111
|
1 /* Subroutines needed for unwinding stack frames for exception handling. */
|
145
|
2 /* Copyright (C) 1997-2020 Free Software Foundation, Inc.
|
111
|
3 Contributed by Jason Merrill <jason@cygnus.com>.
|
|
4
|
|
5 This file is part of GCC.
|
|
6
|
|
7 GCC is free software; you can redistribute it and/or modify it under
|
|
8 the terms of the GNU General Public License as published by the Free
|
|
9 Software Foundation; either version 3, or (at your option) any later
|
|
10 version.
|
|
11
|
|
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
|
|
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
15 for more details.
|
|
16
|
|
17 Under Section 7 of GPL version 3, you are granted additional
|
|
18 permissions described in the GCC Runtime Library Exception, version
|
|
19 3.1, as published by the Free Software Foundation.
|
|
20
|
|
21 You should have received a copy of the GNU General Public License and
|
|
22 a copy of the GCC Runtime Library Exception along with this program;
|
|
23 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
|
24 <http://www.gnu.org/licenses/>. */
|
|
25
|
|
26 #ifndef _Unwind_Find_FDE
|
|
27 #include "tconfig.h"
|
|
28 #include "tsystem.h"
|
|
29 #include "coretypes.h"
|
|
30 #include "tm.h"
|
|
31 #include "libgcc_tm.h"
|
|
32 #include "dwarf2.h"
|
|
33 #include "unwind.h"
|
|
34 #define NO_BASE_OF_ENCODED_VALUE
|
|
35 #include "unwind-pe.h"
|
|
36 #include "unwind-dw2-fde.h"
|
|
37 #include "gthr.h"
|
|
38 #else
|
|
39 #if (defined(__GTHREAD_MUTEX_INIT) || defined(__GTHREAD_MUTEX_INIT_FUNCTION)) \
|
|
40 && defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)
|
|
41 #define ATOMIC_FDE_FAST_PATH 1
|
|
42 #endif
|
|
43 #endif
|
|
44
|
|
45 /* The unseen_objects list contains objects that have been registered
|
|
46 but not yet categorized in any way. The seen_objects list has had
|
|
47 its pc_begin and count fields initialized at minimum, and is sorted
|
|
48 by decreasing value of pc_begin. */
|
|
49 static struct object *unseen_objects;
|
|
50 static struct object *seen_objects;
|
|
51 #ifdef ATOMIC_FDE_FAST_PATH
|
|
52 static int any_objects_registered;
|
|
53 #endif
|
|
54
|
|
55 #ifdef __GTHREAD_MUTEX_INIT
|
|
56 static __gthread_mutex_t object_mutex = __GTHREAD_MUTEX_INIT;
|
|
57 #define init_object_mutex_once()
|
|
58 #else
|
|
59 #ifdef __GTHREAD_MUTEX_INIT_FUNCTION
|
|
60 static __gthread_mutex_t object_mutex;
|
|
61
|
|
62 static void
|
|
63 init_object_mutex (void)
|
|
64 {
|
|
65 __GTHREAD_MUTEX_INIT_FUNCTION (&object_mutex);
|
|
66 }
|
|
67
|
|
68 static void
|
|
69 init_object_mutex_once (void)
|
|
70 {
|
|
71 static __gthread_once_t once = __GTHREAD_ONCE_INIT;
|
|
72 __gthread_once (&once, init_object_mutex);
|
|
73 }
|
|
74 #else
|
|
75 /* ??? Several targets include this file with stubbing parts of gthr.h
|
|
76 and expect no locking to be done. */
|
|
77 #define init_object_mutex_once()
|
|
78 static __gthread_mutex_t object_mutex;
|
|
79 #endif
|
|
80 #endif
|
|
81
|
|
82 /* Called from crtbegin.o to register the unwind info for an object. */
|
|
83
|
|
84 void
|
|
85 __register_frame_info_bases (const void *begin, struct object *ob,
|
|
86 void *tbase, void *dbase)
|
|
87 {
|
|
88 /* If .eh_frame is empty, don't register at all. */
|
|
89 if ((const uword *) begin == 0 || *(const uword *) begin == 0)
|
|
90 return;
|
|
91
|
|
92 ob->pc_begin = (void *)-1;
|
|
93 ob->tbase = tbase;
|
|
94 ob->dbase = dbase;
|
|
95 ob->u.single = begin;
|
|
96 ob->s.i = 0;
|
|
97 ob->s.b.encoding = DW_EH_PE_omit;
|
|
98 #ifdef DWARF2_OBJECT_END_PTR_EXTENSION
|
|
99 ob->fde_end = NULL;
|
|
100 #endif
|
|
101
|
|
102 init_object_mutex_once ();
|
|
103 __gthread_mutex_lock (&object_mutex);
|
|
104
|
|
105 ob->next = unseen_objects;
|
|
106 unseen_objects = ob;
|
|
107 #ifdef ATOMIC_FDE_FAST_PATH
|
|
108 /* Set flag that at least one library has registered FDEs.
|
|
109 Use relaxed MO here, it is up to the app to ensure that the library
|
|
110 loading/initialization happens-before using that library in other
|
|
111 threads (in particular unwinding with that library's functions
|
|
112 appearing in the backtraces). Calling that library's functions
|
|
113 without waiting for the library to initialize would be racy. */
|
|
114 if (!any_objects_registered)
|
|
115 __atomic_store_n (&any_objects_registered, 1, __ATOMIC_RELAXED);
|
|
116 #endif
|
|
117
|
|
118 __gthread_mutex_unlock (&object_mutex);
|
|
119 }
|
|
120
|
|
121 void
|
|
122 __register_frame_info (const void *begin, struct object *ob)
|
|
123 {
|
|
124 __register_frame_info_bases (begin, ob, 0, 0);
|
|
125 }
|
|
126
|
|
127 void
|
|
128 __register_frame (void *begin)
|
|
129 {
|
|
130 struct object *ob;
|
|
131
|
|
132 /* If .eh_frame is empty, don't register at all. */
|
|
133 if (*(uword *) begin == 0)
|
|
134 return;
|
|
135
|
|
136 ob = malloc (sizeof (struct object));
|
|
137 __register_frame_info (begin, ob);
|
|
138 }
|
|
139
|
|
140 /* Similar, but BEGIN is actually a pointer to a table of unwind entries
|
|
141 for different translation units. Called from the file generated by
|
|
142 collect2. */
|
|
143
|
|
144 void
|
|
145 __register_frame_info_table_bases (void *begin, struct object *ob,
|
|
146 void *tbase, void *dbase)
|
|
147 {
|
|
148 ob->pc_begin = (void *)-1;
|
|
149 ob->tbase = tbase;
|
|
150 ob->dbase = dbase;
|
|
151 ob->u.array = begin;
|
|
152 ob->s.i = 0;
|
|
153 ob->s.b.from_array = 1;
|
|
154 ob->s.b.encoding = DW_EH_PE_omit;
|
|
155
|
|
156 init_object_mutex_once ();
|
|
157 __gthread_mutex_lock (&object_mutex);
|
|
158
|
|
159 ob->next = unseen_objects;
|
|
160 unseen_objects = ob;
|
|
161 #ifdef ATOMIC_FDE_FAST_PATH
|
|
162 /* Set flag that at least one library has registered FDEs.
|
|
163 Use relaxed MO here, it is up to the app to ensure that the library
|
|
164 loading/initialization happens-before using that library in other
|
|
165 threads (in particular unwinding with that library's functions
|
|
166 appearing in the backtraces). Calling that library's functions
|
|
167 without waiting for the library to initialize would be racy. */
|
|
168 if (!any_objects_registered)
|
|
169 __atomic_store_n (&any_objects_registered, 1, __ATOMIC_RELAXED);
|
|
170 #endif
|
|
171
|
|
172 __gthread_mutex_unlock (&object_mutex);
|
|
173 }
|
|
174
|
|
175 void
|
|
176 __register_frame_info_table (void *begin, struct object *ob)
|
|
177 {
|
|
178 __register_frame_info_table_bases (begin, ob, 0, 0);
|
|
179 }
|
|
180
|
|
181 void
|
|
182 __register_frame_table (void *begin)
|
|
183 {
|
|
184 struct object *ob = malloc (sizeof (struct object));
|
|
185 __register_frame_info_table (begin, ob);
|
|
186 }
|
|
187
|
|
188 /* Called from crtbegin.o to deregister the unwind info for an object. */
|
|
189 /* ??? Glibc has for a while now exported __register_frame_info and
|
|
190 __deregister_frame_info. If we call __register_frame_info_bases
|
|
191 from crtbegin (wherein it is declared weak), and this object does
|
|
192 not get pulled from libgcc.a for other reasons, then the
|
|
193 invocation of __deregister_frame_info will be resolved from glibc.
|
|
194 Since the registration did not happen there, we'll die.
|
|
195
|
|
196 Therefore, declare a new deregistration entry point that does the
|
|
197 exact same thing, but will resolve to the same library as
|
|
198 implements __register_frame_info_bases. */
|
|
199
|
|
200 void *
|
|
201 __deregister_frame_info_bases (const void *begin)
|
|
202 {
|
|
203 struct object **p;
|
|
204 struct object *ob = 0;
|
|
205
|
|
206 /* If .eh_frame is empty, we haven't registered. */
|
|
207 if ((const uword *) begin == 0 || *(const uword *) begin == 0)
|
|
208 return ob;
|
|
209
|
|
210 init_object_mutex_once ();
|
|
211 __gthread_mutex_lock (&object_mutex);
|
|
212
|
|
213 for (p = &unseen_objects; *p ; p = &(*p)->next)
|
|
214 if ((*p)->u.single == begin)
|
|
215 {
|
|
216 ob = *p;
|
|
217 *p = ob->next;
|
|
218 goto out;
|
|
219 }
|
|
220
|
|
221 for (p = &seen_objects; *p ; p = &(*p)->next)
|
|
222 if ((*p)->s.b.sorted)
|
|
223 {
|
|
224 if ((*p)->u.sort->orig_data == begin)
|
|
225 {
|
|
226 ob = *p;
|
|
227 *p = ob->next;
|
|
228 free (ob->u.sort);
|
|
229 goto out;
|
|
230 }
|
|
231 }
|
|
232 else
|
|
233 {
|
|
234 if ((*p)->u.single == begin)
|
|
235 {
|
|
236 ob = *p;
|
|
237 *p = ob->next;
|
|
238 goto out;
|
|
239 }
|
|
240 }
|
|
241
|
|
242 out:
|
|
243 __gthread_mutex_unlock (&object_mutex);
|
|
244 gcc_assert (ob);
|
|
245 return (void *) ob;
|
|
246 }
|
|
247
|
|
248 void *
|
|
249 __deregister_frame_info (const void *begin)
|
|
250 {
|
|
251 return __deregister_frame_info_bases (begin);
|
|
252 }
|
|
253
|
|
254 void
|
|
255 __deregister_frame (void *begin)
|
|
256 {
|
|
257 /* If .eh_frame is empty, we haven't registered. */
|
|
258 if (*(uword *) begin != 0)
|
|
259 free (__deregister_frame_info (begin));
|
|
260 }
|
|
261
|
|
262
|
|
263 /* Like base_of_encoded_value, but take the base from a struct object
|
|
264 instead of an _Unwind_Context. */
|
|
265
|
|
266 static _Unwind_Ptr
|
|
267 base_from_object (unsigned char encoding, struct object *ob)
|
|
268 {
|
|
269 if (encoding == DW_EH_PE_omit)
|
|
270 return 0;
|
|
271
|
|
272 switch (encoding & 0x70)
|
|
273 {
|
|
274 case DW_EH_PE_absptr:
|
|
275 case DW_EH_PE_pcrel:
|
|
276 case DW_EH_PE_aligned:
|
|
277 return 0;
|
|
278
|
|
279 case DW_EH_PE_textrel:
|
|
280 return (_Unwind_Ptr) ob->tbase;
|
|
281 case DW_EH_PE_datarel:
|
|
282 return (_Unwind_Ptr) ob->dbase;
|
|
283 default:
|
|
284 gcc_unreachable ();
|
|
285 }
|
|
286 }
|
|
287
|
|
288 /* Return the FDE pointer encoding from the CIE. */
|
|
289 /* ??? This is a subset of extract_cie_info from unwind-dw2.c. */
|
|
290
|
|
291 static int
|
|
292 get_cie_encoding (const struct dwarf_cie *cie)
|
|
293 {
|
|
294 const unsigned char *aug, *p;
|
|
295 _Unwind_Ptr dummy;
|
|
296 _uleb128_t utmp;
|
|
297 _sleb128_t stmp;
|
|
298
|
|
299 aug = cie->augmentation;
|
|
300 p = aug + strlen ((const char *)aug) + 1; /* Skip the augmentation string. */
|
|
301 if (__builtin_expect (cie->version >= 4, 0))
|
|
302 {
|
|
303 if (p[0] != sizeof (void *) || p[1] != 0)
|
|
304 return DW_EH_PE_omit; /* We are not prepared to handle unexpected
|
|
305 address sizes or segment selectors. */
|
|
306 p += 2; /* Skip address size and segment size. */
|
|
307 }
|
|
308
|
|
309 if (aug[0] != 'z')
|
|
310 return DW_EH_PE_absptr;
|
|
311
|
|
312 p = read_uleb128 (p, &utmp); /* Skip code alignment. */
|
|
313 p = read_sleb128 (p, &stmp); /* Skip data alignment. */
|
|
314 if (cie->version == 1) /* Skip return address column. */
|
|
315 p++;
|
|
316 else
|
|
317 p = read_uleb128 (p, &utmp);
|
|
318
|
|
319 aug++; /* Skip 'z' */
|
|
320 p = read_uleb128 (p, &utmp); /* Skip augmentation length. */
|
|
321 while (1)
|
|
322 {
|
|
323 /* This is what we're looking for. */
|
|
324 if (*aug == 'R')
|
|
325 return *p;
|
|
326 /* Personality encoding and pointer. */
|
|
327 else if (*aug == 'P')
|
|
328 {
|
|
329 /* ??? Avoid dereferencing indirect pointers, since we're
|
|
330 faking the base address. Gotta keep DW_EH_PE_aligned
|
|
331 intact, however. */
|
|
332 p = read_encoded_value_with_base (*p & 0x7F, 0, p + 1, &dummy);
|
|
333 }
|
|
334 /* LSDA encoding. */
|
|
335 else if (*aug == 'L')
|
|
336 p++;
|
145
|
337 /* aarch64 b-key pointer authentication. */
|
|
338 else if (*aug == 'B')
|
|
339 p++;
|
111
|
340 /* Otherwise end of string, or unknown augmentation. */
|
|
341 else
|
|
342 return DW_EH_PE_absptr;
|
|
343 aug++;
|
|
344 }
|
|
345 }
|
|
346
|
|
347 static inline int
|
|
348 get_fde_encoding (const struct dwarf_fde *f)
|
|
349 {
|
|
350 return get_cie_encoding (get_cie (f));
|
|
351 }
|
|
352
|
|
353
|
|
354 /* Sorting an array of FDEs by address.
|
|
355 (Ideally we would have the linker sort the FDEs so we don't have to do
|
|
356 it at run time. But the linkers are not yet prepared for this.) */
|
|
357
|
|
358 /* Comparison routines. Three variants of increasing complexity. */
|
|
359
|
|
360 static int
|
|
361 fde_unencoded_compare (struct object *ob __attribute__((unused)),
|
|
362 const fde *x, const fde *y)
|
|
363 {
|
|
364 _Unwind_Ptr x_ptr, y_ptr;
|
|
365 memcpy (&x_ptr, x->pc_begin, sizeof (_Unwind_Ptr));
|
|
366 memcpy (&y_ptr, y->pc_begin, sizeof (_Unwind_Ptr));
|
|
367
|
|
368 if (x_ptr > y_ptr)
|
|
369 return 1;
|
|
370 if (x_ptr < y_ptr)
|
|
371 return -1;
|
|
372 return 0;
|
|
373 }
|
|
374
|
|
375 static int
|
|
376 fde_single_encoding_compare (struct object *ob, const fde *x, const fde *y)
|
|
377 {
|
|
378 _Unwind_Ptr base, x_ptr, y_ptr;
|
|
379
|
|
380 base = base_from_object (ob->s.b.encoding, ob);
|
|
381 read_encoded_value_with_base (ob->s.b.encoding, base, x->pc_begin, &x_ptr);
|
|
382 read_encoded_value_with_base (ob->s.b.encoding, base, y->pc_begin, &y_ptr);
|
|
383
|
|
384 if (x_ptr > y_ptr)
|
|
385 return 1;
|
|
386 if (x_ptr < y_ptr)
|
|
387 return -1;
|
|
388 return 0;
|
|
389 }
|
|
390
|
|
391 static int
|
|
392 fde_mixed_encoding_compare (struct object *ob, const fde *x, const fde *y)
|
|
393 {
|
|
394 int x_encoding, y_encoding;
|
|
395 _Unwind_Ptr x_ptr, y_ptr;
|
|
396
|
|
397 x_encoding = get_fde_encoding (x);
|
|
398 read_encoded_value_with_base (x_encoding, base_from_object (x_encoding, ob),
|
|
399 x->pc_begin, &x_ptr);
|
|
400
|
|
401 y_encoding = get_fde_encoding (y);
|
|
402 read_encoded_value_with_base (y_encoding, base_from_object (y_encoding, ob),
|
|
403 y->pc_begin, &y_ptr);
|
|
404
|
|
405 if (x_ptr > y_ptr)
|
|
406 return 1;
|
|
407 if (x_ptr < y_ptr)
|
|
408 return -1;
|
|
409 return 0;
|
|
410 }
|
|
411
|
|
412 typedef int (*fde_compare_t) (struct object *, const fde *, const fde *);
|
|
413
|
|
414
|
|
415 /* This is a special mix of insertion sort and heap sort, optimized for
|
|
416 the data sets that actually occur. They look like
|
|
417 101 102 103 127 128 105 108 110 190 111 115 119 125 160 126 129 130.
|
|
418 I.e. a linearly increasing sequence (coming from functions in the text
|
|
419 section), with additionally a few unordered elements (coming from functions
|
|
420 in gnu_linkonce sections) whose values are higher than the values in the
|
|
421 surrounding linear sequence (but not necessarily higher than the values
|
|
422 at the end of the linear sequence!).
|
|
423 The worst-case total run time is O(N) + O(n log (n)), where N is the
|
|
424 total number of FDEs and n is the number of erratic ones. */
|
|
425
|
|
426 struct fde_accumulator
|
|
427 {
|
|
428 struct fde_vector *linear;
|
|
429 struct fde_vector *erratic;
|
|
430 };
|
|
431
|
|
432 static inline int
|
|
433 start_fde_sort (struct fde_accumulator *accu, size_t count)
|
|
434 {
|
|
435 size_t size;
|
|
436 if (! count)
|
|
437 return 0;
|
|
438
|
|
439 size = sizeof (struct fde_vector) + sizeof (const fde *) * count;
|
|
440 if ((accu->linear = malloc (size)))
|
|
441 {
|
|
442 accu->linear->count = 0;
|
|
443 if ((accu->erratic = malloc (size)))
|
|
444 accu->erratic->count = 0;
|
|
445 return 1;
|
|
446 }
|
|
447 else
|
|
448 return 0;
|
|
449 }
|
|
450
|
|
451 static inline void
|
|
452 fde_insert (struct fde_accumulator *accu, const fde *this_fde)
|
|
453 {
|
|
454 if (accu->linear)
|
|
455 accu->linear->array[accu->linear->count++] = this_fde;
|
|
456 }
|
|
457
|
|
458 /* Split LINEAR into a linear sequence with low values and an erratic
|
|
459 sequence with high values, put the linear one (of longest possible
|
|
460 length) into LINEAR and the erratic one into ERRATIC. This is O(N).
|
|
461
|
|
462 Because the longest linear sequence we are trying to locate within the
|
|
463 incoming LINEAR array can be interspersed with (high valued) erratic
|
|
464 entries. We construct a chain indicating the sequenced entries.
|
|
465 To avoid having to allocate this chain, we overlay it onto the space of
|
|
466 the ERRATIC array during construction. A final pass iterates over the
|
|
467 chain to determine what should be placed in the ERRATIC array, and
|
|
468 what is the linear sequence. This overlay is safe from aliasing. */
|
|
469
|
|
470 static inline void
|
|
471 fde_split (struct object *ob, fde_compare_t fde_compare,
|
|
472 struct fde_vector *linear, struct fde_vector *erratic)
|
|
473 {
|
|
474 static const fde *marker;
|
|
475 size_t count = linear->count;
|
|
476 const fde *const *chain_end = ▮
|
|
477 size_t i, j, k;
|
|
478
|
|
479 /* This should optimize out, but it is wise to make sure this assumption
|
|
480 is correct. Should these have different sizes, we cannot cast between
|
|
481 them and the overlaying onto ERRATIC will not work. */
|
|
482 gcc_assert (sizeof (const fde *) == sizeof (const fde **));
|
|
483
|
|
484 for (i = 0; i < count; i++)
|
|
485 {
|
|
486 const fde *const *probe;
|
|
487
|
|
488 for (probe = chain_end;
|
|
489 probe != &marker && fde_compare (ob, linear->array[i], *probe) < 0;
|
|
490 probe = chain_end)
|
|
491 {
|
|
492 chain_end = (const fde *const*) erratic->array[probe - linear->array];
|
|
493 erratic->array[probe - linear->array] = NULL;
|
|
494 }
|
|
495 erratic->array[i] = (const fde *) chain_end;
|
|
496 chain_end = &linear->array[i];
|
|
497 }
|
|
498
|
|
499 /* Each entry in LINEAR which is part of the linear sequence we have
|
|
500 discovered will correspond to a non-NULL entry in the chain we built in
|
|
501 the ERRATIC array. */
|
|
502 for (i = j = k = 0; i < count; i++)
|
|
503 if (erratic->array[i])
|
|
504 linear->array[j++] = linear->array[i];
|
|
505 else
|
|
506 erratic->array[k++] = linear->array[i];
|
|
507 linear->count = j;
|
|
508 erratic->count = k;
|
|
509 }
|
|
510
|
|
511 #define SWAP(x,y) do { const fde * tmp = x; x = y; y = tmp; } while (0)
|
|
512
|
|
513 /* Convert a semi-heap to a heap. A semi-heap is a heap except possibly
|
|
514 for the first (root) node; push it down to its rightful place. */
|
|
515
|
|
516 static void
|
|
517 frame_downheap (struct object *ob, fde_compare_t fde_compare, const fde **a,
|
|
518 int lo, int hi)
|
|
519 {
|
|
520 int i, j;
|
|
521
|
|
522 for (i = lo, j = 2*i+1;
|
|
523 j < hi;
|
|
524 j = 2*i+1)
|
|
525 {
|
|
526 if (j+1 < hi && fde_compare (ob, a[j], a[j+1]) < 0)
|
|
527 ++j;
|
|
528
|
|
529 if (fde_compare (ob, a[i], a[j]) < 0)
|
|
530 {
|
|
531 SWAP (a[i], a[j]);
|
|
532 i = j;
|
|
533 }
|
|
534 else
|
|
535 break;
|
|
536 }
|
|
537 }
|
|
538
|
|
539 /* This is O(n log(n)). BSD/OS defines heapsort in stdlib.h, so we must
|
|
540 use a name that does not conflict. */
|
|
541
|
|
542 static void
|
|
543 frame_heapsort (struct object *ob, fde_compare_t fde_compare,
|
|
544 struct fde_vector *erratic)
|
|
545 {
|
|
546 /* For a description of this algorithm, see:
|
|
547 Samuel P. Harbison, Guy L. Steele Jr.: C, a reference manual, 2nd ed.,
|
|
548 p. 60-61. */
|
|
549 const fde ** a = erratic->array;
|
|
550 /* A portion of the array is called a "heap" if for all i>=0:
|
|
551 If i and 2i+1 are valid indices, then a[i] >= a[2i+1].
|
|
552 If i and 2i+2 are valid indices, then a[i] >= a[2i+2]. */
|
|
553 size_t n = erratic->count;
|
|
554 int m;
|
|
555
|
|
556 /* Expand our heap incrementally from the end of the array, heapifying
|
|
557 each resulting semi-heap as we go. After each step, a[m] is the top
|
|
558 of a heap. */
|
|
559 for (m = n/2-1; m >= 0; --m)
|
|
560 frame_downheap (ob, fde_compare, a, m, n);
|
|
561
|
|
562 /* Shrink our heap incrementally from the end of the array, first
|
|
563 swapping out the largest element a[0] and then re-heapifying the
|
|
564 resulting semi-heap. After each step, a[0..m) is a heap. */
|
|
565 for (m = n-1; m >= 1; --m)
|
|
566 {
|
|
567 SWAP (a[0], a[m]);
|
|
568 frame_downheap (ob, fde_compare, a, 0, m);
|
|
569 }
|
|
570 #undef SWAP
|
|
571 }
|
|
572
|
|
573 /* Merge V1 and V2, both sorted, and put the result into V1. */
|
|
574 static inline void
|
|
575 fde_merge (struct object *ob, fde_compare_t fde_compare,
|
|
576 struct fde_vector *v1, struct fde_vector *v2)
|
|
577 {
|
|
578 size_t i1, i2;
|
|
579 const fde * fde2;
|
|
580
|
|
581 i2 = v2->count;
|
|
582 if (i2 > 0)
|
|
583 {
|
|
584 i1 = v1->count;
|
|
585 do
|
|
586 {
|
|
587 i2--;
|
|
588 fde2 = v2->array[i2];
|
|
589 while (i1 > 0 && fde_compare (ob, v1->array[i1-1], fde2) > 0)
|
|
590 {
|
|
591 v1->array[i1+i2] = v1->array[i1-1];
|
|
592 i1--;
|
|
593 }
|
|
594 v1->array[i1+i2] = fde2;
|
|
595 }
|
|
596 while (i2 > 0);
|
|
597 v1->count += v2->count;
|
|
598 }
|
|
599 }
|
|
600
|
|
601 static inline void
|
|
602 end_fde_sort (struct object *ob, struct fde_accumulator *accu, size_t count)
|
|
603 {
|
|
604 fde_compare_t fde_compare;
|
|
605
|
|
606 gcc_assert (!accu->linear || accu->linear->count == count);
|
|
607
|
|
608 if (ob->s.b.mixed_encoding)
|
|
609 fde_compare = fde_mixed_encoding_compare;
|
|
610 else if (ob->s.b.encoding == DW_EH_PE_absptr)
|
|
611 fde_compare = fde_unencoded_compare;
|
|
612 else
|
|
613 fde_compare = fde_single_encoding_compare;
|
|
614
|
|
615 if (accu->erratic)
|
|
616 {
|
|
617 fde_split (ob, fde_compare, accu->linear, accu->erratic);
|
|
618 gcc_assert (accu->linear->count + accu->erratic->count == count);
|
|
619 frame_heapsort (ob, fde_compare, accu->erratic);
|
|
620 fde_merge (ob, fde_compare, accu->linear, accu->erratic);
|
|
621 free (accu->erratic);
|
|
622 }
|
|
623 else
|
|
624 {
|
|
625 /* We've not managed to malloc an erratic array,
|
|
626 so heap sort in the linear one. */
|
|
627 frame_heapsort (ob, fde_compare, accu->linear);
|
|
628 }
|
|
629 }
|
|
630
|
|
631
|
|
632 /* Update encoding, mixed_encoding, and pc_begin for OB for the
|
|
633 fde array beginning at THIS_FDE. Return the number of fdes
|
|
634 encountered along the way. */
|
|
635
|
|
636 static size_t
|
|
637 classify_object_over_fdes (struct object *ob, const fde *this_fde)
|
|
638 {
|
|
639 const struct dwarf_cie *last_cie = 0;
|
|
640 size_t count = 0;
|
|
641 int encoding = DW_EH_PE_absptr;
|
|
642 _Unwind_Ptr base = 0;
|
|
643
|
|
644 for (; ! last_fde (ob, this_fde); this_fde = next_fde (this_fde))
|
|
645 {
|
|
646 const struct dwarf_cie *this_cie;
|
|
647 _Unwind_Ptr mask, pc_begin;
|
|
648
|
|
649 /* Skip CIEs. */
|
|
650 if (this_fde->CIE_delta == 0)
|
|
651 continue;
|
|
652
|
|
653 /* Determine the encoding for this FDE. Note mixed encoded
|
|
654 objects for later. */
|
|
655 this_cie = get_cie (this_fde);
|
|
656 if (this_cie != last_cie)
|
|
657 {
|
|
658 last_cie = this_cie;
|
|
659 encoding = get_cie_encoding (this_cie);
|
|
660 if (encoding == DW_EH_PE_omit)
|
|
661 return -1;
|
|
662 base = base_from_object (encoding, ob);
|
|
663 if (ob->s.b.encoding == DW_EH_PE_omit)
|
|
664 ob->s.b.encoding = encoding;
|
|
665 else if (ob->s.b.encoding != encoding)
|
|
666 ob->s.b.mixed_encoding = 1;
|
|
667 }
|
|
668
|
|
669 read_encoded_value_with_base (encoding, base, this_fde->pc_begin,
|
|
670 &pc_begin);
|
|
671
|
|
672 /* Take care to ignore link-once functions that were removed.
|
|
673 In these cases, the function address will be NULL, but if
|
|
674 the encoding is smaller than a pointer a true NULL may not
|
|
675 be representable. Assume 0 in the representable bits is NULL. */
|
|
676 mask = size_of_encoded_value (encoding);
|
|
677 if (mask < sizeof (void *))
|
|
678 mask = (((_Unwind_Ptr) 1) << (mask << 3)) - 1;
|
|
679 else
|
|
680 mask = -1;
|
|
681
|
|
682 if ((pc_begin & mask) == 0)
|
|
683 continue;
|
|
684
|
|
685 count += 1;
|
|
686 if ((void *) pc_begin < ob->pc_begin)
|
|
687 ob->pc_begin = (void *) pc_begin;
|
|
688 }
|
|
689
|
|
690 return count;
|
|
691 }
|
|
692
|
|
693 static void
|
|
694 add_fdes (struct object *ob, struct fde_accumulator *accu, const fde *this_fde)
|
|
695 {
|
|
696 const struct dwarf_cie *last_cie = 0;
|
|
697 int encoding = ob->s.b.encoding;
|
|
698 _Unwind_Ptr base = base_from_object (ob->s.b.encoding, ob);
|
|
699
|
|
700 for (; ! last_fde (ob, this_fde); this_fde = next_fde (this_fde))
|
|
701 {
|
|
702 const struct dwarf_cie *this_cie;
|
|
703
|
|
704 /* Skip CIEs. */
|
|
705 if (this_fde->CIE_delta == 0)
|
|
706 continue;
|
|
707
|
|
708 if (ob->s.b.mixed_encoding)
|
|
709 {
|
|
710 /* Determine the encoding for this FDE. Note mixed encoded
|
|
711 objects for later. */
|
|
712 this_cie = get_cie (this_fde);
|
|
713 if (this_cie != last_cie)
|
|
714 {
|
|
715 last_cie = this_cie;
|
|
716 encoding = get_cie_encoding (this_cie);
|
|
717 base = base_from_object (encoding, ob);
|
|
718 }
|
|
719 }
|
|
720
|
|
721 if (encoding == DW_EH_PE_absptr)
|
|
722 {
|
|
723 _Unwind_Ptr ptr;
|
|
724 memcpy (&ptr, this_fde->pc_begin, sizeof (_Unwind_Ptr));
|
|
725 if (ptr == 0)
|
|
726 continue;
|
|
727 }
|
|
728 else
|
|
729 {
|
|
730 _Unwind_Ptr pc_begin, mask;
|
|
731
|
|
732 read_encoded_value_with_base (encoding, base, this_fde->pc_begin,
|
|
733 &pc_begin);
|
|
734
|
|
735 /* Take care to ignore link-once functions that were removed.
|
|
736 In these cases, the function address will be NULL, but if
|
|
737 the encoding is smaller than a pointer a true NULL may not
|
|
738 be representable. Assume 0 in the representable bits is NULL. */
|
|
739 mask = size_of_encoded_value (encoding);
|
|
740 if (mask < sizeof (void *))
|
|
741 mask = (((_Unwind_Ptr) 1) << (mask << 3)) - 1;
|
|
742 else
|
|
743 mask = -1;
|
|
744
|
|
745 if ((pc_begin & mask) == 0)
|
|
746 continue;
|
|
747 }
|
|
748
|
|
749 fde_insert (accu, this_fde);
|
|
750 }
|
|
751 }
|
|
752
|
|
753 /* Set up a sorted array of pointers to FDEs for a loaded object. We
|
|
754 count up the entries before allocating the array because it's likely to
|
|
755 be faster. We can be called multiple times, should we have failed to
|
|
756 allocate a sorted fde array on a previous occasion. */
|
|
757
|
|
758 static inline void
|
|
759 init_object (struct object* ob)
|
|
760 {
|
|
761 struct fde_accumulator accu;
|
|
762 size_t count;
|
|
763
|
|
764 count = ob->s.b.count;
|
|
765 if (count == 0)
|
|
766 {
|
|
767 if (ob->s.b.from_array)
|
|
768 {
|
|
769 fde **p = ob->u.array;
|
|
770 for (count = 0; *p; ++p)
|
|
771 {
|
|
772 size_t cur_count = classify_object_over_fdes (ob, *p);
|
|
773 if (cur_count == (size_t) -1)
|
|
774 goto unhandled_fdes;
|
|
775 count += cur_count;
|
|
776 }
|
|
777 }
|
|
778 else
|
|
779 {
|
|
780 count = classify_object_over_fdes (ob, ob->u.single);
|
|
781 if (count == (size_t) -1)
|
|
782 {
|
|
783 static const fde terminator;
|
|
784 unhandled_fdes:
|
|
785 ob->s.i = 0;
|
|
786 ob->s.b.encoding = DW_EH_PE_omit;
|
|
787 ob->u.single = &terminator;
|
|
788 return;
|
|
789 }
|
|
790 }
|
|
791
|
|
792 /* The count field we have in the main struct object is somewhat
|
|
793 limited, but should suffice for virtually all cases. If the
|
|
794 counted value doesn't fit, re-write a zero. The worst that
|
|
795 happens is that we re-count next time -- admittedly non-trivial
|
|
796 in that this implies some 2M fdes, but at least we function. */
|
|
797 ob->s.b.count = count;
|
|
798 if (ob->s.b.count != count)
|
|
799 ob->s.b.count = 0;
|
|
800 }
|
|
801
|
|
802 if (!start_fde_sort (&accu, count))
|
|
803 return;
|
|
804
|
|
805 if (ob->s.b.from_array)
|
|
806 {
|
|
807 fde **p;
|
|
808 for (p = ob->u.array; *p; ++p)
|
|
809 add_fdes (ob, &accu, *p);
|
|
810 }
|
|
811 else
|
|
812 add_fdes (ob, &accu, ob->u.single);
|
|
813
|
|
814 end_fde_sort (ob, &accu, count);
|
|
815
|
|
816 /* Save the original fde pointer, since this is the key by which the
|
|
817 DSO will deregister the object. */
|
|
818 accu.linear->orig_data = ob->u.single;
|
|
819 ob->u.sort = accu.linear;
|
|
820
|
|
821 ob->s.b.sorted = 1;
|
|
822 }
|
|
823
|
|
824 /* A linear search through a set of FDEs for the given PC. This is
|
|
825 used when there was insufficient memory to allocate and sort an
|
|
826 array. */
|
|
827
|
|
828 static const fde *
|
|
829 linear_search_fdes (struct object *ob, const fde *this_fde, void *pc)
|
|
830 {
|
|
831 const struct dwarf_cie *last_cie = 0;
|
|
832 int encoding = ob->s.b.encoding;
|
|
833 _Unwind_Ptr base = base_from_object (ob->s.b.encoding, ob);
|
|
834
|
|
835 for (; ! last_fde (ob, this_fde); this_fde = next_fde (this_fde))
|
|
836 {
|
|
837 const struct dwarf_cie *this_cie;
|
|
838 _Unwind_Ptr pc_begin, pc_range;
|
|
839
|
|
840 /* Skip CIEs. */
|
|
841 if (this_fde->CIE_delta == 0)
|
|
842 continue;
|
|
843
|
|
844 if (ob->s.b.mixed_encoding)
|
|
845 {
|
|
846 /* Determine the encoding for this FDE. Note mixed encoded
|
|
847 objects for later. */
|
|
848 this_cie = get_cie (this_fde);
|
|
849 if (this_cie != last_cie)
|
|
850 {
|
|
851 last_cie = this_cie;
|
|
852 encoding = get_cie_encoding (this_cie);
|
|
853 base = base_from_object (encoding, ob);
|
|
854 }
|
|
855 }
|
|
856
|
|
857 if (encoding == DW_EH_PE_absptr)
|
|
858 {
|
|
859 const _Unwind_Ptr *pc_array = (const _Unwind_Ptr *) this_fde->pc_begin;
|
|
860 pc_begin = pc_array[0];
|
|
861 pc_range = pc_array[1];
|
|
862 if (pc_begin == 0)
|
|
863 continue;
|
|
864 }
|
|
865 else
|
|
866 {
|
|
867 _Unwind_Ptr mask;
|
|
868 const unsigned char *p;
|
|
869
|
|
870 p = read_encoded_value_with_base (encoding, base,
|
|
871 this_fde->pc_begin, &pc_begin);
|
|
872 read_encoded_value_with_base (encoding & 0x0F, 0, p, &pc_range);
|
|
873
|
|
874 /* Take care to ignore link-once functions that were removed.
|
|
875 In these cases, the function address will be NULL, but if
|
|
876 the encoding is smaller than a pointer a true NULL may not
|
|
877 be representable. Assume 0 in the representable bits is NULL. */
|
|
878 mask = size_of_encoded_value (encoding);
|
|
879 if (mask < sizeof (void *))
|
|
880 mask = (((_Unwind_Ptr) 1) << (mask << 3)) - 1;
|
|
881 else
|
|
882 mask = -1;
|
|
883
|
|
884 if ((pc_begin & mask) == 0)
|
|
885 continue;
|
|
886 }
|
|
887
|
|
888 if ((_Unwind_Ptr) pc - pc_begin < pc_range)
|
|
889 return this_fde;
|
|
890 }
|
|
891
|
|
892 return NULL;
|
|
893 }
|
|
894
|
|
895 /* Binary search for an FDE containing the given PC. Here are three
|
|
896 implementations of increasing complexity. */
|
|
897
|
|
898 static inline const fde *
|
|
899 binary_search_unencoded_fdes (struct object *ob, void *pc)
|
|
900 {
|
|
901 struct fde_vector *vec = ob->u.sort;
|
|
902 size_t lo, hi;
|
|
903
|
|
904 for (lo = 0, hi = vec->count; lo < hi; )
|
|
905 {
|
|
906 size_t i = (lo + hi) / 2;
|
|
907 const fde *const f = vec->array[i];
|
|
908 void *pc_begin;
|
|
909 uaddr pc_range;
|
|
910 memcpy (&pc_begin, (const void * const *) f->pc_begin, sizeof (void *));
|
|
911 memcpy (&pc_range, (const uaddr *) f->pc_begin + 1, sizeof (uaddr));
|
|
912
|
|
913 if (pc < pc_begin)
|
|
914 hi = i;
|
|
915 else if (pc >= pc_begin + pc_range)
|
|
916 lo = i + 1;
|
|
917 else
|
|
918 return f;
|
|
919 }
|
|
920
|
|
921 return NULL;
|
|
922 }
|
|
923
|
|
924 static inline const fde *
|
|
925 binary_search_single_encoding_fdes (struct object *ob, void *pc)
|
|
926 {
|
|
927 struct fde_vector *vec = ob->u.sort;
|
|
928 int encoding = ob->s.b.encoding;
|
|
929 _Unwind_Ptr base = base_from_object (encoding, ob);
|
|
930 size_t lo, hi;
|
|
931
|
|
932 for (lo = 0, hi = vec->count; lo < hi; )
|
|
933 {
|
|
934 size_t i = (lo + hi) / 2;
|
|
935 const fde *f = vec->array[i];
|
|
936 _Unwind_Ptr pc_begin, pc_range;
|
|
937 const unsigned char *p;
|
|
938
|
|
939 p = read_encoded_value_with_base (encoding, base, f->pc_begin,
|
|
940 &pc_begin);
|
|
941 read_encoded_value_with_base (encoding & 0x0F, 0, p, &pc_range);
|
|
942
|
|
943 if ((_Unwind_Ptr) pc < pc_begin)
|
|
944 hi = i;
|
|
945 else if ((_Unwind_Ptr) pc >= pc_begin + pc_range)
|
|
946 lo = i + 1;
|
|
947 else
|
|
948 return f;
|
|
949 }
|
|
950
|
|
951 return NULL;
|
|
952 }
|
|
953
|
|
954 static inline const fde *
|
|
955 binary_search_mixed_encoding_fdes (struct object *ob, void *pc)
|
|
956 {
|
|
957 struct fde_vector *vec = ob->u.sort;
|
|
958 size_t lo, hi;
|
|
959
|
|
960 for (lo = 0, hi = vec->count; lo < hi; )
|
|
961 {
|
|
962 size_t i = (lo + hi) / 2;
|
|
963 const fde *f = vec->array[i];
|
|
964 _Unwind_Ptr pc_begin, pc_range;
|
|
965 const unsigned char *p;
|
|
966 int encoding;
|
|
967
|
|
968 encoding = get_fde_encoding (f);
|
|
969 p = read_encoded_value_with_base (encoding,
|
|
970 base_from_object (encoding, ob),
|
|
971 f->pc_begin, &pc_begin);
|
|
972 read_encoded_value_with_base (encoding & 0x0F, 0, p, &pc_range);
|
|
973
|
|
974 if ((_Unwind_Ptr) pc < pc_begin)
|
|
975 hi = i;
|
|
976 else if ((_Unwind_Ptr) pc >= pc_begin + pc_range)
|
|
977 lo = i + 1;
|
|
978 else
|
|
979 return f;
|
|
980 }
|
|
981
|
|
982 return NULL;
|
|
983 }
|
|
984
|
|
985 static const fde *
|
|
986 search_object (struct object* ob, void *pc)
|
|
987 {
|
|
988 /* If the data hasn't been sorted, try to do this now. We may have
|
|
989 more memory available than last time we tried. */
|
|
990 if (! ob->s.b.sorted)
|
|
991 {
|
|
992 init_object (ob);
|
|
993
|
|
994 /* Despite the above comment, the normal reason to get here is
|
|
995 that we've not processed this object before. A quick range
|
|
996 check is in order. */
|
|
997 if (pc < ob->pc_begin)
|
|
998 return NULL;
|
|
999 }
|
|
1000
|
|
1001 if (ob->s.b.sorted)
|
|
1002 {
|
|
1003 if (ob->s.b.mixed_encoding)
|
|
1004 return binary_search_mixed_encoding_fdes (ob, pc);
|
|
1005 else if (ob->s.b.encoding == DW_EH_PE_absptr)
|
|
1006 return binary_search_unencoded_fdes (ob, pc);
|
|
1007 else
|
|
1008 return binary_search_single_encoding_fdes (ob, pc);
|
|
1009 }
|
|
1010 else
|
|
1011 {
|
|
1012 /* Long slow laborious linear search, cos we've no memory. */
|
|
1013 if (ob->s.b.from_array)
|
|
1014 {
|
|
1015 fde **p;
|
|
1016 for (p = ob->u.array; *p ; p++)
|
|
1017 {
|
|
1018 const fde *f = linear_search_fdes (ob, *p, pc);
|
|
1019 if (f)
|
|
1020 return f;
|
|
1021 }
|
|
1022 return NULL;
|
|
1023 }
|
|
1024 else
|
|
1025 return linear_search_fdes (ob, ob->u.single, pc);
|
|
1026 }
|
|
1027 }
|
|
1028
|
|
1029 const fde *
|
|
1030 _Unwind_Find_FDE (void *pc, struct dwarf_eh_bases *bases)
|
|
1031 {
|
|
1032 struct object *ob;
|
|
1033 const fde *f = NULL;
|
|
1034
|
|
1035 #ifdef ATOMIC_FDE_FAST_PATH
|
|
1036 /* For targets where unwind info is usually not registered through these
|
|
1037 APIs anymore, avoid taking a global lock.
|
|
1038 Use relaxed MO here, it is up to the app to ensure that the library
|
|
1039 loading/initialization happens-before using that library in other
|
|
1040 threads (in particular unwinding with that library's functions
|
|
1041 appearing in the backtraces). Calling that library's functions
|
|
1042 without waiting for the library to initialize would be racy. */
|
|
1043 if (__builtin_expect (!__atomic_load_n (&any_objects_registered,
|
|
1044 __ATOMIC_RELAXED), 1))
|
|
1045 return NULL;
|
|
1046 #endif
|
|
1047
|
|
1048 init_object_mutex_once ();
|
|
1049 __gthread_mutex_lock (&object_mutex);
|
|
1050
|
|
1051 /* Linear search through the classified objects, to find the one
|
|
1052 containing the pc. Note that pc_begin is sorted descending, and
|
|
1053 we expect objects to be non-overlapping. */
|
|
1054 for (ob = seen_objects; ob; ob = ob->next)
|
|
1055 if (pc >= ob->pc_begin)
|
|
1056 {
|
|
1057 f = search_object (ob, pc);
|
|
1058 if (f)
|
|
1059 goto fini;
|
|
1060 break;
|
|
1061 }
|
|
1062
|
|
1063 /* Classify and search the objects we've not yet processed. */
|
|
1064 while ((ob = unseen_objects))
|
|
1065 {
|
|
1066 struct object **p;
|
|
1067
|
|
1068 unseen_objects = ob->next;
|
|
1069 f = search_object (ob, pc);
|
|
1070
|
|
1071 /* Insert the object into the classified list. */
|
|
1072 for (p = &seen_objects; *p ; p = &(*p)->next)
|
|
1073 if ((*p)->pc_begin < ob->pc_begin)
|
|
1074 break;
|
|
1075 ob->next = *p;
|
|
1076 *p = ob;
|
|
1077
|
|
1078 if (f)
|
|
1079 goto fini;
|
|
1080 }
|
|
1081
|
|
1082 fini:
|
|
1083 __gthread_mutex_unlock (&object_mutex);
|
|
1084
|
|
1085 if (f)
|
|
1086 {
|
|
1087 int encoding;
|
|
1088 _Unwind_Ptr func;
|
|
1089
|
|
1090 bases->tbase = ob->tbase;
|
|
1091 bases->dbase = ob->dbase;
|
|
1092
|
|
1093 encoding = ob->s.b.encoding;
|
|
1094 if (ob->s.b.mixed_encoding)
|
|
1095 encoding = get_fde_encoding (f);
|
|
1096 read_encoded_value_with_base (encoding, base_from_object (encoding, ob),
|
|
1097 f->pc_begin, &func);
|
|
1098 bases->func = (void *) func;
|
|
1099 }
|
|
1100
|
|
1101 return f;
|
|
1102 }
|