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 * Copyright 2007 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25 /*
26 * Copyright 2012 Jason King. All rights reserved.
27 * Use is subject to license terms.
28 */
29
30 /*
31 * Copyright 2020 Joyent, Inc.
32 */
33
34 /*
35 * CTF DWARF conversion theory.
36 *
37 * DWARF data contains a series of compilation units. Each compilation unit
38 * generally refers to an object file or what once was, in the case of linked
39 * binaries and shared objects. Each compilation unit has a series of what DWARF
40 * calls a DIE (Debugging Information Entry). The set of entries that we care
41 * about have type information stored in a series of attributes. Each DIE also
42 * has a tag that identifies the kind of attributes that it has.
43 *
44 * A given DIE may itself have children. For example, a DIE that represents a
45 * structure has children which represent members. Whenever we encounter a DIE
46 * that has children or other values or types associated with it, we recursively
47 * process those children first so that way we can then refer to the generated
48 * CTF type id while processing its parent. This reduces the amount of unknowns
49 * and fixups that we need. It also ensures that we don't accidentally add types
50 * that an overzealous compiler might add to the DWARF data but aren't used by
51 * anything in the system.
52 *
53 * Once we do a conversion, we store a mapping in an AVL tree that goes from the
54 * DWARF's die offset, which is relative to the given compilation unit, to a
55 * ctf_id_t.
56 *
57 * Unfortunately, some compilers actually will emit duplicate entries for a
58 * given type that look similar, but aren't quite. To that end, we go through
59 * and do a variant on a merge once we're done processing a single compilation
60 * unit which deduplicates all of the types that are in the unit.
61 *
62 * Finally, if we encounter an object that has multiple compilation units, then
63 * we'll convert all of the compilation units separately and then do a merge, so
64 * that way we can result in one single ctf_file_t that represents everything
65 * for the object.
66 *
67 * Conversion Steps
68 * ----------------
69 *
70 * Because a given object we've been given to convert may have multiple
71 * compilation units, we break the work into two halves. The first half
72 * processes each compilation unit (potentially in parallel) and then the second
73 * half optionally merges all of the dies in the first half. First, we'll cover
74 * what's involved in converting a single ctf_cu_t's dwarf to CTF. This covers
75 * the work done in ctf_dwarf_convert_one().
76 *
77 * An individual ctf_cu_t, which represents a compilation unit, is converted to
78 * CTF in a series of multiple passes.
79 *
80 * Pass 1: During the first pass we walk all of the top-level dies and if we
81 * find a function, variable, struct, union, enum or typedef, we recursively
82 * transform all of its types. We don't recurse or process everything, because
83 * we don't want to add some of the types that compilers may add which are
84 * effectively unused.
85 *
86 * During pass 1, if we encounter any structures or unions we mark them for
87 * fixing up later. This is necessary because we may not be able to determine
88 * the full size of a structure at the beginning of time. This will happen if
89 * the DWARF attribute DW_AT_byte_size is not present for a member. Because of
90 * this possibility we defer adding members to structures or even converting
91 * them during pass 1 and save that for pass 2. Adding all of the base
92 * structures without any of their members helps deal with any circular
93 * dependencies that we might encounter.
94 *
95 * Pass 2: This pass is used to do the first half of fixing up structures and
96 * unions. Rather than walk the entire type space again, we actually walk the
97 * list of structures and unions that we marked for later fixing up. Here, we
98 * iterate over every structure and add members to the underlying ctf_file_t,
99 * but not to the structs themselves. One might wonder why we don't, and the
100 * main reason is that libctf requires a ctf_update() be done before adding the
101 * members to structures or unions.
102 *
103 * Pass 3: This pass is used to do the second half of fixing up structures and
104 * unions. During this part we always go through and add members to structures
105 * and unions that we added to the container in the previous pass. In addition,
106 * we set the structure and union's actual size, which may have additional
107 * padding added by the compiler, it isn't simply the last offset. DWARF always
108 * guarantees an attribute exists for this. Importantly no ctf_id_t's change
109 * during pass 2.
110 *
111 * Pass 4: The next phase is to add CTF entries for all of the symbols and
112 * variables that are present in this die. During pass 1 we added entries to a
113 * map for each variable and function. During this pass, we iterate over the
114 * symbol table and when we encounter a symbol that we have in our lists of
115 * translated information which matches, we then add it to the ctf_file_t.
116 *
117 * Pass 5: Here we go and look for any weak symbols and functions and see if
118 * they match anything that we recognize. If so, then we add type information
119 * for them at this point based on the matching type.
120 *
121 * Pass 6: This pass is actually a variant on a merge. The traditional merge
122 * process expects there to be no duplicate types. As such, at the end of
123 * conversion, we do a dedup on all of the types in the system. The
124 * deduplication process is described in lib/libctf/common/ctf_merge.c.
125 *
126 * Once pass 6 is done, we've finished processing the individual compilation
127 * unit.
128 *
129 * The following steps reflect the general process of doing a conversion.
130 *
131 * 1) Walk the dwarf section and determine the number of compilation units
132 * 2) Create a ctf_cu_t for each compilation unit
133 * 3) Add all ctf_cu_t's to a workq
134 * 4) Have the workq process each die with ctf_dwarf_convert_one. This itself
135 * is comprised of several steps, which were already enumerated.
136 * 5) If we have multiple cu's, we do a ctf merge of all the dies. The mechanics
137 * of the merge are discussed in lib/libctf/common/ctf_merge.c.
138 * 6) Free everything up and return a ctf_file_t to the user. If we only had a
139 * single compilation unit, then we give that to the user. Otherwise, we
140 * return the merged ctf_file_t.
141 *
142 * Threading
143 * ---------
144 *
145 * The process has been designed to be amenable to threading. Each compilation
146 * unit has its own type stream, therefore the logical place to divide and
147 * conquer is at the compilation unit. Each ctf_cu_t has been built to be able
148 * to be processed independently of the others. It has its own libdwarf handle,
149 * as a given libdwarf handle may only be used by a single thread at a time.
150 * This allows the various ctf_cu_t's to be processed in parallel by different
151 * threads.
152 *
153 * All of the ctf_cu_t's are loaded into a workq which allows for a number of
154 * threads to be specified and used as a thread pool to process all of the
155 * queued work. We set the number of threads to use in the workq equal to the
156 * number of threads that the user has specified.
157 *
158 * After all of the compilation units have been drained, we use the same number
159 * of threads when performing a merge of multiple compilation units, if they
160 * exist.
161 *
162 * While all of these different parts do support and allow for multiple threads,
163 * it's important that when only a single thread is specified, that it be the
164 * calling thread. This allows the conversion routines to be used in a context
165 * that doesn't allow additional threads, such as rtld.
166 *
167 * Common DWARF Mechanics and Notes
168 * --------------------------------
169 *
170 * At this time, we really only support DWARFv2, though support for DWARFv4 is
171 * mostly there. There is no intent to support DWARFv3.
172 *
173 * Generally types for something are stored in the DW_AT_type attribute. For
174 * example, a function's return type will be stored in the local DW_AT_type
175 * attribute while the arguments will be in child DIEs. There are also various
176 * times when we don't have any DW_AT_type. In that case, the lack of a type
177 * implies, at least for C, that its C type is void. Because DWARF doesn't emit
178 * one, we have a synthetic void type that we create and manipulate instead and
179 * pass it off to consumers on an as-needed basis. If nothing has a void type,
180 * it will not be emitted.
181 *
182 * Architecture Specific Parts
183 * ---------------------------
184 *
185 * The CTF tooling encodes various information about the various architectures
186 * in the system. Importantly, the tool assumes that every architecture has a
187 * data model where long and pointer are the same size. This is currently the
188 * case, as the two data models illumos supports are ILP32 and LP64.
189 *
190 * In addition, we encode the mapping of various floating point sizes to various
191 * types for each architecture. If a new architecture is being added, it should
192 * be added to the list. The general design of the ctf conversion tools is to be
193 * architecture independent. eg. any of the tools here should be able to convert
194 * any architecture's DWARF into ctf; however, this has not been rigorously
195 * tested and more importantly, the ctf routines don't currently write out the
196 * data in an endian-aware form, they only use that of the currently running
197 * library.
198 */
199
200 #include <libctf_impl.h>
201 #include <sys/avl.h>
202 #include <sys/debug.h>
203 #include <gelf.h>
204 #include <libdwarf.h>
205 #include <dwarf.h>
206 #include <libgen.h>
207 #include <workq.h>
208 #include <errno.h>
209
210 #define DWARF_VERSION_TWO 2
211 #define DWARF_VARARGS_NAME "..."
212
213 /*
214 * Dwarf may refer recursively to other types that we've already processed. To
215 * see if we've already converted them, we look them up in an AVL tree that's
216 * sorted by the DWARF id.
217 */
218 typedef struct ctf_dwmap {
219 avl_node_t cdm_avl;
220 Dwarf_Off cdm_off;
221 Dwarf_Die cdm_die;
222 ctf_id_t cdm_id;
223 boolean_t cdm_fix;
224 } ctf_dwmap_t;
225
226 typedef struct ctf_dwvar {
227 ctf_list_t cdv_list;
228 char *cdv_name;
229 ctf_id_t cdv_type;
230 boolean_t cdv_global;
231 } ctf_dwvar_t;
232
233 typedef struct ctf_dwfunc {
234 ctf_list_t cdf_list;
235 char *cdf_name;
236 ctf_funcinfo_t cdf_fip;
237 ctf_id_t *cdf_argv;
238 boolean_t cdf_global;
239 } ctf_dwfunc_t;
240
241 typedef struct ctf_dwbitf {
242 ctf_list_t cdb_list;
243 ctf_id_t cdb_base;
244 uint_t cdb_nbits;
245 ctf_id_t cdb_id;
246 } ctf_dwbitf_t;
247
248 /*
249 * The ctf_cu_t represents a single top-level DWARF die unit. While generally,
250 * the typical object file has only a single die, if we're asked to convert
251 * something that's been linked from multiple sources, multiple dies will exist.
252 */
253 typedef struct ctf_die {
254 Elf *cu_elf; /* shared libelf handle */
255 char *cu_name; /* basename of the DIE */
256 ctf_merge_t *cu_cmh; /* merge handle */
257 ctf_list_t cu_vars; /* List of variables */
258 ctf_list_t cu_funcs; /* List of functions */
259 ctf_list_t cu_bitfields; /* Bit field members */
260 Dwarf_Debug cu_dwarf; /* libdwarf handle */
261 Dwarf_Die cu_cu; /* libdwarf compilation unit */
262 Dwarf_Off cu_cuoff; /* cu's offset */
263 Dwarf_Off cu_maxoff; /* maximum offset */
264 ctf_file_t *cu_ctfp; /* output CTF file */
265 avl_tree_t cu_map; /* map die offsets to CTF types */
266 char *cu_errbuf; /* error message buffer */
267 size_t cu_errlen; /* error message buffer length */
268 size_t cu_ptrsz; /* object's pointer size */
269 boolean_t cu_bigend; /* is it big endian */
270 boolean_t cu_doweaks; /* should we convert weak symbols? */
271 uint_t cu_mach; /* machine type */
272 ctf_id_t cu_voidtid; /* void pointer */
273 ctf_id_t cu_longtid; /* id for a 'long' */
274 } ctf_cu_t;
275
276 static int ctf_dwarf_offset(ctf_cu_t *, Dwarf_Die, Dwarf_Off *);
277 static int ctf_dwarf_convert_die(ctf_cu_t *, Dwarf_Die);
278 static int ctf_dwarf_convert_type(ctf_cu_t *, Dwarf_Die, ctf_id_t *, int);
279
280 static int ctf_dwarf_function_count(ctf_cu_t *, Dwarf_Die, ctf_funcinfo_t *,
281 boolean_t);
282 static int ctf_dwarf_convert_fargs(ctf_cu_t *, Dwarf_Die, ctf_funcinfo_t *,
283 ctf_id_t *);
284
285 /*
286 * This is a generic way to set a CTF Conversion backend error depending on what
287 * we were doing. Unless it was one of a specific set of errors that don't
288 * indicate a programming / translation bug, eg. ENOMEM, then we transform it
289 * into a CTF backend error and fill in the error buffer.
290 */
291 static int
292 ctf_dwarf_error(ctf_cu_t *cup, ctf_file_t *cfp, int err, const char *fmt, ...)
293 {
294 va_list ap;
295 int ret;
296 size_t off = 0;
297 ssize_t rem = cup->cu_errlen;
298 if (cfp != NULL)
299 err = ctf_errno(cfp);
300
301 if (err == ENOMEM)
302 return (err);
303
304 ret = snprintf(cup->cu_errbuf, rem, "die %s: ", cup->cu_name);
305 if (ret < 0)
306 goto err;
307 off += ret;
308 rem = MAX(rem - ret, 0);
309
310 va_start(ap, fmt);
311 ret = vsnprintf(cup->cu_errbuf + off, rem, fmt, ap);
312 va_end(ap);
313 if (ret < 0)
314 goto err;
315
316 off += ret;
317 rem = MAX(rem - ret, 0);
318 if (fmt[strlen(fmt) - 1] != '\n') {
319 (void) snprintf(cup->cu_errbuf + off, rem,
320 ": %s\n", ctf_errmsg(err));
321 }
322 va_end(ap);
323 return (ECTF_CONVBKERR);
324
325 err:
326 cup->cu_errbuf[0] = '\0';
327 return (ECTF_CONVBKERR);
328 }
329
330 /*
331 * DWARF often opts to put no explicit type to describe a void type. eg. if we
332 * have a reference type whose DW_AT_type member doesn't exist, then we should
333 * instead assume it points to void. Because this isn't represented, we
334 * instead cause it to come into existence.
335 */
336 static ctf_id_t
337 ctf_dwarf_void(ctf_cu_t *cup)
338 {
339 if (cup->cu_voidtid == CTF_ERR) {
340 ctf_encoding_t enc = { CTF_INT_SIGNED, 0, 0 };
341 cup->cu_voidtid = ctf_add_integer(cup->cu_ctfp, CTF_ADD_ROOT,
342 "void", &enc);
343 if (cup->cu_voidtid == CTF_ERR) {
344 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
345 "failed to create void type: %s\n",
346 ctf_errmsg(ctf_errno(cup->cu_ctfp)));
347 }
348 }
349
350 return (cup->cu_voidtid);
351 }
352
353 /*
354 * There are many different forms that an array index may take. However, we just
355 * always force it to be of a type long no matter what. Therefore we use this to
356 * have a single instance of long across everything.
357 */
358 static ctf_id_t
359 ctf_dwarf_long(ctf_cu_t *cup)
360 {
361 if (cup->cu_longtid == CTF_ERR) {
362 ctf_encoding_t enc;
363
364 enc.cte_format = CTF_INT_SIGNED;
365 enc.cte_offset = 0;
366 /* All illumos systems are LP */
367 enc.cte_bits = cup->cu_ptrsz * 8;
368 cup->cu_longtid = ctf_add_integer(cup->cu_ctfp, CTF_ADD_NONROOT,
369 "long", &enc);
370 if (cup->cu_longtid == CTF_ERR) {
371 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
372 "failed to create long type: %s\n",
373 ctf_errmsg(ctf_errno(cup->cu_ctfp)));
374 }
375
376 }
377
378 return (cup->cu_longtid);
379 }
380
381 static int
382 ctf_dwmap_comp(const void *a, const void *b)
383 {
384 const ctf_dwmap_t *ca = a;
385 const ctf_dwmap_t *cb = b;
386
387 if (ca->cdm_off > cb->cdm_off)
388 return (1);
389 if (ca->cdm_off < cb->cdm_off)
390 return (-1);
391 return (0);
392 }
393
394 static int
395 ctf_dwmap_add(ctf_cu_t *cup, ctf_id_t id, Dwarf_Die die, boolean_t fix)
396 {
397 int ret;
398 avl_index_t index;
399 ctf_dwmap_t *dwmap;
400 Dwarf_Off off;
401
402 VERIFY(id > 0 && id < CTF_MAX_TYPE);
403
404 if ((ret = ctf_dwarf_offset(cup, die, &off)) != 0)
405 return (ret);
406
407 if ((dwmap = ctf_alloc(sizeof (ctf_dwmap_t))) == NULL)
408 return (ENOMEM);
409
410 dwmap->cdm_die = die;
411 dwmap->cdm_off = off;
412 dwmap->cdm_id = id;
413 dwmap->cdm_fix = fix;
414
415 ctf_dprintf("dwmap: %p %" DW_PR_DUx "->%d\n", dwmap, off, id);
416 VERIFY(avl_find(&cup->cu_map, dwmap, &index) == NULL);
417 avl_insert(&cup->cu_map, dwmap, index);
418 return (0);
419 }
420
421 static int
422 ctf_dwarf_attribute(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Half name,
423 Dwarf_Attribute *attrp)
424 {
425 int ret;
426 Dwarf_Error derr;
427
428 if ((ret = dwarf_attr(die, name, attrp, &derr)) == DW_DLV_OK)
429 return (0);
430 if (ret == DW_DLV_NO_ENTRY) {
431 *attrp = NULL;
432 return (ENOENT);
433 }
434 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
435 "failed to get attribute for type: %s\n",
436 dwarf_errmsg(derr));
437 return (ECTF_CONVBKERR);
438 }
439
440 static int
441 ctf_dwarf_ref(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Half name, Dwarf_Off *refp)
442 {
443 int ret;
444 Dwarf_Attribute attr;
445 Dwarf_Error derr;
446
447 if ((ret = ctf_dwarf_attribute(cup, die, name, &attr)) != 0)
448 return (ret);
449
450 if (dwarf_formref(attr, refp, &derr) == DW_DLV_OK) {
451 dwarf_dealloc(cup->cu_dwarf, attr, DW_DLA_ATTR);
452 return (0);
453 }
454
455 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
456 "failed to get unsigned attribute for type: %s\n",
457 dwarf_errmsg(derr));
458 return (ECTF_CONVBKERR);
459 }
460
461 static int
462 ctf_dwarf_refdie(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Half name,
463 Dwarf_Die *diep)
464 {
465 int ret;
466 Dwarf_Off off;
467 Dwarf_Error derr;
468
469 if ((ret = ctf_dwarf_ref(cup, die, name, &off)) != 0)
470 return (ret);
471
472 off += cup->cu_cuoff;
473 if ((ret = dwarf_offdie(cup->cu_dwarf, off, diep, &derr)) !=
474 DW_DLV_OK) {
475 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
476 "failed to get die from offset %" DW_PR_DUu ": %s\n",
477 off, dwarf_errmsg(derr));
478 return (ECTF_CONVBKERR);
479 }
480
481 return (0);
482 }
483
484 static int
485 ctf_dwarf_signed(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Half name,
486 Dwarf_Signed *valp)
487 {
488 int ret;
489 Dwarf_Attribute attr;
490 Dwarf_Error derr;
491
492 if ((ret = ctf_dwarf_attribute(cup, die, name, &attr)) != 0)
493 return (ret);
494
495 if (dwarf_formsdata(attr, valp, &derr) == DW_DLV_OK) {
496 dwarf_dealloc(cup->cu_dwarf, attr, DW_DLA_ATTR);
497 return (0);
498 }
499
500 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
501 "failed to get unsigned attribute for type: %s\n",
502 dwarf_errmsg(derr));
503 return (ECTF_CONVBKERR);
504 }
505
506 static int
507 ctf_dwarf_unsigned(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Half name,
508 Dwarf_Unsigned *valp)
509 {
510 int ret;
511 Dwarf_Attribute attr;
512 Dwarf_Error derr;
513
514 if ((ret = ctf_dwarf_attribute(cup, die, name, &attr)) != 0)
515 return (ret);
516
517 if (dwarf_formudata(attr, valp, &derr) == DW_DLV_OK) {
518 dwarf_dealloc(cup->cu_dwarf, attr, DW_DLA_ATTR);
519 return (0);
520 }
521
522 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
523 "failed to get unsigned attribute for type: %s\n",
524 dwarf_errmsg(derr));
525 return (ECTF_CONVBKERR);
526 }
527
528 static int
529 ctf_dwarf_boolean(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Half name,
530 Dwarf_Bool *val)
531 {
532 int ret;
533 Dwarf_Attribute attr;
534 Dwarf_Error derr;
535
536 if ((ret = ctf_dwarf_attribute(cup, die, name, &attr)) != 0)
537 return (ret);
538
539 if (dwarf_formflag(attr, val, &derr) == DW_DLV_OK) {
540 dwarf_dealloc(cup->cu_dwarf, attr, DW_DLA_ATTR);
541 return (0);
542 }
543
544 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
545 "failed to get boolean attribute for type: %s\n",
546 dwarf_errmsg(derr));
547
548 return (ECTF_CONVBKERR);
549 }
550
551 static int
552 ctf_dwarf_string(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Half name, char **strp)
553 {
554 int ret;
555 char *s;
556 Dwarf_Attribute attr;
557 Dwarf_Error derr;
558
559 *strp = NULL;
560 if ((ret = ctf_dwarf_attribute(cup, die, name, &attr)) != 0)
561 return (ret);
562
563 if (dwarf_formstring(attr, &s, &derr) == DW_DLV_OK) {
564 if ((*strp = ctf_strdup(s)) == NULL)
565 ret = ENOMEM;
566 else
567 ret = 0;
568 dwarf_dealloc(cup->cu_dwarf, attr, DW_DLA_ATTR);
569 return (ret);
570 }
571
572 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
573 "failed to get string attribute for type: %s\n",
574 dwarf_errmsg(derr));
575 return (ECTF_CONVBKERR);
576 }
577
578 static int
579 ctf_dwarf_member_location(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Unsigned *valp)
580 {
581 int ret;
582 Dwarf_Error derr;
583 Dwarf_Attribute attr;
584 Dwarf_Locdesc *loc;
585 Dwarf_Signed locnum;
586
587 if ((ret = ctf_dwarf_attribute(cup, die, DW_AT_data_member_location,
588 &attr)) != 0)
589 return (ret);
590
591 if (dwarf_loclist(attr, &loc, &locnum, &derr) != DW_DLV_OK) {
592 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
593 "failed to obtain location list for member offset: %s",
594 dwarf_errmsg(derr));
595 dwarf_dealloc(cup->cu_dwarf, attr, DW_DLA_ATTR);
596 return (ECTF_CONVBKERR);
597 }
598 dwarf_dealloc(cup->cu_dwarf, attr, DW_DLA_ATTR);
599
600 if (locnum != 1 || loc->ld_s->lr_atom != DW_OP_plus_uconst) {
601 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
602 "failed to parse location structure for member");
603 dwarf_dealloc(cup->cu_dwarf, loc->ld_s, DW_DLA_LOC_BLOCK);
604 dwarf_dealloc(cup->cu_dwarf, loc, DW_DLA_LOCDESC);
605 return (ECTF_CONVBKERR);
606 }
607
608 *valp = loc->ld_s->lr_number;
609
610 dwarf_dealloc(cup->cu_dwarf, loc->ld_s, DW_DLA_LOC_BLOCK);
611 dwarf_dealloc(cup->cu_dwarf, loc, DW_DLA_LOCDESC);
612 return (0);
613 }
614
615
616 static int
617 ctf_dwarf_offset(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Off *offsetp)
618 {
619 Dwarf_Error derr;
620
621 if (dwarf_dieoffset(die, offsetp, &derr) == DW_DLV_OK)
622 return (0);
623
624 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
625 "failed to get die offset: %s\n",
626 dwarf_errmsg(derr));
627 return (ECTF_CONVBKERR);
628 }
629
630 /* simpler variant for debugging output */
631 static Dwarf_Off
632 ctf_die_offset(Dwarf_Die die)
633 {
634 Dwarf_Off off = -1;
635 Dwarf_Error derr;
636
637 (void) dwarf_dieoffset(die, &off, &derr);
638 return (off);
639 }
640
641 static int
642 ctf_dwarf_tag(ctf_cu_t *cup, Dwarf_Die die, Dwarf_Half *tagp)
643 {
644 Dwarf_Error derr;
645
646 if (dwarf_tag(die, tagp, &derr) == DW_DLV_OK)
647 return (0);
648
649 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
650 "failed to get tag type: %s\n",
651 dwarf_errmsg(derr));
652 return (ECTF_CONVBKERR);
653 }
654
655 static int
656 ctf_dwarf_sib(ctf_cu_t *cup, Dwarf_Die base, Dwarf_Die *sibp)
657 {
658 Dwarf_Error derr;
659 int ret;
660
661 *sibp = NULL;
662 ret = dwarf_siblingof(cup->cu_dwarf, base, sibp, &derr);
663 if (ret == DW_DLV_OK || ret == DW_DLV_NO_ENTRY)
664 return (0);
665
666 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
667 "failed to sibling from die: %s\n",
668 dwarf_errmsg(derr));
669 return (ECTF_CONVBKERR);
670 }
671
672 static int
673 ctf_dwarf_child(ctf_cu_t *cup, Dwarf_Die base, Dwarf_Die *childp)
674 {
675 Dwarf_Error derr;
676 int ret;
677
678 *childp = NULL;
679 ret = dwarf_child(base, childp, &derr);
680 if (ret == DW_DLV_OK || ret == DW_DLV_NO_ENTRY)
681 return (0);
682
683 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
684 "failed to child from die: %s\n",
685 dwarf_errmsg(derr));
686 return (ECTF_CONVBKERR);
687 }
688
689 /*
690 * Compilers disagree on what to do to determine if something has global
691 * visiblity. Traditionally gcc has used DW_AT_external to indicate this while
692 * Studio has used DW_AT_visibility. We check DW_AT_visibility first and then
693 * fall back to DW_AT_external. Lack of DW_AT_external implies that it is not.
694 */
695 static int
696 ctf_dwarf_isglobal(ctf_cu_t *cup, Dwarf_Die die, boolean_t *igp)
697 {
698 int ret;
699 Dwarf_Signed vis;
700 Dwarf_Bool ext;
701
702 if ((ret = ctf_dwarf_signed(cup, die, DW_AT_visibility, &vis)) == 0) {
703 *igp = vis == DW_VIS_exported;
704 return (0);
705 } else if (ret != ENOENT) {
706 return (ret);
707 }
708
709 if ((ret = ctf_dwarf_boolean(cup, die, DW_AT_external, &ext)) != 0) {
710 if (ret == ENOENT) {
711 *igp = B_FALSE;
712 return (0);
713 }
714 return (ret);
715 }
716 *igp = ext != 0 ? B_TRUE : B_FALSE;
717 return (0);
718 }
719
720 static int
721 ctf_dwarf_die_elfenc(Elf *elf, ctf_cu_t *cup, char *errbuf, size_t errlen)
722 {
723 GElf_Ehdr ehdr;
724
725 if (gelf_getehdr(elf, &ehdr) == NULL) {
726 (void) snprintf(errbuf, errlen,
727 "failed to get ELF header: %s\n",
728 elf_errmsg(elf_errno()));
729 return (ECTF_CONVBKERR);
730 }
731
732 cup->cu_mach = ehdr.e_machine;
733
734 if (ehdr.e_ident[EI_CLASS] == ELFCLASS32) {
735 cup->cu_ptrsz = 4;
736 VERIFY(ctf_setmodel(cup->cu_ctfp, CTF_MODEL_ILP32) == 0);
737 } else if (ehdr.e_ident[EI_CLASS] == ELFCLASS64) {
738 cup->cu_ptrsz = 8;
739 VERIFY(ctf_setmodel(cup->cu_ctfp, CTF_MODEL_LP64) == 0);
740 } else {
741 (void) snprintf(errbuf, errlen,
742 "unknown ELF class %d", ehdr.e_ident[EI_CLASS]);
743 return (ECTF_CONVBKERR);
744 }
745
746 if (ehdr.e_ident[EI_DATA] == ELFDATA2LSB) {
747 cup->cu_bigend = B_FALSE;
748 } else if (ehdr.e_ident[EI_DATA] == ELFDATA2MSB) {
749 cup->cu_bigend = B_TRUE;
750 } else {
751 (void) snprintf(errbuf, errlen,
752 "unknown ELF data encoding: %hhu", ehdr.e_ident[EI_DATA]);
753 return (ECTF_CONVBKERR);
754 }
755
756 return (0);
757 }
758
759 typedef struct ctf_dwarf_fpent {
760 size_t cdfe_size;
761 uint_t cdfe_enc[3];
762 } ctf_dwarf_fpent_t;
763
764 typedef struct ctf_dwarf_fpmap {
765 uint_t cdf_mach;
766 ctf_dwarf_fpent_t cdf_ents[4];
767 } ctf_dwarf_fpmap_t;
768
769 static const ctf_dwarf_fpmap_t ctf_dwarf_fpmaps[] = {
770 { EM_SPARC, {
771 { 4, { CTF_FP_SINGLE, CTF_FP_CPLX, CTF_FP_IMAGRY } },
772 { 8, { CTF_FP_DOUBLE, CTF_FP_DCPLX, CTF_FP_DIMAGRY } },
773 { 16, { CTF_FP_LDOUBLE, CTF_FP_LDCPLX, CTF_FP_LDIMAGRY } },
774 { 0, { 0 } }
775 } },
776 { EM_SPARC32PLUS, {
777 { 4, { CTF_FP_SINGLE, CTF_FP_CPLX, CTF_FP_IMAGRY } },
778 { 8, { CTF_FP_DOUBLE, CTF_FP_DCPLX, CTF_FP_DIMAGRY } },
779 { 16, { CTF_FP_LDOUBLE, CTF_FP_LDCPLX, CTF_FP_LDIMAGRY } },
780 { 0, { 0 } }
781 } },
782 { EM_SPARCV9, {
783 { 4, { CTF_FP_SINGLE, CTF_FP_CPLX, CTF_FP_IMAGRY } },
784 { 8, { CTF_FP_DOUBLE, CTF_FP_DCPLX, CTF_FP_DIMAGRY } },
785 { 16, { CTF_FP_LDOUBLE, CTF_FP_LDCPLX, CTF_FP_LDIMAGRY } },
786 { 0, { 0 } }
787 } },
788 { EM_386, {
789 { 4, { CTF_FP_SINGLE, CTF_FP_CPLX, CTF_FP_IMAGRY } },
790 { 8, { CTF_FP_DOUBLE, CTF_FP_DCPLX, CTF_FP_DIMAGRY } },
791 { 12, { CTF_FP_LDOUBLE, CTF_FP_LDCPLX, CTF_FP_LDIMAGRY } },
792 { 0, { 0 } }
793 } },
794 { EM_X86_64, {
795 { 4, { CTF_FP_SINGLE, CTF_FP_CPLX, CTF_FP_IMAGRY } },
796 { 8, { CTF_FP_DOUBLE, CTF_FP_DCPLX, CTF_FP_DIMAGRY } },
797 { 16, { CTF_FP_LDOUBLE, CTF_FP_LDCPLX, CTF_FP_LDIMAGRY } },
798 { 0, { 0 } }
799 } },
800 { EM_NONE }
801 };
802
803 static int
804 ctf_dwarf_float_base(ctf_cu_t *cup, Dwarf_Signed type, ctf_encoding_t *enc)
805 {
806 const ctf_dwarf_fpmap_t *map = &ctf_dwarf_fpmaps[0];
807 const ctf_dwarf_fpent_t *ent;
808 uint_t col = 0, mult = 1;
809
810 for (map = &ctf_dwarf_fpmaps[0]; map->cdf_mach != EM_NONE; map++) {
811 if (map->cdf_mach == cup->cu_mach)
812 break;
813 }
814
815 if (map->cdf_mach == EM_NONE) {
816 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
817 "Unsupported machine type: %d\n", cup->cu_mach);
818 return (ENOTSUP);
819 }
820
821 if (type == DW_ATE_complex_float) {
822 mult = 2;
823 col = 1;
824 } else if (type == DW_ATE_imaginary_float ||
825 type == DW_ATE_SUN_imaginary_float) {
826 col = 2;
827 }
828
829 ent = &map->cdf_ents[0];
830 for (ent = &map->cdf_ents[0]; ent->cdfe_size != 0; ent++) {
831 if (ent->cdfe_size * mult * 8 == enc->cte_bits) {
832 enc->cte_format = ent->cdfe_enc[col];
833 return (0);
834 }
835 }
836
837 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
838 "failed to find valid fp mapping for encoding %d, size %d bits\n",
839 type, enc->cte_bits);
840 return (EINVAL);
841 }
842
843 static int
844 ctf_dwarf_dwarf_base(ctf_cu_t *cup, Dwarf_Die die, int *kindp,
845 ctf_encoding_t *enc)
846 {
847 int ret;
848 Dwarf_Signed type;
849
850 if ((ret = ctf_dwarf_signed(cup, die, DW_AT_encoding, &type)) != 0)
851 return (ret);
852
853 switch (type) {
854 case DW_ATE_unsigned:
855 case DW_ATE_address:
856 *kindp = CTF_K_INTEGER;
857 enc->cte_format = 0;
858 break;
859 case DW_ATE_unsigned_char:
860 *kindp = CTF_K_INTEGER;
861 enc->cte_format = CTF_INT_CHAR;
862 break;
863 case DW_ATE_signed:
864 *kindp = CTF_K_INTEGER;
865 enc->cte_format = CTF_INT_SIGNED;
866 break;
867 case DW_ATE_signed_char:
868 *kindp = CTF_K_INTEGER;
869 enc->cte_format = CTF_INT_SIGNED | CTF_INT_CHAR;
870 break;
871 case DW_ATE_boolean:
872 *kindp = CTF_K_INTEGER;
873 enc->cte_format = CTF_INT_SIGNED | CTF_INT_BOOL;
874 break;
875 case DW_ATE_float:
876 case DW_ATE_complex_float:
877 case DW_ATE_imaginary_float:
878 case DW_ATE_SUN_imaginary_float:
879 case DW_ATE_SUN_interval_float:
880 *kindp = CTF_K_FLOAT;
881 if ((ret = ctf_dwarf_float_base(cup, type, enc)) != 0)
882 return (ret);
883 break;
884 default:
885 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
886 "encountered unknown DWARF encoding: %d", type);
887 return (ECTF_CONVBKERR);
888 }
889
890 return (0);
891 }
892
893 /*
894 * Different compilers (at least GCC and Studio) use different names for types.
895 * This parses the types and attempts to unify them. If this fails, we just fall
896 * back to using the DWARF itself.
897 */
898 static int
899 ctf_dwarf_parse_base(const char *name, int *kindp, ctf_encoding_t *enc,
900 char **newnamep)
901 {
902 char buf[256];
903 char *base, *c, *last;
904 int nlong = 0, nshort = 0, nchar = 0, nint = 0;
905 int sign = 1;
906
907 if (strlen(name) + 1 > sizeof (buf))
908 return (EINVAL);
909
910 (void) strlcpy(buf, name, sizeof (buf));
911 for (c = strtok_r(buf, " ", &last); c != NULL;
912 c = strtok_r(NULL, " ", &last)) {
913 if (strcmp(c, "signed") == 0) {
914 sign = 1;
915 } else if (strcmp(c, "unsigned") == 0) {
916 sign = 0;
917 } else if (strcmp(c, "long") == 0) {
918 nlong++;
919 } else if (strcmp(c, "char") == 0) {
920 nchar++;
921 } else if (strcmp(c, "short") == 0) {
922 nshort++;
923 } else if (strcmp(c, "int") == 0) {
924 nint++;
925 } else {
926 /*
927 * If we don't recognize any of the tokens, we'll tell
928 * the caller to fall back to the dwarf-provided
929 * encoding information.
930 */
931 return (EINVAL);
932 }
933 }
934
935 if (nchar > 1 || nshort > 1 || nint > 1 || nlong > 2)
936 return (EINVAL);
937
938 if (nchar > 0) {
939 if (nlong > 0 || nshort > 0 || nint > 0)
940 return (EINVAL);
941 base = "char";
942 } else if (nshort > 0) {
943 if (nlong > 0)
944 return (EINVAL);
945 base = "short";
946 } else if (nlong > 0) {
947 base = "long";
948 } else {
949 base = "int";
950 }
951
952 if (nchar > 0)
953 enc->cte_format = CTF_INT_CHAR;
954 else
955 enc->cte_format = 0;
956
957 if (sign > 0)
958 enc->cte_format |= CTF_INT_SIGNED;
959
960 (void) snprintf(buf, sizeof (buf), "%s%s%s",
961 (sign ? "" : "unsigned "),
962 (nlong > 1 ? "long " : ""),
963 base);
964
965 *newnamep = ctf_strdup(buf);
966 if (*newnamep == NULL)
967 return (ENOMEM);
968 *kindp = CTF_K_INTEGER;
969 return (0);
970 }
971
972 static int
973 ctf_dwarf_create_base(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp, int isroot,
974 Dwarf_Off off)
975 {
976 int ret;
977 char *name, *nname;
978 Dwarf_Unsigned sz;
979 int kind;
980 ctf_encoding_t enc;
981 ctf_id_t id;
982
983 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0)
984 return (ret);
985 if ((ret = ctf_dwarf_unsigned(cup, die, DW_AT_byte_size, &sz)) != 0) {
986 goto out;
987 }
988 ctf_dprintf("Creating base type %s from off %llu, size: %d\n", name,
989 off, sz);
990
991 bzero(&enc, sizeof (ctf_encoding_t));
992 enc.cte_bits = sz * 8;
993 if ((ret = ctf_dwarf_parse_base(name, &kind, &enc, &nname)) == 0) {
994 ctf_free(name, strlen(name) + 1);
995 name = nname;
996 } else {
997 if (ret != EINVAL)
998 return (ret);
999 ctf_dprintf("falling back to dwarf for base type %s\n", name);
1000 if ((ret = ctf_dwarf_dwarf_base(cup, die, &kind, &enc)) != 0)
1001 return (ret);
1002 }
1003
1004 id = ctf_add_encoded(cup->cu_ctfp, isroot, name, &enc, kind);
1005 if (id == CTF_ERR) {
1006 ret = ctf_errno(cup->cu_ctfp);
1007 } else {
1008 *idp = id;
1009 ret = ctf_dwmap_add(cup, id, die, B_FALSE);
1010 }
1011 out:
1012 ctf_free(name, strlen(name) + 1);
1013 return (ret);
1014 }
1015
1016 /*
1017 * Getting a member's offset is a surprisingly intricate dance. It works as
1018 * follows:
1019 *
1020 * 1) If we're in DWARFv4, then we either have a DW_AT_data_bit_offset or we
1021 * have a DW_AT_data_member_location. We won't have both. Thus we check first
1022 * for DW_AT_data_bit_offset, and if it exists, we're set.
1023 *
1024 * Next, if we have a bitfield and we don't have a DW_AT_data_bit_offset, then
1025 * we have to grab the data location and use the following dance:
1026 *
1027 * 2) Gather the set of DW_AT_byte_size, DW_AT_bit_offset, and DW_AT_bit_size.
1028 * Of course, the DW_AT_byte_size may be omitted, even though it isn't always.
1029 * When it's been omitted, we then have to say that the size is that of the
1030 * underlying type, which forces that to be after a ctf_update(). Here, we have
1031 * to do different things based on whether or not we're using big endian or
1032 * little endian to obtain the proper offset.
1033 */
1034 static int
1035 ctf_dwarf_member_offset(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t mid,
1036 ulong_t *offp)
1037 {
1038 int ret;
1039 Dwarf_Unsigned loc, bitsz, bytesz;
1040 Dwarf_Signed bitoff;
1041 size_t off;
1042 ssize_t tsz;
1043
1044 if ((ret = ctf_dwarf_unsigned(cup, die, DW_AT_data_bit_offset,
1045 &loc)) == 0) {
1046 *offp = loc;
1047 return (0);
1048 } else if (ret != ENOENT) {
1049 return (ret);
1050 }
1051
1052 if ((ret = ctf_dwarf_member_location(cup, die, &loc)) != 0)
1053 return (ret);
1054 off = loc * 8;
1055
1056 if ((ret = ctf_dwarf_signed(cup, die, DW_AT_bit_offset,
1057 &bitoff)) != 0) {
1058 if (ret != ENOENT)
1059 return (ret);
1060 *offp = off;
1061 return (0);
1062 }
1063
1064 /* At this point we have to have DW_AT_bit_size */
1065 if ((ret = ctf_dwarf_unsigned(cup, die, DW_AT_bit_size, &bitsz)) != 0)
1066 return (ret);
1067
1068 if ((ret = ctf_dwarf_unsigned(cup, die, DW_AT_byte_size,
1069 &bytesz)) != 0) {
1070 if (ret != ENOENT)
1071 return (ret);
1072 if ((tsz = ctf_type_size(cup->cu_ctfp, mid)) == CTF_ERR) {
1073 int e = ctf_errno(cup->cu_ctfp);
1074 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1075 "failed to get type size: %s", ctf_errmsg(e));
1076 return (ECTF_CONVBKERR);
1077 }
1078 } else {
1079 tsz = bytesz;
1080 }
1081 tsz *= 8;
1082 if (cup->cu_bigend == B_TRUE) {
1083 *offp = off + bitoff;
1084 } else {
1085 *offp = off + tsz - bitoff - bitsz;
1086 }
1087
1088 return (0);
1089 }
1090
1091 /*
1092 * We need to determine if the member in question is a bitfield. If it is, then
1093 * we need to go through and create a new type that's based on the actual base
1094 * type, but has a different size. We also rename the type as a result to help
1095 * deal with future collisions.
1096 *
1097 * Here we need to look and see if we have a DW_AT_bit_size value. If we have a
1098 * bit size member and it does not equal the byte size member, then we need to
1099 * create a bitfield type based on this.
1100 *
1101 * Note: When we support DWARFv4, there may be a chance that we need to also
1102 * search for the DW_AT_byte_size if we don't have a DW_AT_bit_size member.
1103 */
1104 static int
1105 ctf_dwarf_member_bitfield(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp)
1106 {
1107 int ret;
1108 Dwarf_Unsigned bitsz;
1109 ctf_encoding_t e;
1110 ctf_dwbitf_t *cdb;
1111 ctf_dtdef_t *dtd;
1112 ctf_id_t base = *idp;
1113 int kind;
1114
1115 if ((ret = ctf_dwarf_unsigned(cup, die, DW_AT_bit_size, &bitsz)) != 0) {
1116 if (ret == ENOENT)
1117 return (0);
1118 return (ret);
1119 }
1120
1121 ctf_dprintf("Trying to deal with bitfields on %d:%d\n", base, bitsz);
1122 /*
1123 * Given that we now have a bitsize, time to go do something about it.
1124 * We're going to create a new type based on the current one, but first
1125 * we need to find the base type. This means we need to traverse any
1126 * typedef's, consts, and volatiles until we get to what should be
1127 * something of type integer or enumeration.
1128 */
1129 VERIFY(bitsz < UINT32_MAX);
1130 dtd = ctf_dtd_lookup(cup->cu_ctfp, base);
1131 VERIFY(dtd != NULL);
1132 kind = CTF_INFO_KIND(dtd->dtd_data.ctt_info);
1133 while (kind == CTF_K_TYPEDEF || kind == CTF_K_CONST ||
1134 kind == CTF_K_VOLATILE) {
1135 dtd = ctf_dtd_lookup(cup->cu_ctfp, dtd->dtd_data.ctt_type);
1136 VERIFY(dtd != NULL);
1137 kind = CTF_INFO_KIND(dtd->dtd_data.ctt_info);
1138 }
1139 ctf_dprintf("got kind %d\n", kind);
1140 VERIFY(kind == CTF_K_INTEGER || kind == CTF_K_ENUM);
1141
1142 /*
1143 * As surprising as it may be, it is strictly possible to create a
1144 * bitfield that is based on an enum. Of course, the C standard leaves
1145 * enums sizing as an ABI concern more or less. To that effect, today on
1146 * all illumos platforms the size of an enum is generally that of an
1147 * int as our supported data models and ABIs all agree on that. So what
1148 * we'll do is fake up a CTF encoding here to use. In this case, we'll
1149 * treat it as an unsigned value of whatever size the underlying enum
1150 * currently has (which is in the ctt_size member of its dynamic type
1151 * data).
1152 */
1153 if (kind == CTF_K_INTEGER) {
1154 e = dtd->dtd_u.dtu_enc;
1155 } else {
1156 bzero(&e, sizeof (ctf_encoding_t));
1157 e.cte_bits = dtd->dtd_data.ctt_size * NBBY;
1158 }
1159
1160 for (cdb = ctf_list_next(&cup->cu_bitfields); cdb != NULL;
1161 cdb = ctf_list_next(cdb)) {
1162 if (cdb->cdb_base == base && cdb->cdb_nbits == bitsz)
1163 break;
1164 }
1165
1166 /*
1167 * Create a new type if none exists. We name all types in a way that is
1168 * guaranteed not to conflict with the corresponding C type. We do this
1169 * by using the ':' operator.
1170 */
1171 if (cdb == NULL) {
1172 size_t namesz;
1173 char *name;
1174
1175 e.cte_bits = bitsz;
1176 namesz = snprintf(NULL, 0, "%s:%d", dtd->dtd_name,
1177 (uint32_t)bitsz);
1178 name = ctf_alloc(namesz + 1);
1179 if (name == NULL)
1180 return (ENOMEM);
1181 cdb = ctf_alloc(sizeof (ctf_dwbitf_t));
1182 if (cdb == NULL) {
1183 ctf_free(name, namesz + 1);
1184 return (ENOMEM);
1185 }
1186 (void) snprintf(name, namesz + 1, "%s:%d", dtd->dtd_name,
1187 (uint32_t)bitsz);
1188
1189 cdb->cdb_base = base;
1190 cdb->cdb_nbits = bitsz;
1191 cdb->cdb_id = ctf_add_integer(cup->cu_ctfp, CTF_ADD_NONROOT,
1192 name, &e);
1193 if (cdb->cdb_id == CTF_ERR) {
1194 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1195 "failed to get add bitfield type %s: %s", name,
1196 ctf_errmsg(ctf_errno(cup->cu_ctfp)));
1197 ctf_free(name, namesz + 1);
1198 ctf_free(cdb, sizeof (ctf_dwbitf_t));
1199 return (ECTF_CONVBKERR);
1200 }
1201 ctf_free(name, namesz + 1);
1202 ctf_list_append(&cup->cu_bitfields, cdb);
1203 }
1204
1205 *idp = cdb->cdb_id;
1206
1207 return (0);
1208 }
1209
1210 static int
1211 ctf_dwarf_fixup_sou(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t base, boolean_t add)
1212 {
1213 int ret, kind;
1214 Dwarf_Die child, memb;
1215 Dwarf_Unsigned size;
1216
1217 kind = ctf_type_kind(cup->cu_ctfp, base);
1218 VERIFY(kind != CTF_ERR);
1219 VERIFY(kind == CTF_K_STRUCT || kind == CTF_K_UNION);
1220
1221 /*
1222 * Members are in children. However, gcc also allows empty ones.
1223 */
1224 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1225 return (ret);
1226 if (child == NULL)
1227 return (0);
1228
1229 memb = child;
1230 while (memb != NULL) {
1231 Dwarf_Die sib, tdie;
1232 Dwarf_Half tag;
1233 ctf_id_t mid;
1234 char *mname;
1235 ulong_t memboff = 0;
1236
1237 if ((ret = ctf_dwarf_tag(cup, memb, &tag)) != 0)
1238 return (ret);
1239
1240 if (tag != DW_TAG_member)
1241 continue;
1242
1243 if ((ret = ctf_dwarf_refdie(cup, memb, DW_AT_type, &tdie)) != 0)
1244 return (ret);
1245
1246 if ((ret = ctf_dwarf_convert_type(cup, tdie, &mid,
1247 CTF_ADD_NONROOT)) != 0)
1248 return (ret);
1249 ctf_dprintf("Got back type id: %d\n", mid);
1250
1251 /*
1252 * If we're not adding a member, just go ahead and return.
1253 */
1254 if (add == B_FALSE) {
1255 if ((ret = ctf_dwarf_member_bitfield(cup, memb,
1256 &mid)) != 0)
1257 return (ret);
1258 goto next;
1259 }
1260
1261 if ((ret = ctf_dwarf_string(cup, memb, DW_AT_name,
1262 &mname)) != 0 && ret != ENOENT)
1263 return (ret);
1264 if (ret == ENOENT)
1265 mname = NULL;
1266
1267 if (kind == CTF_K_UNION) {
1268 memboff = 0;
1269 } else if ((ret = ctf_dwarf_member_offset(cup, memb, mid,
1270 &memboff)) != 0) {
1271 if (mname != NULL)
1272 ctf_free(mname, strlen(mname) + 1);
1273 return (ret);
1274 }
1275
1276 if ((ret = ctf_dwarf_member_bitfield(cup, memb, &mid)) != 0)
1277 return (ret);
1278
1279 ret = ctf_add_member(cup->cu_ctfp, base, mname, mid, memboff);
1280 if (ret == CTF_ERR) {
1281 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1282 "failed to add member %s: %s",
1283 mname, ctf_errmsg(ctf_errno(cup->cu_ctfp)));
1284 if (mname != NULL)
1285 ctf_free(mname, strlen(mname) + 1);
1286 return (ECTF_CONVBKERR);
1287 }
1288
1289 if (mname != NULL)
1290 ctf_free(mname, strlen(mname) + 1);
1291
1292 next:
1293 if ((ret = ctf_dwarf_sib(cup, memb, &sib)) != 0)
1294 return (ret);
1295 memb = sib;
1296 }
1297
1298 /*
1299 * If we're not adding members, then we don't know the final size of the
1300 * structure, so end here.
1301 */
1302 if (add == B_FALSE)
1303 return (0);
1304
1305 /* Finally set the size of the structure to the actual byte size */
1306 if ((ret = ctf_dwarf_unsigned(cup, die, DW_AT_byte_size, &size)) != 0)
1307 return (ret);
1308 if ((ctf_set_size(cup->cu_ctfp, base, size)) == CTF_ERR) {
1309 int e = ctf_errno(cup->cu_ctfp);
1310 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1311 "failed to set type size for %d to 0x%x: %s", base,
1312 (uint32_t)size, ctf_errmsg(e));
1313 return (ECTF_CONVBKERR);
1314 }
1315
1316 return (0);
1317 }
1318
1319 static int
1320 ctf_dwarf_create_sou(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp,
1321 int kind, int isroot)
1322 {
1323 int ret;
1324 char *name;
1325 ctf_id_t base;
1326 Dwarf_Die child;
1327 Dwarf_Bool decl;
1328
1329 /*
1330 * Deal with the terribly annoying case of anonymous structs and unions.
1331 * If they don't have a name, set the name to the empty string.
1332 */
1333 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0 &&
1334 ret != ENOENT)
1335 return (ret);
1336 if (ret == ENOENT)
1337 name = NULL;
1338
1339 /*
1340 * We need to check if we just have a declaration here. If we do, then
1341 * instead of creating an actual structure or union, we're just going to
1342 * go ahead and create a forward. During a dedup or merge, the forward
1343 * will be replaced with the real thing.
1344 */
1345 if ((ret = ctf_dwarf_boolean(cup, die, DW_AT_declaration,
1346 &decl)) != 0) {
1347 if (ret != ENOENT)
1348 return (ret);
1349 decl = 0;
1350 }
1351
1352 if (decl != 0) {
1353 base = ctf_add_forward(cup->cu_ctfp, isroot, name, kind);
1354 } else if (kind == CTF_K_STRUCT) {
1355 base = ctf_add_struct(cup->cu_ctfp, isroot, name);
1356 } else {
1357 base = ctf_add_union(cup->cu_ctfp, isroot, name);
1358 }
1359 ctf_dprintf("added sou %s (%d) (%d)\n", name, kind, base);
1360 if (name != NULL)
1361 ctf_free(name, strlen(name) + 1);
1362 if (base == CTF_ERR)
1363 return (ctf_errno(cup->cu_ctfp));
1364 *idp = base;
1365
1366 /*
1367 * If it's just a declaration, we're not going to mark it for fix up or
1368 * do anything else.
1369 */
1370 if (decl == B_TRUE)
1371 return (ctf_dwmap_add(cup, base, die, B_FALSE));
1372 if ((ret = ctf_dwmap_add(cup, base, die, B_TRUE)) != 0)
1373 return (ret);
1374
1375 /*
1376 * Members are in children. However, gcc also allows empty ones.
1377 */
1378 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1379 return (ret);
1380 if (child == NULL)
1381 return (0);
1382
1383 return (0);
1384 }
1385
1386 static int
1387 ctf_dwarf_create_array_range(ctf_cu_t *cup, Dwarf_Die range, ctf_id_t *idp,
1388 ctf_id_t base, int isroot)
1389 {
1390 int ret;
1391 Dwarf_Die sib;
1392 Dwarf_Unsigned val;
1393 Dwarf_Signed sval;
1394 ctf_arinfo_t ar;
1395
1396 ctf_dprintf("creating array range\n");
1397
1398 if ((ret = ctf_dwarf_sib(cup, range, &sib)) != 0)
1399 return (ret);
1400 if (sib != NULL) {
1401 ctf_id_t id;
1402 if ((ret = ctf_dwarf_create_array_range(cup, sib, &id,
1403 base, CTF_ADD_NONROOT)) != 0)
1404 return (ret);
1405 ar.ctr_contents = id;
1406 } else {
1407 ar.ctr_contents = base;
1408 }
1409
1410 if ((ar.ctr_index = ctf_dwarf_long(cup)) == CTF_ERR)
1411 return (ctf_errno(cup->cu_ctfp));
1412
1413 /*
1414 * Array bounds can be signed or unsigned, but there are several kinds
1415 * of signless forms (data1, data2, etc) that take their sign from the
1416 * routine that is trying to interpret them. That is, data1 can be
1417 * either signed or unsigned, depending on whether you use the signed or
1418 * unsigned accessor function. GCC will use the signless forms to store
1419 * unsigned values which have their high bit set, so we need to try to
1420 * read them first as unsigned to get positive values. We could also
1421 * try signed first, falling back to unsigned if we got a negative
1422 * value.
1423 */
1424 if ((ret = ctf_dwarf_unsigned(cup, range, DW_AT_upper_bound,
1425 &val)) == 0) {
1426 ar.ctr_nelems = val + 1;
1427 } else if (ret != ENOENT) {
1428 return (ret);
1429 } else if ((ret = ctf_dwarf_signed(cup, range, DW_AT_upper_bound,
1430 &sval)) == 0) {
1431 ar.ctr_nelems = sval + 1;
1432 } else if (ret != ENOENT) {
1433 return (ret);
1434 } else {
1435 ar.ctr_nelems = 0;
1436 }
1437
1438 if ((*idp = ctf_add_array(cup->cu_ctfp, isroot, &ar)) == CTF_ERR)
1439 return (ctf_errno(cup->cu_ctfp));
1440
1441 return (0);
1442 }
1443
1444 /*
1445 * Try and create an array type. First, the kind of the array is specified in
1446 * the DW_AT_type entry. Next, the number of entries is stored in a more
1447 * complicated form, we should have a child that has the DW_TAG_subrange type.
1448 */
1449 static int
1450 ctf_dwarf_create_array(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp, int isroot)
1451 {
1452 int ret;
1453 Dwarf_Die tdie, rdie;
1454 ctf_id_t tid;
1455 Dwarf_Half rtag;
1456
1457 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) != 0)
1458 return (ret);
1459 if ((ret = ctf_dwarf_convert_type(cup, tdie, &tid,
1460 CTF_ADD_NONROOT)) != 0)
1461 return (ret);
1462
1463 if ((ret = ctf_dwarf_child(cup, die, &rdie)) != 0)
1464 return (ret);
1465 if ((ret = ctf_dwarf_tag(cup, rdie, &rtag)) != 0)
1466 return (ret);
1467 if (rtag != DW_TAG_subrange_type) {
1468 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1469 "encountered array without DW_TAG_subrange_type child\n");
1470 return (ECTF_CONVBKERR);
1471 }
1472
1473 /*
1474 * The compiler may opt to describe a multi-dimensional array as one
1475 * giant array or it may opt to instead encode it as a series of
1476 * subranges. If it's the latter, then for each subrange we introduce a
1477 * type. We can always use the base type.
1478 */
1479 if ((ret = ctf_dwarf_create_array_range(cup, rdie, idp, tid,
1480 isroot)) != 0)
1481 return (ret);
1482 ctf_dprintf("Got back id %d\n", *idp);
1483 return (ctf_dwmap_add(cup, *idp, die, B_FALSE));
1484 }
1485
1486 /*
1487 * Given "const int const_array3[11]", GCC7 at least will create a DIE tree of
1488 * DW_TAG_const_type:DW_TAG_array_type:DW_Tag_const_type:<member_type>.
1489 *
1490 * Given C's syntax, this renders out as "const const int const_array3[11]". To
1491 * get closer to round-tripping (and make the unit tests work), we'll peek for
1492 * this case, and avoid adding the extraneous qualifier if we see that the
1493 * underlying array referent already has the same qualifier.
1494 *
1495 * This is unfortunately less trivial than it could be: this issue applies to
1496 * qualifier sets like "const volatile", as well as multi-dimensional arrays, so
1497 * we need to descend down those.
1498 *
1499 * Returns CTF_ERR on error, or a boolean value otherwise.
1500 */
1501 static int
1502 needed_array_qualifier(ctf_cu_t *cup, int kind, ctf_id_t ref_id)
1503 {
1504 const ctf_type_t *t;
1505 ctf_arinfo_t arinfo;
1506 int akind;
1507
1508 if (kind != CTF_K_CONST && kind != CTF_K_VOLATILE &&
1509 kind != CTF_K_RESTRICT)
1510 return (1);
1511
1512 if ((t = ctf_dyn_lookup_by_id(cup->cu_ctfp, ref_id)) == NULL)
1513 return (CTF_ERR);
1514
1515 if (LCTF_INFO_KIND(cup->cu_ctfp, t->ctt_info) != CTF_K_ARRAY)
1516 return (1);
1517
1518 if (ctf_dyn_array_info(cup->cu_ctfp, ref_id, &arinfo) != 0)
1519 return (CTF_ERR);
1520
1521 ctf_id_t id = arinfo.ctr_contents;
1522
1523 for (;;) {
1524 if ((t = ctf_dyn_lookup_by_id(cup->cu_ctfp, id)) == NULL)
1525 return (CTF_ERR);
1526
1527 akind = LCTF_INFO_KIND(cup->cu_ctfp, t->ctt_info);
1528
1529 if (akind == kind)
1530 break;
1531
1532 if (akind == CTF_K_ARRAY) {
1533 if (ctf_dyn_array_info(cup->cu_ctfp,
1534 id, &arinfo) != 0)
1535 return (CTF_ERR);
1536 id = arinfo.ctr_contents;
1537 continue;
1538 }
1539
1540 if (akind != CTF_K_CONST && akind != CTF_K_VOLATILE &&
1541 akind != CTF_K_RESTRICT)
1542 break;
1543
1544 id = t->ctt_type;
1545 }
1546
1547 if (kind == akind) {
1548 ctf_dprintf("ignoring extraneous %s qualifier for array %d\n",
1549 ctf_kind_name(cup->cu_ctfp, kind), ref_id);
1550 }
1551
1552 return (kind != akind);
1553 }
1554
1555 static int
1556 ctf_dwarf_create_reference(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp,
1557 int kind, int isroot)
1558 {
1559 int ret;
1560 ctf_id_t id;
1561 Dwarf_Die tdie;
1562 char *name;
1563 size_t namelen;
1564
1565 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0 &&
1566 ret != ENOENT)
1567 return (ret);
1568 if (ret == ENOENT) {
1569 name = NULL;
1570 namelen = 0;
1571 } else {
1572 namelen = strlen(name);
1573 }
1574
1575 ctf_dprintf("reference kind %d %s\n", kind, name != NULL ? name : "<>");
1576
1577 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) != 0) {
1578 if (ret != ENOENT) {
1579 ctf_free(name, namelen);
1580 return (ret);
1581 }
1582 if ((id = ctf_dwarf_void(cup)) == CTF_ERR) {
1583 ctf_free(name, namelen);
1584 return (ctf_errno(cup->cu_ctfp));
1585 }
1586 } else {
1587 if ((ret = ctf_dwarf_convert_type(cup, tdie, &id,
1588 CTF_ADD_NONROOT)) != 0) {
1589 ctf_free(name, namelen);
1590 return (ret);
1591 }
1592 }
1593
1594 if ((ret = needed_array_qualifier(cup, kind, id)) <= 0) {
1595 if (ret != 0) {
1596 ret = (ctf_errno(cup->cu_ctfp));
1597 } else {
1598 *idp = id;
1599 }
1600
1601 ctf_free(name, namelen);
1602 return (ret);
1603 }
1604
1605 if ((*idp = ctf_add_reftype(cup->cu_ctfp, isroot, name, id, kind)) ==
1606 CTF_ERR) {
1607 ctf_free(name, namelen);
1608 return (ctf_errno(cup->cu_ctfp));
1609 }
1610
1611 ctf_free(name, namelen);
1612 return (ctf_dwmap_add(cup, *idp, die, B_FALSE));
1613 }
1614
1615 /*
1616 * Get the size of the type of a particular die. Note that this is a simple
1617 * version that doesn't attempt to traverse further than expecting a single
1618 * sized type reference (so no qualifiers etc.). Nor does it attempt to do as
1619 * much as ctf_type_size() - which we cannot use here as that doesn't look up
1620 * dynamic types, and we don't yet want to do a ctf_update().
1621 */
1622 static int
1623 ctf_dwarf_get_type_size(ctf_cu_t *cup, Dwarf_Die die, size_t *sizep)
1624 {
1625 const ctf_type_t *t;
1626 Dwarf_Die tdie;
1627 ctf_id_t tid;
1628 int ret;
1629
1630 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) != 0)
1631 return (ret);
1632
1633 if ((ret = ctf_dwarf_convert_type(cup, tdie, &tid,
1634 CTF_ADD_NONROOT)) != 0)
1635 return (ret);
1636
1637 if ((t = ctf_dyn_lookup_by_id(cup->cu_ctfp, tid)) == NULL)
1638 return (ENOENT);
1639
1640 *sizep = ctf_get_ctt_size(cup->cu_ctfp, t, NULL, NULL);
1641 return (0);
1642 }
1643
1644 static int
1645 ctf_dwarf_create_enum(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp, int isroot)
1646 {
1647 size_t size = 0;
1648 Dwarf_Die child;
1649 ctf_id_t id;
1650 char *name;
1651 int ret;
1652
1653 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0 &&
1654 ret != ENOENT)
1655 return (ret);
1656 if (ret == ENOENT)
1657 name = NULL;
1658
1659 (void) ctf_dwarf_get_type_size(cup, die, &size);
1660
1661 id = ctf_add_enum(cup->cu_ctfp, isroot, name, size);
1662 ctf_dprintf("added enum %s (%d)\n", name, id);
1663 if (name != NULL)
1664 ctf_free(name, strlen(name) + 1);
1665 if (id == CTF_ERR)
1666 return (ctf_errno(cup->cu_ctfp));
1667 *idp = id;
1668 if ((ret = ctf_dwmap_add(cup, id, die, B_FALSE)) != 0)
1669 return (ret);
1670
1671 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0) {
1672 if (ret == ENOENT)
1673 ret = 0;
1674 return (ret);
1675 }
1676
1677 while (child != NULL) {
1678 Dwarf_Half tag;
1679 Dwarf_Signed sval;
1680 Dwarf_Unsigned uval;
1681 Dwarf_Die arg = child;
1682 int eval;
1683
1684 if ((ret = ctf_dwarf_sib(cup, arg, &child)) != 0)
1685 return (ret);
1686
1687 if ((ret = ctf_dwarf_tag(cup, arg, &tag)) != 0)
1688 return (ret);
1689
1690 if (tag != DW_TAG_enumerator) {
1691 if ((ret = ctf_dwarf_convert_type(cup, arg, NULL,
1692 CTF_ADD_NONROOT)) != 0)
1693 return (ret);
1694 continue;
1695 }
1696
1697 /*
1698 * DWARF v4 section 5.7 tells us we'll always have names.
1699 */
1700 if ((ret = ctf_dwarf_string(cup, arg, DW_AT_name, &name)) != 0)
1701 return (ret);
1702
1703 /*
1704 * We have to be careful here: newer GCCs generate DWARF where
1705 * an unsigned value will happily pass ctf_dwarf_signed().
1706 * Since negative values will fail ctf_dwarf_unsigned(), we try
1707 * that first to make sure we get the right value.
1708 */
1709 if ((ret = ctf_dwarf_unsigned(cup, arg, DW_AT_const_value,
1710 &uval)) == 0) {
1711 eval = (int)uval;
1712 } else if ((ret = ctf_dwarf_signed(cup, arg, DW_AT_const_value,
1713 &sval)) == 0) {
1714 eval = sval;
1715 }
1716
1717 if (ret != 0) {
1718 if (ret != ENOENT)
1719 return (ret);
1720
1721 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1722 "encountered enumeration without constant value\n");
1723 return (ECTF_CONVBKERR);
1724 }
1725
1726 ret = ctf_add_enumerator(cup->cu_ctfp, id, name, eval);
1727 if (ret == CTF_ERR) {
1728 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1729 "failed to add enumarator %s (%d) to %d\n",
1730 name, eval, id);
1731 ctf_free(name, strlen(name) + 1);
1732 return (ctf_errno(cup->cu_ctfp));
1733 }
1734 ctf_free(name, strlen(name) + 1);
1735 }
1736
1737 return (0);
1738 }
1739
1740 /*
1741 * For a function pointer, walk over and process all of its children, unless we
1742 * encounter one that's just a declaration. In which case, we error on it.
1743 */
1744 static int
1745 ctf_dwarf_create_fptr(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp, int isroot)
1746 {
1747 int ret;
1748 Dwarf_Bool b;
1749 ctf_funcinfo_t fi;
1750 Dwarf_Die retdie;
1751 ctf_id_t *argv = NULL;
1752
1753 bzero(&fi, sizeof (ctf_funcinfo_t));
1754
1755 if ((ret = ctf_dwarf_boolean(cup, die, DW_AT_declaration, &b)) != 0) {
1756 if (ret != ENOENT)
1757 return (ret);
1758 } else {
1759 if (b != 0)
1760 return (EPROTOTYPE);
1761 }
1762
1763 /*
1764 * Return type is in DW_AT_type, if none, it returns void.
1765 */
1766 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &retdie)) != 0) {
1767 if (ret != ENOENT)
1768 return (ret);
1769 if ((fi.ctc_return = ctf_dwarf_void(cup)) == CTF_ERR)
1770 return (ctf_errno(cup->cu_ctfp));
1771 } else {
1772 if ((ret = ctf_dwarf_convert_type(cup, retdie, &fi.ctc_return,
1773 CTF_ADD_NONROOT)) != 0)
1774 return (ret);
1775 }
1776
1777 if ((ret = ctf_dwarf_function_count(cup, die, &fi, B_TRUE)) != 0) {
1778 return (ret);
1779 }
1780
1781 if (fi.ctc_argc != 0) {
1782 argv = ctf_alloc(sizeof (ctf_id_t) * fi.ctc_argc);
1783 if (argv == NULL)
1784 return (ENOMEM);
1785
1786 if ((ret = ctf_dwarf_convert_fargs(cup, die, &fi, argv)) != 0) {
1787 ctf_free(argv, sizeof (ctf_id_t) * fi.ctc_argc);
1788 return (ret);
1789 }
1790 }
1791
1792 if ((*idp = ctf_add_funcptr(cup->cu_ctfp, isroot, &fi, argv)) ==
1793 CTF_ERR) {
1794 ctf_free(argv, sizeof (ctf_id_t) * fi.ctc_argc);
1795 return (ctf_errno(cup->cu_ctfp));
1796 }
1797
1798 ctf_free(argv, sizeof (ctf_id_t) * fi.ctc_argc);
1799 return (ctf_dwmap_add(cup, *idp, die, B_FALSE));
1800 }
1801
1802 static int
1803 ctf_dwarf_convert_type(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp,
1804 int isroot)
1805 {
1806 int ret;
1807 Dwarf_Off offset;
1808 Dwarf_Half tag;
1809 ctf_dwmap_t lookup, *map;
1810 ctf_id_t id;
1811
1812 if (idp == NULL)
1813 idp = &id;
1814
1815 if ((ret = ctf_dwarf_offset(cup, die, &offset)) != 0)
1816 return (ret);
1817
1818 if (offset > cup->cu_maxoff) {
1819 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1820 "die offset %llu beyond maximum for header %llu\n",
1821 offset, cup->cu_maxoff);
1822 return (ECTF_CONVBKERR);
1823 }
1824
1825 /*
1826 * If we've already added an entry for this offset, then we're done.
1827 */
1828 lookup.cdm_off = offset;
1829 if ((map = avl_find(&cup->cu_map, &lookup, NULL)) != NULL) {
1830 *idp = map->cdm_id;
1831 return (0);
1832 }
1833
1834 if ((ret = ctf_dwarf_tag(cup, die, &tag)) != 0)
1835 return (ret);
1836
1837 ret = ENOTSUP;
1838 switch (tag) {
1839 case DW_TAG_base_type:
1840 ctf_dprintf("base\n");
1841 ret = ctf_dwarf_create_base(cup, die, idp, isroot, offset);
1842 break;
1843 case DW_TAG_array_type:
1844 ctf_dprintf("array\n");
1845 ret = ctf_dwarf_create_array(cup, die, idp, isroot);
1846 break;
1847 case DW_TAG_enumeration_type:
1848 ctf_dprintf("enum\n");
1849 ret = ctf_dwarf_create_enum(cup, die, idp, isroot);
1850 break;
1851 case DW_TAG_pointer_type:
1852 ctf_dprintf("pointer\n");
1853 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_POINTER,
1854 isroot);
1855 break;
1856 case DW_TAG_structure_type:
1857 ctf_dprintf("struct\n");
1858 ret = ctf_dwarf_create_sou(cup, die, idp, CTF_K_STRUCT,
1859 isroot);
1860 break;
1861 case DW_TAG_subroutine_type:
1862 ctf_dprintf("fptr\n");
1863 ret = ctf_dwarf_create_fptr(cup, die, idp, isroot);
1864 break;
1865 case DW_TAG_typedef:
1866 ctf_dprintf("typedef\n");
1867 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_TYPEDEF,
1868 isroot);
1869 break;
1870 case DW_TAG_union_type:
1871 ctf_dprintf("union\n");
1872 ret = ctf_dwarf_create_sou(cup, die, idp, CTF_K_UNION,
1873 isroot);
1874 break;
1875 case DW_TAG_const_type:
1876 ctf_dprintf("const\n");
1877 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_CONST,
1878 isroot);
1879 break;
1880 case DW_TAG_volatile_type:
1881 ctf_dprintf("volatile\n");
1882 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_VOLATILE,
1883 isroot);
1884 break;
1885 case DW_TAG_restrict_type:
1886 ctf_dprintf("restrict\n");
1887 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_RESTRICT,
1888 isroot);
1889 break;
1890 default:
1891 ctf_dprintf("ignoring tag type %x\n", tag);
1892 *idp = CTF_ERR;
1893 ret = 0;
1894 break;
1895 }
1896 ctf_dprintf("ctf_dwarf_convert_type tag specific handler returned %d\n",
1897 ret);
1898
1899 return (ret);
1900 }
1901
1902 static int
1903 ctf_dwarf_walk_lexical(ctf_cu_t *cup, Dwarf_Die die)
1904 {
1905 int ret;
1906 Dwarf_Die child;
1907
1908 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1909 return (ret);
1910
1911 if (child == NULL)
1912 return (0);
1913
1914 return (ctf_dwarf_convert_die(cup, die));
1915 }
1916
1917 static int
1918 ctf_dwarf_function_count(ctf_cu_t *cup, Dwarf_Die die, ctf_funcinfo_t *fip,
1919 boolean_t fptr)
1920 {
1921 int ret;
1922 Dwarf_Die child, sib, arg;
1923
1924 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1925 return (ret);
1926
1927 arg = child;
1928 while (arg != NULL) {
1929 Dwarf_Half tag;
1930
1931 if ((ret = ctf_dwarf_tag(cup, arg, &tag)) != 0)
1932 return (ret);
1933
1934 /*
1935 * We have to check for a varargs type declaration. This will
1936 * happen in one of two ways. If we have a function pointer
1937 * type, then it'll be done with a tag of type
1938 * DW_TAG_unspecified_parameters. However, it only means we have
1939 * a variable number of arguments, if we have more than one
1940 * argument found so far. Otherwise, when we have a function
1941 * type, it instead uses a formal parameter whose name is '...'
1942 * to indicate a variable arguments member.
1943 *
1944 * Also, if we have a function pointer, then we have to expect
1945 * that we might not get a name at all.
1946 */
1947 if (tag == DW_TAG_formal_parameter && fptr == B_FALSE) {
1948 char *name;
1949 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name,
1950 &name)) != 0)
1951 return (ret);
1952 if (strcmp(name, DWARF_VARARGS_NAME) == 0)
1953 fip->ctc_flags |= CTF_FUNC_VARARG;
1954 else
1955 fip->ctc_argc++;
1956 ctf_free(name, strlen(name) + 1);
1957 } else if (tag == DW_TAG_formal_parameter) {
1958 fip->ctc_argc++;
1959 } else if (tag == DW_TAG_unspecified_parameters &&
1960 fip->ctc_argc > 0) {
1961 fip->ctc_flags |= CTF_FUNC_VARARG;
1962 }
1963 if ((ret = ctf_dwarf_sib(cup, arg, &sib)) != 0)
1964 return (ret);
1965 arg = sib;
1966 }
1967
1968 return (0);
1969 }
1970
1971 static int
1972 ctf_dwarf_convert_fargs(ctf_cu_t *cup, Dwarf_Die die, ctf_funcinfo_t *fip,
1973 ctf_id_t *argv)
1974 {
1975 int ret;
1976 int i = 0;
1977 Dwarf_Die child, sib, arg;
1978
1979 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1980 return (ret);
1981
1982 arg = child;
1983 while (arg != NULL) {
1984 Dwarf_Half tag;
1985
1986 if ((ret = ctf_dwarf_tag(cup, arg, &tag)) != 0)
1987 return (ret);
1988 if (tag == DW_TAG_formal_parameter) {
1989 Dwarf_Die tdie;
1990
1991 if ((ret = ctf_dwarf_refdie(cup, arg, DW_AT_type,
1992 &tdie)) != 0)
1993 return (ret);
1994
1995 if ((ret = ctf_dwarf_convert_type(cup, tdie, &argv[i],
1996 CTF_ADD_ROOT)) != 0)
1997 return (ret);
1998 i++;
1999
2000 /*
2001 * Once we hit argc entries, we're done. This ensures we
2002 * don't accidentally hit a varargs which should be the
2003 * last entry.
2004 */
2005 if (i == fip->ctc_argc)
2006 break;
2007 }
2008
2009 if ((ret = ctf_dwarf_sib(cup, arg, &sib)) != 0)
2010 return (ret);
2011 arg = sib;
2012 }
2013
2014 return (0);
2015 }
2016
2017 static int
2018 ctf_dwarf_convert_function(ctf_cu_t *cup, Dwarf_Die die)
2019 {
2020 ctf_dwfunc_t *cdf;
2021 Dwarf_Die tdie;
2022 Dwarf_Bool b;
2023 char *name;
2024 int ret;
2025
2026 /*
2027 * Functions that don't have a name are generally functions that have
2028 * been inlined and thus most information about them has been lost. If
2029 * we can't get a name, then instead of returning ENOENT, we silently
2030 * swallow the error.
2031 */
2032 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0) {
2033 if (ret == ENOENT)
2034 return (0);
2035 return (ret);
2036 }
2037
2038 ctf_dprintf("beginning work on function %s (die %llx)\n",
2039 name, ctf_die_offset(die));
2040
2041 if ((ret = ctf_dwarf_boolean(cup, die, DW_AT_declaration, &b)) != 0) {
2042 if (ret != ENOENT)
2043 return (ret);
2044 } else if (b != 0) {
2045 /*
2046 * GCC7 at least creates empty DW_AT_declarations for functions
2047 * defined in headers. As they lack details on the function
2048 * prototype, we need to ignore them. If we later actually
2049 * see the relevant function's definition, we will see another
2050 * DW_TAG_subprogram that is more complete.
2051 */
2052 ctf_dprintf("ignoring declaration of function %s (die %llx)\n",
2053 name, ctf_die_offset(die));
2054 return (0);
2055 }
2056
2057 if ((cdf = ctf_alloc(sizeof (ctf_dwfunc_t))) == NULL) {
2058 ctf_free(name, strlen(name) + 1);
2059 return (ENOMEM);
2060 }
2061 bzero(cdf, sizeof (ctf_dwfunc_t));
2062 cdf->cdf_name = name;
2063
2064 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) == 0) {
2065 if ((ret = ctf_dwarf_convert_type(cup, tdie,
2066 &(cdf->cdf_fip.ctc_return), CTF_ADD_ROOT)) != 0) {
2067 ctf_free(name, strlen(name) + 1);
2068 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2069 return (ret);
2070 }
2071 } else if (ret != ENOENT) {
2072 ctf_free(name, strlen(name) + 1);
2073 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2074 return (ret);
2075 } else {
2076 if ((cdf->cdf_fip.ctc_return = ctf_dwarf_void(cup)) ==
2077 CTF_ERR) {
2078 ctf_free(name, strlen(name) + 1);
2079 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2080 return (ctf_errno(cup->cu_ctfp));
2081 }
2082 }
2083
2084 /*
2085 * A function has a number of children, some of which may not be ones we
2086 * care about. Children that we care about have a type of
2087 * DW_TAG_formal_parameter. We're going to do two passes, the first to
2088 * count the arguments, the second to process them. Afterwards, we
2089 * should be good to go ahead and add this function.
2090 *
2091 * Note, we already got the return type by going in and grabbing it out
2092 * of the DW_AT_type.
2093 */
2094 if ((ret = ctf_dwarf_function_count(cup, die, &cdf->cdf_fip,
2095 B_FALSE)) != 0) {
2096 ctf_free(name, strlen(name) + 1);
2097 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2098 return (ret);
2099 }
2100
2101 ctf_dprintf("beginning to convert function arguments %s\n", name);
2102 if (cdf->cdf_fip.ctc_argc != 0) {
2103 uint_t argc = cdf->cdf_fip.ctc_argc;
2104 cdf->cdf_argv = ctf_alloc(sizeof (ctf_id_t) * argc);
2105 if (cdf->cdf_argv == NULL) {
2106 ctf_free(name, strlen(name) + 1);
2107 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2108 return (ENOMEM);
2109 }
2110 if ((ret = ctf_dwarf_convert_fargs(cup, die,
2111 &cdf->cdf_fip, cdf->cdf_argv)) != 0) {
2112 ctf_free(cdf->cdf_argv, sizeof (ctf_id_t) * argc);
2113 ctf_free(name, strlen(name) + 1);
2114 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2115 return (ret);
2116 }
2117 } else {
2118 cdf->cdf_argv = NULL;
2119 }
2120
2121 if ((ret = ctf_dwarf_isglobal(cup, die, &cdf->cdf_global)) != 0) {
2122 ctf_free(cdf->cdf_argv, sizeof (ctf_id_t) *
2123 cdf->cdf_fip.ctc_argc);
2124 ctf_free(name, strlen(name) + 1);
2125 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2126 return (ret);
2127 }
2128
2129 ctf_list_append(&cup->cu_funcs, cdf);
2130 return (ret);
2131 }
2132
2133 /*
2134 * Convert variables, but only if they're not prototypes and have names.
2135 */
2136 static int
2137 ctf_dwarf_convert_variable(ctf_cu_t *cup, Dwarf_Die die)
2138 {
2139 int ret;
2140 char *name;
2141 Dwarf_Bool b;
2142 Dwarf_Die tdie;
2143 ctf_id_t id;
2144 ctf_dwvar_t *cdv;
2145
2146 /* Skip "Non-Defining Declarations" */
2147 if ((ret = ctf_dwarf_boolean(cup, die, DW_AT_declaration, &b)) == 0) {
2148 if (b != 0)
2149 return (0);
2150 } else if (ret != ENOENT) {
2151 return (ret);
2152 }
2153
2154 /*
2155 * If we find a DIE of "Declarations Completing Non-Defining
2156 * Declarations", we will use the referenced type's DIE. This isn't
2157 * quite correct, e.g. DW_AT_decl_line will be the forward declaration
2158 * not this site. It's sufficient for what we need, however: in
2159 * particular, we should find DW_AT_external as needed there.
2160 */
2161 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_specification,
2162 &tdie)) == 0) {
2163 Dwarf_Off offset;
2164 if ((ret = ctf_dwarf_offset(cup, tdie, &offset)) != 0)
2165 return (ret);
2166 ctf_dprintf("die 0x%llx DW_AT_specification -> die 0x%llx\n",
2167 ctf_die_offset(die), ctf_die_offset(tdie));
2168 die = tdie;
2169 } else if (ret != ENOENT) {
2170 return (ret);
2171 }
2172
2173 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0 &&
2174 ret != ENOENT)
2175 return (ret);
2176 if (ret == ENOENT)
2177 return (0);
2178
2179 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) != 0) {
2180 ctf_free(name, strlen(name) + 1);
2181 return (ret);
2182 }
2183
2184 if ((ret = ctf_dwarf_convert_type(cup, tdie, &id,
2185 CTF_ADD_ROOT)) != 0)
2186 return (ret);
2187
2188 if ((cdv = ctf_alloc(sizeof (ctf_dwvar_t))) == NULL) {
2189 ctf_free(name, strlen(name) + 1);
2190 return (ENOMEM);
2191 }
2192
2193 cdv->cdv_name = name;
2194 cdv->cdv_type = id;
2195
2196 if ((ret = ctf_dwarf_isglobal(cup, die, &cdv->cdv_global)) != 0) {
2197 ctf_free(cdv, sizeof (ctf_dwvar_t));
2198 ctf_free(name, strlen(name) + 1);
2199 return (ret);
2200 }
2201
2202 ctf_list_append(&cup->cu_vars, cdv);
2203 return (0);
2204 }
2205
2206 /*
2207 * Walk through our set of top-level types and process them.
2208 */
2209 static int
2210 ctf_dwarf_walk_toplevel(ctf_cu_t *cup, Dwarf_Die die)
2211 {
2212 int ret;
2213 Dwarf_Off offset;
2214 Dwarf_Half tag;
2215
2216 if ((ret = ctf_dwarf_offset(cup, die, &offset)) != 0)
2217 return (ret);
2218
2219 if (offset > cup->cu_maxoff) {
2220 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
2221 "die offset %llu beyond maximum for header %llu\n",
2222 offset, cup->cu_maxoff);
2223 return (ECTF_CONVBKERR);
2224 }
2225
2226 if ((ret = ctf_dwarf_tag(cup, die, &tag)) != 0)
2227 return (ret);
2228
2229 ret = 0;
2230 switch (tag) {
2231 case DW_TAG_subprogram:
2232 ctf_dprintf("top level func\n");
2233 ret = ctf_dwarf_convert_function(cup, die);
2234 break;
2235 case DW_TAG_variable:
2236 ctf_dprintf("top level var\n");
2237 ret = ctf_dwarf_convert_variable(cup, die);
2238 break;
2239 case DW_TAG_lexical_block:
2240 ctf_dprintf("top level block\n");
2241 ret = ctf_dwarf_walk_lexical(cup, die);
2242 break;
2243 case DW_TAG_enumeration_type:
2244 case DW_TAG_structure_type:
2245 case DW_TAG_typedef:
2246 case DW_TAG_union_type:
2247 ctf_dprintf("top level type\n");
2248 ret = ctf_dwarf_convert_type(cup, die, NULL, B_TRUE);
2249 break;
2250 default:
2251 break;
2252 }
2253
2254 return (ret);
2255 }
2256
2257
2258 /*
2259 * We're given a node. At this node we need to convert it and then proceed to
2260 * convert any siblings that are associaed with this die.
2261 */
2262 static int
2263 ctf_dwarf_convert_die(ctf_cu_t *cup, Dwarf_Die die)
2264 {
2265 while (die != NULL) {
2266 int ret;
2267 Dwarf_Die sib;
2268
2269 if ((ret = ctf_dwarf_walk_toplevel(cup, die)) != 0)
2270 return (ret);
2271
2272 if ((ret = ctf_dwarf_sib(cup, die, &sib)) != 0)
2273 return (ret);
2274 die = sib;
2275 }
2276 return (0);
2277 }
2278
2279 static int
2280 ctf_dwarf_fixup_die(ctf_cu_t *cup, boolean_t addpass)
2281 {
2282 ctf_dwmap_t *map;
2283
2284 for (map = avl_first(&cup->cu_map); map != NULL;
2285 map = AVL_NEXT(&cup->cu_map, map)) {
2286 int ret;
2287 if (map->cdm_fix == B_FALSE)
2288 continue;
2289 if ((ret = ctf_dwarf_fixup_sou(cup, map->cdm_die, map->cdm_id,
2290 addpass)) != 0)
2291 return (ret);
2292 }
2293
2294 return (0);
2295 }
2296
2297 /*
2298 * The DWARF information about a symbol and the information in the symbol table
2299 * may not be the same due to symbol reduction that is performed by ld due to a
2300 * mapfile or other such directive. We process weak symbols at a later time.
2301 *
2302 * The following are the rules that we employ:
2303 *
2304 * 1. A DWARF function that is considered exported matches STB_GLOBAL entries
2305 * with the same name.
2306 *
2307 * 2. A DWARF function that is considered exported matches STB_LOCAL entries
2308 * with the same name and the same file. This case may happen due to mapfile
2309 * reduction.
2310 *
2311 * 3. A DWARF function that is not considered exported matches STB_LOCAL entries
2312 * with the same name and the same file.
2313 *
2314 * 4. A DWARF function that has the same name as the symbol table entry, but the
2315 * files do not match. This is considered a 'fuzzy' match. This may also happen
2316 * due to a mapfile reduction. Fuzzy matching is only used when we know that the
2317 * file in question refers to the primary object. This is because when a symbol
2318 * is reduced in a mapfile, it's always going to be tagged as a local value in
2319 * the generated output and it is considered as to belong to the primary file
2320 * which is the first STT_FILE symbol we see.
2321 */
2322 static boolean_t
2323 ctf_dwarf_symbol_match(const char *symtab_file, const char *symtab_name,
2324 uint_t symtab_bind, const char *dwarf_file, const char *dwarf_name,
2325 boolean_t dwarf_global, boolean_t *is_fuzzy)
2326 {
2327 *is_fuzzy = B_FALSE;
2328
2329 if (symtab_bind != STB_LOCAL && symtab_bind != STB_GLOBAL) {
2330 return (B_FALSE);
2331 }
2332
2333 if (strcmp(symtab_name, dwarf_name) != 0) {
2334 return (B_FALSE);
2335 }
2336
2337 if (symtab_bind == STB_GLOBAL) {
2338 return (dwarf_global);
2339 }
2340
2341 if (strcmp(symtab_file, dwarf_file) == 0) {
2342 return (B_TRUE);
2343 }
2344
2345 if (dwarf_global) {
2346 *is_fuzzy = B_TRUE;
2347 return (B_TRUE);
2348 }
2349
2350 return (B_FALSE);
2351 }
2352
2353 static ctf_dwfunc_t *
2354 ctf_dwarf_match_func(ctf_cu_t *cup, const char *file, const char *name,
2355 uint_t bind, boolean_t primary)
2356 {
2357 ctf_dwfunc_t *cdf, *fuzzy = NULL;
2358
2359 if (bind == STB_WEAK)
2360 return (NULL);
2361
2362 if (bind == STB_LOCAL && (file == NULL || cup->cu_name == NULL))
2363 return (NULL);
2364
2365 for (cdf = ctf_list_next(&cup->cu_funcs); cdf != NULL;
2366 cdf = ctf_list_next(cdf)) {
2367 boolean_t is_fuzzy = B_FALSE;
2368
2369 if (ctf_dwarf_symbol_match(file, name, bind, cup->cu_name,
2370 cdf->cdf_name, cdf->cdf_global, &is_fuzzy)) {
2371 if (is_fuzzy) {
2372 if (primary) {
2373 fuzzy = cdf;
2374 }
2375 continue;
2376 } else {
2377 return (cdf);
2378 }
2379 }
2380 }
2381
2382 return (fuzzy);
2383 }
2384
2385 static ctf_dwvar_t *
2386 ctf_dwarf_match_var(ctf_cu_t *cup, const char *file, const char *name,
2387 uint_t bind, boolean_t primary)
2388 {
2389 ctf_dwvar_t *cdv, *fuzzy = NULL;
2390
2391 if (bind == STB_WEAK)
2392 return (NULL);
2393
2394 if (bind == STB_LOCAL && (file == NULL || cup->cu_name == NULL))
2395 return (NULL);
2396
2397 for (cdv = ctf_list_next(&cup->cu_vars); cdv != NULL;
2398 cdv = ctf_list_next(cdv)) {
2399 boolean_t is_fuzzy = B_FALSE;
2400
2401 if (ctf_dwarf_symbol_match(file, name, bind, cup->cu_name,
2402 cdv->cdv_name, cdv->cdv_global, &is_fuzzy)) {
2403 if (is_fuzzy) {
2404 if (primary) {
2405 fuzzy = cdv;
2406 }
2407 } else {
2408 return (cdv);
2409 }
2410 }
2411 }
2412
2413 return (fuzzy);
2414 }
2415
2416 static int
2417 ctf_dwarf_conv_funcvars_cb(const Elf64_Sym *symp, ulong_t idx,
2418 const char *file, const char *name, boolean_t primary, void *arg)
2419 {
2420 int ret;
2421 uint_t bind, type;
2422 ctf_cu_t *cup = arg;
2423
2424 bind = GELF_ST_BIND(symp->st_info);
2425 type = GELF_ST_TYPE(symp->st_info);
2426
2427 /*
2428 * Come back to weak symbols in another pass
2429 */
2430 if (bind == STB_WEAK)
2431 return (0);
2432
2433 if (type == STT_OBJECT) {
2434 ctf_dwvar_t *cdv = ctf_dwarf_match_var(cup, file, name,
2435 bind, primary);
2436 if (cdv == NULL)
2437 return (0);
2438 ret = ctf_add_object(cup->cu_ctfp, idx, cdv->cdv_type);
2439 ctf_dprintf("added object %s->%ld\n", name, cdv->cdv_type);
2440 } else {
2441 ctf_dwfunc_t *cdf = ctf_dwarf_match_func(cup, file, name,
2442 bind, primary);
2443 if (cdf == NULL)
2444 return (0);
2445 ret = ctf_add_function(cup->cu_ctfp, idx, &cdf->cdf_fip,
2446 cdf->cdf_argv);
2447 ctf_dprintf("added function %s\n", name);
2448 }
2449
2450 if (ret == CTF_ERR) {
2451 return (ctf_errno(cup->cu_ctfp));
2452 }
2453
2454 return (0);
2455 }
2456
2457 static int
2458 ctf_dwarf_conv_funcvars(ctf_cu_t *cup)
2459 {
2460 return (ctf_symtab_iter(cup->cu_ctfp, ctf_dwarf_conv_funcvars_cb, cup));
2461 }
2462
2463 /*
2464 * If we have a weak symbol, attempt to find the strong symbol it will resolve
2465 * to. Note: the code where this actually happens is in sym_process() in
2466 * cmd/sgs/libld/common/syms.c
2467 *
2468 * Finding the matching symbol is unfortunately not trivial. For a symbol to be
2469 * a candidate, it must:
2470 *
2471 * - have the same type (function, object)
2472 * - have the same value (address)
2473 * - have the same size
2474 * - not be another weak symbol
2475 * - belong to the same section (checked via section index)
2476 *
2477 * To perform this check, we first iterate over the symbol table. For each weak
2478 * symbol that we encounter, we then do a second walk over the symbol table,
2479 * calling ctf_dwarf_conv_check_weak(). If a symbol matches the above, then it's
2480 * either a local or global symbol. If we find a global symbol then we go with
2481 * it and stop searching for additional matches.
2482 *
2483 * If instead, we find a local symbol, things are more complicated. The first
2484 * thing we do is to try and see if we have file information about both symbols
2485 * (STT_FILE). If they both have file information and it matches, then we treat
2486 * that as a good match and stop searching for additional matches.
2487 *
2488 * Otherwise, this means we have a non-matching file and a local symbol. We
2489 * treat this as a candidate and if we find a better match (one of the two cases
2490 * above), use that instead. There are two different ways this can happen.
2491 * Either this is a completely different symbol, or it's a once-global symbol
2492 * that was scoped to local via a mapfile. In the former case, curfile is
2493 * likely inaccurate since the linker does not preserve the needed curfile in
2494 * the order of the symbol table (see the comments about locally scoped symbols
2495 * in libld's update_osym()). As we can't tell this case from the former one,
2496 * we use this symbol iff no other matching symbol is found.
2497 *
2498 * What we really need here is a SUNW section containing weak<->strong mappings
2499 * that we can consume.
2500 */
2501 typedef struct ctf_dwarf_weak_arg {
2502 const Elf64_Sym *cweak_symp;
2503 const char *cweak_file;
2504 boolean_t cweak_candidate;
2505 ulong_t cweak_idx;
2506 } ctf_dwarf_weak_arg_t;
2507
2508 static int
2509 ctf_dwarf_conv_check_weak(const Elf64_Sym *symp, ulong_t idx, const char *file,
2510 const char *name, boolean_t primary, void *arg)
2511 {
2512 ctf_dwarf_weak_arg_t *cweak = arg;
2513
2514 const Elf64_Sym *wsymp = cweak->cweak_symp;
2515
2516 ctf_dprintf("comparing weak to %s\n", name);
2517
2518 if (GELF_ST_BIND(symp->st_info) == STB_WEAK) {
2519 return (0);
2520 }
2521
2522 if (GELF_ST_TYPE(wsymp->st_info) != GELF_ST_TYPE(symp->st_info)) {
2523 return (0);
2524 }
2525
2526 if (wsymp->st_value != symp->st_value) {
2527 return (0);
2528 }
2529
2530 if (wsymp->st_size != symp->st_size) {
2531 return (0);
2532 }
2533
2534 if (wsymp->st_shndx != symp->st_shndx) {
2535 return (0);
2536 }
2537
2538 /*
2539 * Check if it's a weak candidate.
2540 */
2541 if (GELF_ST_BIND(symp->st_info) == STB_LOCAL &&
2542 (file == NULL || cweak->cweak_file == NULL ||
2543 strcmp(file, cweak->cweak_file) != 0)) {
2544 cweak->cweak_candidate = B_TRUE;
2545 cweak->cweak_idx = idx;
2546 return (0);
2547 }
2548
2549 /*
2550 * Found a match, break.
2551 */
2552 cweak->cweak_idx = idx;
2553 return (1);
2554 }
2555
2556 static int
2557 ctf_dwarf_duplicate_sym(ctf_cu_t *cup, ulong_t idx, ulong_t matchidx)
2558 {
2559 ctf_id_t id = ctf_lookup_by_symbol(cup->cu_ctfp, matchidx);
2560
2561 /*
2562 * If we matched something that for some reason didn't have type data,
2563 * we don't consider that a fatal error and silently swallow it.
2564 */
2565 if (id == CTF_ERR) {
2566 if (ctf_errno(cup->cu_ctfp) == ECTF_NOTYPEDAT)
2567 return (0);
2568 else
2569 return (ctf_errno(cup->cu_ctfp));
2570 }
2571
2572 if (ctf_add_object(cup->cu_ctfp, idx, id) == CTF_ERR)
2573 return (ctf_errno(cup->cu_ctfp));
2574
2575 return (0);
2576 }
2577
2578 static int
2579 ctf_dwarf_duplicate_func(ctf_cu_t *cup, ulong_t idx, ulong_t matchidx)
2580 {
2581 int ret;
2582 ctf_funcinfo_t fip;
2583 ctf_id_t *args = NULL;
2584
2585 if (ctf_func_info(cup->cu_ctfp, matchidx, &fip) == CTF_ERR) {
2586 if (ctf_errno(cup->cu_ctfp) == ECTF_NOFUNCDAT)
2587 return (0);
2588 else
2589 return (ctf_errno(cup->cu_ctfp));
2590 }
2591
2592 if (fip.ctc_argc != 0) {
2593 args = ctf_alloc(sizeof (ctf_id_t) * fip.ctc_argc);
2594 if (args == NULL)
2595 return (ENOMEM);
2596
2597 if (ctf_func_args(cup->cu_ctfp, matchidx, fip.ctc_argc, args) ==
2598 CTF_ERR) {
2599 ctf_free(args, sizeof (ctf_id_t) * fip.ctc_argc);
2600 return (ctf_errno(cup->cu_ctfp));
2601 }
2602 }
2603
2604 ret = ctf_add_function(cup->cu_ctfp, idx, &fip, args);
2605 if (args != NULL)
2606 ctf_free(args, sizeof (ctf_id_t) * fip.ctc_argc);
2607 if (ret == CTF_ERR)
2608 return (ctf_errno(cup->cu_ctfp));
2609
2610 return (0);
2611 }
2612
2613 static int
2614 ctf_dwarf_conv_weaks_cb(const Elf64_Sym *symp, ulong_t idx, const char *file,
2615 const char *name, boolean_t primary, void *arg)
2616 {
2617 int ret, type;
2618 ctf_dwarf_weak_arg_t cweak;
2619 ctf_cu_t *cup = arg;
2620
2621 /*
2622 * We only care about weak symbols.
2623 */
2624 if (GELF_ST_BIND(symp->st_info) != STB_WEAK)
2625 return (0);
2626
2627 type = GELF_ST_TYPE(symp->st_info);
2628 ASSERT(type == STT_OBJECT || type == STT_FUNC);
2629
2630 /*
2631 * For each weak symbol we encounter, we need to do a second iteration
2632 * to try and find a match. We should probably think about other
2633 * techniques to try and save us time in the future.
2634 */
2635 cweak.cweak_symp = symp;
2636 cweak.cweak_file = file;
2637 cweak.cweak_candidate = B_FALSE;
2638 cweak.cweak_idx = 0;
2639
2640 ctf_dprintf("Trying to find weak equiv for %s\n", name);
2641
2642 ret = ctf_symtab_iter(cup->cu_ctfp, ctf_dwarf_conv_check_weak, &cweak);
2643 VERIFY(ret == 0 || ret == 1);
2644
2645 /*
2646 * Nothing was ever found, we're not going to add anything for this
2647 * entry.
2648 */
2649 if (ret == 0 && cweak.cweak_candidate == B_FALSE) {
2650 ctf_dprintf("found no weak match for %s\n", name);
2651 return (0);
2652 }
2653
2654 /*
2655 * Now, finally go and add the type based on the match.
2656 */
2657 ctf_dprintf("matched weak symbol %lu to %lu\n", idx, cweak.cweak_idx);
2658 if (type == STT_OBJECT) {
2659 ret = ctf_dwarf_duplicate_sym(cup, idx, cweak.cweak_idx);
2660 } else {
2661 ret = ctf_dwarf_duplicate_func(cup, idx, cweak.cweak_idx);
2662 }
2663
2664 return (ret);
2665 }
2666
2667 static int
2668 ctf_dwarf_conv_weaks(ctf_cu_t *cup)
2669 {
2670 return (ctf_symtab_iter(cup->cu_ctfp, ctf_dwarf_conv_weaks_cb, cup));
2671 }
2672
2673 /* ARGSUSED */
2674 static int
2675 ctf_dwarf_convert_one(void *arg, void *unused)
2676 {
2677 int ret;
2678 ctf_file_t *dedup;
2679 ctf_cu_t *cup = arg;
2680
2681 ctf_dprintf("converting die: %s\n", cup->cu_name);
2682 ctf_dprintf("max offset: %x\n", cup->cu_maxoff);
2683 VERIFY(cup != NULL);
2684
2685 ret = ctf_dwarf_convert_die(cup, cup->cu_cu);
2686 ctf_dprintf("ctf_dwarf_convert_die (%s) returned %d\n", cup->cu_name,
2687 ret);
2688 if (ret != 0) {
2689 return (ret);
2690 }
2691 if (ctf_update(cup->cu_ctfp) != 0) {
2692 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2693 "failed to update output ctf container"));
2694 }
2695
2696 ret = ctf_dwarf_fixup_die(cup, B_FALSE);
2697 ctf_dprintf("ctf_dwarf_fixup_die (%s) returned %d\n", cup->cu_name,
2698 ret);
2699 if (ret != 0) {
2700 return (ret);
2701 }
2702 if (ctf_update(cup->cu_ctfp) != 0) {
2703 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2704 "failed to update output ctf container"));
2705 }
2706
2707 ret = ctf_dwarf_fixup_die(cup, B_TRUE);
2708 ctf_dprintf("ctf_dwarf_fixup_die (%s) returned %d\n", cup->cu_name,
2709 ret);
2710 if (ret != 0) {
2711 return (ret);
2712 }
2713 if (ctf_update(cup->cu_ctfp) != 0) {
2714 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2715 "failed to update output ctf container"));
2716 }
2717
2718
2719 if ((ret = ctf_dwarf_conv_funcvars(cup)) != 0) {
2720 return (ctf_dwarf_error(cup, NULL, ret,
2721 "failed to convert strong functions and variables"));
2722 }
2723
2724 if (ctf_update(cup->cu_ctfp) != 0) {
2725 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2726 "failed to update output ctf container"));
2727 }
2728
2729 if (cup->cu_doweaks == B_TRUE) {
2730 if ((ret = ctf_dwarf_conv_weaks(cup)) != 0) {
2731 return (ctf_dwarf_error(cup, NULL, ret,
2732 "failed to convert weak functions and variables"));
2733 }
2734
2735 if (ctf_update(cup->cu_ctfp) != 0) {
2736 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2737 "failed to update output ctf container"));
2738 }
2739 }
2740
2741 ctf_phase_dump(cup->cu_ctfp, "pre-dwarf-dedup", cup->cu_name);
2742 ctf_dprintf("adding inputs for dedup\n");
2743 if ((ret = ctf_merge_add(cup->cu_cmh, cup->cu_ctfp)) != 0) {
2744 return (ctf_dwarf_error(cup, NULL, ret,
2745 "failed to add inputs for merge"));
2746 }
2747
2748 ctf_dprintf("starting dedup of %s\n", cup->cu_name);
2749 if ((ret = ctf_merge_dedup(cup->cu_cmh, &dedup)) != 0) {
2750 return (ctf_dwarf_error(cup, NULL, ret,
2751 "failed to deduplicate die"));
2752 }
2753 ctf_close(cup->cu_ctfp);
2754 cup->cu_ctfp = dedup;
2755 ctf_phase_dump(cup->cu_ctfp, "post-dwarf-dedup", cup->cu_name);
2756
2757 return (0);
2758 }
2759
2760 /*
2761 * Note, we expect that if we're returning a ctf_file_t from one of the dies,
2762 * say in the single node case, it's been saved and the entry here has been set
2763 * to NULL, which ctf_close happily ignores.
2764 */
2765 static void
2766 ctf_dwarf_free_die(ctf_cu_t *cup)
2767 {
2768 ctf_dwfunc_t *cdf, *ndf;
2769 ctf_dwvar_t *cdv, *ndv;
2770 ctf_dwbitf_t *cdb, *ndb;
2771 ctf_dwmap_t *map;
2772 void *cookie;
2773 Dwarf_Error derr;
2774
2775 ctf_dprintf("Beginning to free die: %p\n", cup);
2776 cup->cu_elf = NULL;
2777 ctf_dprintf("Trying to free name: %p\n", cup->cu_name);
2778 if (cup->cu_name != NULL)
2779 ctf_free(cup->cu_name, strlen(cup->cu_name) + 1);
2780 ctf_dprintf("Trying to free merge handle: %p\n", cup->cu_cmh);
2781 if (cup->cu_cmh != NULL) {
2782 ctf_merge_fini(cup->cu_cmh);
2783 cup->cu_cmh = NULL;
2784 }
2785
2786 ctf_dprintf("Trying to free functions\n");
2787 for (cdf = ctf_list_next(&cup->cu_funcs); cdf != NULL; cdf = ndf) {
2788 ndf = ctf_list_next(cdf);
2789 ctf_free(cdf->cdf_name, strlen(cdf->cdf_name) + 1);
2790 if (cdf->cdf_fip.ctc_argc != 0) {
2791 ctf_free(cdf->cdf_argv,
2792 sizeof (ctf_id_t) * cdf->cdf_fip.ctc_argc);
2793 }
2794 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2795 }
2796
2797 ctf_dprintf("Trying to free variables\n");
2798 for (cdv = ctf_list_next(&cup->cu_vars); cdv != NULL; cdv = ndv) {
2799 ndv = ctf_list_next(cdv);
2800 ctf_free(cdv->cdv_name, strlen(cdv->cdv_name) + 1);
2801 ctf_free(cdv, sizeof (ctf_dwvar_t));
2802 }
2803
2804 ctf_dprintf("Trying to free bitfields\n");
2805 for (cdb = ctf_list_next(&cup->cu_bitfields); cdb != NULL; cdb = ndb) {
2806 ndb = ctf_list_next(cdb);
2807 ctf_free(cdb, sizeof (ctf_dwbitf_t));
2808 }
2809
2810 ctf_dprintf("Trying to clean up dwarf_t: %p\n", cup->cu_dwarf);
2811 if (cup->cu_dwarf != NULL)
2812 (void) dwarf_finish(cup->cu_dwarf, &derr);
2813 cup->cu_dwarf = NULL;
2814 ctf_close(cup->cu_ctfp);
2815
2816 cookie = NULL;
2817 while ((map = avl_destroy_nodes(&cup->cu_map, &cookie)) != NULL) {
2818 ctf_free(map, sizeof (ctf_dwmap_t));
2819 }
2820 avl_destroy(&cup->cu_map);
2821 cup->cu_errbuf = NULL;
2822 }
2823
2824 static void
2825 ctf_dwarf_free_dies(ctf_cu_t *cdies, int ndies)
2826 {
2827 int i;
2828
2829 ctf_dprintf("Beginning to free dies\n");
2830 for (i = 0; i < ndies; i++) {
2831 ctf_dwarf_free_die(&cdies[i]);
2832 }
2833
2834 ctf_free(cdies, sizeof (ctf_cu_t) * ndies);
2835 }
2836
2837 static int
2838 ctf_dwarf_count_dies(Dwarf_Debug dw, Dwarf_Error *derr, int *ndies,
2839 char *errbuf, size_t errlen)
2840 {
2841 int ret;
2842 Dwarf_Half vers;
2843 Dwarf_Unsigned nexthdr;
2844
2845 while ((ret = dwarf_next_cu_header(dw, NULL, &vers, NULL, NULL,
2846 &nexthdr, derr)) != DW_DLV_NO_ENTRY) {
2847 if (ret != DW_DLV_OK) {
2848 (void) snprintf(errbuf, errlen,
2849 "file does not contain valid DWARF data: %s\n",
2850 dwarf_errmsg(*derr));
2851 return (ECTF_CONVBKERR);
2852 }
2853
2854 if (vers != DWARF_VERSION_TWO) {
2855 (void) snprintf(errbuf, errlen,
2856 "unsupported DWARF version: %d\n", vers);
2857 return (ECTF_CONVBKERR);
2858 }
2859 *ndies = *ndies + 1;
2860 }
2861
2862 return (0);
2863 }
2864
2865 static int
2866 ctf_dwarf_init_die(int fd, Elf *elf, ctf_cu_t *cup, int ndie, char *errbuf,
2867 size_t errlen)
2868 {
2869 int ret;
2870 Dwarf_Unsigned hdrlen, abboff, nexthdr;
2871 Dwarf_Half addrsz;
2872 Dwarf_Unsigned offset = 0;
2873 Dwarf_Error derr;
2874
2875 while ((ret = dwarf_next_cu_header(cup->cu_dwarf, &hdrlen, NULL,
2876 &abboff, &addrsz, &nexthdr, &derr)) != DW_DLV_NO_ENTRY) {
2877 char *name;
2878 Dwarf_Die cu, child;
2879
2880 /* Based on the counting above, we should be good to go */
2881 VERIFY(ret == DW_DLV_OK);
2882 if (ndie > 0) {
2883 ndie--;
2884 offset = nexthdr;
2885 continue;
2886 }
2887
2888 /*
2889 * Compilers are apparently inconsistent. Some emit no DWARF for
2890 * empty files and others emit empty compilation unit.
2891 */
2892 cup->cu_voidtid = CTF_ERR;
2893 cup->cu_longtid = CTF_ERR;
2894 cup->cu_elf = elf;
2895 cup->cu_maxoff = nexthdr - 1;
2896 cup->cu_ctfp = ctf_fdcreate(fd, &ret);
2897 if (cup->cu_ctfp == NULL)
2898 return (ret);
2899
2900 avl_create(&cup->cu_map, ctf_dwmap_comp, sizeof (ctf_dwmap_t),
2901 offsetof(ctf_dwmap_t, cdm_avl));
2902 cup->cu_errbuf = errbuf;
2903 cup->cu_errlen = errlen;
2904 bzero(&cup->cu_vars, sizeof (ctf_list_t));
2905 bzero(&cup->cu_funcs, sizeof (ctf_list_t));
2906 bzero(&cup->cu_bitfields, sizeof (ctf_list_t));
2907
2908 if ((ret = ctf_dwarf_die_elfenc(elf, cup, errbuf,
2909 errlen)) != 0)
2910 return (ret);
2911
2912 if ((ret = ctf_dwarf_sib(cup, NULL, &cu)) != 0)
2913 return (ret);
2914
2915 if (cu == NULL) {
2916 (void) snprintf(errbuf, errlen,
2917 "file does not contain DWARF data");
2918 return (ECTF_CONVNODEBUG);
2919 }
2920
2921 if ((ret = ctf_dwarf_child(cup, cu, &child)) != 0)
2922 return (ret);
2923
2924 if (child == NULL) {
2925 (void) snprintf(errbuf, errlen,
2926 "file does not contain DWARF data");
2927 return (ECTF_CONVNODEBUG);
2928 }
2929
2930 cup->cu_cuoff = offset;
2931 cup->cu_cu = child;
2932
2933 if ((cup->cu_cmh = ctf_merge_init(fd, &ret)) == NULL)
2934 return (ret);
2935
2936 if (ctf_dwarf_string(cup, cu, DW_AT_name, &name) == 0) {
2937 size_t len = strlen(name) + 1;
2938 char *b = basename(name);
2939 cup->cu_name = strdup(b);
2940 ctf_free(name, len);
2941 }
2942 break;
2943 }
2944
2945 return (0);
2946 }
2947
2948 /*
2949 * This is our only recourse to identify a C source file that is missing debug
2950 * info: it will be mentioned as an STT_FILE, but not have a compile unit entry.
2951 * (A traditional ctfmerge works on individual files, so can identify missing
2952 * DWARF more directly, via ctf_has_c_source() on the .o file.)
2953 *
2954 * As we operate on basenames, this can of course miss some cases, but it's
2955 * better than not checking at all.
2956 *
2957 * We explicitly whitelist some CRT components. Failing that, there's always
2958 * the -m option.
2959 */
2960 static boolean_t
2961 c_source_has_debug(const char *file, ctf_cu_t *cus, size_t nr_cus)
2962 {
2963 const char *basename = strrchr(file, '/');
2964
2965 if (basename == NULL)
2966 basename = file;
2967 else
2968 basename++;
2969
2970 if (strcmp(basename, "common-crt.c") == 0 ||
2971 strcmp(basename, "gmon.c") == 0 ||
2972 strcmp(basename, "dlink_init.c") == 0 ||
2973 strcmp(basename, "dlink_common.c") == 0 ||
2974 strncmp(basename, "crt", strlen("crt")) == 0 ||
2975 strncmp(basename, "values-", strlen("values-")) == 0)
2976 return (B_TRUE);
2977
2978 for (size_t i = 0; i < nr_cus; i++) {
2979 if (strcmp(basename, cus[i].cu_name) == 0)
2980 return (B_TRUE);
2981 }
2982
2983 return (B_FALSE);
2984 }
2985
2986 static int
2987 ctf_dwarf_check_missing(ctf_cu_t *cus, size_t nr_cus, Elf *elf,
2988 char *errmsg, size_t errlen)
2989 {
2990 Elf_Scn *scn, *strscn;
2991 Elf_Data *data, *strdata;
2992 GElf_Shdr shdr;
2993 ulong_t i;
2994
2995 scn = NULL;
2996 while ((scn = elf_nextscn(elf, scn)) != NULL) {
2997 if (gelf_getshdr(scn, &shdr) == NULL) {
2998 (void) snprintf(errmsg, errlen,
2999 "failed to get section header: %s\n",
3000 elf_errmsg(elf_errno()));
3001 return (EINVAL);
3002 }
3003
3004 if (shdr.sh_type == SHT_SYMTAB)
3005 break;
3006 }
3007
3008 if (scn == NULL)
3009 return (0);
3010
3011 if ((strscn = elf_getscn(elf, shdr.sh_link)) == NULL) {
3012 (void) snprintf(errmsg, errlen,
3013 "failed to get str section: %s\n",
3014 elf_errmsg(elf_errno()));
3015 return (EINVAL);
3016 }
3017
3018 if ((data = elf_getdata(scn, NULL)) == NULL) {
3019 (void) snprintf(errmsg, errlen, "failed to read section: %s\n",
3020 elf_errmsg(elf_errno()));
3021 return (EINVAL);
3022 }
3023
3024 if ((strdata = elf_getdata(strscn, NULL)) == NULL) {
3025 (void) snprintf(errmsg, errlen,
3026 "failed to read string table: %s\n",
3027 elf_errmsg(elf_errno()));
3028 return (EINVAL);
3029 }
3030
3031 for (i = 0; i < shdr.sh_size / shdr.sh_entsize; i++) {
3032 GElf_Sym sym;
3033 const char *file;
3034 size_t len;
3035
3036 if (gelf_getsym(data, i, &sym) == NULL) {
3037 (void) snprintf(errmsg, errlen,
3038 "failed to read sym %lu: %s\n",
3039 i, elf_errmsg(elf_errno()));
3040 return (EINVAL);
3041 }
3042
3043 if (GELF_ST_TYPE(sym.st_info) != STT_FILE)
3044 continue;
3045
3046 file = (const char *)((uintptr_t)strdata->d_buf + sym.st_name);
3047 len = strlen(file);
3048 if (len < 2 || strncmp(".c", &file[len - 2], 2) != 0)
3049 continue;
3050
3051 if (!c_source_has_debug(file, cus, nr_cus)) {
3052 (void) snprintf(errmsg, errlen,
3053 "file %s is missing debug info\n", file);
3054 return (ECTF_CONVNODEBUG);
3055 }
3056 }
3057
3058 return (0);
3059 }
3060
3061 int
3062 ctf_dwarf_convert(int fd, Elf *elf, uint_t nthrs, uint_t flags,
3063 ctf_file_t **fpp, char *errbuf, size_t errlen)
3064 {
3065 int err, ret, ndies, i;
3066 Dwarf_Debug dw;
3067 Dwarf_Error derr;
3068 ctf_cu_t *cdies = NULL, *cup;
3069 workq_t *wqp = NULL;
3070
3071 *fpp = NULL;
3072
3073 ret = dwarf_elf_init(elf, DW_DLC_READ, NULL, NULL, &dw, &derr);
3074 if (ret != DW_DLV_OK) {
3075 if (ret == DW_DLV_NO_ENTRY ||
3076 dwarf_errno(derr) == DW_DLE_DEBUG_INFO_NULL) {
3077 (void) snprintf(errbuf, errlen,
3078 "file does not contain DWARF data\n");
3079 return (ECTF_CONVNODEBUG);
3080 }
3081
3082 (void) snprintf(errbuf, errlen,
3083 "dwarf_elf_init() failed: %s\n", dwarf_errmsg(derr));
3084 return (ECTF_CONVBKERR);
3085 }
3086
3087 /*
3088 * Iterate over all of the compilation units and create a ctf_cu_t for
3089 * each of them. This is used to determine if we have zero, one, or
3090 * multiple dies to convert. If we have zero, that's an error. If
3091 * there's only one die, that's the simple case. No merge needed and
3092 * only a single Dwarf_Debug as well.
3093 */
3094 ndies = 0;
3095 err = ctf_dwarf_count_dies(dw, &derr, &ndies, errbuf, errlen);
3096
3097 ctf_dprintf("found %d DWARF CUs\n", ndies);
3098
3099 if (ndies == 0) {
3100 (void) snprintf(errbuf, errlen,
3101 "file does not contain DWARF data\n");
3102 return (ECTF_CONVNODEBUG);
3103 }
3104
3105 (void) dwarf_finish(dw, &derr);
3106 cdies = ctf_alloc(sizeof (ctf_cu_t) * ndies);
3107 if (cdies == NULL) {
3108 return (ENOMEM);
3109 }
3110
3111 bzero(cdies, sizeof (ctf_cu_t) * ndies);
3112
3113 for (i = 0; i < ndies; i++) {
3114 cup = &cdies[i];
3115 ret = dwarf_elf_init(elf, DW_DLC_READ, NULL, NULL,
3116 &cup->cu_dwarf, &derr);
3117 if (ret != 0) {
3118 ctf_free(cdies, sizeof (ctf_cu_t) * ndies);
3119 (void) snprintf(errbuf, errlen,
3120 "failed to initialize DWARF: %s\n",
3121 dwarf_errmsg(derr));
3122 return (ECTF_CONVBKERR);
3123 }
3124
3125 err = ctf_dwarf_init_die(fd, elf, cup, i, errbuf, errlen);
3126 if (err != 0)
3127 goto out;
3128
3129 cup->cu_doweaks = ndies > 1 ? B_FALSE : B_TRUE;
3130 }
3131
3132 if (!(flags & CTF_ALLOW_MISSING_DEBUG) &&
3133 (err = ctf_dwarf_check_missing(cdies, ndies,
3134 elf, errbuf, errlen)) != 0)
3135 goto out;
3136
3137 /*
3138 * If we only have one compilation unit, there's no reason to use
3139 * multiple threads, even if the user requested them. After all, they
3140 * just gave us an upper bound.
3141 */
3142 if (ndies == 1)
3143 nthrs = 1;
3144
3145 if (workq_init(&wqp, nthrs) == -1) {
3146 err = errno;
3147 goto out;
3148 }
3149
3150 for (i = 0; i < ndies; i++) {
3151 cup = &cdies[i];
3152 ctf_dprintf("adding cu %s: %p, %x %x\n", cup->cu_name,
3153 cup->cu_cu, cup->cu_cuoff, cup->cu_maxoff);
3154 if (workq_add(wqp, cup) == -1) {
3155 err = errno;
3156 goto out;
3157 }
3158 }
3159
3160 ret = workq_work(wqp, ctf_dwarf_convert_one, NULL, &err);
3161 if (ret == WORKQ_ERROR) {
3162 err = errno;
3163 goto out;
3164 } else if (ret == WORKQ_UERROR) {
3165 ctf_dprintf("internal convert failed: %s\n",
3166 ctf_errmsg(err));
3167 goto out;
3168 }
3169
3170 ctf_dprintf("Determining next phase: have %d CUs\n", ndies);
3171 if (ndies != 1) {
3172 ctf_merge_t *cmp;
3173
3174 cmp = ctf_merge_init(fd, &err);
3175 if (cmp == NULL)
3176 goto out;
3177
3178 ctf_dprintf("setting threads\n");
3179 if ((err = ctf_merge_set_nthreads(cmp, nthrs)) != 0) {
3180 ctf_merge_fini(cmp);
3181 goto out;
3182 }
3183
3184 for (i = 0; i < ndies; i++) {
3185 cup = &cdies[i];
3186 if ((err = ctf_merge_add(cmp, cup->cu_ctfp)) != 0) {
3187 ctf_merge_fini(cmp);
3188 goto out;
3189 }
3190 }
3191
3192 ctf_dprintf("performing merge\n");
3193 err = ctf_merge_merge(cmp, fpp);
3194 if (err != 0) {
3195 ctf_dprintf("failed merge!\n");
3196 *fpp = NULL;
3197 ctf_merge_fini(cmp);
3198 goto out;
3199 }
3200 ctf_merge_fini(cmp);
3201 err = 0;
3202 ctf_dprintf("successfully converted!\n");
3203 } else {
3204 err = 0;
3205 *fpp = cdies->cu_ctfp;
3206 cdies->cu_ctfp = NULL;
3207 ctf_dprintf("successfully converted!\n");
3208 }
3209
3210 out:
3211 workq_fini(wqp);
3212 ctf_dwarf_free_dies(cdies, ndies);
3213 return (err);
3214 }