] != eclass) {
1297 		return (dt_link_error(dtp, elf, fd, bufs,
1298 		    "incorrect ELF class for object file: %s", obj));
1299 	}
1300 
1301 	if (ehdr.e_machine != emachine1 && ehdr.e_machine != emachine2) {
1302 		return (dt_link_error(dtp, elf, fd, bufs,
1303 		    "incorrect ELF machine type for object file: %s", obj));
1304 	}
1305 
1306 	/*
1307 	 * We use this token as a relatively unique handle for this file on the
1308 	 * system in order to disambiguate potential conflicts between files of
1309 	 * the same name which contain identially named local symbols.
1310 	 */
1311 	if ((objkey = ftok(obj, 0)) == (key_t)-1) {
1312 		return (dt_link_error(dtp, elf, fd, bufs,
1313 		    "failed to generate unique key for object file: %s", obj));
1314 	}
1315 
1316 	scn_rel = NULL;
1317 	while ((scn_rel = elf_nextscn(elf, scn_rel)) != NULL) {
1318 		if (gelf_getshdr(scn_rel, &shdr_rel) == NULL)
1319 			goto err;
1320 
1321 		/*
1322 		 * Skip any non-relocation sections.
1323 		 */
1324 		if (shdr_rel.sh_type != SHT_RELA && shdr_rel.sh_type != SHT_REL)
1325 			continue;
1326 
1327 		if ((data_rel = elf_getdata(scn_rel, NULL)) == NULL)
1328 			goto err;
1329 
1330 		/*
1331 		 * Grab the section, section header and section data for the
1332 		 * symbol table that this relocation section references.
1333 		 */
1334 		if ((scn_sym = elf_getscn(elf, shdr_rel.sh_link)) == NULL ||
1335 		    gelf_getshdr(scn_sym, &shdr_sym) == NULL ||
1336 		    (data_sym = elf_getdata(scn_sym, NULL)) == NULL)
1337 			goto err;
1338 
1339 		/*
1340 		 * Ditto for that symbol table's string table.
1341 		 */
1342 		if ((scn_str = elf_getscn(elf, shdr_sym.sh_link)) == NULL ||
1343 		    gelf_getshdr(scn_str, &shdr_str) == NULL ||
1344 		    (data_str = elf_getdata(scn_str, NULL)) == NULL)
1345 			goto err;
1346 
1347 		/*
1348 		 * Grab the section, section header and section data for the
1349 		 * target section for the relocations. For the relocations
1350 		 * we're looking for -- this will typically be the text of the
1351 		 * object file.
1352 		 */
1353 		if ((scn_tgt = elf_getscn(elf, shdr_rel.sh_info)) == NULL ||
1354 		    gelf_getshdr(scn_tgt, &shdr_tgt) == NULL ||
1355 		    (data_tgt = elf_getdata(scn_tgt, NULL)) == NULL)
1356 			goto err;
1357 
1358 		/*
1359 		 * We're looking for relocations to symbols matching this form:
1360 		 *
1361 		 *   __dtrace[enabled]_<prov>___<probe>
1362 		 *
1363 		 * For the generated object, we need to record the location
1364 		 * identified by the relocation, and create a new relocation
1365 		 * in the generated object that will be resolved at link time
1366 		 * to the location of the function in which the probe is
1367 		 * embedded. In the target object, we change the matched symbol
1368 		 * so that it will be ignored at link time, and we modify the
1369 		 * target (text) section to replace the call instruction with
1370 		 * one or more nops.
1371 		 *
1372 		 * To avoid runtime overhead, the relocations added to the
1373 		 * generated object should be resolved at static link time. We
1374 		 * therefore create aliases for the functions that contain
1375 		 * probes. An alias is global (so that the relocation from the
1376 		 * generated object can be resolved), and hidden (so that its
1377 		 * address is known at static link time). Such aliases have this
1378 		 * form:
1379 		 *
1380 		 *   $dtrace<key>.<function>
1381 		 *
1382 		 * We take a first pass through all the relocations to
1383 		 * populate our string table and count the number of extra
1384 		 * symbols we'll require.
1385 		 */
1386 		strtab = dt_strtab_create(1);
1387 		nsym = 0;
1388 		isym = data_sym->d_size / symsize;
1389 		istr = data_str->d_size;
1390 
1391 		for (i = 0; i < shdr_rel.sh_size / shdr_rel.sh_entsize; i++) {
1392 
1393 			if (shdr_rel.sh_type == SHT_RELA) {
1394 				if (gelf_getrela(data_rel, i, &rela) == NULL)
1395 					continue;
1396 			} else {
1397 				GElf_Rel rel;
1398 				if (gelf_getrel(data_rel, i, &rel) == NULL)
1399 					continue;
1400 				rela.r_offset = rel.r_offset;
1401 				rela.r_info = rel.r_info;
1402 				rela.r_addend = 0;
1403 			}
1404 
1405 			if (gelf_getsym(data_sym, GELF_R_SYM(rela.r_info),
1406 			    &rsym) == NULL) {
1407 				dt_strtab_destroy(strtab);
1408 				goto err;
1409 			}
1410 
1411 			s = (char *)data_str->d_buf + rsym.st_name;
1412 
1413 			if (strncmp(s, dt_prefix, sizeof (dt_prefix) - 1) != 0)
1414 				continue;
1415 
1416 			if (dt_symtab_lookup(data_sym, 0, isym, rela.r_offset,
1417 			    shdr_rel.sh_info, &fsym, (emachine1 == EM_PPC64),
1418 			    elf) != 0) {
1419 				dt_strtab_destroy(strtab);
1420 				goto err;
1421 			}
1422 
1423 			if (fsym.st_name > data_str->d_size) {
1424 				dt_strtab_destroy(strtab);
1425 				goto err;
1426 			}
1427 
1428 			s = (char *)data_str->d_buf + fsym.st_name;
1429 
1430 			/*
1431 			 * If this symbol isn't of type function, we've really
1432 			 * driven off the rails or the object file is corrupt.
1433 			 */
1434 			if (GELF_ST_TYPE(fsym.st_info) != STT_FUNC) {
1435 				dt_strtab_destroy(strtab);
1436 				return (dt_link_error(dtp, elf, fd, bufs,
1437 				    "expected %s to be of type function", s));
1438 			}
1439 
1440 			len = snprintf(NULL, 0, dt_symfmt, dt_symprefix,
1441 			    objkey, s) + 1;
1442 			if ((p = dt_alloc(dtp, len)) == NULL) {
1443 				dt_strtab_destroy(strtab);
1444 				goto err;
1445 			}
1446 			(void) snprintf(p, len, dt_symfmt, dt_symprefix,
1447 			    objkey, s);
1448 
1449 			if (dt_strtab_index(strtab, p) == -1) {
1450 				nsym++;
1451 				(void) dt_strtab_insert(strtab, p);
1452 			}
1453 
1454 			dt_free(dtp, p);
1455 		}
1456 
1457 		/*
1458 		 * If any probes were found, allocate the additional space for
1459 		 * the symbol table and string table, copying the old data into
1460 		 * the new buffers, and marking the buffers as dirty. We inject
1461 		 * those newly allocated buffers into the libelf data
1462 		 * structures, but are still responsible for freeing them once
1463 		 * we're done with the elf handle.
1464 		 */
1465 		if (nsym > 0) {
1466 			/*
1467 			 * The first byte of the string table is reserved for
1468 			 * the \0 entry.
1469 			 */
1470 			len = dt_strtab_size(strtab) - 1;
1471 
1472 			assert(len > 0);
1473 			assert(dt_strtab_index(strtab, "") == 0);
1474 
1475 			dt_strtab_destroy(strtab);
1476 
1477 			if ((pair = dt_alloc(dtp, sizeof (*pair))) == NULL)
1478 				goto err;
1479 
1480 			if ((pair->dlp_str = dt_alloc(dtp, data_str->d_size +
1481 			    len)) == NULL) {
1482 				dt_free(dtp, pair);
1483 				goto err;
1484 			}
1485 
1486 			if ((pair->dlp_sym = dt_alloc(dtp, data_sym->d_size +
1487 			    nsym * symsize)) == NULL) {
1488 				dt_free(dtp, pair->dlp_str);
1489 				dt_free(dtp, pair);
1490 				goto err;
1491 			}
1492 
1493 			pair->dlp_next = bufs;
1494 			bufs = pair;
1495 
1496 			bcopy(data_str->d_buf, pair->dlp_str, data_str->d_size);
1497 			tmp = data_str->d_buf;
1498 			data_str->d_buf = pair->dlp_str;
1499 			pair->dlp_str = tmp;
1500 			data_str->d_size += len;
1501 			(void) elf_flagdata(data_str, ELF_C_SET, ELF_F_DIRTY);
1502 
1503 			shdr_str.sh_size += len;
1504 			(void) gelf_update_shdr(scn_str, &shdr_str);
1505 
1506 			bcopy(data_sym->d_buf, pair->dlp_sym, data_sym->d_size);
1507 			tmp = data_sym->d_buf;
1508 			data_sym->d_buf = pair->dlp_sym;
1509 			pair->dlp_sym = tmp;
1510 			data_sym->d_size += nsym * symsize;
1511 			(void) elf_flagdata(data_sym, ELF_C_SET, ELF_F_DIRTY);
1512 
1513 			shdr_sym.sh_size += nsym * symsize;
1514 			(void) gelf_update_shdr(scn_sym, &shdr_sym);
1515 
1516 			osym = isym;
1517 			nsym += isym;
1518 		} else {
1519 			dt_strtab_destroy(strtab);
1520 			continue;
1521 		}
1522 
1523 		/*
1524 		 * Now that the tables have been allocated, perform the
1525 		 * modifications described above.
1526 		 */
1527 		for (i = 0; i < shdr_rel.sh_size / shdr_rel.sh_entsize; i++) {
1528 
1529 			if (shdr_rel.sh_type == SHT_RELA) {
1530 				if (gelf_getrela(data_rel, i, &rela) == NULL)
1531 					continue;
1532 			} else {
1533 				GElf_Rel rel;
1534 				if (gelf_getrel(data_rel, i, &rel) == NULL)
1535 					continue;
1536 				rela.r_offset = rel.r_offset;
1537 				rela.r_info = rel.r_info;
1538 				rela.r_addend = 0;
1539 			}
1540 
1541 			ndx = GELF_R_SYM(rela.r_info);
1542 
1543 			if (gelf_getsym(data_sym, ndx, &rsym) == NULL ||
1544 			    rsym.st_name > data_str->d_size)
1545 				goto err;
1546 
1547 			s = (char *)data_str->d_buf + rsym.st_name;
1548 
1549 			if (strncmp(s, dt_prefix, sizeof (dt_prefix) - 1) != 0)
1550 				continue;
1551 
1552 			s += sizeof (dt_prefix) - 1;
1553 
1554 			/*
1555 			 * Check to see if this is an 'is-enabled' check as
1556 			 * opposed to a normal probe.
1557 			 */
1558 			if (strncmp(s, dt_enabled,
1559 			    sizeof (dt_enabled) - 1) == 0) {
1560 				s += sizeof (dt_enabled) - 1;
1561 				eprobe = 1;
1562 				*eprobesp = 1;
1563 				dt_dprintf("is-enabled probe\n");
1564 			} else {
1565 				eprobe = 0;
1566 				dt_dprintf("normal probe\n");
1567 			}
1568 
1569 			if (*s++ != '_')
1570 				goto err;
1571 
1572 			if ((p = strstr(s, "___")) == NULL ||
1573 			    p - s >= sizeof (pname))
1574 				goto err;
1575 
1576 			bcopy(s, pname, p - s);
1577 			pname[p - s] = '\0';
1578 
1579 			if (dt_symtab_lookup(data_sym, osym, isym,
1580 			    rela.r_offset, shdr_rel.sh_info, &fsym,
1581 			    (emachine1 == EM_PPC64), elf) == 0) {
1582 				if (fsym.st_name > data_str->d_size)
1583 					goto err;
1584 
1585 				r = s = (char *) data_str->d_buf + fsym.st_name;
1586 				assert(strstr(s, dt_symprefix) == s);
1587 				s = strchr(s, '.') + 1;
1588 			} else if (dt_symtab_lookup(data_sym, 0, osym,
1589 			    rela.r_offset, shdr_rel.sh_info, &fsym,
1590 			    (emachine1 == EM_PPC64), elf) == 0) {
1591 				u_int bind;
1592 
1593 				bind = GELF_ST_BIND(fsym.st_info) == STB_WEAK ?
1594 				    STB_WEAK : STB_GLOBAL;
1595 
1596 				/*
1597 				 * Emit an alias for the symbol. It needs to be
1598 				 * non-preemptible so that .SUNW_dof relocations
1599 				 * may be resolved at static link time. Aliases
1600 				 * of weak symbols are given a non-unique name
1601 				 * so that they may be merged by the linker.
1602 				 */
1603 				dsym = fsym;
1604 				dsym.st_name = istr;
1605 				dsym.st_info = GELF_ST_INFO(bind, STT_FUNC);
1606 				dsym.st_other = GELF_ST_VISIBILITY(STV_HIDDEN);
1607 				(void) gelf_update_sym(data_sym, isym, &dsym);
1608 				r = (char *) data_str->d_buf + istr;
1609 				s = (char *) data_str->d_buf + fsym.st_name;
1610 				if (bind == STB_WEAK)
1611 					istr += sprintf(r, dt_weaksymfmt,
1612 					    dt_symprefix, s);
1613 				else
1614 					istr += sprintf(r, dt_symfmt,
1615 					    dt_symprefix, objkey, s);
1616 				istr++;
1617 				isym++;
1618 				assert(isym <= nsym);
1619 			} else
1620 				goto err;
1621 
1622 			if ((pvp = dt_provider_lookup(dtp, pname)) == NULL) {
1623 				return (dt_link_error(dtp, elf, fd, bufs,
1624 				    "no such provider %s", pname));
1625 			}
1626 
1627 			if (strlcpy(probename, p + 3, sizeof (probename)) >=
1628 			    sizeof (probename))
1629 				return (dt_link_error(dtp, elf, fd, bufs,
1630 				    "invalid probe name %s", probename));
1631 			(void) strhyphenate(probename);
1632 			if ((prp = dt_probe_lookup(pvp, probename)) == NULL)
1633 				return (dt_link_error(dtp, elf, fd, bufs,
1634 				    "no such probe %s", probename));
1635 
1636 			assert(fsym.st_value <= rela.r_offset);
1637 
1638 			off = rela.r_offset - fsym.st_value;
1639 			if (dt_modtext(dtp, data_tgt->d_buf, eprobe,
1640 			    &rela, &off) != 0)
1641 				goto err;
1642 
1643 			if (dt_probe_define(pvp, prp, s, r, off, eprobe) != 0) {
1644 				return (dt_link_error(dtp, elf, fd, bufs,
1645 				    "failed to allocate space for probe"));
1646 			}
1647 #ifndef illumos
1648 			/*
1649 			 * Our linker doesn't understand the SUNW_IGNORE ndx and
1650 			 * will try to use this relocation when we build the
1651 			 * final executable. Since we are done processing this
1652 			 * relocation, mark it as inexistant and let libelf
1653 			 * remove it from the file.
1654 			 * If this wasn't done, we would have garbage added to
1655 			 * the executable file as the symbol is going to be
1656 			 * change from UND to ABS.
1657 			 */
1658 			if (shdr_rel.sh_type == SHT_RELA) {
1659 				rela.r_offset = 0;
1660 				rela.r_info  = 0;
1661 				rela.r_addend = 0;
1662 				(void) gelf_update_rela(data_rel, i, &rela);
1663 			} else {
1664 				GElf_Rel rel;
1665 				rel.r_offset = 0;
1666 				rel.r_info = 0;
1667 				(void) gelf_update_rel(data_rel, i, &rel);
1668 			}
1669 #endif
1670 
1671 			mod = 1;
1672 			(void) elf_flagdata(data_tgt, ELF_C_SET, ELF_F_DIRTY);
1673 
1674 			/*
1675 			 * This symbol may already have been marked to
1676 			 * be ignored by another relocation referencing
1677 			 * the same symbol or if this object file has
1678 			 * already been processed by an earlier link
1679 			 * invocation.
1680 			 */
1681 			if (rsym.st_shndx != SHN_ABS) {
1682 				rsym.st_shndx = SHN_ABS;
1683 				(void) gelf_update_sym(data_sym, ndx, &rsym);
1684 			}
1685 		}
1686 	}
1687 
1688 	if (mod && elf_update(elf, ELF_C_WRITE) == -1)
1689 		goto err;
1690 
1691 	(void) elf_end(elf);
1692 	(void) close(fd);
1693 
1694 	while ((pair = bufs) != NULL) {
1695 		bufs = pair->dlp_next;
1696 		dt_free(dtp, pair->dlp_str);
1697 		dt_free(dtp, pair->dlp_sym);
1698 		dt_free(dtp, pair);
1699 	}
1700 
1701 	return (0);
1702 
1703 err:
1704 	return (dt_link_error(dtp, elf, fd, bufs,
1705 	    "an error was encountered while processing %s", obj));
1706 }
1707 
1708 int
dtrace_program_link(dtrace_hdl_t * dtp,dtrace_prog_t * pgp,uint_t dflags,const char * file,int objc,char * const objv[])1709 dtrace_program_link(dtrace_hdl_t *dtp, dtrace_prog_t *pgp, uint_t dflags,
1710     const char *file, int objc, char *const objv[])
1711 {
1712 #ifndef illumos
1713 	char tfile[PATH_MAX];
1714 #endif
1715 	char drti[PATH_MAX];
1716 	dof_hdr_t *dof;
1717 	int fd, status, i, cur;
1718 	char *cmd, tmp;
1719 	size_t len;
1720 	int eprobes = 0, ret = 0;
1721 
1722 #ifndef illumos
1723 	if (access(file, R_OK) == 0) {
1724 		fprintf(stderr, "dtrace: target object (%s) already exists. "
1725 		    "Please remove the target\ndtrace: object and rebuild all "
1726 		    "the source objects if you wish to run the DTrace\n"
1727 		    "dtrace: linking process again\n", file);
1728 		/*
1729 		 * Several build infrastructures run DTrace twice (e.g.
1730 		 * postgres) and we don't want the build to fail. Return
1731 		 * 0 here since this isn't really a fatal error.
1732 		 */
1733 		return (0);
1734 	}
1735 #endif
1736 
1737 	/*
1738 	 * A NULL program indicates a special use in which we just link
1739 	 * together a bunch of object files specified in objv and then
1740 	 * unlink(2) those object files.
1741 	 */
1742 	if (pgp == NULL) {
1743 		const char *fmt = "%s -o %s -r";
1744 
1745 		len = snprintf(&tmp, 1, fmt, dtp->dt_ld_path, file) + 1;
1746 
1747 		for (i = 0; i < objc; i++)
1748 			len += strlen(objv[i]) + 1;
1749 
1750 		cmd = alloca(len);
1751 
1752 		cur = snprintf(cmd, len, fmt, dtp->dt_ld_path, file);
1753 
1754 		for (i = 0; i < objc; i++)
1755 			cur += snprintf(cmd + cur, len - cur, " %s", objv[i]);
1756 
1757 		if ((status = system(cmd)) == -1) {
1758 			return (dt_link_error(dtp, NULL, -1, NULL,
1759 			    "failed to run %s: %s", dtp->dt_ld_path,
1760 			    strerror(errno)));
1761 		}
1762 
1763 		if (WIFSIGNALED(status)) {
1764 			return (dt_link_error(dtp, NULL, -1, NULL,
1765 			    "failed to link %s: %s failed due to signal %d",
1766 			    file, dtp->dt_ld_path, WTERMSIG(status)));
1767 		}
1768 
1769 		if (WEXITSTATUS(status) != 0) {
1770 			return (dt_link_error(dtp, NULL, -1, NULL,
1771 			    "failed to link %s: %s exited with status %d\n",
1772 			    file, dtp->dt_ld_path, WEXITSTATUS(status)));
1773 		}
1774 
1775 		for (i = 0; i < objc; i++) {
1776 			if (strcmp(objv[i], file) != 0)
1777 				(void) unlink(objv[i]);
1778 		}
1779 
1780 		return (0);
1781 	}
1782 
1783 	for (i = 0; i < objc; i++) {
1784 		if (process_obj(dtp, objv[i], &eprobes) != 0)
1785 			return (-1); /* errno is set for us */
1786 	}
1787 
1788 	/*
1789 	 * If there are is-enabled probes then we need to force use of DOF
1790 	 * version 2.
1791 	 */
1792 	if (eprobes && pgp->dp_dofversion < DOF_VERSION_2)
1793 		pgp->dp_dofversion = DOF_VERSION_2;
1794 
1795 	if ((dof = dtrace_dof_create(dtp, pgp, dflags)) == NULL)
1796 		return (-1); /* errno is set for us */
1797 
1798 #ifdef illumos
1799 	/*
1800 	 * Create a temporary file and then unlink it if we're going to
1801 	 * combine it with drti.o later.  We can still refer to it in child
1802 	 * processes as /dev/fd/<fd>.
1803 	 */
1804 	if ((fd = open64(file, O_RDWR | O_CREAT | O_TRUNC, 0666)) == -1) {
1805 		return (dt_link_error(dtp, NULL, -1, NULL,
1806 		    "failed to open %s: %s", file, strerror(errno)));
1807 	}
1808 #else
1809 	snprintf(tfile, sizeof(tfile), "%s.XXXXXX", file);
1810 	if ((fd = mkostemp(tfile, O_CLOEXEC)) == -1)
1811 		return (dt_link_error(dtp, NULL, -1, NULL,
1812 		    "failed to create temporary file %s: %s",
1813 		    tfile, strerror(errno)));
1814 #endif
1815 
1816 	/*
1817 	 * If -xlinktype=DOF has been selected, just write out the DOF.
1818 	 * Otherwise proceed to the default of generating and linking ELF.
1819 	 */
1820 	switch (dtp->dt_linktype) {
1821 	case DT_LTYP_DOF:
1822 		if (dt_write(dtp, fd, dof, dof->dofh_filesz) < dof->dofh_filesz)
1823 			ret = errno;
1824 
1825 		if (close(fd) != 0 && ret == 0)
1826 			ret = errno;
1827 
1828 		if (ret != 0) {
1829 			return (dt_link_error(dtp, NULL, -1, NULL,
1830 			    "failed to write %s: %s", file, strerror(ret)));
1831 		}
1832 
1833 		return (0);
1834 
1835 	case DT_LTYP_ELF:
1836 		break; /* fall through to the rest of dtrace_program_link() */
1837 
1838 	default:
1839 		return (dt_link_error(dtp, NULL, -1, NULL,
1840 		    "invalid link type %u\n", dtp->dt_linktype));
1841 	}
1842 
1843 
1844 #ifdef illumos
1845 	if (!dtp->dt_lazyload)
1846 		(void) unlink(file);
1847 #endif
1848 
1849 	if (dtp->dt_oflags & DTRACE_O_LP64)
1850 		status = dump_elf64(dtp, dof, fd);
1851 	else
1852 		status = dump_elf32(dtp, dof, fd);
1853 
1854 #ifdef illumos
1855 	if (status != 0 || lseek(fd, 0, SEEK_SET) != 0) {
1856 		return (dt_link_error(dtp, NULL, -1, NULL,
1857 		    "failed to write %s: %s", file, strerror(errno)));
1858 	}
1859 #else
1860 	if (status != 0)
1861 		return (dt_link_error(dtp, NULL, -1, NULL,
1862 		    "failed to write %s: %s", tfile,
1863 		    strerror(dtrace_errno(dtp))));
1864 #endif
1865 
1866 	if (!dtp->dt_lazyload) {
1867 #ifdef illumos
1868 		const char *fmt = "%s -o %s -r -Blocal -Breduce /dev/fd/%d %s";
1869 
1870 		if (dtp->dt_oflags & DTRACE_O_LP64) {
1871 			(void) snprintf(drti, sizeof (drti),
1872 			    "%s/64/drti.o", _dtrace_libdir);
1873 		} else {
1874 			(void) snprintf(drti, sizeof (drti),
1875 			    "%s/drti.o", _dtrace_libdir);
1876 		}
1877 
1878 		len = snprintf(&tmp, 1, fmt, dtp->dt_ld_path, file, fd,
1879 		    drti) + 1;
1880 
1881 		cmd = alloca(len);
1882 
1883 		(void) snprintf(cmd, len, fmt, dtp->dt_ld_path, file, fd, drti);
1884 #else
1885 		const char *fmt = "%s -o %s -r %s %s";
1886 		dt_dirpath_t *dp = dt_list_next(&dtp->dt_lib_path);
1887 
1888 		(void) snprintf(drti, sizeof (drti), "%s/drti.o", dp->dir_path);
1889 
1890 		len = snprintf(&tmp, 1, fmt, dtp->dt_ld_path, file, tfile,
1891 		    drti) + 1;
1892 
1893 		cmd = alloca(len);
1894 
1895 		(void) snprintf(cmd, len, fmt, dtp->dt_ld_path, file, tfile,
1896 		    drti);
1897 #endif
1898 		if ((status = system(cmd)) == -1) {
1899 			ret = dt_link_error(dtp, NULL, fd, NULL,
1900 			    "failed to run %s: %s", dtp->dt_ld_path,
1901 			    strerror(errno));
1902 			goto done;
1903 		}
1904 
1905 		if (WIFSIGNALED(status)) {
1906 			ret = dt_link_error(dtp, NULL, fd, NULL,
1907 			    "failed to link %s: %s failed due to signal %d",
1908 			    file, dtp->dt_ld_path, WTERMSIG(status));
1909 			goto done;
1910 		}
1911 
1912 		if (WEXITSTATUS(status) != 0) {
1913 			ret = dt_link_error(dtp, NULL, fd, NULL,
1914 			    "failed to link %s: %s exited with status %d\n",
1915 			    file, dtp->dt_ld_path, WEXITSTATUS(status));
1916 			goto done;
1917 		}
1918 		(void) close(fd); /* release temporary file */
1919 
1920 #if defined(__FreeBSD__) || defined(__NetBSD__)
1921 		/*
1922 		 * Now that we've linked drti.o, reduce the global __SUNW_dof
1923 		 * symbol to a local symbol. This is needed to so that multiple
1924 		 * generated object files (for different providers, for
1925 		 * instance) can be linked together. This is accomplished using
1926 		 * the -Blocal flag with Sun's linker, but GNU ld doesn't appear
1927 		 * to have an equivalent option.
1928 		 */
1929 		asprintf(&cmd, "%s --localize-hidden %s", dtp->dt_objcopy_path,
1930 		    file);
1931 		if ((status = system(cmd)) == -1) {
1932 			ret = dt_link_error(dtp, NULL, -1, NULL,
1933 			    "failed to run %s: %s", dtp->dt_objcopy_path,
1934 			    strerror(errno));
1935 			free(cmd);
1936 			goto done;
1937 		}
1938 		free(cmd);
1939 
1940 		if (WIFSIGNALED(status)) {
1941 			ret = dt_link_error(dtp, NULL, -1, NULL,
1942 			    "failed to link %s: %s failed due to signal %d",
1943 			    file, dtp->dt_objcopy_path, WTERMSIG(status));
1944 			goto done;
1945 		}
1946 
1947 		if (WEXITSTATUS(status) != 0) {
1948 			ret = dt_link_error(dtp, NULL, -1, NULL,
1949 			    "failed to link %s: %s exited with status %d\n",
1950 			    file, dtp->dt_objcopy_path, WEXITSTATUS(status));
1951 			goto done;
1952 		}
1953 #endif
1954 	} else {
1955 #if defined(__FreeBSD__) || defined(__NetBSD__)
1956 		if (rename(tfile, file) != 0) {
1957 			ret = dt_link_error(dtp, NULL, fd, NULL,
1958 			    "failed to rename %s to %s: %s", tfile, file,
1959 			    strerror(errno));
1960 			goto done;
1961 		}
1962 #endif
1963 		(void) close(fd);
1964 	}
1965 
1966 done:
1967 	dtrace_dof_destroy(dtp, dof);
1968 
1969 #if defined(__FreeBSD__) || defined(__NetBSD__)
1970 	if (!dtp->dt_lazyload)
1971 		(void) unlink(tfile);
1972 #endif
1973 	return (ret);
1974 }
1975