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 2019 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 ulong_t nsz;
1217
1218 kind = ctf_type_kind(cup->cu_ctfp, base);
1219 VERIFY(kind != CTF_ERR);
1220 VERIFY(kind == CTF_K_STRUCT || kind == CTF_K_UNION);
1221
1222 /*
1223 * Members are in children. However, gcc also allows empty ones.
1224 */
1225 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1226 return (ret);
1227 if (child == NULL)
1228 return (0);
1229
1230 memb = child;
1231 while (memb != NULL) {
1232 Dwarf_Die sib, tdie;
1233 Dwarf_Half tag;
1234 ctf_id_t mid;
1235 char *mname;
1236 ulong_t memboff = 0;
1237
1238 if ((ret = ctf_dwarf_tag(cup, memb, &tag)) != 0)
1239 return (ret);
1240
1241 if (tag != DW_TAG_member)
1242 continue;
1243
1244 if ((ret = ctf_dwarf_refdie(cup, memb, DW_AT_type, &tdie)) != 0)
1245 return (ret);
1246
1247 if ((ret = ctf_dwarf_convert_type(cup, tdie, &mid,
1248 CTF_ADD_NONROOT)) != 0)
1249 return (ret);
1250 ctf_dprintf("Got back type id: %d\n", mid);
1251
1252 /*
1253 * If we're not adding a member, just go ahead and return.
1254 */
1255 if (add == B_FALSE) {
1256 if ((ret = ctf_dwarf_member_bitfield(cup, memb,
1257 &mid)) != 0)
1258 return (ret);
1259 goto next;
1260 }
1261
1262 if ((ret = ctf_dwarf_string(cup, memb, DW_AT_name,
1263 &mname)) != 0 && ret != ENOENT)
1264 return (ret);
1265 if (ret == ENOENT)
1266 mname = NULL;
1267
1268 if (kind == CTF_K_UNION) {
1269 memboff = 0;
1270 } else if ((ret = ctf_dwarf_member_offset(cup, memb, mid,
1271 &memboff)) != 0) {
1272 if (mname != NULL)
1273 ctf_free(mname, strlen(mname) + 1);
1274 return (ret);
1275 }
1276
1277 if ((ret = ctf_dwarf_member_bitfield(cup, memb, &mid)) != 0)
1278 return (ret);
1279
1280 ret = ctf_add_member(cup->cu_ctfp, base, mname, mid, memboff);
1281 if (ret == CTF_ERR) {
1282 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1283 "failed to add member %s: %s",
1284 mname, ctf_errmsg(ctf_errno(cup->cu_ctfp)));
1285 if (mname != NULL)
1286 ctf_free(mname, strlen(mname) + 1);
1287 return (ECTF_CONVBKERR);
1288 }
1289
1290 if (mname != NULL)
1291 ctf_free(mname, strlen(mname) + 1);
1292
1293 next:
1294 if ((ret = ctf_dwarf_sib(cup, memb, &sib)) != 0)
1295 return (ret);
1296 memb = sib;
1297 }
1298
1299 /*
1300 * If we're not adding members, then we don't know the final size of the
1301 * structure, so end here.
1302 */
1303 if (add == B_FALSE)
1304 return (0);
1305
1306 /* Finally set the size of the structure to the actual byte size */
1307 if ((ret = ctf_dwarf_unsigned(cup, die, DW_AT_byte_size, &size)) != 0)
1308 return (ret);
1309 nsz = size;
1310 if ((ctf_set_size(cup->cu_ctfp, base, nsz)) == CTF_ERR) {
1311 int e = ctf_errno(cup->cu_ctfp);
1312 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1313 "failed to set type size for %d to 0x%x: %s", base,
1314 (uint32_t)size, ctf_errmsg(e));
1315 return (ECTF_CONVBKERR);
1316 }
1317
1318 return (0);
1319 }
1320
1321 static int
1322 ctf_dwarf_create_sou(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp,
1323 int kind, int isroot)
1324 {
1325 int ret;
1326 char *name;
1327 ctf_id_t base;
1328 Dwarf_Die child;
1329 Dwarf_Bool decl;
1330
1331 /*
1332 * Deal with the terribly annoying case of anonymous structs and unions.
1333 * If they don't have a name, set the name to the empty string.
1334 */
1335 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0 &&
1336 ret != ENOENT)
1337 return (ret);
1338 if (ret == ENOENT)
1339 name = NULL;
1340
1341 /*
1342 * We need to check if we just have a declaration here. If we do, then
1343 * instead of creating an actual structure or union, we're just going to
1344 * go ahead and create a forward. During a dedup or merge, the forward
1345 * will be replaced with the real thing.
1346 */
1347 if ((ret = ctf_dwarf_boolean(cup, die, DW_AT_declaration,
1348 &decl)) != 0) {
1349 if (ret != ENOENT)
1350 return (ret);
1351 decl = 0;
1352 }
1353
1354 if (decl != 0) {
1355 base = ctf_add_forward(cup->cu_ctfp, isroot, name, kind);
1356 } else if (kind == CTF_K_STRUCT) {
1357 base = ctf_add_struct(cup->cu_ctfp, isroot, name);
1358 } else {
1359 base = ctf_add_union(cup->cu_ctfp, isroot, name);
1360 }
1361 ctf_dprintf("added sou %s (%d) (%d)\n", name, kind, base);
1362 if (name != NULL)
1363 ctf_free(name, strlen(name) + 1);
1364 if (base == CTF_ERR)
1365 return (ctf_errno(cup->cu_ctfp));
1366 *idp = base;
1367
1368 /*
1369 * If it's just a declaration, we're not going to mark it for fix up or
1370 * do anything else.
1371 */
1372 if (decl == B_TRUE)
1373 return (ctf_dwmap_add(cup, base, die, B_FALSE));
1374 if ((ret = ctf_dwmap_add(cup, base, die, B_TRUE)) != 0)
1375 return (ret);
1376
1377 /*
1378 * Members are in children. However, gcc also allows empty ones.
1379 */
1380 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1381 return (ret);
1382 if (child == NULL)
1383 return (0);
1384
1385 return (0);
1386 }
1387
1388 static int
1389 ctf_dwarf_create_array_range(ctf_cu_t *cup, Dwarf_Die range, ctf_id_t *idp,
1390 ctf_id_t base, int isroot)
1391 {
1392 int ret;
1393 Dwarf_Die sib;
1394 Dwarf_Unsigned val;
1395 Dwarf_Signed sval;
1396 ctf_arinfo_t ar;
1397
1398 ctf_dprintf("creating array range\n");
1399
1400 if ((ret = ctf_dwarf_sib(cup, range, &sib)) != 0)
1401 return (ret);
1402 if (sib != NULL) {
1403 ctf_id_t id;
1404 if ((ret = ctf_dwarf_create_array_range(cup, sib, &id,
1405 base, CTF_ADD_NONROOT)) != 0)
1406 return (ret);
1407 ar.ctr_contents = id;
1408 } else {
1409 ar.ctr_contents = base;
1410 }
1411
1412 if ((ar.ctr_index = ctf_dwarf_long(cup)) == CTF_ERR)
1413 return (ctf_errno(cup->cu_ctfp));
1414
1415 /*
1416 * Array bounds can be signed or unsigned, but there are several kinds
1417 * of signless forms (data1, data2, etc) that take their sign from the
1418 * routine that is trying to interpret them. That is, data1 can be
1419 * either signed or unsigned, depending on whether you use the signed or
1420 * unsigned accessor function. GCC will use the signless forms to store
1421 * unsigned values which have their high bit set, so we need to try to
1422 * read them first as unsigned to get positive values. We could also
1423 * try signed first, falling back to unsigned if we got a negative
1424 * value.
1425 */
1426 if ((ret = ctf_dwarf_unsigned(cup, range, DW_AT_upper_bound,
1427 &val)) == 0) {
1428 ar.ctr_nelems = val + 1;
1429 } else if (ret != ENOENT) {
1430 return (ret);
1431 } else if ((ret = ctf_dwarf_signed(cup, range, DW_AT_upper_bound,
1432 &sval)) == 0) {
1433 ar.ctr_nelems = sval + 1;
1434 } else if (ret != ENOENT) {
1435 return (ret);
1436 } else {
1437 ar.ctr_nelems = 0;
1438 }
1439
1440 if ((*idp = ctf_add_array(cup->cu_ctfp, isroot, &ar)) == CTF_ERR)
1441 return (ctf_errno(cup->cu_ctfp));
1442
1443 return (0);
1444 }
1445
1446 /*
1447 * Try and create an array type. First, the kind of the array is specified in
1448 * the DW_AT_type entry. Next, the number of entries is stored in a more
1449 * complicated form, we should have a child that has the DW_TAG_subrange type.
1450 */
1451 static int
1452 ctf_dwarf_create_array(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp, int isroot)
1453 {
1454 int ret;
1455 Dwarf_Die tdie, rdie;
1456 ctf_id_t tid;
1457 Dwarf_Half rtag;
1458
1459 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) != 0)
1460 return (ret);
1461 if ((ret = ctf_dwarf_convert_type(cup, tdie, &tid,
1462 CTF_ADD_NONROOT)) != 0)
1463 return (ret);
1464
1465 if ((ret = ctf_dwarf_child(cup, die, &rdie)) != 0)
1466 return (ret);
1467 if ((ret = ctf_dwarf_tag(cup, rdie, &rtag)) != 0)
1468 return (ret);
1469 if (rtag != DW_TAG_subrange_type) {
1470 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1471 "encountered array without DW_TAG_subrange_type child\n");
1472 return (ECTF_CONVBKERR);
1473 }
1474
1475 /*
1476 * The compiler may opt to describe a multi-dimensional array as one
1477 * giant array or it may opt to instead encode it as a series of
1478 * subranges. If it's the latter, then for each subrange we introduce a
1479 * type. We can always use the base type.
1480 */
1481 if ((ret = ctf_dwarf_create_array_range(cup, rdie, idp, tid,
1482 isroot)) != 0)
1483 return (ret);
1484 ctf_dprintf("Got back id %d\n", *idp);
1485 return (ctf_dwmap_add(cup, *idp, die, B_FALSE));
1486 }
1487
1488 static int
1489 ctf_dwarf_create_reference(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp,
1490 int kind, int isroot)
1491 {
1492 int ret;
1493 ctf_id_t id;
1494 Dwarf_Die tdie;
1495 char *name;
1496 size_t namelen;
1497
1498 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0 &&
1499 ret != ENOENT)
1500 return (ret);
1501 if (ret == ENOENT) {
1502 name = NULL;
1503 namelen = 0;
1504 } else {
1505 namelen = strlen(name);
1506 }
1507
1508 ctf_dprintf("reference kind %d %s\n", kind, name != NULL ? name : "<>");
1509
1510 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) != 0) {
1511 if (ret != ENOENT) {
1512 ctf_free(name, namelen);
1513 return (ret);
1514 }
1515 if ((id = ctf_dwarf_void(cup)) == CTF_ERR) {
1516 ctf_free(name, namelen);
1517 return (ctf_errno(cup->cu_ctfp));
1518 }
1519 } else {
1520 if ((ret = ctf_dwarf_convert_type(cup, tdie, &id,
1521 CTF_ADD_NONROOT)) != 0) {
1522 ctf_free(name, namelen);
1523 return (ret);
1524 }
1525 }
1526
1527 if ((*idp = ctf_add_reftype(cup->cu_ctfp, isroot, name, id, kind)) ==
1528 CTF_ERR) {
1529 ctf_free(name, namelen);
1530 return (ctf_errno(cup->cu_ctfp));
1531 }
1532
1533 ctf_free(name, namelen);
1534 return (ctf_dwmap_add(cup, *idp, die, B_FALSE));
1535 }
1536
1537 static int
1538 ctf_dwarf_create_enum(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp, int isroot)
1539 {
1540 int ret;
1541 ctf_id_t id;
1542 Dwarf_Die child;
1543 char *name;
1544
1545 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0 &&
1546 ret != ENOENT)
1547 return (ret);
1548 if (ret == ENOENT)
1549 name = NULL;
1550 id = ctf_add_enum(cup->cu_ctfp, isroot, name);
1551 ctf_dprintf("added enum %s (%d)\n", name, id);
1552 if (name != NULL)
1553 ctf_free(name, strlen(name) + 1);
1554 if (id == CTF_ERR)
1555 return (ctf_errno(cup->cu_ctfp));
1556 *idp = id;
1557 if ((ret = ctf_dwmap_add(cup, id, die, B_FALSE)) != 0)
1558 return (ret);
1559
1560 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0) {
1561 if (ret == ENOENT)
1562 ret = 0;
1563 return (ret);
1564 }
1565
1566 while (child != NULL) {
1567 Dwarf_Half tag;
1568 Dwarf_Signed sval;
1569 Dwarf_Unsigned uval;
1570 Dwarf_Die arg = child;
1571 int eval;
1572
1573 if ((ret = ctf_dwarf_sib(cup, arg, &child)) != 0)
1574 return (ret);
1575
1576 if ((ret = ctf_dwarf_tag(cup, arg, &tag)) != 0)
1577 return (ret);
1578
1579 if (tag != DW_TAG_enumerator) {
1580 if ((ret = ctf_dwarf_convert_type(cup, arg, NULL,
1581 CTF_ADD_NONROOT)) != 0)
1582 return (ret);
1583 continue;
1584 }
1585
1586 /*
1587 * DWARF v4 section 5.7 tells us we'll always have names.
1588 */
1589 if ((ret = ctf_dwarf_string(cup, arg, DW_AT_name, &name)) != 0)
1590 return (ret);
1591
1592 /*
1593 * We have to be careful here: newer GCCs generate DWARF where
1594 * an unsigned value will happily pass ctf_dwarf_signed().
1595 * Since negative values will fail ctf_dwarf_unsigned(), we try
1596 * that first to make sure we get the right value.
1597 */
1598 if ((ret = ctf_dwarf_unsigned(cup, arg, DW_AT_const_value,
1599 &uval)) == 0) {
1600 eval = (int)uval;
1601 } else if ((ret = ctf_dwarf_signed(cup, arg, DW_AT_const_value,
1602 &sval)) == 0) {
1603 eval = sval;
1604 }
1605
1606 if (ret != 0) {
1607 if (ret != ENOENT)
1608 return (ret);
1609
1610 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1611 "encountered enumeration without constant value\n");
1612 return (ECTF_CONVBKERR);
1613 }
1614
1615 ret = ctf_add_enumerator(cup->cu_ctfp, id, name, eval);
1616 if (ret == CTF_ERR) {
1617 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1618 "failed to add enumarator %s (%d) to %d\n",
1619 name, eval, id);
1620 ctf_free(name, strlen(name) + 1);
1621 return (ctf_errno(cup->cu_ctfp));
1622 }
1623 ctf_free(name, strlen(name) + 1);
1624 }
1625
1626 return (0);
1627 }
1628
1629 /*
1630 * For a function pointer, walk over and process all of its children, unless we
1631 * encounter one that's just a declaration. In which case, we error on it.
1632 */
1633 static int
1634 ctf_dwarf_create_fptr(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp, int isroot)
1635 {
1636 int ret;
1637 Dwarf_Bool b;
1638 ctf_funcinfo_t fi;
1639 Dwarf_Die retdie;
1640 ctf_id_t *argv = NULL;
1641
1642 bzero(&fi, sizeof (ctf_funcinfo_t));
1643
1644 if ((ret = ctf_dwarf_boolean(cup, die, DW_AT_declaration, &b)) != 0) {
1645 if (ret != ENOENT)
1646 return (ret);
1647 } else {
1648 if (b != 0)
1649 return (EPROTOTYPE);
1650 }
1651
1652 /*
1653 * Return type is in DW_AT_type, if none, it returns void.
1654 */
1655 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &retdie)) != 0) {
1656 if (ret != ENOENT)
1657 return (ret);
1658 if ((fi.ctc_return = ctf_dwarf_void(cup)) == CTF_ERR)
1659 return (ctf_errno(cup->cu_ctfp));
1660 } else {
1661 if ((ret = ctf_dwarf_convert_type(cup, retdie, &fi.ctc_return,
1662 CTF_ADD_NONROOT)) != 0)
1663 return (ret);
1664 }
1665
1666 if ((ret = ctf_dwarf_function_count(cup, die, &fi, B_TRUE)) != 0) {
1667 return (ret);
1668 }
1669
1670 if (fi.ctc_argc != 0) {
1671 argv = ctf_alloc(sizeof (ctf_id_t) * fi.ctc_argc);
1672 if (argv == NULL)
1673 return (ENOMEM);
1674
1675 if ((ret = ctf_dwarf_convert_fargs(cup, die, &fi, argv)) != 0) {
1676 ctf_free(argv, sizeof (ctf_id_t) * fi.ctc_argc);
1677 return (ret);
1678 }
1679 }
1680
1681 if ((*idp = ctf_add_funcptr(cup->cu_ctfp, isroot, &fi, argv)) ==
1682 CTF_ERR) {
1683 ctf_free(argv, sizeof (ctf_id_t) * fi.ctc_argc);
1684 return (ctf_errno(cup->cu_ctfp));
1685 }
1686
1687 ctf_free(argv, sizeof (ctf_id_t) * fi.ctc_argc);
1688 return (ctf_dwmap_add(cup, *idp, die, B_FALSE));
1689 }
1690
1691 static int
1692 ctf_dwarf_convert_type(ctf_cu_t *cup, Dwarf_Die die, ctf_id_t *idp,
1693 int isroot)
1694 {
1695 int ret;
1696 Dwarf_Off offset;
1697 Dwarf_Half tag;
1698 ctf_dwmap_t lookup, *map;
1699 ctf_id_t id;
1700
1701 if (idp == NULL)
1702 idp = &id;
1703
1704 if ((ret = ctf_dwarf_offset(cup, die, &offset)) != 0)
1705 return (ret);
1706
1707 if (offset > cup->cu_maxoff) {
1708 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
1709 "die offset %llu beyond maximum for header %llu\n",
1710 offset, cup->cu_maxoff);
1711 return (ECTF_CONVBKERR);
1712 }
1713
1714 /*
1715 * If we've already added an entry for this offset, then we're done.
1716 */
1717 lookup.cdm_off = offset;
1718 if ((map = avl_find(&cup->cu_map, &lookup, NULL)) != NULL) {
1719 *idp = map->cdm_id;
1720 return (0);
1721 }
1722
1723 if ((ret = ctf_dwarf_tag(cup, die, &tag)) != 0)
1724 return (ret);
1725
1726 ret = ENOTSUP;
1727 switch (tag) {
1728 case DW_TAG_base_type:
1729 ctf_dprintf("base\n");
1730 ret = ctf_dwarf_create_base(cup, die, idp, isroot, offset);
1731 break;
1732 case DW_TAG_array_type:
1733 ctf_dprintf("array\n");
1734 ret = ctf_dwarf_create_array(cup, die, idp, isroot);
1735 break;
1736 case DW_TAG_enumeration_type:
1737 ctf_dprintf("enum\n");
1738 ret = ctf_dwarf_create_enum(cup, die, idp, isroot);
1739 break;
1740 case DW_TAG_pointer_type:
1741 ctf_dprintf("pointer\n");
1742 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_POINTER,
1743 isroot);
1744 break;
1745 case DW_TAG_structure_type:
1746 ctf_dprintf("struct\n");
1747 ret = ctf_dwarf_create_sou(cup, die, idp, CTF_K_STRUCT,
1748 isroot);
1749 break;
1750 case DW_TAG_subroutine_type:
1751 ctf_dprintf("fptr\n");
1752 ret = ctf_dwarf_create_fptr(cup, die, idp, isroot);
1753 break;
1754 case DW_TAG_typedef:
1755 ctf_dprintf("typedef\n");
1756 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_TYPEDEF,
1757 isroot);
1758 break;
1759 case DW_TAG_union_type:
1760 ctf_dprintf("union\n");
1761 ret = ctf_dwarf_create_sou(cup, die, idp, CTF_K_UNION,
1762 isroot);
1763 break;
1764 case DW_TAG_const_type:
1765 ctf_dprintf("const\n");
1766 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_CONST,
1767 isroot);
1768 break;
1769 case DW_TAG_volatile_type:
1770 ctf_dprintf("volatile\n");
1771 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_VOLATILE,
1772 isroot);
1773 break;
1774 case DW_TAG_restrict_type:
1775 ctf_dprintf("restrict\n");
1776 ret = ctf_dwarf_create_reference(cup, die, idp, CTF_K_RESTRICT,
1777 isroot);
1778 break;
1779 default:
1780 ctf_dprintf("ignoring tag type %x\n", tag);
1781 *idp = CTF_ERR;
1782 ret = 0;
1783 break;
1784 }
1785 ctf_dprintf("ctf_dwarf_convert_type tag specific handler returned %d\n",
1786 ret);
1787
1788 return (ret);
1789 }
1790
1791 static int
1792 ctf_dwarf_walk_lexical(ctf_cu_t *cup, Dwarf_Die die)
1793 {
1794 int ret;
1795 Dwarf_Die child;
1796
1797 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1798 return (ret);
1799
1800 if (child == NULL)
1801 return (0);
1802
1803 return (ctf_dwarf_convert_die(cup, die));
1804 }
1805
1806 static int
1807 ctf_dwarf_function_count(ctf_cu_t *cup, Dwarf_Die die, ctf_funcinfo_t *fip,
1808 boolean_t fptr)
1809 {
1810 int ret;
1811 Dwarf_Die child, sib, arg;
1812
1813 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1814 return (ret);
1815
1816 arg = child;
1817 while (arg != NULL) {
1818 Dwarf_Half tag;
1819
1820 if ((ret = ctf_dwarf_tag(cup, arg, &tag)) != 0)
1821 return (ret);
1822
1823 /*
1824 * We have to check for a varargs type decleration. This will
1825 * happen in one of two ways. If we have a function pointer
1826 * type, then it'll be done with a tag of type
1827 * DW_TAG_unspecified_parameters. However, it only means we have
1828 * a variable number of arguments, if we have more than one
1829 * argument found so far. Otherwise, when we have a function
1830 * type, it instead uses a formal parameter whose name is '...'
1831 * to indicate a variable arguments member.
1832 *
1833 * Also, if we have a function pointer, then we have to expect
1834 * that we might not get a name at all.
1835 */
1836 if (tag == DW_TAG_formal_parameter && fptr == B_FALSE) {
1837 char *name;
1838 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name,
1839 &name)) != 0)
1840 return (ret);
1841 if (strcmp(name, DWARF_VARARGS_NAME) == 0)
1842 fip->ctc_flags |= CTF_FUNC_VARARG;
1843 else
1844 fip->ctc_argc++;
1845 ctf_free(name, strlen(name) + 1);
1846 } else if (tag == DW_TAG_formal_parameter) {
1847 fip->ctc_argc++;
1848 } else if (tag == DW_TAG_unspecified_parameters &&
1849 fip->ctc_argc > 0) {
1850 fip->ctc_flags |= CTF_FUNC_VARARG;
1851 }
1852 if ((ret = ctf_dwarf_sib(cup, arg, &sib)) != 0)
1853 return (ret);
1854 arg = sib;
1855 }
1856
1857 return (0);
1858 }
1859
1860 static int
1861 ctf_dwarf_convert_fargs(ctf_cu_t *cup, Dwarf_Die die, ctf_funcinfo_t *fip,
1862 ctf_id_t *argv)
1863 {
1864 int ret;
1865 int i = 0;
1866 Dwarf_Die child, sib, arg;
1867
1868 if ((ret = ctf_dwarf_child(cup, die, &child)) != 0)
1869 return (ret);
1870
1871 arg = child;
1872 while (arg != NULL) {
1873 Dwarf_Half tag;
1874
1875 if ((ret = ctf_dwarf_tag(cup, arg, &tag)) != 0)
1876 return (ret);
1877 if (tag == DW_TAG_formal_parameter) {
1878 Dwarf_Die tdie;
1879
1880 if ((ret = ctf_dwarf_refdie(cup, arg, DW_AT_type,
1881 &tdie)) != 0)
1882 return (ret);
1883
1884 if ((ret = ctf_dwarf_convert_type(cup, tdie, &argv[i],
1885 CTF_ADD_ROOT)) != 0)
1886 return (ret);
1887 i++;
1888
1889 /*
1890 * Once we hit argc entries, we're done. This ensures we
1891 * don't accidentally hit a varargs which should be the
1892 * last entry.
1893 */
1894 if (i == fip->ctc_argc)
1895 break;
1896 }
1897
1898 if ((ret = ctf_dwarf_sib(cup, arg, &sib)) != 0)
1899 return (ret);
1900 arg = sib;
1901 }
1902
1903 return (0);
1904 }
1905
1906 static int
1907 ctf_dwarf_convert_function(ctf_cu_t *cup, Dwarf_Die die)
1908 {
1909 int ret;
1910 char *name;
1911 ctf_dwfunc_t *cdf;
1912 Dwarf_Die tdie;
1913
1914 /*
1915 * Functions that don't have a name are generally functions that have
1916 * been inlined and thus most information about them has been lost. If
1917 * we can't get a name, then instead of returning ENOENT, we silently
1918 * swallow the error.
1919 */
1920 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0) {
1921 if (ret == ENOENT)
1922 return (0);
1923 return (ret);
1924 }
1925
1926 ctf_dprintf("beginning work on function %s\n", name);
1927 if ((cdf = ctf_alloc(sizeof (ctf_dwfunc_t))) == NULL) {
1928 ctf_free(name, strlen(name) + 1);
1929 return (ENOMEM);
1930 }
1931 bzero(cdf, sizeof (ctf_dwfunc_t));
1932 cdf->cdf_name = name;
1933
1934 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) == 0) {
1935 if ((ret = ctf_dwarf_convert_type(cup, tdie,
1936 &(cdf->cdf_fip.ctc_return), CTF_ADD_ROOT)) != 0) {
1937 ctf_free(name, strlen(name) + 1);
1938 ctf_free(cdf, sizeof (ctf_dwfunc_t));
1939 return (ret);
1940 }
1941 } else if (ret != ENOENT) {
1942 ctf_free(name, strlen(name) + 1);
1943 ctf_free(cdf, sizeof (ctf_dwfunc_t));
1944 return (ret);
1945 } else {
1946 if ((cdf->cdf_fip.ctc_return = ctf_dwarf_void(cup)) ==
1947 CTF_ERR) {
1948 ctf_free(name, strlen(name) + 1);
1949 ctf_free(cdf, sizeof (ctf_dwfunc_t));
1950 return (ctf_errno(cup->cu_ctfp));
1951 }
1952 }
1953
1954 /*
1955 * A function has a number of children, some of which may not be ones we
1956 * care about. Children that we care about have a type of
1957 * DW_TAG_formal_parameter. We're going to do two passes, the first to
1958 * count the arguments, the second to process them. Afterwards, we
1959 * should be good to go ahead and add this function.
1960 *
1961 * Note, we already got the return type by going in and grabbing it out
1962 * of the DW_AT_type.
1963 */
1964 if ((ret = ctf_dwarf_function_count(cup, die, &cdf->cdf_fip,
1965 B_FALSE)) != 0) {
1966 ctf_free(name, strlen(name) + 1);
1967 ctf_free(cdf, sizeof (ctf_dwfunc_t));
1968 return (ret);
1969 }
1970
1971 ctf_dprintf("beginning to convert function arguments %s\n", name);
1972 if (cdf->cdf_fip.ctc_argc != 0) {
1973 uint_t argc = cdf->cdf_fip.ctc_argc;
1974 cdf->cdf_argv = ctf_alloc(sizeof (ctf_id_t) * argc);
1975 if (cdf->cdf_argv == NULL) {
1976 ctf_free(name, strlen(name) + 1);
1977 ctf_free(cdf, sizeof (ctf_dwfunc_t));
1978 return (ENOMEM);
1979 }
1980 if ((ret = ctf_dwarf_convert_fargs(cup, die,
1981 &cdf->cdf_fip, cdf->cdf_argv)) != 0) {
1982 ctf_free(cdf->cdf_argv, sizeof (ctf_id_t) * argc);
1983 ctf_free(name, strlen(name) + 1);
1984 ctf_free(cdf, sizeof (ctf_dwfunc_t));
1985 return (ret);
1986 }
1987 } else {
1988 cdf->cdf_argv = NULL;
1989 }
1990
1991 if ((ret = ctf_dwarf_isglobal(cup, die, &cdf->cdf_global)) != 0) {
1992 ctf_free(cdf->cdf_argv, sizeof (ctf_id_t) *
1993 cdf->cdf_fip.ctc_argc);
1994 ctf_free(name, strlen(name) + 1);
1995 ctf_free(cdf, sizeof (ctf_dwfunc_t));
1996 return (ret);
1997 }
1998
1999 ctf_list_append(&cup->cu_funcs, cdf);
2000 return (ret);
2001 }
2002
2003 /*
2004 * Convert variables, but only if they're not prototypes and have names.
2005 */
2006 static int
2007 ctf_dwarf_convert_variable(ctf_cu_t *cup, Dwarf_Die die)
2008 {
2009 int ret;
2010 char *name;
2011 Dwarf_Bool b;
2012 Dwarf_Die tdie;
2013 ctf_id_t id;
2014 ctf_dwvar_t *cdv;
2015
2016 /* Skip "Non-Defining Declarations" */
2017 if ((ret = ctf_dwarf_boolean(cup, die, DW_AT_declaration, &b)) == 0) {
2018 if (b != 0)
2019 return (0);
2020 } else if (ret != ENOENT) {
2021 return (ret);
2022 }
2023
2024 /*
2025 * If we find a DIE of "Declarations Completing Non-Defining
2026 * Declarations", we will use the referenced type's DIE. This isn't
2027 * quite correct, e.g. DW_AT_decl_line will be the forward declaration
2028 * not this site. It's sufficient for what we need, however: in
2029 * particular, we should find DW_AT_external as needed there.
2030 */
2031 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_specification,
2032 &tdie)) == 0) {
2033 Dwarf_Off offset;
2034 if ((ret = ctf_dwarf_offset(cup, tdie, &offset)) != 0)
2035 return (ret);
2036 ctf_dprintf("die 0x%llx DW_AT_specification -> die 0x%llx\n",
2037 ctf_die_offset(die), ctf_die_offset(tdie));
2038 die = tdie;
2039 } else if (ret != ENOENT) {
2040 return (ret);
2041 }
2042
2043 if ((ret = ctf_dwarf_string(cup, die, DW_AT_name, &name)) != 0 &&
2044 ret != ENOENT)
2045 return (ret);
2046 if (ret == ENOENT)
2047 return (0);
2048
2049 if ((ret = ctf_dwarf_refdie(cup, die, DW_AT_type, &tdie)) != 0) {
2050 ctf_free(name, strlen(name) + 1);
2051 return (ret);
2052 }
2053
2054 if ((ret = ctf_dwarf_convert_type(cup, tdie, &id,
2055 CTF_ADD_ROOT)) != 0)
2056 return (ret);
2057
2058 if ((cdv = ctf_alloc(sizeof (ctf_dwvar_t))) == NULL) {
2059 ctf_free(name, strlen(name) + 1);
2060 return (ENOMEM);
2061 }
2062
2063 cdv->cdv_name = name;
2064 cdv->cdv_type = id;
2065
2066 if ((ret = ctf_dwarf_isglobal(cup, die, &cdv->cdv_global)) != 0) {
2067 ctf_free(cdv, sizeof (ctf_dwvar_t));
2068 ctf_free(name, strlen(name) + 1);
2069 return (ret);
2070 }
2071
2072 ctf_list_append(&cup->cu_vars, cdv);
2073 return (0);
2074 }
2075
2076 /*
2077 * Walk through our set of top-level types and process them.
2078 */
2079 static int
2080 ctf_dwarf_walk_toplevel(ctf_cu_t *cup, Dwarf_Die die)
2081 {
2082 int ret;
2083 Dwarf_Off offset;
2084 Dwarf_Half tag;
2085
2086 if ((ret = ctf_dwarf_offset(cup, die, &offset)) != 0)
2087 return (ret);
2088
2089 if (offset > cup->cu_maxoff) {
2090 (void) snprintf(cup->cu_errbuf, cup->cu_errlen,
2091 "die offset %llu beyond maximum for header %llu\n",
2092 offset, cup->cu_maxoff);
2093 return (ECTF_CONVBKERR);
2094 }
2095
2096 if ((ret = ctf_dwarf_tag(cup, die, &tag)) != 0)
2097 return (ret);
2098
2099 ret = 0;
2100 switch (tag) {
2101 case DW_TAG_subprogram:
2102 ctf_dprintf("top level func\n");
2103 ret = ctf_dwarf_convert_function(cup, die);
2104 break;
2105 case DW_TAG_variable:
2106 ctf_dprintf("top level var\n");
2107 ret = ctf_dwarf_convert_variable(cup, die);
2108 break;
2109 case DW_TAG_lexical_block:
2110 ctf_dprintf("top level block\n");
2111 ret = ctf_dwarf_walk_lexical(cup, die);
2112 break;
2113 case DW_TAG_enumeration_type:
2114 case DW_TAG_structure_type:
2115 case DW_TAG_typedef:
2116 case DW_TAG_union_type:
2117 ctf_dprintf("top level type\n");
2118 ret = ctf_dwarf_convert_type(cup, die, NULL, B_TRUE);
2119 break;
2120 default:
2121 break;
2122 }
2123
2124 return (ret);
2125 }
2126
2127
2128 /*
2129 * We're given a node. At this node we need to convert it and then proceed to
2130 * convert any siblings that are associaed with this die.
2131 */
2132 static int
2133 ctf_dwarf_convert_die(ctf_cu_t *cup, Dwarf_Die die)
2134 {
2135 while (die != NULL) {
2136 int ret;
2137 Dwarf_Die sib;
2138
2139 if ((ret = ctf_dwarf_walk_toplevel(cup, die)) != 0)
2140 return (ret);
2141
2142 if ((ret = ctf_dwarf_sib(cup, die, &sib)) != 0)
2143 return (ret);
2144 die = sib;
2145 }
2146 return (0);
2147 }
2148
2149 static int
2150 ctf_dwarf_fixup_die(ctf_cu_t *cup, boolean_t addpass)
2151 {
2152 ctf_dwmap_t *map;
2153
2154 for (map = avl_first(&cup->cu_map); map != NULL;
2155 map = AVL_NEXT(&cup->cu_map, map)) {
2156 int ret;
2157 if (map->cdm_fix == B_FALSE)
2158 continue;
2159 if ((ret = ctf_dwarf_fixup_sou(cup, map->cdm_die, map->cdm_id,
2160 addpass)) != 0)
2161 return (ret);
2162 }
2163
2164 return (0);
2165 }
2166
2167 /*
2168 * The DWARF information about a symbol and the information in the symbol table
2169 * may not be the same due to symbol reduction that is performed by ld due to a
2170 * mapfile or other such directive. We process weak symbols at a later time.
2171 *
2172 * The following are the rules that we employ:
2173 *
2174 * 1. A DWARF function that is considered exported matches STB_GLOBAL entries
2175 * with the same name.
2176 *
2177 * 2. A DWARF function that is considered exported matches STB_LOCAL entries
2178 * with the same name and the same file. This case may happen due to mapfile
2179 * reduction.
2180 *
2181 * 3. A DWARF function that is not considered exported matches STB_LOCAL entries
2182 * with the same name and the same file.
2183 *
2184 * 4. A DWARF function that has the same name as the symbol table entry, but the
2185 * files do not match. This is considered a 'fuzzy' match. This may also happen
2186 * due to a mapfile reduction. Fuzzy matching is only used when we know that the
2187 * file in question refers to the primary object. This is because when a symbol
2188 * is reduced in a mapfile, it's always going to be tagged as a local value in
2189 * the generated output and it is considered as to belong to the primary file
2190 * which is the first STT_FILE symbol we see.
2191 */
2192 static boolean_t
2193 ctf_dwarf_symbol_match(const char *symtab_file, const char *symtab_name,
2194 uint_t symtab_bind, const char *dwarf_file, const char *dwarf_name,
2195 boolean_t dwarf_global, boolean_t *is_fuzzy)
2196 {
2197 *is_fuzzy = B_FALSE;
2198
2199 if (symtab_bind != STB_LOCAL && symtab_bind != STB_GLOBAL) {
2200 return (B_FALSE);
2201 }
2202
2203 if (strcmp(symtab_name, dwarf_name) != 0) {
2204 return (B_FALSE);
2205 }
2206
2207 if (symtab_bind == STB_GLOBAL) {
2208 return (dwarf_global);
2209 }
2210
2211 if (strcmp(symtab_file, dwarf_file) == 0) {
2212 return (B_TRUE);
2213 }
2214
2215 if (dwarf_global) {
2216 *is_fuzzy = B_TRUE;
2217 return (B_TRUE);
2218 }
2219
2220 return (B_FALSE);
2221 }
2222
2223 static ctf_dwfunc_t *
2224 ctf_dwarf_match_func(ctf_cu_t *cup, const char *file, const char *name,
2225 uint_t bind, boolean_t primary)
2226 {
2227 ctf_dwfunc_t *cdf, *fuzzy = NULL;
2228
2229 if (bind == STB_WEAK)
2230 return (NULL);
2231
2232 if (bind == STB_LOCAL && (file == NULL || cup->cu_name == NULL))
2233 return (NULL);
2234
2235 for (cdf = ctf_list_next(&cup->cu_funcs); cdf != NULL;
2236 cdf = ctf_list_next(cdf)) {
2237 boolean_t is_fuzzy = B_FALSE;
2238
2239 if (ctf_dwarf_symbol_match(file, name, bind, cup->cu_name,
2240 cdf->cdf_name, cdf->cdf_global, &is_fuzzy)) {
2241 if (is_fuzzy) {
2242 if (primary) {
2243 fuzzy = cdf;
2244 }
2245 continue;
2246 } else {
2247 return (cdf);
2248 }
2249 }
2250 }
2251
2252 return (fuzzy);
2253 }
2254
2255 static ctf_dwvar_t *
2256 ctf_dwarf_match_var(ctf_cu_t *cup, const char *file, const char *name,
2257 uint_t bind, boolean_t primary)
2258 {
2259 ctf_dwvar_t *cdv, *fuzzy = NULL;
2260
2261 if (bind == STB_WEAK)
2262 return (NULL);
2263
2264 if (bind == STB_LOCAL && (file == NULL || cup->cu_name == NULL))
2265 return (NULL);
2266
2267 for (cdv = ctf_list_next(&cup->cu_vars); cdv != NULL;
2268 cdv = ctf_list_next(cdv)) {
2269 boolean_t is_fuzzy = B_FALSE;
2270
2271 if (ctf_dwarf_symbol_match(file, name, bind, cup->cu_name,
2272 cdv->cdv_name, cdv->cdv_global, &is_fuzzy)) {
2273 if (is_fuzzy) {
2274 if (primary) {
2275 fuzzy = cdv;
2276 }
2277 } else {
2278 return (cdv);
2279 }
2280 }
2281 }
2282
2283 return (fuzzy);
2284 }
2285
2286 static int
2287 ctf_dwarf_conv_funcvars_cb(const Elf64_Sym *symp, ulong_t idx,
2288 const char *file, const char *name, boolean_t primary, void *arg)
2289 {
2290 int ret;
2291 uint_t bind, type;
2292 ctf_cu_t *cup = arg;
2293
2294 bind = GELF_ST_BIND(symp->st_info);
2295 type = GELF_ST_TYPE(symp->st_info);
2296
2297 /*
2298 * Come back to weak symbols in another pass
2299 */
2300 if (bind == STB_WEAK)
2301 return (0);
2302
2303 if (type == STT_OBJECT) {
2304 ctf_dwvar_t *cdv = ctf_dwarf_match_var(cup, file, name,
2305 bind, primary);
2306 if (cdv == NULL)
2307 return (0);
2308 ret = ctf_add_object(cup->cu_ctfp, idx, cdv->cdv_type);
2309 ctf_dprintf("added object %s->%ld\n", name, cdv->cdv_type);
2310 } else {
2311 ctf_dwfunc_t *cdf = ctf_dwarf_match_func(cup, file, name,
2312 bind, primary);
2313 if (cdf == NULL)
2314 return (0);
2315 ret = ctf_add_function(cup->cu_ctfp, idx, &cdf->cdf_fip,
2316 cdf->cdf_argv);
2317 ctf_dprintf("added function %s\n", name);
2318 }
2319
2320 if (ret == CTF_ERR) {
2321 return (ctf_errno(cup->cu_ctfp));
2322 }
2323
2324 return (0);
2325 }
2326
2327 static int
2328 ctf_dwarf_conv_funcvars(ctf_cu_t *cup)
2329 {
2330 return (ctf_symtab_iter(cup->cu_ctfp, ctf_dwarf_conv_funcvars_cb, cup));
2331 }
2332
2333 /*
2334 * If we have a weak symbol, attempt to find the strong symbol it will resolve
2335 * to. Note: the code where this actually happens is in sym_process() in
2336 * cmd/sgs/libld/common/syms.c
2337 *
2338 * Finding the matching symbol is unfortunately not trivial. For a symbol to be
2339 * a candidate, it must:
2340 *
2341 * - have the same type (function, object)
2342 * - have the same value (address)
2343 * - have the same size
2344 * - not be another weak symbol
2345 * - belong to the same section (checked via section index)
2346 *
2347 * To perform this check, we first iterate over the symbol table. For each weak
2348 * symbol that we encounter, we then do a second walk over the symbol table,
2349 * calling ctf_dwarf_conv_check_weak(). If a symbol matches the above, then it's
2350 * either a local or global symbol. If we find a global symbol then we go with
2351 * it and stop searching for additional matches.
2352 *
2353 * If instead, we find a local symbol, things are more complicated. The first
2354 * thing we do is to try and see if we have file information about both symbols
2355 * (STT_FILE). If they both have file information and it matches, then we treat
2356 * that as a good match and stop searching for additional matches.
2357 *
2358 * Otherwise, this means we have a non-matching file and a local symbol. We
2359 * treat this as a candidate and if we find a better match (one of the two cases
2360 * above), use that instead. There are two different ways this can happen.
2361 * Either this is a completely different symbol, or it's a once-global symbol
2362 * that was scoped to local via a mapfile. In the former case, curfile is
2363 * likely inaccurate since the linker does not preserve the needed curfile in
2364 * the order of the symbol table (see the comments about locally scoped symbols
2365 * in libld's update_osym()). As we can't tell this case from the former one,
2366 * we use this symbol iff no other matching symbol is found.
2367 *
2368 * What we really need here is a SUNW section containing weak<->strong mappings
2369 * that we can consume.
2370 */
2371 typedef struct ctf_dwarf_weak_arg {
2372 const Elf64_Sym *cweak_symp;
2373 const char *cweak_file;
2374 boolean_t cweak_candidate;
2375 ulong_t cweak_idx;
2376 } ctf_dwarf_weak_arg_t;
2377
2378 static int
2379 ctf_dwarf_conv_check_weak(const Elf64_Sym *symp, ulong_t idx, const char *file,
2380 const char *name, boolean_t primary, void *arg)
2381 {
2382 ctf_dwarf_weak_arg_t *cweak = arg;
2383
2384 const Elf64_Sym *wsymp = cweak->cweak_symp;
2385
2386 ctf_dprintf("comparing weak to %s\n", name);
2387
2388 if (GELF_ST_BIND(symp->st_info) == STB_WEAK) {
2389 return (0);
2390 }
2391
2392 if (GELF_ST_TYPE(wsymp->st_info) != GELF_ST_TYPE(symp->st_info)) {
2393 return (0);
2394 }
2395
2396 if (wsymp->st_value != symp->st_value) {
2397 return (0);
2398 }
2399
2400 if (wsymp->st_size != symp->st_size) {
2401 return (0);
2402 }
2403
2404 if (wsymp->st_shndx != symp->st_shndx) {
2405 return (0);
2406 }
2407
2408 /*
2409 * Check if it's a weak candidate.
2410 */
2411 if (GELF_ST_BIND(symp->st_info) == STB_LOCAL &&
2412 (file == NULL || cweak->cweak_file == NULL ||
2413 strcmp(file, cweak->cweak_file) != 0)) {
2414 cweak->cweak_candidate = B_TRUE;
2415 cweak->cweak_idx = idx;
2416 return (0);
2417 }
2418
2419 /*
2420 * Found a match, break.
2421 */
2422 cweak->cweak_idx = idx;
2423 return (1);
2424 }
2425
2426 static int
2427 ctf_dwarf_duplicate_sym(ctf_cu_t *cup, ulong_t idx, ulong_t matchidx)
2428 {
2429 ctf_id_t id = ctf_lookup_by_symbol(cup->cu_ctfp, matchidx);
2430
2431 /*
2432 * If we matched something that for some reason didn't have type data,
2433 * we don't consider that a fatal error and silently swallow it.
2434 */
2435 if (id == CTF_ERR) {
2436 if (ctf_errno(cup->cu_ctfp) == ECTF_NOTYPEDAT)
2437 return (0);
2438 else
2439 return (ctf_errno(cup->cu_ctfp));
2440 }
2441
2442 if (ctf_add_object(cup->cu_ctfp, idx, id) == CTF_ERR)
2443 return (ctf_errno(cup->cu_ctfp));
2444
2445 return (0);
2446 }
2447
2448 static int
2449 ctf_dwarf_duplicate_func(ctf_cu_t *cup, ulong_t idx, ulong_t matchidx)
2450 {
2451 int ret;
2452 ctf_funcinfo_t fip;
2453 ctf_id_t *args = NULL;
2454
2455 if (ctf_func_info(cup->cu_ctfp, matchidx, &fip) == CTF_ERR) {
2456 if (ctf_errno(cup->cu_ctfp) == ECTF_NOFUNCDAT)
2457 return (0);
2458 else
2459 return (ctf_errno(cup->cu_ctfp));
2460 }
2461
2462 if (fip.ctc_argc != 0) {
2463 args = ctf_alloc(sizeof (ctf_id_t) * fip.ctc_argc);
2464 if (args == NULL)
2465 return (ENOMEM);
2466
2467 if (ctf_func_args(cup->cu_ctfp, matchidx, fip.ctc_argc, args) ==
2468 CTF_ERR) {
2469 ctf_free(args, sizeof (ctf_id_t) * fip.ctc_argc);
2470 return (ctf_errno(cup->cu_ctfp));
2471 }
2472 }
2473
2474 ret = ctf_add_function(cup->cu_ctfp, idx, &fip, args);
2475 if (args != NULL)
2476 ctf_free(args, sizeof (ctf_id_t) * fip.ctc_argc);
2477 if (ret == CTF_ERR)
2478 return (ctf_errno(cup->cu_ctfp));
2479
2480 return (0);
2481 }
2482
2483 static int
2484 ctf_dwarf_conv_weaks_cb(const Elf64_Sym *symp, ulong_t idx, const char *file,
2485 const char *name, boolean_t primary, void *arg)
2486 {
2487 int ret, type;
2488 ctf_dwarf_weak_arg_t cweak;
2489 ctf_cu_t *cup = arg;
2490
2491 /*
2492 * We only care about weak symbols.
2493 */
2494 if (GELF_ST_BIND(symp->st_info) != STB_WEAK)
2495 return (0);
2496
2497 type = GELF_ST_TYPE(symp->st_info);
2498 ASSERT(type == STT_OBJECT || type == STT_FUNC);
2499
2500 /*
2501 * For each weak symbol we encounter, we need to do a second iteration
2502 * to try and find a match. We should probably think about other
2503 * techniques to try and save us time in the future.
2504 */
2505 cweak.cweak_symp = symp;
2506 cweak.cweak_file = file;
2507 cweak.cweak_candidate = B_FALSE;
2508 cweak.cweak_idx = 0;
2509
2510 ctf_dprintf("Trying to find weak equiv for %s\n", name);
2511
2512 ret = ctf_symtab_iter(cup->cu_ctfp, ctf_dwarf_conv_check_weak, &cweak);
2513 VERIFY(ret == 0 || ret == 1);
2514
2515 /*
2516 * Nothing was ever found, we're not going to add anything for this
2517 * entry.
2518 */
2519 if (ret == 0 && cweak.cweak_candidate == B_FALSE) {
2520 ctf_dprintf("found no weak match for %s\n", name);
2521 return (0);
2522 }
2523
2524 /*
2525 * Now, finally go and add the type based on the match.
2526 */
2527 ctf_dprintf("matched weak symbol %lu to %lu\n", idx, cweak.cweak_idx);
2528 if (type == STT_OBJECT) {
2529 ret = ctf_dwarf_duplicate_sym(cup, idx, cweak.cweak_idx);
2530 } else {
2531 ret = ctf_dwarf_duplicate_func(cup, idx, cweak.cweak_idx);
2532 }
2533
2534 return (ret);
2535 }
2536
2537 static int
2538 ctf_dwarf_conv_weaks(ctf_cu_t *cup)
2539 {
2540 return (ctf_symtab_iter(cup->cu_ctfp, ctf_dwarf_conv_weaks_cb, cup));
2541 }
2542
2543 /* ARGSUSED */
2544 static int
2545 ctf_dwarf_convert_one(void *arg, void *unused)
2546 {
2547 int ret;
2548 ctf_file_t *dedup;
2549 ctf_cu_t *cup = arg;
2550
2551 ctf_dprintf("converting die: %s\n", cup->cu_name);
2552 ctf_dprintf("max offset: %x\n", cup->cu_maxoff);
2553 VERIFY(cup != NULL);
2554
2555 ret = ctf_dwarf_convert_die(cup, cup->cu_cu);
2556 ctf_dprintf("ctf_dwarf_convert_die (%s) returned %d\n", cup->cu_name,
2557 ret);
2558 if (ret != 0) {
2559 return (ret);
2560 }
2561 if (ctf_update(cup->cu_ctfp) != 0) {
2562 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2563 "failed to update output ctf container"));
2564 }
2565
2566 ret = ctf_dwarf_fixup_die(cup, B_FALSE);
2567 ctf_dprintf("ctf_dwarf_fixup_die (%s) returned %d\n", cup->cu_name,
2568 ret);
2569 if (ret != 0) {
2570 return (ret);
2571 }
2572 if (ctf_update(cup->cu_ctfp) != 0) {
2573 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2574 "failed to update output ctf container"));
2575 }
2576
2577 ret = ctf_dwarf_fixup_die(cup, B_TRUE);
2578 ctf_dprintf("ctf_dwarf_fixup_die (%s) returned %d\n", cup->cu_name,
2579 ret);
2580 if (ret != 0) {
2581 return (ret);
2582 }
2583 if (ctf_update(cup->cu_ctfp) != 0) {
2584 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2585 "failed to update output ctf container"));
2586 }
2587
2588
2589 if ((ret = ctf_dwarf_conv_funcvars(cup)) != 0) {
2590 return (ctf_dwarf_error(cup, NULL, ret,
2591 "failed to convert strong functions and variables"));
2592 }
2593
2594 if (ctf_update(cup->cu_ctfp) != 0) {
2595 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2596 "failed to update output ctf container"));
2597 }
2598
2599 if (cup->cu_doweaks == B_TRUE) {
2600 if ((ret = ctf_dwarf_conv_weaks(cup)) != 0) {
2601 return (ctf_dwarf_error(cup, NULL, ret,
2602 "failed to convert weak functions and variables"));
2603 }
2604
2605 if (ctf_update(cup->cu_ctfp) != 0) {
2606 return (ctf_dwarf_error(cup, cup->cu_ctfp, 0,
2607 "failed to update output ctf container"));
2608 }
2609 }
2610
2611 ctf_phase_dump(cup->cu_ctfp, "pre-dwarf-dedup", cup->cu_name);
2612 ctf_dprintf("adding inputs for dedup\n");
2613 if ((ret = ctf_merge_add(cup->cu_cmh, cup->cu_ctfp)) != 0) {
2614 return (ctf_dwarf_error(cup, NULL, ret,
2615 "failed to add inputs for merge"));
2616 }
2617
2618 ctf_dprintf("starting dedup of %s\n", cup->cu_name);
2619 if ((ret = ctf_merge_dedup(cup->cu_cmh, &dedup)) != 0) {
2620 return (ctf_dwarf_error(cup, NULL, ret,
2621 "failed to deduplicate die"));
2622 }
2623 ctf_close(cup->cu_ctfp);
2624 cup->cu_ctfp = dedup;
2625 ctf_phase_dump(cup->cu_ctfp, "post-dwarf-dedup", cup->cu_name);
2626
2627 return (0);
2628 }
2629
2630 /*
2631 * Note, we expect that if we're returning a ctf_file_t from one of the dies,
2632 * say in the single node case, it's been saved and the entry here has been set
2633 * to NULL, which ctf_close happily ignores.
2634 */
2635 static void
2636 ctf_dwarf_free_die(ctf_cu_t *cup)
2637 {
2638 ctf_dwfunc_t *cdf, *ndf;
2639 ctf_dwvar_t *cdv, *ndv;
2640 ctf_dwbitf_t *cdb, *ndb;
2641 ctf_dwmap_t *map;
2642 void *cookie;
2643 Dwarf_Error derr;
2644
2645 ctf_dprintf("Beginning to free die: %p\n", cup);
2646 cup->cu_elf = NULL;
2647 ctf_dprintf("Trying to free name: %p\n", cup->cu_name);
2648 if (cup->cu_name != NULL)
2649 ctf_free(cup->cu_name, strlen(cup->cu_name) + 1);
2650 ctf_dprintf("Trying to free merge handle: %p\n", cup->cu_cmh);
2651 if (cup->cu_cmh != NULL) {
2652 ctf_merge_fini(cup->cu_cmh);
2653 cup->cu_cmh = NULL;
2654 }
2655
2656 ctf_dprintf("Trying to free functions\n");
2657 for (cdf = ctf_list_next(&cup->cu_funcs); cdf != NULL; cdf = ndf) {
2658 ndf = ctf_list_next(cdf);
2659 ctf_free(cdf->cdf_name, strlen(cdf->cdf_name) + 1);
2660 if (cdf->cdf_fip.ctc_argc != 0) {
2661 ctf_free(cdf->cdf_argv,
2662 sizeof (ctf_id_t) * cdf->cdf_fip.ctc_argc);
2663 }
2664 ctf_free(cdf, sizeof (ctf_dwfunc_t));
2665 }
2666
2667 ctf_dprintf("Trying to free variables\n");
2668 for (cdv = ctf_list_next(&cup->cu_vars); cdv != NULL; cdv = ndv) {
2669 ndv = ctf_list_next(cdv);
2670 ctf_free(cdv->cdv_name, strlen(cdv->cdv_name) + 1);
2671 ctf_free(cdv, sizeof (ctf_dwvar_t));
2672 }
2673
2674 ctf_dprintf("Trying to free bitfields\n");
2675 for (cdb = ctf_list_next(&cup->cu_bitfields); cdb != NULL; cdb = ndb) {
2676 ndb = ctf_list_next(cdb);
2677 ctf_free(cdb, sizeof (ctf_dwbitf_t));
2678 }
2679
2680 ctf_dprintf("Trying to clean up dwarf_t: %p\n", cup->cu_dwarf);
2681 if (cup->cu_dwarf != NULL)
2682 (void) dwarf_finish(cup->cu_dwarf, &derr);
2683 cup->cu_dwarf = NULL;
2684 ctf_close(cup->cu_ctfp);
2685
2686 cookie = NULL;
2687 while ((map = avl_destroy_nodes(&cup->cu_map, &cookie)) != NULL) {
2688 ctf_free(map, sizeof (ctf_dwmap_t));
2689 }
2690 avl_destroy(&cup->cu_map);
2691 cup->cu_errbuf = NULL;
2692 }
2693
2694 static void
2695 ctf_dwarf_free_dies(ctf_cu_t *cdies, int ndies)
2696 {
2697 int i;
2698
2699 ctf_dprintf("Beginning to free dies\n");
2700 for (i = 0; i < ndies; i++) {
2701 ctf_dwarf_free_die(&cdies[i]);
2702 }
2703
2704 ctf_free(cdies, sizeof (ctf_cu_t) * ndies);
2705 }
2706
2707 static int
2708 ctf_dwarf_count_dies(Dwarf_Debug dw, Dwarf_Error *derr, int *ndies,
2709 char *errbuf, size_t errlen)
2710 {
2711 int ret;
2712 Dwarf_Half vers;
2713 Dwarf_Unsigned nexthdr;
2714
2715 while ((ret = dwarf_next_cu_header(dw, NULL, &vers, NULL, NULL,
2716 &nexthdr, derr)) != DW_DLV_NO_ENTRY) {
2717 if (ret != DW_DLV_OK) {
2718 (void) snprintf(errbuf, errlen,
2719 "file does not contain valid DWARF data: %s\n",
2720 dwarf_errmsg(*derr));
2721 return (ECTF_CONVBKERR);
2722 }
2723
2724 if (vers != DWARF_VERSION_TWO) {
2725 (void) snprintf(errbuf, errlen,
2726 "unsupported DWARF version: %d\n", vers);
2727 return (ECTF_CONVBKERR);
2728 }
2729 *ndies = *ndies + 1;
2730 }
2731
2732 return (0);
2733 }
2734
2735 static int
2736 ctf_dwarf_init_die(int fd, Elf *elf, ctf_cu_t *cup, int ndie, char *errbuf,
2737 size_t errlen)
2738 {
2739 int ret;
2740 Dwarf_Unsigned hdrlen, abboff, nexthdr;
2741 Dwarf_Half addrsz;
2742 Dwarf_Unsigned offset = 0;
2743 Dwarf_Error derr;
2744
2745 while ((ret = dwarf_next_cu_header(cup->cu_dwarf, &hdrlen, NULL,
2746 &abboff, &addrsz, &nexthdr, &derr)) != DW_DLV_NO_ENTRY) {
2747 char *name;
2748 Dwarf_Die cu, child;
2749
2750 /* Based on the counting above, we should be good to go */
2751 VERIFY(ret == DW_DLV_OK);
2752 if (ndie > 0) {
2753 ndie--;
2754 offset = nexthdr;
2755 continue;
2756 }
2757
2758 /*
2759 * Compilers are apparently inconsistent. Some emit no DWARF for
2760 * empty files and others emit empty compilation unit.
2761 */
2762 cup->cu_voidtid = CTF_ERR;
2763 cup->cu_longtid = CTF_ERR;
2764 cup->cu_elf = elf;
2765 cup->cu_maxoff = nexthdr - 1;
2766 cup->cu_ctfp = ctf_fdcreate(fd, &ret);
2767 if (cup->cu_ctfp == NULL)
2768 return (ret);
2769
2770 avl_create(&cup->cu_map, ctf_dwmap_comp, sizeof (ctf_dwmap_t),
2771 offsetof(ctf_dwmap_t, cdm_avl));
2772 cup->cu_errbuf = errbuf;
2773 cup->cu_errlen = errlen;
2774 bzero(&cup->cu_vars, sizeof (ctf_list_t));
2775 bzero(&cup->cu_funcs, sizeof (ctf_list_t));
2776 bzero(&cup->cu_bitfields, sizeof (ctf_list_t));
2777
2778 if ((ret = ctf_dwarf_die_elfenc(elf, cup, errbuf,
2779 errlen)) != 0)
2780 return (ret);
2781
2782 if ((ret = ctf_dwarf_sib(cup, NULL, &cu)) != 0)
2783 return (ret);
2784
2785 if (cu == NULL) {
2786 (void) snprintf(errbuf, errlen,
2787 "file does not contain DWARF data");
2788 return (ECTF_CONVNODEBUG);
2789 }
2790
2791 if ((ret = ctf_dwarf_child(cup, cu, &child)) != 0)
2792 return (ret);
2793
2794 if (child == NULL) {
2795 (void) snprintf(errbuf, errlen,
2796 "file does not contain DWARF data");
2797 return (ECTF_CONVNODEBUG);
2798 }
2799
2800 cup->cu_cuoff = offset;
2801 cup->cu_cu = child;
2802
2803 if ((cup->cu_cmh = ctf_merge_init(fd, &ret)) == NULL)
2804 return (ret);
2805
2806 if (ctf_dwarf_string(cup, cu, DW_AT_name, &name) == 0) {
2807 size_t len = strlen(name) + 1;
2808 char *b = basename(name);
2809 cup->cu_name = strdup(b);
2810 ctf_free(name, len);
2811 }
2812 break;
2813 }
2814
2815 return (0);
2816 }
2817
2818 /*
2819 * This is our only recourse to identify a C source file that is missing debug
2820 * info: it will be mentioned as an STT_FILE, but not have a compile unit entry.
2821 * (A traditional ctfmerge works on individual files, so can identify missing
2822 * DWARF more directly, via ctf_has_c_source() on the .o file.)
2823 *
2824 * As we operate on basenames, this can of course miss some cases, but it's
2825 * better than not checking at all.
2826 *
2827 * We explicitly whitelist some CRT components. Failing that, there's always
2828 * the -m option.
2829 */
2830 static boolean_t
2831 c_source_has_debug(const char *file, ctf_cu_t *cus, size_t nr_cus)
2832 {
2833 const char *basename = strrchr(file, '/');
2834
2835 if (basename == NULL)
2836 basename = file;
2837 else
2838 basename++;
2839
2840 if (strcmp(basename, "common-crt.c") == 0 ||
2841 strcmp(basename, "gmon.c") == 0 ||
2842 strcmp(basename, "dlink_init.c") == 0 ||
2843 strcmp(basename, "dlink_common.c") == 0 ||
2844 strncmp(basename, "crt", strlen("crt")) == 0 ||
2845 strncmp(basename, "values-", strlen("values-")) == 0)
2846 return (B_TRUE);
2847
2848 for (size_t i = 0; i < nr_cus; i++) {
2849 if (strcmp(basename, cus[i].cu_name) == 0)
2850 return (B_TRUE);
2851 }
2852
2853 return (B_FALSE);
2854 }
2855
2856 static int
2857 ctf_dwarf_check_missing(ctf_cu_t *cus, size_t nr_cus, Elf *elf,
2858 char *errmsg, size_t errlen)
2859 {
2860 Elf_Scn *scn, *strscn;
2861 Elf_Data *data, *strdata;
2862 GElf_Shdr shdr;
2863 ulong_t i;
2864
2865 scn = NULL;
2866 while ((scn = elf_nextscn(elf, scn)) != NULL) {
2867 if (gelf_getshdr(scn, &shdr) == NULL) {
2868 (void) snprintf(errmsg, errlen,
2869 "failed to get section header: %s\n",
2870 elf_errmsg(elf_errno()));
2871 return (EINVAL);
2872 }
2873
2874 if (shdr.sh_type == SHT_SYMTAB)
2875 break;
2876 }
2877
2878 if (scn == NULL)
2879 return (0);
2880
2881 if ((strscn = elf_getscn(elf, shdr.sh_link)) == NULL) {
2882 (void) snprintf(errmsg, errlen,
2883 "failed to get str section: %s\n",
2884 elf_errmsg(elf_errno()));
2885 return (EINVAL);
2886 }
2887
2888 if ((data = elf_getdata(scn, NULL)) == NULL) {
2889 (void) snprintf(errmsg, errlen, "failed to read section: %s\n",
2890 elf_errmsg(elf_errno()));
2891 return (EINVAL);
2892 }
2893
2894 if ((strdata = elf_getdata(strscn, NULL)) == NULL) {
2895 (void) snprintf(errmsg, errlen,
2896 "failed to read string table: %s\n",
2897 elf_errmsg(elf_errno()));
2898 return (EINVAL);
2899 }
2900
2901 for (i = 0; i < shdr.sh_size / shdr.sh_entsize; i++) {
2902 GElf_Sym sym;
2903 const char *file;
2904 size_t len;
2905
2906 if (gelf_getsym(data, i, &sym) == NULL) {
2907 (void) snprintf(errmsg, errlen,
2908 "failed to read sym %lu: %s\n",
2909 i, elf_errmsg(elf_errno()));
2910 return (EINVAL);
2911 }
2912
2913 if (GELF_ST_TYPE(sym.st_info) != STT_FILE)
2914 continue;
2915
2916 file = (const char *)((uintptr_t)strdata->d_buf + sym.st_name);
2917 len = strlen(file);
2918 if (len < 2 || strncmp(".c", &file[len - 2], 2) != 0)
2919 continue;
2920
2921 if (!c_source_has_debug(file, cus, nr_cus)) {
2922 (void) snprintf(errmsg, errlen,
2923 "file %s is missing debug info\n", file);
2924 return (ECTF_CONVNODEBUG);
2925 }
2926 }
2927
2928 return (0);
2929 }
2930
2931 int
2932 ctf_dwarf_convert(int fd, Elf *elf, uint_t nthrs, uint_t flags,
2933 ctf_file_t **fpp, char *errbuf, size_t errlen)
2934 {
2935 int err, ret, ndies, i;
2936 Dwarf_Debug dw;
2937 Dwarf_Error derr;
2938 ctf_cu_t *cdies = NULL, *cup;
2939 workq_t *wqp = NULL;
2940
2941 *fpp = NULL;
2942
2943 ret = dwarf_elf_init(elf, DW_DLC_READ, NULL, NULL, &dw, &derr);
2944 if (ret != DW_DLV_OK) {
2945 if (ret == DW_DLV_NO_ENTRY ||
2946 dwarf_errno(derr) == DW_DLE_DEBUG_INFO_NULL) {
2947 (void) snprintf(errbuf, errlen,
2948 "file does not contain DWARF data\n");
2949 return (ECTF_CONVNODEBUG);
2950 }
2951
2952 (void) snprintf(errbuf, errlen,
2953 "dwarf_elf_init() failed: %s\n", dwarf_errmsg(derr));
2954 return (ECTF_CONVBKERR);
2955 }
2956
2957 /*
2958 * Iterate over all of the compilation units and create a ctf_cu_t for
2959 * each of them. This is used to determine if we have zero, one, or
2960 * multiple dies to convert. If we have zero, that's an error. If
2961 * there's only one die, that's the simple case. No merge needed and
2962 * only a single Dwarf_Debug as well.
2963 */
2964 ndies = 0;
2965 err = ctf_dwarf_count_dies(dw, &derr, &ndies, errbuf, errlen);
2966
2967 ctf_dprintf("found %d DWARF CUs\n", ndies);
2968
2969 if (ndies == 0) {
2970 (void) snprintf(errbuf, errlen,
2971 "file does not contain DWARF data\n");
2972 return (ECTF_CONVNODEBUG);
2973 }
2974
2975 (void) dwarf_finish(dw, &derr);
2976 cdies = ctf_alloc(sizeof (ctf_cu_t) * ndies);
2977 if (cdies == NULL) {
2978 return (ENOMEM);
2979 }
2980
2981 bzero(cdies, sizeof (ctf_cu_t) * ndies);
2982
2983 for (i = 0; i < ndies; i++) {
2984 cup = &cdies[i];
2985 ret = dwarf_elf_init(elf, DW_DLC_READ, NULL, NULL,
2986 &cup->cu_dwarf, &derr);
2987 if (ret != 0) {
2988 ctf_free(cdies, sizeof (ctf_cu_t) * ndies);
2989 (void) snprintf(errbuf, errlen,
2990 "failed to initialize DWARF: %s\n",
2991 dwarf_errmsg(derr));
2992 return (ECTF_CONVBKERR);
2993 }
2994
2995 err = ctf_dwarf_init_die(fd, elf, cup, i, errbuf, errlen);
2996 if (err != 0)
2997 goto out;
2998
2999 cup->cu_doweaks = ndies > 1 ? B_FALSE : B_TRUE;
3000 }
3001
3002 if (!(flags & CTF_ALLOW_MISSING_DEBUG) &&
3003 (err = ctf_dwarf_check_missing(cdies, ndies,
3004 elf, errbuf, errlen)) != 0)
3005 goto out;
3006
3007 /*
3008 * If we only have one compilation unit, there's no reason to use
3009 * multiple threads, even if the user requested them. After all, they
3010 * just gave us an upper bound.
3011 */
3012 if (ndies == 1)
3013 nthrs = 1;
3014
3015 if (workq_init(&wqp, nthrs) == -1) {
3016 err = errno;
3017 goto out;
3018 }
3019
3020 for (i = 0; i < ndies; i++) {
3021 cup = &cdies[i];
3022 ctf_dprintf("adding cu %s: %p, %x %x\n", cup->cu_name,
3023 cup->cu_cu, cup->cu_cuoff, cup->cu_maxoff);
3024 if (workq_add(wqp, cup) == -1) {
3025 err = errno;
3026 goto out;
3027 }
3028 }
3029
3030 ret = workq_work(wqp, ctf_dwarf_convert_one, NULL, &err);
3031 if (ret == WORKQ_ERROR) {
3032 err = errno;
3033 goto out;
3034 } else if (ret == WORKQ_UERROR) {
3035 ctf_dprintf("internal convert failed: %s\n",
3036 ctf_errmsg(err));
3037 goto out;
3038 }
3039
3040 ctf_dprintf("Determining next phase: have %d CUs\n", ndies);
3041 if (ndies != 1) {
3042 ctf_merge_t *cmp;
3043
3044 cmp = ctf_merge_init(fd, &err);
3045 if (cmp == NULL)
3046 goto out;
3047
3048 ctf_dprintf("setting threads\n");
3049 if ((err = ctf_merge_set_nthreads(cmp, nthrs)) != 0) {
3050 ctf_merge_fini(cmp);
3051 goto out;
3052 }
3053
3054 for (i = 0; i < ndies; i++) {
3055 cup = &cdies[i];
3056 if ((err = ctf_merge_add(cmp, cup->cu_ctfp)) != 0) {
3057 ctf_merge_fini(cmp);
3058 goto out;
3059 }
3060 }
3061
3062 ctf_dprintf("performing merge\n");
3063 err = ctf_merge_merge(cmp, fpp);
3064 if (err != 0) {
3065 ctf_dprintf("failed merge!\n");
3066 *fpp = NULL;
3067 ctf_merge_fini(cmp);
3068 goto out;
3069 }
3070 ctf_merge_fini(cmp);
3071 err = 0;
3072 ctf_dprintf("successfully converted!\n");
3073 } else {
3074 err = 0;
3075 *fpp = cdies->cu_ctfp;
3076 cdies->cu_ctfp = NULL;
3077 ctf_dprintf("successfully converted!\n");
3078 }
3079
3080 out:
3081 workq_fini(wqp);
3082 ctf_dwarf_free_dies(cdies, ndies);
3083 return (err);
3084 }