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  * Copyright 2016 Mark Johnston.
26  */
27 
28 #define	ELF_TARGET_ALL
29 #include <elf.h>
30 
31 #include <sys/types.h>
32 #include <sys/sysmacros.h>
33 
34 #include <unistd.h>
35 #include <strings.h>
36 #include <alloca.h>
37 #include <limits.h>
38 #include <stddef.h>
39 #include <stdlib.h>
40 #include <stdio.h>
41 #include <fcntl.h>
42 #include <errno.h>
43 #include <wait.h>
44 #include <assert.h>
45 #include <sys/ipc.h>
46 
47 #include <dt_impl.h>
48 #include <dt_provider.h>
49 #include <dt_program.h>
50 #include <dt_string.h>
51 
52 #define	ESHDR_NULL	0
53 #define	ESHDR_SHSTRTAB	1
54 #define	ESHDR_DOF	2
55 #define	ESHDR_STRTAB	3
56 #define	ESHDR_SYMTAB	4
57 #define	ESHDR_REL	5
58 #define	ESHDR_NUM	6
59 
60 #define	PWRITE_SCN(index, data) \
61 	(lseek64(fd, (off64_t)elf_file.shdr[(index)].sh_offset, SEEK_SET) != \
62 	(off64_t)elf_file.shdr[(index)].sh_offset || \
63 	dt_write(dtp, fd, (data), elf_file.shdr[(index)].sh_size) != \
64 	elf_file.shdr[(index)].sh_size)
65 
66 static const char DTRACE_SHSTRTAB32[] = "\0"
67 ".shstrtab\0"		/* 1 */
68 ".SUNW_dof\0"		/* 11 */
69 ".strtab\0"		/* 21 */
70 ".symtab\0"		/* 29 */
71 #ifdef __sparc
72 ".rela.SUNW_dof";	/* 37 */
73 #else
74 ".rel.SUNW_dof";	/* 37 */
75 #endif
76 
77 static const char DTRACE_SHSTRTAB64[] = "\0"
78 ".shstrtab\0"		/* 1 */
79 ".SUNW_dof\0"		/* 11 */
80 ".strtab\0"		/* 21 */
81 ".symtab\0"		/* 29 */
82 ".rela.SUNW_dof";	/* 37 */
83 
84 static const char DOFSTR[] = "__SUNW_dof";
85 static const char DOFLAZYSTR[] = "___SUNW_dof";
86 
87 typedef struct dt_link_pair {
88 	struct dt_link_pair *dlp_next;	/* next pair in linked list */
89 	void *dlp_str;			/* buffer for string table */
90 	void *dlp_sym;			/* buffer for symbol table */
91 } dt_link_pair_t;
92 
93 typedef struct dof_elf32 {
94 	uint32_t de_nrel;		/* relocation count */
95 #ifdef __sparc
96 	Elf32_Rela *de_rel;		/* array of relocations for sparc */
97 #else
98 	Elf32_Rel *de_rel;		/* array of relocations for x86 */
99 #endif
100 	uint32_t de_nsym;		/* symbol count */
101 	Elf32_Sym *de_sym;		/* array of symbols */
102 	uint32_t de_strlen;		/* size of of string table */
103 	char *de_strtab;		/* string table */
104 	uint32_t de_global;		/* index of the first global symbol */
105 } dof_elf32_t;
106 
107 static int
prepare_elf32(dtrace_hdl_t * dtp,const dof_hdr_t * dof,dof_elf32_t * dep)108 prepare_elf32(dtrace_hdl_t *dtp, const dof_hdr_t *dof, dof_elf32_t *dep)
109 {
110 	dof_sec_t *dofs, *s;
111 	dof_relohdr_t *dofrh;
112 	dof_relodesc_t *dofr;
113 	char *strtab;
114 	int i, j, nrel;
115 	size_t strtabsz = 1;
116 	uint32_t count = 0;
117 	size_t base;
118 	Elf32_Sym *sym;
119 #ifdef __sparc
120 	Elf32_Rela *rel;
121 #else
122 	Elf32_Rel *rel;
123 #endif
124 
125 	/*LINTED*/
126 	dofs = (dof_sec_t *)((char *)dof + dof->dofh_secoff);
127 
128 	/*
129 	 * First compute the size of the string table and the number of
130 	 * relocations present in the DOF.
131 	 */
132 	for (i = 0; i < dof->dofh_secnum; i++) {
133 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
134 			continue;
135 
136 		/*LINTED*/
137 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
138 
139 		s = &dofs[dofrh->dofr_strtab];
140 		strtab = (char *)dof + s->dofs_offset;
141 		assert(strtab[0] == '\0');
142 		strtabsz += s->dofs_size - 1;
143 
144 		s = &dofs[dofrh->dofr_relsec];
145 		/*LINTED*/
146 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
147 		count += s->dofs_size / s->dofs_entsize;
148 	}
149 
150 	dep->de_strlen = strtabsz;
151 	dep->de_nrel = count;
152 	dep->de_nsym = count + 1; /* the first symbol is always null */
153 
154 	if (dtp->dt_lazyload) {
155 		dep->de_strlen += sizeof (DOFLAZYSTR);
156 		dep->de_nsym++;
157 	} else {
158 		dep->de_strlen += sizeof (DOFSTR);
159 		dep->de_nsym++;
160 	}
161 
162 	if ((dep->de_rel = calloc(dep->de_nrel,
163 	    sizeof (dep->de_rel[0]))) == NULL) {
164 		return (dt_set_errno(dtp, EDT_NOMEM));
165 	}
166 
167 	if ((dep->de_sym = calloc(dep->de_nsym, sizeof (Elf32_Sym))) == NULL) {
168 		free(dep->de_rel);
169 		return (dt_set_errno(dtp, EDT_NOMEM));
170 	}
171 
172 	if ((dep->de_strtab = calloc(dep->de_strlen, 1)) == NULL) {
173 		free(dep->de_rel);
174 		free(dep->de_sym);
175 		return (dt_set_errno(dtp, EDT_NOMEM));
176 	}
177 
178 	count = 0;
179 	strtabsz = 1;
180 	dep->de_strtab[0] = '\0';
181 	rel = dep->de_rel;
182 	sym = dep->de_sym;
183 	dep->de_global = 1;
184 
185 	/*
186 	 * The first symbol table entry must be zeroed and is always ignored.
187 	 */
188 	bzero(sym, sizeof (Elf32_Sym));
189 	sym++;
190 
191 	/*
192 	 * Take a second pass through the DOF sections filling in the
193 	 * memory we allocated.
194 	 */
195 	for (i = 0; i < dof->dofh_secnum; i++) {
196 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
197 			continue;
198 
199 		/*LINTED*/
200 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
201 
202 		s = &dofs[dofrh->dofr_strtab];
203 		strtab = (char *)dof + s->dofs_offset;
204 		bcopy(strtab + 1, dep->de_strtab + strtabsz, s->dofs_size);
205 		base = strtabsz;
206 		strtabsz += s->dofs_size - 1;
207 
208 		s = &dofs[dofrh->dofr_relsec];
209 		/*LINTED*/
210 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
211 		nrel = s->dofs_size / s->dofs_entsize;
212 
213 		s = &dofs[dofrh->dofr_tgtsec];
214 
215 		for (j = 0; j < nrel; j++) {
216 #if defined(__i386) || defined(__amd64)
217 			rel->r_offset = s->dofs_offset +
218 			    dofr[j].dofr_offset;
219 			rel->r_info = ELF32_R_INFO(count + dep->de_global,
220 			    R_386_32);
221 #elif defined(__sparc)
222 			/*
223 			 * Add 4 bytes to hit the low half of this 64-bit
224 			 * big-endian address.
225 			 */
226 			rel->r_offset = s->dofs_offset +
227 			    dofr[j].dofr_offset + 4;
228 			rel->r_info = ELF32_R_INFO(count + dep->de_global,
229 			    R_SPARC_32);
230 #else
231 #error unknown ISA
232 #endif
233 
234 			sym->st_name = base + dofr[j].dofr_name - 1;
235 			sym->st_value = 0;
236 			sym->st_size = 0;
237 			sym->st_info = ELF32_ST_INFO(STB_GLOBAL, STT_FUNC);
238 			sym->st_other = 0;
239 			sym->st_shndx = SHN_UNDEF;
240 
241 			rel++;
242 			sym++;
243 			count++;
244 		}
245 	}
246 
247 	/*
248 	 * Add a symbol for the DOF itself. We use a different symbol for
249 	 * lazily and actively loaded DOF to make them easy to distinguish.
250 	 */
251 	sym->st_name = strtabsz;
252 	sym->st_value = 0;
253 	sym->st_size = dof->dofh_filesz;
254 	sym->st_info = ELF32_ST_INFO(STB_GLOBAL, STT_OBJECT);
255 	sym->st_other = 0;
256 	sym->st_shndx = ESHDR_DOF;
257 	sym++;
258 
259 	if (dtp->dt_lazyload) {
260 		bcopy(DOFLAZYSTR, dep->de_strtab + strtabsz,
261 		    sizeof (DOFLAZYSTR));
262 		strtabsz += sizeof (DOFLAZYSTR);
263 	} else {
264 		bcopy(DOFSTR, dep->de_strtab + strtabsz, sizeof (DOFSTR));
265 		strtabsz += sizeof (DOFSTR);
266 	}
267 
268 	assert(count == dep->de_nrel);
269 	assert(strtabsz == dep->de_strlen);
270 
271 	return (0);
272 }
273 
274 
275 typedef struct dof_elf64 {
276 	uint32_t de_nrel;
277 	Elf64_Rela *de_rel;
278 	uint32_t de_nsym;
279 	Elf64_Sym *de_sym;
280 
281 	uint32_t de_strlen;
282 	char *de_strtab;
283 
284 	uint32_t de_global;
285 } dof_elf64_t;
286 
287 static int
prepare_elf64(dtrace_hdl_t * dtp,const dof_hdr_t * dof,dof_elf64_t * dep)288 prepare_elf64(dtrace_hdl_t *dtp, const dof_hdr_t *dof, dof_elf64_t *dep)
289 {
290 	dof_sec_t *dofs, *s;
291 	dof_relohdr_t *dofrh;
292 	dof_relodesc_t *dofr;
293 	char *strtab;
294 	int i, j, nrel;
295 	size_t strtabsz = 1;
296 	uint32_t count = 0;
297 	size_t base;
298 	Elf64_Sym *sym;
299 	Elf64_Rela *rel;
300 
301 	/*LINTED*/
302 	dofs = (dof_sec_t *)((char *)dof + dof->dofh_secoff);
303 
304 	/*
305 	 * First compute the size of the string table and the number of
306 	 * relocations present in the DOF.
307 	 */
308 	for (i = 0; i < dof->dofh_secnum; i++) {
309 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
310 			continue;
311 
312 		/*LINTED*/
313 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
314 
315 		s = &dofs[dofrh->dofr_strtab];
316 		strtab = (char *)dof + s->dofs_offset;
317 		assert(strtab[0] == '\0');
318 		strtabsz += s->dofs_size - 1;
319 
320 		s = &dofs[dofrh->dofr_relsec];
321 		/*LINTED*/
322 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
323 		count += s->dofs_size / s->dofs_entsize;
324 	}
325 
326 	dep->de_strlen = strtabsz;
327 	dep->de_nrel = count;
328 	dep->de_nsym = count + 1; /* the first symbol is always null */
329 
330 	if (dtp->dt_lazyload) {
331 		dep->de_strlen += sizeof (DOFLAZYSTR);
332 		dep->de_nsym++;
333 	} else {
334 		dep->de_strlen += sizeof (DOFSTR);
335 		dep->de_nsym++;
336 	}
337 
338 	if ((dep->de_rel = calloc(dep->de_nrel,
339 	    sizeof (dep->de_rel[0]))) == NULL) {
340 		return (dt_set_errno(dtp, EDT_NOMEM));
341 	}
342 
343 	if ((dep->de_sym = calloc(dep->de_nsym, sizeof (Elf64_Sym))) == NULL) {
344 		free(dep->de_rel);
345 		return (dt_set_errno(dtp, EDT_NOMEM));
346 	}
347 
348 	if ((dep->de_strtab = calloc(dep->de_strlen, 1)) == NULL) {
349 		free(dep->de_rel);
350 		free(dep->de_sym);
351 		return (dt_set_errno(dtp, EDT_NOMEM));
352 	}
353 
354 	count = 0;
355 	strtabsz = 1;
356 	dep->de_strtab[0] = '\0';
357 	rel = dep->de_rel;
358 	sym = dep->de_sym;
359 	dep->de_global = 1;
360 
361 	/*
362 	 * The first symbol table entry must be zeroed and is always ignored.
363 	 */
364 	bzero(sym, sizeof (Elf64_Sym));
365 	sym++;
366 
367 	/*
368 	 * Take a second pass through the DOF sections filling in the
369 	 * memory we allocated.
370 	 */
371 	for (i = 0; i < dof->dofh_secnum; i++) {
372 		if (dofs[i].dofs_type != DOF_SECT_URELHDR)
373 			continue;
374 
375 		/*LINTED*/
376 		dofrh = (dof_relohdr_t *)((char *)dof + dofs[i].dofs_offset);
377 
378 		s = &dofs[dofrh->dofr_strtab];
379 		strtab = (char *)dof + s->dofs_offset;
380 		bcopy(strtab + 1, dep->de_strtab + strtabsz, s->dofs_size);
381 		base = strtabsz;
382 		strtabsz += s->dofs_size - 1;
383 
384 		s = &dofs[dofrh->dofr_relsec];
385 		/*LINTED*/
386 		dofr = (dof_relodesc_t *)((char *)dof + s->dofs_offset);
387 		nrel = s->dofs_size / s->dofs_entsize;
388 
389 		s = &dofs[dofrh->dofr_tgtsec];
390 
391 		for (j = 0; j < nrel; j++) {
392 #if defined(__i386) || defined(__amd64)
393 			rel->r_offset = s->dofs_offset +
394 			    dofr[j].dofr_offset;
395 			rel->r_info = ELF64_R_INFO(count + dep->de_global,
396 			    R_AMD64_64);
397 #elif defined(__sparc)
398 			rel->r_offset = s->dofs_offset +
399 			    dofr[j].dofr_offset;
400 			rel->r_info = ELF64_R_INFO(count + dep->de_global,
401 			    R_SPARC_64);
402 #else
403 #error unknown ISA
404 #endif
405 
406 			sym->st_name = base + dofr[j].dofr_name - 1;
407 			sym->st_value = 0;
408 			sym->st_size = 0;
409 			sym->st_info = GELF_ST_INFO(STB_GLOBAL, STT_FUNC);
410 			sym->st_other = 0;
411 			sym->st_shndx = SHN_UNDEF;
412 
413 			rel++;
414 			sym++;
415 			count++;
416 		}
417 	}
418 
419 	/*
420 	 * Add a symbol for the DOF itself. We use a different symbol for
421 	 * lazily and actively loaded DOF to make them easy to distinguish.
422 	 */
423 	sym->st_name = strtabsz;
424 	sym->st_value = 0;
425 	sym->st_size = dof->dofh_filesz;
426 	sym->st_info = GELF_ST_INFO(STB_GLOBAL, STT_OBJECT);
427 	sym->st_other = 0;
428 	sym->st_shndx = ESHDR_DOF;
429 	sym++;
430 
431 	if (dtp->dt_lazyload) {
432 		bcopy(DOFLAZYSTR, dep->de_strtab + strtabsz,
433 		    sizeof (DOFLAZYSTR));
434 		strtabsz += sizeof (DOFLAZYSTR);
435 	} else {
436 		bcopy(DOFSTR, dep->de_strtab + strtabsz, sizeof (DOFSTR));
437 		strtabsz += sizeof (DOFSTR);
438 	}
439 
440 	assert(count == dep->de_nrel);
441 	assert(strtabsz == dep->de_strlen);
442 
443 	return (0);
444 }
445 
446 /*
447  * Write out an ELF32 file prologue consisting of a header, section headers,
448  * and a section header string table.  The DOF data will follow this prologue
449  * and complete the contents of the given ELF file.
450  */
451 static int
dump_elf32(dtrace_hdl_t * dtp,const dof_hdr_t * dof,int fd)452 dump_elf32(dtrace_hdl_t *dtp, const dof_hdr_t *dof, int fd)
453 {
454 	struct {
455 		Elf32_Ehdr ehdr;
456 		Elf32_Shdr shdr[ESHDR_NUM];
457 	} elf_file;
458 
459 	Elf32_Shdr *shp;
460 	Elf32_Off off;
461 	dof_elf32_t de;
462 	int ret = 0;
463 	uint_t nshdr;
464 
465 	if (prepare_elf32(dtp, dof, &de) != 0)
466 		return (-1); /* errno is set for us */
467 
468 	/*
469 	 * If there are no relocations, we only need enough sections for
470 	 * the shstrtab and the DOF.
471 	 */
472 	nshdr = de.de_nrel == 0 ? ESHDR_SYMTAB + 1 : ESHDR_NUM;
473 
474 	bzero(&elf_file, sizeof (elf_file));
475 
476 	elf_file.ehdr.e_ident[EI_MAG0] = ELFMAG0;
477 	elf_file.ehdr.e_ident[EI_MAG1] = ELFMAG1;
478 	elf_file.ehdr.e_ident[EI_MAG2] = ELFMAG2;
479 	elf_file.ehdr.e_ident[EI_MAG3] = ELFMAG3;
480 	elf_file.ehdr.e_ident[EI_VERSION] = EV_CURRENT;
481 	elf_file.ehdr.e_ident[EI_CLASS] = ELFCLASS32;
482 #if defined(_BIG_ENDIAN)
483 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2MSB;
484 #elif defined(_LITTLE_ENDIAN)
485 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2LSB;
486 #endif
487 	elf_file.ehdr.e_type = ET_REL;
488 #if defined(__sparc)
489 	elf_file.ehdr.e_machine = EM_SPARC;
490 #elif defined(__i386) || defined(__amd64)
491 	elf_file.ehdr.e_machine = EM_386;
492 #endif
493 	elf_file.ehdr.e_version = EV_CURRENT;
494 	elf_file.ehdr.e_shoff = sizeof (Elf32_Ehdr);
495 	elf_file.ehdr.e_ehsize = sizeof (Elf32_Ehdr);
496 	elf_file.ehdr.e_phentsize = sizeof (Elf32_Phdr);
497 	elf_file.ehdr.e_shentsize = sizeof (Elf32_Shdr);
498 	elf_file.ehdr.e_shnum = nshdr;
499 	elf_file.ehdr.e_shstrndx = ESHDR_SHSTRTAB;
500 	off = sizeof (elf_file) + nshdr * sizeof (Elf32_Shdr);
501 
502 	shp = &elf_file.shdr[ESHDR_SHSTRTAB];
503 	shp->sh_name = 1; /* DTRACE_SHSTRTAB32[1] = ".shstrtab" */
504 	shp->sh_type = SHT_STRTAB;
505 	shp->sh_offset = off;
506 	shp->sh_size = sizeof (DTRACE_SHSTRTAB32);
507 	shp->sh_addralign = sizeof (char);
508 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
509 
510 	shp = &elf_file.shdr[ESHDR_DOF];
511 	shp->sh_name = 11; /* DTRACE_SHSTRTAB32[11] = ".SUNW_dof" */
512 	shp->sh_flags = SHF_ALLOC;
513 	shp->sh_type = SHT_SUNW_dof;
514 	shp->sh_offset = off;
515 	shp->sh_size = dof->dofh_filesz;
516 	shp->sh_addralign = 8;
517 	off = shp->sh_offset + shp->sh_size;
518 
519 	shp = &elf_file.shdr[ESHDR_STRTAB];
520 	shp->sh_name = 21; /* DTRACE_SHSTRTAB32[21] = ".strtab" */
521 	shp->sh_flags = SHF_ALLOC;
522 	shp->sh_type = SHT_STRTAB;
523 	shp->sh_offset = off;
524 	shp->sh_size = de.de_strlen;
525 	shp->sh_addralign = sizeof (char);
526 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 4);
527 
528 	shp = &elf_file.shdr[ESHDR_SYMTAB];
529 	shp->sh_name = 29; /* DTRACE_SHSTRTAB32[29] = ".symtab" */
530 	shp->sh_flags = SHF_ALLOC;
531 	shp->sh_type = SHT_SYMTAB;
532 	shp->sh_entsize = sizeof (Elf32_Sym);
533 	shp->sh_link = ESHDR_STRTAB;
534 	shp->sh_offset = off;
535 	shp->sh_info = de.de_global;
536 	shp->sh_size = de.de_nsym * sizeof (Elf32_Sym);
537 	shp->sh_addralign = 4;
538 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 4);
539 
540 	if (de.de_nrel == 0) {
541 		if (dt_write(dtp, fd, &elf_file,
542 		    sizeof (elf_file)) != sizeof (elf_file) ||
543 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB32) ||
544 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
545 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
546 		    PWRITE_SCN(ESHDR_DOF, dof)) {
547 			ret = dt_set_errno(dtp, errno);
548 		}
549 	} else {
550 		shp = &elf_file.shdr[ESHDR_REL];
551 		shp->sh_name = 37; /* DTRACE_SHSTRTAB32[37] = ".rel.SUNW_dof" */
552 		shp->sh_flags = SHF_ALLOC;
553 #ifdef __sparc
554 		shp->sh_type = SHT_RELA;
555 #else
556 		shp->sh_type = SHT_REL;
557 #endif
558 		shp->sh_entsize = sizeof (de.de_rel[0]);
559 		shp->sh_link = ESHDR_SYMTAB;
560 		shp->sh_info = ESHDR_DOF;
561 		shp->sh_offset = off;
562 		shp->sh_size = de.de_nrel * sizeof (de.de_rel[0]);
563 		shp->sh_addralign = 4;
564 
565 		if (dt_write(dtp, fd, &elf_file,
566 		    sizeof (elf_file)) != sizeof (elf_file) ||
567 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB32) ||
568 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
569 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
570 		    PWRITE_SCN(ESHDR_REL, de.de_rel) ||
571 		    PWRITE_SCN(ESHDR_DOF, dof)) {
572 			ret = dt_set_errno(dtp, errno);
573 		}
574 	}
575 
576 	free(de.de_strtab);
577 	free(de.de_sym);
578 	free(de.de_rel);
579 
580 	return (ret);
581 }
582 
583 /*
584  * Write out an ELF64 file prologue consisting of a header, section headers,
585  * and a section header string table.  The DOF data will follow this prologue
586  * and complete the contents of the given ELF file.
587  */
588 static int
dump_elf64(dtrace_hdl_t * dtp,const dof_hdr_t * dof,int fd)589 dump_elf64(dtrace_hdl_t *dtp, const dof_hdr_t *dof, int fd)
590 {
591 	struct {
592 		Elf64_Ehdr ehdr;
593 		Elf64_Shdr shdr[ESHDR_NUM];
594 	} elf_file;
595 
596 	Elf64_Shdr *shp;
597 	Elf64_Off off;
598 	dof_elf64_t de;
599 	int ret = 0;
600 	uint_t nshdr;
601 
602 	if (prepare_elf64(dtp, dof, &de) != 0)
603 		return (-1); /* errno is set for us */
604 
605 	/*
606 	 * If there are no relocations, we only need enough sections for
607 	 * the shstrtab and the DOF.
608 	 */
609 	nshdr = de.de_nrel == 0 ? ESHDR_SYMTAB + 1 : ESHDR_NUM;
610 
611 	bzero(&elf_file, sizeof (elf_file));
612 
613 	elf_file.ehdr.e_ident[EI_MAG0] = ELFMAG0;
614 	elf_file.ehdr.e_ident[EI_MAG1] = ELFMAG1;
615 	elf_file.ehdr.e_ident[EI_MAG2] = ELFMAG2;
616 	elf_file.ehdr.e_ident[EI_MAG3] = ELFMAG3;
617 	elf_file.ehdr.e_ident[EI_VERSION] = EV_CURRENT;
618 	elf_file.ehdr.e_ident[EI_CLASS] = ELFCLASS64;
619 #if defined(_BIG_ENDIAN)
620 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2MSB;
621 #elif defined(_LITTLE_ENDIAN)
622 	elf_file.ehdr.e_ident[EI_DATA] = ELFDATA2LSB;
623 #endif
624 	elf_file.ehdr.e_type = ET_REL;
625 #if defined(__sparc)
626 	elf_file.ehdr.e_machine = EM_SPARCV9;
627 #elif defined(__i386) || defined(__amd64)
628 	elf_file.ehdr.e_machine = EM_AMD64;
629 #endif
630 	elf_file.ehdr.e_version = EV_CURRENT;
631 	elf_file.ehdr.e_shoff = sizeof (Elf64_Ehdr);
632 	elf_file.ehdr.e_ehsize = sizeof (Elf64_Ehdr);
633 	elf_file.ehdr.e_phentsize = sizeof (Elf64_Phdr);
634 	elf_file.ehdr.e_shentsize = sizeof (Elf64_Shdr);
635 	elf_file.ehdr.e_shnum = nshdr;
636 	elf_file.ehdr.e_shstrndx = ESHDR_SHSTRTAB;
637 	off = sizeof (elf_file) + nshdr * sizeof (Elf64_Shdr);
638 
639 	shp = &elf_file.shdr[ESHDR_SHSTRTAB];
640 	shp->sh_name = 1; /* DTRACE_SHSTRTAB64[1] = ".shstrtab" */
641 	shp->sh_type = SHT_STRTAB;
642 	shp->sh_offset = off;
643 	shp->sh_size = sizeof (DTRACE_SHSTRTAB64);
644 	shp->sh_addralign = sizeof (char);
645 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
646 
647 	shp = &elf_file.shdr[ESHDR_DOF];
648 	shp->sh_name = 11; /* DTRACE_SHSTRTAB64[11] = ".SUNW_dof" */
649 	shp->sh_flags = SHF_ALLOC;
650 	shp->sh_type = SHT_SUNW_dof;
651 	shp->sh_offset = off;
652 	shp->sh_size = dof->dofh_filesz;
653 	shp->sh_addralign = 8;
654 	off = shp->sh_offset + shp->sh_size;
655 
656 	shp = &elf_file.shdr[ESHDR_STRTAB];
657 	shp->sh_name = 21; /* DTRACE_SHSTRTAB64[21] = ".strtab" */
658 	shp->sh_flags = SHF_ALLOC;
659 	shp->sh_type = SHT_STRTAB;
660 	shp->sh_offset = off;
661 	shp->sh_size = de.de_strlen;
662 	shp->sh_addralign = sizeof (char);
663 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
664 
665 	shp = &elf_file.shdr[ESHDR_SYMTAB];
666 	shp->sh_name = 29; /* DTRACE_SHSTRTAB64[29] = ".symtab" */
667 	shp->sh_flags = SHF_ALLOC;
668 	shp->sh_type = SHT_SYMTAB;
669 	shp->sh_entsize = sizeof (Elf64_Sym);
670 	shp->sh_link = ESHDR_STRTAB;
671 	shp->sh_offset = off;
672 	shp->sh_info = de.de_global;
673 	shp->sh_size = de.de_nsym * sizeof (Elf64_Sym);
674 	shp->sh_addralign = 8;
675 	off = P2ROUNDUP(shp->sh_offset + shp->sh_size, 8);
676 
677 	if (de.de_nrel == 0) {
678 		if (dt_write(dtp, fd, &elf_file,
679 		    sizeof (elf_file)) != sizeof (elf_file) ||
680 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB64) ||
681 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
682 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
683 		    PWRITE_SCN(ESHDR_DOF, dof)) {
684 			ret = dt_set_errno(dtp, errno);
685 		}
686 	} else {
687 		shp = &elf_file.shdr[ESHDR_REL];
688 		shp->sh_name = 37; /* DTRACE_SHSTRTAB64[37] = ".rel.SUNW_dof" */
689 		shp->sh_flags = SHF_ALLOC;
690 		shp->sh_type = SHT_RELA;
691 		shp->sh_entsize = sizeof (de.de_rel[0]);
692 		shp->sh_link = ESHDR_SYMTAB;
693 		shp->sh_info = ESHDR_DOF;
694 		shp->sh_offset = off;
695 		shp->sh_size = de.de_nrel * sizeof (de.de_rel[0]);
696 		shp->sh_addralign = 8;
697 
698 		if (dt_write(dtp, fd, &elf_file,
699 		    sizeof (elf_file)) != sizeof (elf_file) ||
700 		    PWRITE_SCN(ESHDR_SHSTRTAB, DTRACE_SHSTRTAB64) ||
701 		    PWRITE_SCN(ESHDR_STRTAB, de.de_strtab) ||
702 		    PWRITE_SCN(ESHDR_SYMTAB, de.de_sym) ||
703 		    PWRITE_SCN(ESHDR_REL, de.de_rel) ||
704 		    PWRITE_SCN(ESHDR_DOF, dof)) {
705 			ret = dt_set_errno(dtp, errno);
706 		}
707 	}
708 
709 	free(de.de_strtab);
710 	free(de.de_sym);
711 	free(de.de_rel);
712 
713 	return (ret);
714 }
715 
716 static int
dt_symtab_lookup(Elf_Data * data_sym,int nsym,uintptr_t addr,uint_t shn,GElf_Sym * sym)717 dt_symtab_lookup(Elf_Data *data_sym, int nsym, uintptr_t addr, uint_t shn,
718     GElf_Sym *sym)
719 {
720 	int i, ret = -1;
721 	GElf_Sym s;
722 
723 	for (i = 0; i < nsym && gelf_getsym(data_sym, i, sym) != NULL; i++) {
724 		if (GELF_ST_TYPE(sym->st_info) == STT_FUNC &&
725 		    shn == sym->st_shndx &&
726 		    sym->st_value <= addr &&
727 		    addr < sym->st_value + sym->st_size) {
728 			if (GELF_ST_BIND(sym->st_info) == STB_GLOBAL)
729 				return (0);
730 
731 			ret = 0;
732 			s = *sym;
733 		}
734 	}
735 
736 	if (ret == 0)
737 		*sym = s;
738 	return (ret);
739 }
740 
741 #if defined(__sparc)
742 
743 #define	DT_OP_RET		0x81c7e008
744 #define	DT_OP_NOP		0x01000000
745 #define	DT_OP_CALL		0x40000000
746 #define	DT_OP_CLR_O0		0x90102000
747 
748 #define	DT_IS_MOV_O7(inst)	(((inst) & 0xffffe000) == 0x9e100000)
749 #define	DT_IS_RESTORE(inst)	(((inst) & 0xc1f80000) == 0x81e80000)
750 #define	DT_IS_RETL(inst)	(((inst) & 0xfff83fff) == 0x81c02008)
751 
752 #define	DT_RS2(inst)		((inst) & 0x1f)
753 #define	DT_MAKE_RETL(reg)	(0x81c02008 | ((reg) << 14))
754 
755 /*ARGSUSED*/
756 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)757 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
758     uint32_t *off)
759 {
760 	uint32_t *ip;
761 
762 	if ((rela->r_offset & (sizeof (uint32_t) - 1)) != 0)
763 		return (-1);
764 
765 	/*LINTED*/
766 	ip = (uint32_t *)(p + rela->r_offset);
767 
768 	/*
769 	 * We only know about some specific relocation types.
770 	 */
771 	if (GELF_R_TYPE(rela->r_info) != R_SPARC_WDISP30 &&
772 	    GELF_R_TYPE(rela->r_info) != R_SPARC_WPLT30)
773 		return (-1);
774 
775 	/*
776 	 * We may have already processed this object file in an earlier linker
777 	 * invocation. Check to see if the present instruction sequence matches
778 	 * the one we would install below.
779 	 */
780 	if (isenabled) {
781 		if (ip[0] == DT_OP_NOP) {
782 			(*off) += sizeof (ip[0]);
783 			return (0);
784 		}
785 	} else {
786 		if (DT_IS_RESTORE(ip[1])) {
787 			if (ip[0] == DT_OP_RET) {
788 				(*off) += sizeof (ip[0]);
789 				return (0);
790 			}
791 		} else if (DT_IS_MOV_O7(ip[1])) {
792 			if (DT_IS_RETL(ip[0]))
793 				return (0);
794 		} else {
795 			if (ip[0] == DT_OP_NOP) {
796 				(*off) += sizeof (ip[0]);
797 				return (0);
798 			}
799 		}
800 	}
801 
802 	/*
803 	 * We only expect call instructions with a displacement of 0.
804 	 */
805 	if (ip[0] != DT_OP_CALL) {
806 		dt_dprintf("found %x instead of a call instruction at %llx\n",
807 		    ip[0], (u_longlong_t)rela->r_offset);
808 		return (-1);
809 	}
810 
811 	if (isenabled) {
812 		/*
813 		 * It would necessarily indicate incorrect usage if an is-
814 		 * enabled probe were tail-called so flag that as an error.
815 		 * It's also potentially (very) tricky to handle gracefully,
816 		 * but could be done if this were a desired use scenario.
817 		 */
818 		if (DT_IS_RESTORE(ip[1]) || DT_IS_MOV_O7(ip[1])) {
819 			dt_dprintf("tail call to is-enabled probe at %llx\n",
820 			    (u_longlong_t)rela->r_offset);
821 			return (-1);
822 		}
823 
824 
825 		/*
826 		 * On SPARC, we take advantage of the fact that the first
827 		 * argument shares the same register as for the return value.
828 		 * The macro handles the work of zeroing that register so we
829 		 * don't need to do anything special here. We instrument the
830 		 * instruction in the delay slot as we'll need to modify the
831 		 * return register after that instruction has been emulated.
832 		 */
833 		ip[0] = DT_OP_NOP;
834 		(*off) += sizeof (ip[0]);
835 	} else {
836 		/*
837 		 * If the call is followed by a restore, it's a tail call so
838 		 * change the call to a ret. If the call if followed by a mov
839 		 * of a register into %o7, it's a tail call in leaf context
840 		 * so change the call to a retl-like instruction that returns
841 		 * to that register value + 8 (rather than the typical %o7 +
842 		 * 8); the delay slot instruction is left, but should have no
843 		 * effect. Otherwise we change the call to be a nop. We
844 		 * identify the subsequent instruction as the probe point in
845 		 * all but the leaf tail-call case to ensure that arguments to
846 		 * the probe are complete and consistent. An astute, though
847 		 * largely hypothetical, observer would note that there is the
848 		 * possibility of a false-positive probe firing if the function
849 		 * contained a branch to the instruction in the delay slot of
850 		 * the call. Fixing this would require significant in-kernel
851 		 * modifications, and isn't worth doing until we see it in the
852 		 * wild.
853 		 */
854 		if (DT_IS_RESTORE(ip[1])) {
855 			ip[0] = DT_OP_RET;
856 			(*off) += sizeof (ip[0]);
857 		} else if (DT_IS_MOV_O7(ip[1])) {
858 			ip[0] = DT_MAKE_RETL(DT_RS2(ip[1]));
859 		} else {
860 			ip[0] = DT_OP_NOP;
861 			(*off) += sizeof (ip[0]);
862 		}
863 	}
864 
865 	return (0);
866 }
867 
868 #elif defined(__i386) || defined(__amd64)
869 
870 #define	DT_OP_NOP		0x90
871 #define	DT_OP_RET		0xc3
872 #define	DT_OP_CALL		0xe8
873 #define	DT_OP_JMP32		0xe9
874 #define	DT_OP_REX_RAX		0x48
875 #define	DT_OP_XOR_EAX_0		0x33
876 #define	DT_OP_XOR_EAX_1		0xc0
877 
878 static int
dt_modtext(dtrace_hdl_t * dtp,char * p,int isenabled,GElf_Rela * rela,uint32_t * off)879 dt_modtext(dtrace_hdl_t *dtp, char *p, int isenabled, GElf_Rela *rela,
880     uint32_t *off)
881 {
882 	uint8_t *ip = (uint8_t *)(p + rela->r_offset - 1);
883 	uint8_t ret;
884 
885 	/*
886 	 * On x86, the first byte of the instruction is the call opcode and
887 	 * the next four bytes are the 32-bit address; the relocation is for
888 	 * the address operand. We back up the offset to the first byte of
889 	 * the instruction. For is-enabled probes, we later advance the offset
890 	 * so that it hits the first nop in the instruction sequence.
891 	 */
892 	(*off) -= 1;
893 
894 	/*
895 	 * We only know about some specific relocation types. Luckily
896 	 * these types have the same values on both 32-bit and 64-bit
897 	 * x86 architectures.
898 	 */
899 	if (GELF_R_TYPE(rela->r_info) != R_386_PC32 &&
900 	    GELF_R_TYPE(rela->r_info) != R_386_PLT32)
901 		return (-1);
902 
903 	/*
904 	 * We may have already processed this object file in an earlier linker
905 	 * invocation. Check to see if the present instruction sequence matches
906 	 * the one we would install. For is-enabled probes, we advance the
907 	 * offset to the first nop instruction in the sequence to match the
908 	 * text modification code below.
909 	 */
910 	if (!isenabled) {
911 		if ((ip[0] == DT_OP_NOP || ip[0] == DT_OP_RET) &&
912 		    ip[1] == DT_OP_NOP && ip[2] == DT_OP_NOP &&
913 		    ip[3] == DT_OP_NOP && ip[4] == DT_OP_NOP)
914 			return (0);
915 	} else if (dtp->dt_oflags & DTRACE_O_LP64) {
916 		if (ip[0] == DT_OP_REX_RAX &&
917 		    ip[1] == DT_OP_XOR_EAX_0 && ip[2] == DT_OP_XOR_EAX_1 &&
918 		    (ip[3] == DT_OP_NOP || ip[3] == DT_OP_RET) &&
919 		    ip[4] == DT_OP_NOP) {
920 			(*off) += 3;
921 			return (0);
922 		}
923 	} else {
924 		if (ip[0] == DT_OP_XOR_EAX_0 && ip[1] == DT_OP_XOR_EAX_1 &&
925 		    (ip[2] == DT_OP_NOP || ip[2] == DT_OP_RET) &&
926 		    ip[3] == DT_OP_NOP && ip[4] == DT_OP_NOP) {
927 			(*off) += 2;
928 			return (0);
929 		}
930 	}
931 
932 	/*
933 	 * We expect either a call instrution with a 32-bit displacement or a
934 	 * jmp instruction with a 32-bit displacement acting as a tail-call.
935 	 */
936 	if (ip[0] != DT_OP_CALL && ip[0] != DT_OP_JMP32) {
937 		dt_dprintf("found %x instead of a call or jmp instruction at "
938 		    "%llx\n", ip[0], (u_longlong_t)rela->r_offset);
939 		return (-1);
940 	}
941 
942 	ret = (ip[0] == DT_OP_JMP32) ? DT_OP_RET : DT_OP_NOP;
943 
944 	/*
945 	 * Establish the instruction sequence -- all nops for probes, and an
946 	 * instruction to clear the return value register (%eax/%rax) followed
947 	 * by nops for is-enabled probes. For is-enabled probes, we advance
948 	 * the offset to the first nop. This isn't stricly necessary but makes
949 	 * for more readable disassembly when the probe is enabled.
950 	 */
951 	if (!isenabled) {
952 		ip[0] = ret;
953 		ip[1] = DT_OP_NOP;
954 		ip[2] = DT_OP_NOP;
955 		ip[3] = DT_OP_NOP;
956 		ip[4] = DT_OP_NOP;
957 	} else if (dtp->dt_oflags & DTRACE_O_LP64) {
958 		ip[0] = DT_OP_REX_RAX;
959 		ip[1] = DT_OP_XOR_EAX_0;
960 		ip[2] = DT_OP_XOR_EAX_1;
961 		ip[3] = ret;
962 		ip[4] = DT_OP_NOP;
963 		(*off) += 3;
964 	} else {
965 		ip[0] = DT_OP_XOR_EAX_0;
966 		ip[1] = DT_OP_XOR_EAX_1;
967 		ip[2] = ret;
968 		ip[3] = DT_OP_NOP;
969 		ip[4] = DT_OP_NOP;
970 		(*off) += 2;
971 	}
972 
973 	return (0);
974 }
975 
976 #else
977 #error unknown ISA
978 #endif
979 
980 /*PRINTFLIKE5*/
981 static int
dt_link_error(dtrace_hdl_t * dtp,Elf * elf,int fd,dt_link_pair_t * bufs,const char * format,...)982 dt_link_error(dtrace_hdl_t *dtp, Elf *elf, int fd, dt_link_pair_t *bufs,
983     const char *format, ...)
984 {
985 	va_list ap;
986 	dt_link_pair_t *pair;
987 
988 	va_start(ap, format);
989 	dt_set_errmsg(dtp, NULL, NULL, NULL, 0, format, ap);
990 	va_end(ap);
991 
992 	if (elf != NULL)
993 		(void) elf_end(elf);
994 
995 	if (fd >= 0)
996 		(void) close(fd);
997 
998 	while ((pair = bufs) != NULL) {
999 		bufs = pair->dlp_next;
1000 		dt_free(dtp, pair->dlp_str);
1001 		dt_free(dtp, pair->dlp_sym);
1002 		dt_free(dtp, pair);
1003 	}
1004 
1005 	return (dt_set_errno(dtp, EDT_COMPILER));
1006 }
1007 
1008 static int
process_obj(dtrace_hdl_t * dtp,const char * obj,int * eprobesp)1009 process_obj(dtrace_hdl_t *dtp, const char *obj, int *eprobesp)
1010 {
1011 	static const char dt_prefix[] = "__dtrace";
1012 	static const char dt_enabled[] = "enabled";
1013 	static const char dt_symprefix[] = "$dtrace";
1014 	static const char dt_symfmt[] = "%s%d.%s";
1015 	char probename[DTRACE_NAMELEN];
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 			if (dt_symtab_lookup(data_sym, isym, rela.r_offset,
1359 			    shdr_rel.sh_info, &fsym) != 0)
1360 				goto err;
1361 
1362 			if (fsym.st_name > data_str->d_size)
1363 				goto err;
1364 
1365 			assert(GELF_ST_TYPE(fsym.st_info) == STT_FUNC);
1366 
1367 			/*
1368 			 * If a NULL relocation name is passed to
1369 			 * dt_probe_define(), the function name is used for the
1370 			 * relocation. The relocation needs to use a mangled
1371 			 * name if the symbol is locally scoped; the function
1372 			 * name may need to change if we've found the global
1373 			 * alias for the locally scoped symbol (we prefer
1374 			 * global symbols to locals in dt_symtab_lookup()).
1375 			 */
1376 			s = (char *)data_str->d_buf + fsym.st_name;
1377 			r = NULL;
1378 
1379 			if (GELF_ST_BIND(fsym.st_info) == STB_LOCAL) {
1380 				dsym = fsym;
1381 				dsym.st_name = istr;
1382 				dsym.st_info = GELF_ST_INFO(STB_GLOBAL,
1383 				    STT_FUNC);
1384 				dsym.st_other =
1385 				    ELF64_ST_VISIBILITY(STV_ELIMINATE);
1386 				(void) gelf_update_sym(data_sym, isym, &dsym);
1387 
1388 				r = (char *)data_str->d_buf + istr;
1389 				istr += 1 + sprintf(r, dt_symfmt,
1390 				    dt_symprefix, objkey, s);
1391 				isym++;
1392 				assert(isym <= nsym);
1393 
1394 			} else if (strncmp(s, dt_symprefix,
1395 			    strlen(dt_symprefix)) == 0) {
1396 				r = s;
1397 				if ((s = strchr(s, '.')) == NULL)
1398 					goto err;
1399 				s++;
1400 			}
1401 
1402 			if ((pvp = dt_provider_lookup(dtp, pname)) == NULL) {
1403 				return (dt_link_error(dtp, elf, fd, bufs,
1404 				    "no such provider %s", pname));
1405 			}
1406 
1407 			/* strlen("___") */
1408 			if (strlcpy(probename, p + 3, sizeof (probename)) >=
1409 			    sizeof (probename))
1410 				return (dt_link_error(dtp, elf, fd, bufs,
1411 				    "probe name too long %s", probename));
1412 
1413 			(void) strhyphenate(probename);
1414 
1415 			if ((prp = dt_probe_lookup(pvp, probename)) == NULL) {
1416 				return (dt_link_error(dtp, elf, fd, bufs,
1417 				    "no such probe %s", probename));
1418 			}
1419 
1420 			assert(fsym.st_value <= rela.r_offset);
1421 
1422 			off = rela.r_offset - fsym.st_value;
1423 			if (dt_modtext(dtp, data_tgt->d_buf, eprobe,
1424 			    &rela, &off) != 0) {
1425 				goto err;
1426 			}
1427 
1428 			if (dt_probe_define(pvp, prp, s, r, off, eprobe) != 0) {
1429 				return (dt_link_error(dtp, elf, fd, bufs,
1430 				    "failed to allocate space for probe"));
1431 			}
1432 
1433 			mod = 1;
1434 			(void) elf_flagdata(data_tgt, ELF_C_SET, ELF_F_DIRTY);
1435 
1436 			/*
1437 			 * This symbol may already have been marked to
1438 			 * be ignored by another relocation referencing
1439 			 * the same symbol or if this object file has
1440 			 * already been processed by an earlier link
1441 			 * invocation.
1442 			 */
1443 			if (rsym.st_shndx != SHN_SUNW_IGNORE) {
1444 				rsym.st_shndx = SHN_SUNW_IGNORE;
1445 				(void) gelf_update_sym(data_sym, ndx, &rsym);
1446 			}
1447 		}
1448 	}
1449 
1450 	if (mod && elf_update(elf, ELF_C_WRITE) == -1)
1451 		goto err;
1452 
1453 	(void) elf_end(elf);
1454 	(void) close(fd);
1455 
1456 	while ((pair = bufs) != NULL) {
1457 		bufs = pair->dlp_next;
1458 		dt_free(dtp, pair->dlp_str);
1459 		dt_free(dtp, pair->dlp_sym);
1460 		dt_free(dtp, pair);
1461 	}
1462 
1463 	return (0);
1464 
1465 err:
1466 	return (dt_link_error(dtp, elf, fd, bufs,
1467 	    "an error was encountered while processing %s", obj));
1468 }
1469 
1470 int
dtrace_program_link(dtrace_hdl_t * dtp,dtrace_prog_t * pgp,uint_t dflags,const char * file,int objc,char * const objv[])1471 dtrace_program_link(dtrace_hdl_t *dtp, dtrace_prog_t *pgp, uint_t dflags,
1472     const char *file, int objc, char *const objv[])
1473 {
1474 	char drti[PATH_MAX];
1475 	dof_hdr_t *dof;
1476 	int fd, status, i, cur;
1477 	char *cmd, tmp;
1478 	size_t len;
1479 	int eprobes = 0, ret = 0;
1480 
1481 	/*
1482 	 * A NULL program indicates a special use in which we just link
1483 	 * together a bunch of object files specified in objv and then
1484 	 * unlink(2) those object files.
1485 	 */
1486 	if (pgp == NULL) {
1487 		const char *fmt = "%s -o %s -r";
1488 
1489 		len = snprintf(&tmp, 1, fmt, dtp->dt_ld_path, file) + 1;
1490 
1491 		for (i = 0; i < objc; i++)
1492 			len += strlen(objv[i]) + 1;
1493 
1494 		cmd = alloca(len);
1495 
1496 		cur = snprintf(cmd, len, fmt, dtp->dt_ld_path, file);
1497 
1498 		for (i = 0; i < objc; i++)
1499 			cur += snprintf(cmd + cur, len - cur, " %s", objv[i]);
1500 
1501 		if ((status = system(cmd)) == -1) {
1502 			return (dt_link_error(dtp, NULL, -1, NULL,
1503 			    "failed to run %s: %s", dtp->dt_ld_path,
1504 			    strerror(errno)));
1505 		}
1506 
1507 		if (WIFSIGNALED(status)) {
1508 			return (dt_link_error(dtp, NULL, -1, NULL,
1509 			    "failed to link %s: %s failed due to signal %d",
1510 			    file, dtp->dt_ld_path, WTERMSIG(status)));
1511 		}
1512 
1513 		if (WEXITSTATUS(status) != 0) {
1514 			return (dt_link_error(dtp, NULL, -1, NULL,
1515 			    "failed to link %s: %s exited with status %d\n",
1516 			    file, dtp->dt_ld_path, WEXITSTATUS(status)));
1517 		}
1518 
1519 		for (i = 0; i < objc; i++) {
1520 			if (strcmp(objv[i], file) != 0)
1521 				(void) unlink(objv[i]);
1522 		}
1523 
1524 		return (0);
1525 	}
1526 
1527 	for (i = 0; i < objc; i++) {
1528 		if (process_obj(dtp, objv[i], &eprobes) != 0)
1529 			return (-1); /* errno is set for us */
1530 	}
1531 
1532 	/*
1533 	 * If there are is-enabled probes then we need to force use of DOF
1534 	 * version 2.
1535 	 */
1536 	if (eprobes && pgp->dp_dofversion < DOF_VERSION_2)
1537 		pgp->dp_dofversion = DOF_VERSION_2;
1538 
1539 	if ((dof = dtrace_dof_create(dtp, pgp, dflags)) == NULL)
1540 		return (-1); /* errno is set for us */
1541 
1542 	/*
1543 	 * Create a temporary file and then unlink it if we're going to
1544 	 * combine it with drti.o later.  We can still refer to it in child
1545 	 * processes as /dev/fd/<fd>.
1546 	 */
1547 	if ((fd = open64(file, O_RDWR | O_CREAT | O_TRUNC, 0666)) == -1) {
1548 		return (dt_link_error(dtp, NULL, -1, NULL,
1549 		    "failed to open %s: %s", file, strerror(errno)));
1550 	}
1551 
1552 	/*
1553 	 * If -xlinktype=DOF has been selected, just write out the DOF.
1554 	 * Otherwise proceed to the default of generating and linking ELF.
1555 	 */
1556 	switch (dtp->dt_linktype) {
1557 	case DT_LTYP_DOF:
1558 		if (dt_write(dtp, fd, dof, dof->dofh_filesz) < dof->dofh_filesz)
1559 			ret = errno;
1560 
1561 		if (close(fd) != 0 && ret == 0)
1562 			ret = errno;
1563 
1564 		if (ret != 0) {
1565 			return (dt_link_error(dtp, NULL, -1, NULL,
1566 			    "failed to write %s: %s", file, strerror(ret)));
1567 		}
1568 
1569 		return (0);
1570 
1571 	case DT_LTYP_ELF:
1572 		break; /* fall through to the rest of dtrace_program_link() */
1573 
1574 	default:
1575 		return (dt_link_error(dtp, NULL, -1, NULL,
1576 		    "invalid link type %u\n", dtp->dt_linktype));
1577 	}
1578 
1579 
1580 	if (!dtp->dt_lazyload)
1581 		(void) unlink(file);
1582 
1583 	if (dtp->dt_oflags & DTRACE_O_LP64)
1584 		status = dump_elf64(dtp, dof, fd);
1585 	else
1586 		status = dump_elf32(dtp, dof, fd);
1587 
1588 	if (status != 0 || lseek(fd, 0, SEEK_SET) != 0) {
1589 		return (dt_link_error(dtp, NULL, -1, NULL,
1590 		    "failed to write %s: %s", file, strerror(errno)));
1591 	}
1592 
1593 	if (!dtp->dt_lazyload) {
1594 		const char *fmt = "%s -o %s -r -Blocal -Breduce /dev/fd/%d %s";
1595 
1596 		if (dtp->dt_oflags & DTRACE_O_LP64) {
1597 			(void) snprintf(drti, sizeof (drti),
1598 			    "%s/64/drti.o", _dtrace_libdir);
1599 		} else {
1600 			(void) snprintf(drti, sizeof (drti),
1601 			    "%s/drti.o", _dtrace_libdir);
1602 		}
1603 
1604 		len = snprintf(&tmp, 1, fmt, dtp->dt_ld_path, file, fd,
1605 		    drti) + 1;
1606 
1607 		cmd = alloca(len);
1608 
1609 		(void) snprintf(cmd, len, fmt, dtp->dt_ld_path, file, fd, drti);
1610 
1611 		if ((status = system(cmd)) == -1) {
1612 			ret = dt_link_error(dtp, NULL, -1, NULL,
1613 			    "failed to run %s: %s", dtp->dt_ld_path,
1614 			    strerror(errno));
1615 			goto done;
1616 		}
1617 
1618 		(void) close(fd); /* release temporary file */
1619 
1620 		if (WIFSIGNALED(status)) {
1621 			ret = dt_link_error(dtp, NULL, -1, NULL,
1622 			    "failed to link %s: %s failed due to signal %d",
1623 			    file, dtp->dt_ld_path, WTERMSIG(status));
1624 			goto done;
1625 		}
1626 
1627 		if (WEXITSTATUS(status) != 0) {
1628 			ret = dt_link_error(dtp, NULL, -1, NULL,
1629 			    "failed to link %s: %s exited with status %d\n",
1630 			    file, dtp->dt_ld_path, WEXITSTATUS(status));
1631 			goto done;
1632 		}
1633 	} else {
1634 		(void) close(fd);
1635 	}
1636 
1637 done:
1638 	dtrace_dof_destroy(dtp, dof);
1639 	return (ret);
1640 }
1641