Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2#include <linux/module.h>
3#include <linux/sort.h>
4#include <asm/ptrace.h>
5#include <asm/stacktrace.h>
6#include <asm/unwind.h>
7#include <asm/orc_types.h>
8#include <asm/orc_lookup.h>
9
10#define orc_warn(fmt, ...) \
11 printk_deferred_once(KERN_WARNING pr_fmt("WARNING: " fmt), ##__VA_ARGS__)
12
13extern int __start_orc_unwind_ip[];
14extern int __stop_orc_unwind_ip[];
15extern struct orc_entry __start_orc_unwind[];
16extern struct orc_entry __stop_orc_unwind[];
17
18static DEFINE_MUTEX(sort_mutex);
19int *cur_orc_ip_table = __start_orc_unwind_ip;
20struct orc_entry *cur_orc_table = __start_orc_unwind;
21
22unsigned int lookup_num_blocks;
23bool orc_init;
24
25static inline unsigned long orc_ip(const int *ip)
26{
27 return (unsigned long)ip + *ip;
28}
29
30static struct orc_entry *__orc_find(int *ip_table, struct orc_entry *u_table,
31 unsigned int num_entries, unsigned long ip)
32{
33 int *first = ip_table;
34 int *last = ip_table + num_entries - 1;
35 int *mid = first, *found = first;
36
37 if (!num_entries)
38 return NULL;
39
40 /*
41 * Do a binary range search to find the rightmost duplicate of a given
42 * starting address. Some entries are section terminators which are
43 * "weak" entries for ensuring there are no gaps. They should be
44 * ignored when they conflict with a real entry.
45 */
46 while (first <= last) {
47 mid = first + ((last - first) / 2);
48
49 if (orc_ip(mid) <= ip) {
50 found = mid;
51 first = mid + 1;
52 } else
53 last = mid - 1;
54 }
55
56 return u_table + (found - ip_table);
57}
58
59#ifdef CONFIG_MODULES
60static struct orc_entry *orc_module_find(unsigned long ip)
61{
62 struct module *mod;
63
64 mod = __module_address(ip);
65 if (!mod || !mod->arch.orc_unwind || !mod->arch.orc_unwind_ip)
66 return NULL;
67 return __orc_find(mod->arch.orc_unwind_ip, mod->arch.orc_unwind,
68 mod->arch.num_orcs, ip);
69}
70#else
71static struct orc_entry *orc_module_find(unsigned long ip)
72{
73 return NULL;
74}
75#endif
76
77#ifdef CONFIG_DYNAMIC_FTRACE
78static struct orc_entry *orc_find(unsigned long ip);
79
80/*
81 * Ftrace dynamic trampolines do not have orc entries of their own.
82 * But they are copies of the ftrace entries that are static and
83 * defined in ftrace_*.S, which do have orc entries.
84 *
85 * If the unwinder comes across a ftrace trampoline, then find the
86 * ftrace function that was used to create it, and use that ftrace
87 * function's orc entry, as the placement of the return code in
88 * the stack will be identical.
89 */
90static struct orc_entry *orc_ftrace_find(unsigned long ip)
91{
92 struct ftrace_ops *ops;
93 unsigned long caller;
94
95 ops = ftrace_ops_trampoline(ip);
96 if (!ops)
97 return NULL;
98
99 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS)
100 caller = (unsigned long)ftrace_regs_call;
101 else
102 caller = (unsigned long)ftrace_call;
103
104 /* Prevent unlikely recursion */
105 if (ip == caller)
106 return NULL;
107
108 return orc_find(caller);
109}
110#else
111static struct orc_entry *orc_ftrace_find(unsigned long ip)
112{
113 return NULL;
114}
115#endif
116
117/*
118 * If we crash with IP==0, the last successfully executed instruction
119 * was probably an indirect function call with a NULL function pointer,
120 * and we don't have unwind information for NULL.
121 * This hardcoded ORC entry for IP==0 allows us to unwind from a NULL function
122 * pointer into its parent and then continue normally from there.
123 */
124static struct orc_entry null_orc_entry = {
125 .sp_offset = sizeof(long),
126 .sp_reg = ORC_REG_SP,
127 .bp_reg = ORC_REG_UNDEFINED,
128 .type = ORC_TYPE_CALL
129};
130
131/* Fake frame pointer entry -- used as a fallback for generated code */
132static struct orc_entry orc_fp_entry = {
133 .type = ORC_TYPE_CALL,
134 .sp_reg = ORC_REG_BP,
135 .sp_offset = 16,
136 .bp_reg = ORC_REG_PREV_SP,
137 .bp_offset = -16,
138 .end = 0,
139};
140
141static struct orc_entry *orc_find(unsigned long ip)
142{
143 static struct orc_entry *orc;
144
145 if (!orc_init)
146 return NULL;
147
148 if (ip == 0)
149 return &null_orc_entry;
150
151 /* For non-init vmlinux addresses, use the fast lookup table: */
152 if (ip >= LOOKUP_START_IP && ip < LOOKUP_STOP_IP) {
153 unsigned int idx, start, stop;
154
155 idx = (ip - LOOKUP_START_IP) / LOOKUP_BLOCK_SIZE;
156
157 if (unlikely((idx >= lookup_num_blocks-1))) {
158 orc_warn("WARNING: bad lookup idx: idx=%u num=%u ip=%pB\n",
159 idx, lookup_num_blocks, (void *)ip);
160 return NULL;
161 }
162
163 start = orc_lookup[idx];
164 stop = orc_lookup[idx + 1] + 1;
165
166 if (unlikely((__start_orc_unwind + start >= __stop_orc_unwind) ||
167 (__start_orc_unwind + stop > __stop_orc_unwind))) {
168 orc_warn("WARNING: bad lookup value: idx=%u num=%u start=%u stop=%u ip=%pB\n",
169 idx, lookup_num_blocks, start, stop, (void *)ip);
170 return NULL;
171 }
172
173 return __orc_find(__start_orc_unwind_ip + start,
174 __start_orc_unwind + start, stop - start, ip);
175 }
176
177 /* vmlinux .init slow lookup: */
178 if (init_kernel_text(ip))
179 return __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
180 __stop_orc_unwind_ip - __start_orc_unwind_ip, ip);
181
182 /* Module lookup: */
183 orc = orc_module_find(ip);
184 if (orc)
185 return orc;
186
187 return orc_ftrace_find(ip);
188}
189
190static void orc_sort_swap(void *_a, void *_b, int size)
191{
192 struct orc_entry *orc_a, *orc_b;
193 struct orc_entry orc_tmp;
194 int *a = _a, *b = _b, tmp;
195 int delta = _b - _a;
196
197 /* Swap the .orc_unwind_ip entries: */
198 tmp = *a;
199 *a = *b + delta;
200 *b = tmp - delta;
201
202 /* Swap the corresponding .orc_unwind entries: */
203 orc_a = cur_orc_table + (a - cur_orc_ip_table);
204 orc_b = cur_orc_table + (b - cur_orc_ip_table);
205 orc_tmp = *orc_a;
206 *orc_a = *orc_b;
207 *orc_b = orc_tmp;
208}
209
210static int orc_sort_cmp(const void *_a, const void *_b)
211{
212 struct orc_entry *orc_a;
213 const int *a = _a, *b = _b;
214 unsigned long a_val = orc_ip(a);
215 unsigned long b_val = orc_ip(b);
216
217 if (a_val > b_val)
218 return 1;
219 if (a_val < b_val)
220 return -1;
221
222 /*
223 * The "weak" section terminator entries need to always be on the left
224 * to ensure the lookup code skips them in favor of real entries.
225 * These terminator entries exist to handle any gaps created by
226 * whitelisted .o files which didn't get objtool generation.
227 */
228 orc_a = cur_orc_table + (a - cur_orc_ip_table);
229 return orc_a->sp_reg == ORC_REG_UNDEFINED && !orc_a->end ? -1 : 1;
230}
231
232#ifdef CONFIG_MODULES
233void unwind_module_init(struct module *mod, void *_orc_ip, size_t orc_ip_size,
234 void *_orc, size_t orc_size)
235{
236 int *orc_ip = _orc_ip;
237 struct orc_entry *orc = _orc;
238 unsigned int num_entries = orc_ip_size / sizeof(int);
239
240 WARN_ON_ONCE(orc_ip_size % sizeof(int) != 0 ||
241 orc_size % sizeof(*orc) != 0 ||
242 num_entries != orc_size / sizeof(*orc));
243
244 /*
245 * The 'cur_orc_*' globals allow the orc_sort_swap() callback to
246 * associate an .orc_unwind_ip table entry with its corresponding
247 * .orc_unwind entry so they can both be swapped.
248 */
249 mutex_lock(&sort_mutex);
250 cur_orc_ip_table = orc_ip;
251 cur_orc_table = orc;
252 sort(orc_ip, num_entries, sizeof(int), orc_sort_cmp, orc_sort_swap);
253 mutex_unlock(&sort_mutex);
254
255 mod->arch.orc_unwind_ip = orc_ip;
256 mod->arch.orc_unwind = orc;
257 mod->arch.num_orcs = num_entries;
258}
259#endif
260
261void __init unwind_init(void)
262{
263 size_t orc_ip_size = (void *)__stop_orc_unwind_ip - (void *)__start_orc_unwind_ip;
264 size_t orc_size = (void *)__stop_orc_unwind - (void *)__start_orc_unwind;
265 size_t num_entries = orc_ip_size / sizeof(int);
266 struct orc_entry *orc;
267 int i;
268
269 if (!num_entries || orc_ip_size % sizeof(int) != 0 ||
270 orc_size % sizeof(struct orc_entry) != 0 ||
271 num_entries != orc_size / sizeof(struct orc_entry)) {
272 orc_warn("WARNING: Bad or missing .orc_unwind table. Disabling unwinder.\n");
273 return;
274 }
275
276 /* Sort the .orc_unwind and .orc_unwind_ip tables: */
277 sort(__start_orc_unwind_ip, num_entries, sizeof(int), orc_sort_cmp,
278 orc_sort_swap);
279
280 /* Initialize the fast lookup table: */
281 lookup_num_blocks = orc_lookup_end - orc_lookup;
282 for (i = 0; i < lookup_num_blocks-1; i++) {
283 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
284 num_entries,
285 LOOKUP_START_IP + (LOOKUP_BLOCK_SIZE * i));
286 if (!orc) {
287 orc_warn("WARNING: Corrupt .orc_unwind table. Disabling unwinder.\n");
288 return;
289 }
290
291 orc_lookup[i] = orc - __start_orc_unwind;
292 }
293
294 /* Initialize the ending block: */
295 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind, num_entries,
296 LOOKUP_STOP_IP);
297 if (!orc) {
298 orc_warn("WARNING: Corrupt .orc_unwind table. Disabling unwinder.\n");
299 return;
300 }
301 orc_lookup[lookup_num_blocks-1] = orc - __start_orc_unwind;
302
303 orc_init = true;
304}
305
306unsigned long unwind_get_return_address(struct unwind_state *state)
307{
308 if (unwind_done(state))
309 return 0;
310
311 return __kernel_text_address(state->ip) ? state->ip : 0;
312}
313EXPORT_SYMBOL_GPL(unwind_get_return_address);
314
315unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
316{
317 if (unwind_done(state))
318 return NULL;
319
320 if (state->regs)
321 return &state->regs->ip;
322
323 if (state->sp)
324 return (unsigned long *)state->sp - 1;
325
326 return NULL;
327}
328
329static bool stack_access_ok(struct unwind_state *state, unsigned long _addr,
330 size_t len)
331{
332 struct stack_info *info = &state->stack_info;
333 void *addr = (void *)_addr;
334
335 if (!on_stack(info, addr, len) &&
336 (get_stack_info(addr, state->task, info, &state->stack_mask)))
337 return false;
338
339 return true;
340}
341
342static bool deref_stack_reg(struct unwind_state *state, unsigned long addr,
343 unsigned long *val)
344{
345 if (!stack_access_ok(state, addr, sizeof(long)))
346 return false;
347
348 *val = READ_ONCE_NOCHECK(*(unsigned long *)addr);
349 return true;
350}
351
352static bool deref_stack_regs(struct unwind_state *state, unsigned long addr,
353 unsigned long *ip, unsigned long *sp)
354{
355 struct pt_regs *regs = (struct pt_regs *)addr;
356
357 /* x86-32 support will be more complicated due to the ®s->sp hack */
358 BUILD_BUG_ON(IS_ENABLED(CONFIG_X86_32));
359
360 if (!stack_access_ok(state, addr, sizeof(struct pt_regs)))
361 return false;
362
363 *ip = regs->ip;
364 *sp = regs->sp;
365 return true;
366}
367
368static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr,
369 unsigned long *ip, unsigned long *sp)
370{
371 struct pt_regs *regs = (void *)addr - IRET_FRAME_OFFSET;
372
373 if (!stack_access_ok(state, addr, IRET_FRAME_SIZE))
374 return false;
375
376 *ip = regs->ip;
377 *sp = regs->sp;
378 return true;
379}
380
381bool unwind_next_frame(struct unwind_state *state)
382{
383 unsigned long ip_p, sp, orig_ip = state->ip, prev_sp = state->sp;
384 enum stack_type prev_type = state->stack_info.type;
385 struct orc_entry *orc;
386 bool indirect = false;
387
388 if (unwind_done(state))
389 return false;
390
391 /* Don't let modules unload while we're reading their ORC data. */
392 preempt_disable();
393
394 /* End-of-stack check for user tasks: */
395 if (state->regs && user_mode(state->regs))
396 goto the_end;
397
398 /*
399 * Find the orc_entry associated with the text address.
400 *
401 * Decrement call return addresses by one so they work for sibling
402 * calls and calls to noreturn functions.
403 */
404 orc = orc_find(state->signal ? state->ip : state->ip - 1);
405 if (!orc) {
406 /*
407 * As a fallback, try to assume this code uses a frame pointer.
408 * This is useful for generated code, like BPF, which ORC
409 * doesn't know about. This is just a guess, so the rest of
410 * the unwind is no longer considered reliable.
411 */
412 orc = &orc_fp_entry;
413 state->error = true;
414 }
415
416 /* End-of-stack check for kernel threads: */
417 if (orc->sp_reg == ORC_REG_UNDEFINED) {
418 if (!orc->end)
419 goto err;
420
421 goto the_end;
422 }
423
424 /* Find the previous frame's stack: */
425 switch (orc->sp_reg) {
426 case ORC_REG_SP:
427 sp = state->sp + orc->sp_offset;
428 break;
429
430 case ORC_REG_BP:
431 sp = state->bp + orc->sp_offset;
432 break;
433
434 case ORC_REG_SP_INDIRECT:
435 sp = state->sp + orc->sp_offset;
436 indirect = true;
437 break;
438
439 case ORC_REG_BP_INDIRECT:
440 sp = state->bp + orc->sp_offset;
441 indirect = true;
442 break;
443
444 case ORC_REG_R10:
445 if (!state->regs || !state->full_regs) {
446 orc_warn("missing regs for base reg R10 at ip %pB\n",
447 (void *)state->ip);
448 goto err;
449 }
450 sp = state->regs->r10;
451 break;
452
453 case ORC_REG_R13:
454 if (!state->regs || !state->full_regs) {
455 orc_warn("missing regs for base reg R13 at ip %pB\n",
456 (void *)state->ip);
457 goto err;
458 }
459 sp = state->regs->r13;
460 break;
461
462 case ORC_REG_DI:
463 if (!state->regs || !state->full_regs) {
464 orc_warn("missing regs for base reg DI at ip %pB\n",
465 (void *)state->ip);
466 goto err;
467 }
468 sp = state->regs->di;
469 break;
470
471 case ORC_REG_DX:
472 if (!state->regs || !state->full_regs) {
473 orc_warn("missing regs for base reg DX at ip %pB\n",
474 (void *)state->ip);
475 goto err;
476 }
477 sp = state->regs->dx;
478 break;
479
480 default:
481 orc_warn("unknown SP base reg %d for ip %pB\n",
482 orc->sp_reg, (void *)state->ip);
483 goto err;
484 }
485
486 if (indirect) {
487 if (!deref_stack_reg(state, sp, &sp))
488 goto err;
489 }
490
491 /* Find IP, SP and possibly regs: */
492 switch (orc->type) {
493 case ORC_TYPE_CALL:
494 ip_p = sp - sizeof(long);
495
496 if (!deref_stack_reg(state, ip_p, &state->ip))
497 goto err;
498
499 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
500 state->ip, (void *)ip_p);
501
502 state->sp = sp;
503 state->regs = NULL;
504 state->signal = false;
505 break;
506
507 case ORC_TYPE_REGS:
508 if (!deref_stack_regs(state, sp, &state->ip, &state->sp)) {
509 orc_warn("can't dereference registers at %p for ip %pB\n",
510 (void *)sp, (void *)orig_ip);
511 goto err;
512 }
513
514 state->regs = (struct pt_regs *)sp;
515 state->full_regs = true;
516 state->signal = true;
517 break;
518
519 case ORC_TYPE_REGS_IRET:
520 if (!deref_stack_iret_regs(state, sp, &state->ip, &state->sp)) {
521 orc_warn("can't dereference iret registers at %p for ip %pB\n",
522 (void *)sp, (void *)orig_ip);
523 goto err;
524 }
525
526 state->regs = (void *)sp - IRET_FRAME_OFFSET;
527 state->full_regs = false;
528 state->signal = true;
529 break;
530
531 default:
532 orc_warn("unknown .orc_unwind entry type %d for ip %pB\n",
533 orc->type, (void *)orig_ip);
534 break;
535 }
536
537 /* Find BP: */
538 switch (orc->bp_reg) {
539 case ORC_REG_UNDEFINED:
540 if (state->regs && state->full_regs)
541 state->bp = state->regs->bp;
542 break;
543
544 case ORC_REG_PREV_SP:
545 if (!deref_stack_reg(state, sp + orc->bp_offset, &state->bp))
546 goto err;
547 break;
548
549 case ORC_REG_BP:
550 if (!deref_stack_reg(state, state->bp + orc->bp_offset, &state->bp))
551 goto err;
552 break;
553
554 default:
555 orc_warn("unknown BP base reg %d for ip %pB\n",
556 orc->bp_reg, (void *)orig_ip);
557 goto err;
558 }
559
560 /* Prevent a recursive loop due to bad ORC data: */
561 if (state->stack_info.type == prev_type &&
562 on_stack(&state->stack_info, (void *)state->sp, sizeof(long)) &&
563 state->sp <= prev_sp) {
564 orc_warn("stack going in the wrong direction? ip=%pB\n",
565 (void *)orig_ip);
566 goto err;
567 }
568
569 preempt_enable();
570 return true;
571
572err:
573 state->error = true;
574
575the_end:
576 preempt_enable();
577 state->stack_info.type = STACK_TYPE_UNKNOWN;
578 return false;
579}
580EXPORT_SYMBOL_GPL(unwind_next_frame);
581
582void __unwind_start(struct unwind_state *state, struct task_struct *task,
583 struct pt_regs *regs, unsigned long *first_frame)
584{
585 memset(state, 0, sizeof(*state));
586 state->task = task;
587
588 /*
589 * Refuse to unwind the stack of a task while it's executing on another
590 * CPU. This check is racy, but that's ok: the unwinder has other
591 * checks to prevent it from going off the rails.
592 */
593 if (task_on_another_cpu(task))
594 goto done;
595
596 if (regs) {
597 if (user_mode(regs))
598 goto done;
599
600 state->ip = regs->ip;
601 state->sp = regs->sp;
602 state->bp = regs->bp;
603 state->regs = regs;
604 state->full_regs = true;
605 state->signal = true;
606
607 } else if (task == current) {
608 asm volatile("lea (%%rip), %0\n\t"
609 "mov %%rsp, %1\n\t"
610 "mov %%rbp, %2\n\t"
611 : "=r" (state->ip), "=r" (state->sp),
612 "=r" (state->bp));
613
614 } else {
615 struct inactive_task_frame *frame = (void *)task->thread.sp;
616
617 state->sp = task->thread.sp;
618 state->bp = READ_ONCE_NOCHECK(frame->bp);
619 state->ip = READ_ONCE_NOCHECK(frame->ret_addr);
620 }
621
622 if (get_stack_info((unsigned long *)state->sp, state->task,
623 &state->stack_info, &state->stack_mask)) {
624 /*
625 * We weren't on a valid stack. It's possible that
626 * we overflowed a valid stack into a guard page.
627 * See if the next page up is valid so that we can
628 * generate some kind of backtrace if this happens.
629 */
630 void *next_page = (void *)PAGE_ALIGN((unsigned long)state->sp);
631 if (get_stack_info(next_page, state->task, &state->stack_info,
632 &state->stack_mask))
633 return;
634 }
635
636 /*
637 * The caller can provide the address of the first frame directly
638 * (first_frame) or indirectly (regs->sp) to indicate which stack frame
639 * to start unwinding at. Skip ahead until we reach it.
640 */
641
642 /* When starting from regs, skip the regs frame: */
643 if (regs) {
644 unwind_next_frame(state);
645 return;
646 }
647
648 /* Otherwise, skip ahead to the user-specified starting frame: */
649 while (!unwind_done(state) &&
650 (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
651 state->sp <= (unsigned long)first_frame))
652 unwind_next_frame(state);
653
654 return;
655
656done:
657 state->stack_info.type = STACK_TYPE_UNKNOWN;
658 return;
659}
660EXPORT_SYMBOL_GPL(__unwind_start);
1// SPDX-License-Identifier: GPL-2.0-only
2#include <linux/objtool.h>
3#include <linux/module.h>
4#include <linux/sort.h>
5#include <asm/ptrace.h>
6#include <asm/stacktrace.h>
7#include <asm/unwind.h>
8#include <asm/orc_types.h>
9#include <asm/orc_lookup.h>
10
11#define orc_warn(fmt, ...) \
12 printk_deferred_once(KERN_WARNING "WARNING: " fmt, ##__VA_ARGS__)
13
14#define orc_warn_current(args...) \
15({ \
16 if (state->task == current && !state->error) \
17 orc_warn(args); \
18})
19
20extern int __start_orc_unwind_ip[];
21extern int __stop_orc_unwind_ip[];
22extern struct orc_entry __start_orc_unwind[];
23extern struct orc_entry __stop_orc_unwind[];
24
25static bool orc_init __ro_after_init;
26static unsigned int lookup_num_blocks __ro_after_init;
27
28static inline unsigned long orc_ip(const int *ip)
29{
30 return (unsigned long)ip + *ip;
31}
32
33static struct orc_entry *__orc_find(int *ip_table, struct orc_entry *u_table,
34 unsigned int num_entries, unsigned long ip)
35{
36 int *first = ip_table;
37 int *last = ip_table + num_entries - 1;
38 int *mid = first, *found = first;
39
40 if (!num_entries)
41 return NULL;
42
43 /*
44 * Do a binary range search to find the rightmost duplicate of a given
45 * starting address. Some entries are section terminators which are
46 * "weak" entries for ensuring there are no gaps. They should be
47 * ignored when they conflict with a real entry.
48 */
49 while (first <= last) {
50 mid = first + ((last - first) / 2);
51
52 if (orc_ip(mid) <= ip) {
53 found = mid;
54 first = mid + 1;
55 } else
56 last = mid - 1;
57 }
58
59 return u_table + (found - ip_table);
60}
61
62#ifdef CONFIG_MODULES
63static struct orc_entry *orc_module_find(unsigned long ip)
64{
65 struct module *mod;
66
67 mod = __module_address(ip);
68 if (!mod || !mod->arch.orc_unwind || !mod->arch.orc_unwind_ip)
69 return NULL;
70 return __orc_find(mod->arch.orc_unwind_ip, mod->arch.orc_unwind,
71 mod->arch.num_orcs, ip);
72}
73#else
74static struct orc_entry *orc_module_find(unsigned long ip)
75{
76 return NULL;
77}
78#endif
79
80#ifdef CONFIG_DYNAMIC_FTRACE
81static struct orc_entry *orc_find(unsigned long ip);
82
83/*
84 * Ftrace dynamic trampolines do not have orc entries of their own.
85 * But they are copies of the ftrace entries that are static and
86 * defined in ftrace_*.S, which do have orc entries.
87 *
88 * If the unwinder comes across a ftrace trampoline, then find the
89 * ftrace function that was used to create it, and use that ftrace
90 * function's orc entry, as the placement of the return code in
91 * the stack will be identical.
92 */
93static struct orc_entry *orc_ftrace_find(unsigned long ip)
94{
95 struct ftrace_ops *ops;
96 unsigned long tramp_addr, offset;
97
98 ops = ftrace_ops_trampoline(ip);
99 if (!ops)
100 return NULL;
101
102 /* Set tramp_addr to the start of the code copied by the trampoline */
103 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS)
104 tramp_addr = (unsigned long)ftrace_regs_caller;
105 else
106 tramp_addr = (unsigned long)ftrace_caller;
107
108 /* Now place tramp_addr to the location within the trampoline ip is at */
109 offset = ip - ops->trampoline;
110 tramp_addr += offset;
111
112 /* Prevent unlikely recursion */
113 if (ip == tramp_addr)
114 return NULL;
115
116 return orc_find(tramp_addr);
117}
118#else
119static struct orc_entry *orc_ftrace_find(unsigned long ip)
120{
121 return NULL;
122}
123#endif
124
125/*
126 * If we crash with IP==0, the last successfully executed instruction
127 * was probably an indirect function call with a NULL function pointer,
128 * and we don't have unwind information for NULL.
129 * This hardcoded ORC entry for IP==0 allows us to unwind from a NULL function
130 * pointer into its parent and then continue normally from there.
131 */
132static struct orc_entry null_orc_entry = {
133 .sp_offset = sizeof(long),
134 .sp_reg = ORC_REG_SP,
135 .bp_reg = ORC_REG_UNDEFINED,
136 .type = UNWIND_HINT_TYPE_CALL
137};
138
139#ifdef CONFIG_CALL_THUNKS
140static struct orc_entry *orc_callthunk_find(unsigned long ip)
141{
142 if (!is_callthunk((void *)ip))
143 return NULL;
144
145 return &null_orc_entry;
146}
147#else
148static struct orc_entry *orc_callthunk_find(unsigned long ip)
149{
150 return NULL;
151}
152#endif
153
154/* Fake frame pointer entry -- used as a fallback for generated code */
155static struct orc_entry orc_fp_entry = {
156 .type = UNWIND_HINT_TYPE_CALL,
157 .sp_reg = ORC_REG_BP,
158 .sp_offset = 16,
159 .bp_reg = ORC_REG_PREV_SP,
160 .bp_offset = -16,
161 .end = 0,
162};
163
164static struct orc_entry *orc_find(unsigned long ip)
165{
166 static struct orc_entry *orc;
167
168 if (ip == 0)
169 return &null_orc_entry;
170
171 /* For non-init vmlinux addresses, use the fast lookup table: */
172 if (ip >= LOOKUP_START_IP && ip < LOOKUP_STOP_IP) {
173 unsigned int idx, start, stop;
174
175 idx = (ip - LOOKUP_START_IP) / LOOKUP_BLOCK_SIZE;
176
177 if (unlikely((idx >= lookup_num_blocks-1))) {
178 orc_warn("WARNING: bad lookup idx: idx=%u num=%u ip=%pB\n",
179 idx, lookup_num_blocks, (void *)ip);
180 return NULL;
181 }
182
183 start = orc_lookup[idx];
184 stop = orc_lookup[idx + 1] + 1;
185
186 if (unlikely((__start_orc_unwind + start >= __stop_orc_unwind) ||
187 (__start_orc_unwind + stop > __stop_orc_unwind))) {
188 orc_warn("WARNING: bad lookup value: idx=%u num=%u start=%u stop=%u ip=%pB\n",
189 idx, lookup_num_blocks, start, stop, (void *)ip);
190 return NULL;
191 }
192
193 return __orc_find(__start_orc_unwind_ip + start,
194 __start_orc_unwind + start, stop - start, ip);
195 }
196
197 /* vmlinux .init slow lookup: */
198 if (is_kernel_inittext(ip))
199 return __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
200 __stop_orc_unwind_ip - __start_orc_unwind_ip, ip);
201
202 /* Module lookup: */
203 orc = orc_module_find(ip);
204 if (orc)
205 return orc;
206
207 orc = orc_ftrace_find(ip);
208 if (orc)
209 return orc;
210
211 return orc_callthunk_find(ip);
212}
213
214#ifdef CONFIG_MODULES
215
216static DEFINE_MUTEX(sort_mutex);
217static int *cur_orc_ip_table = __start_orc_unwind_ip;
218static struct orc_entry *cur_orc_table = __start_orc_unwind;
219
220static void orc_sort_swap(void *_a, void *_b, int size)
221{
222 struct orc_entry *orc_a, *orc_b;
223 struct orc_entry orc_tmp;
224 int *a = _a, *b = _b, tmp;
225 int delta = _b - _a;
226
227 /* Swap the .orc_unwind_ip entries: */
228 tmp = *a;
229 *a = *b + delta;
230 *b = tmp - delta;
231
232 /* Swap the corresponding .orc_unwind entries: */
233 orc_a = cur_orc_table + (a - cur_orc_ip_table);
234 orc_b = cur_orc_table + (b - cur_orc_ip_table);
235 orc_tmp = *orc_a;
236 *orc_a = *orc_b;
237 *orc_b = orc_tmp;
238}
239
240static int orc_sort_cmp(const void *_a, const void *_b)
241{
242 struct orc_entry *orc_a;
243 const int *a = _a, *b = _b;
244 unsigned long a_val = orc_ip(a);
245 unsigned long b_val = orc_ip(b);
246
247 if (a_val > b_val)
248 return 1;
249 if (a_val < b_val)
250 return -1;
251
252 /*
253 * The "weak" section terminator entries need to always be on the left
254 * to ensure the lookup code skips them in favor of real entries.
255 * These terminator entries exist to handle any gaps created by
256 * whitelisted .o files which didn't get objtool generation.
257 */
258 orc_a = cur_orc_table + (a - cur_orc_ip_table);
259 return orc_a->sp_reg == ORC_REG_UNDEFINED && !orc_a->end ? -1 : 1;
260}
261
262void unwind_module_init(struct module *mod, void *_orc_ip, size_t orc_ip_size,
263 void *_orc, size_t orc_size)
264{
265 int *orc_ip = _orc_ip;
266 struct orc_entry *orc = _orc;
267 unsigned int num_entries = orc_ip_size / sizeof(int);
268
269 WARN_ON_ONCE(orc_ip_size % sizeof(int) != 0 ||
270 orc_size % sizeof(*orc) != 0 ||
271 num_entries != orc_size / sizeof(*orc));
272
273 /*
274 * The 'cur_orc_*' globals allow the orc_sort_swap() callback to
275 * associate an .orc_unwind_ip table entry with its corresponding
276 * .orc_unwind entry so they can both be swapped.
277 */
278 mutex_lock(&sort_mutex);
279 cur_orc_ip_table = orc_ip;
280 cur_orc_table = orc;
281 sort(orc_ip, num_entries, sizeof(int), orc_sort_cmp, orc_sort_swap);
282 mutex_unlock(&sort_mutex);
283
284 mod->arch.orc_unwind_ip = orc_ip;
285 mod->arch.orc_unwind = orc;
286 mod->arch.num_orcs = num_entries;
287}
288#endif
289
290void __init unwind_init(void)
291{
292 size_t orc_ip_size = (void *)__stop_orc_unwind_ip - (void *)__start_orc_unwind_ip;
293 size_t orc_size = (void *)__stop_orc_unwind - (void *)__start_orc_unwind;
294 size_t num_entries = orc_ip_size / sizeof(int);
295 struct orc_entry *orc;
296 int i;
297
298 if (!num_entries || orc_ip_size % sizeof(int) != 0 ||
299 orc_size % sizeof(struct orc_entry) != 0 ||
300 num_entries != orc_size / sizeof(struct orc_entry)) {
301 orc_warn("WARNING: Bad or missing .orc_unwind table. Disabling unwinder.\n");
302 return;
303 }
304
305 /*
306 * Note, the orc_unwind and orc_unwind_ip tables were already
307 * sorted at build time via the 'sorttable' tool.
308 * It's ready for binary search straight away, no need to sort it.
309 */
310
311 /* Initialize the fast lookup table: */
312 lookup_num_blocks = orc_lookup_end - orc_lookup;
313 for (i = 0; i < lookup_num_blocks-1; i++) {
314 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
315 num_entries,
316 LOOKUP_START_IP + (LOOKUP_BLOCK_SIZE * i));
317 if (!orc) {
318 orc_warn("WARNING: Corrupt .orc_unwind table. Disabling unwinder.\n");
319 return;
320 }
321
322 orc_lookup[i] = orc - __start_orc_unwind;
323 }
324
325 /* Initialize the ending block: */
326 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind, num_entries,
327 LOOKUP_STOP_IP);
328 if (!orc) {
329 orc_warn("WARNING: Corrupt .orc_unwind table. Disabling unwinder.\n");
330 return;
331 }
332 orc_lookup[lookup_num_blocks-1] = orc - __start_orc_unwind;
333
334 orc_init = true;
335}
336
337unsigned long unwind_get_return_address(struct unwind_state *state)
338{
339 if (unwind_done(state))
340 return 0;
341
342 return __kernel_text_address(state->ip) ? state->ip : 0;
343}
344EXPORT_SYMBOL_GPL(unwind_get_return_address);
345
346unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
347{
348 if (unwind_done(state))
349 return NULL;
350
351 if (state->regs)
352 return &state->regs->ip;
353
354 if (state->sp)
355 return (unsigned long *)state->sp - 1;
356
357 return NULL;
358}
359
360static bool stack_access_ok(struct unwind_state *state, unsigned long _addr,
361 size_t len)
362{
363 struct stack_info *info = &state->stack_info;
364 void *addr = (void *)_addr;
365
366 if (on_stack(info, addr, len))
367 return true;
368
369 return !get_stack_info(addr, state->task, info, &state->stack_mask) &&
370 on_stack(info, addr, len);
371}
372
373static bool deref_stack_reg(struct unwind_state *state, unsigned long addr,
374 unsigned long *val)
375{
376 if (!stack_access_ok(state, addr, sizeof(long)))
377 return false;
378
379 *val = READ_ONCE_NOCHECK(*(unsigned long *)addr);
380 return true;
381}
382
383static bool deref_stack_regs(struct unwind_state *state, unsigned long addr,
384 unsigned long *ip, unsigned long *sp)
385{
386 struct pt_regs *regs = (struct pt_regs *)addr;
387
388 /* x86-32 support will be more complicated due to the ®s->sp hack */
389 BUILD_BUG_ON(IS_ENABLED(CONFIG_X86_32));
390
391 if (!stack_access_ok(state, addr, sizeof(struct pt_regs)))
392 return false;
393
394 *ip = READ_ONCE_NOCHECK(regs->ip);
395 *sp = READ_ONCE_NOCHECK(regs->sp);
396 return true;
397}
398
399static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr,
400 unsigned long *ip, unsigned long *sp)
401{
402 struct pt_regs *regs = (void *)addr - IRET_FRAME_OFFSET;
403
404 if (!stack_access_ok(state, addr, IRET_FRAME_SIZE))
405 return false;
406
407 *ip = READ_ONCE_NOCHECK(regs->ip);
408 *sp = READ_ONCE_NOCHECK(regs->sp);
409 return true;
410}
411
412/*
413 * If state->regs is non-NULL, and points to a full pt_regs, just get the reg
414 * value from state->regs.
415 *
416 * Otherwise, if state->regs just points to IRET regs, and the previous frame
417 * had full regs, it's safe to get the value from the previous regs. This can
418 * happen when early/late IRQ entry code gets interrupted by an NMI.
419 */
420static bool get_reg(struct unwind_state *state, unsigned int reg_off,
421 unsigned long *val)
422{
423 unsigned int reg = reg_off/8;
424
425 if (!state->regs)
426 return false;
427
428 if (state->full_regs) {
429 *val = READ_ONCE_NOCHECK(((unsigned long *)state->regs)[reg]);
430 return true;
431 }
432
433 if (state->prev_regs) {
434 *val = READ_ONCE_NOCHECK(((unsigned long *)state->prev_regs)[reg]);
435 return true;
436 }
437
438 return false;
439}
440
441bool unwind_next_frame(struct unwind_state *state)
442{
443 unsigned long ip_p, sp, tmp, orig_ip = state->ip, prev_sp = state->sp;
444 enum stack_type prev_type = state->stack_info.type;
445 struct orc_entry *orc;
446 bool indirect = false;
447
448 if (unwind_done(state))
449 return false;
450
451 /* Don't let modules unload while we're reading their ORC data. */
452 preempt_disable();
453
454 /* End-of-stack check for user tasks: */
455 if (state->regs && user_mode(state->regs))
456 goto the_end;
457
458 /*
459 * Find the orc_entry associated with the text address.
460 *
461 * For a call frame (as opposed to a signal frame), state->ip points to
462 * the instruction after the call. That instruction's stack layout
463 * could be different from the call instruction's layout, for example
464 * if the call was to a noreturn function. So get the ORC data for the
465 * call instruction itself.
466 */
467 orc = orc_find(state->signal ? state->ip : state->ip - 1);
468 if (!orc) {
469 /*
470 * As a fallback, try to assume this code uses a frame pointer.
471 * This is useful for generated code, like BPF, which ORC
472 * doesn't know about. This is just a guess, so the rest of
473 * the unwind is no longer considered reliable.
474 */
475 orc = &orc_fp_entry;
476 state->error = true;
477 }
478
479 /* End-of-stack check for kernel threads: */
480 if (orc->sp_reg == ORC_REG_UNDEFINED) {
481 if (!orc->end)
482 goto err;
483
484 goto the_end;
485 }
486
487 /* Find the previous frame's stack: */
488 switch (orc->sp_reg) {
489 case ORC_REG_SP:
490 sp = state->sp + orc->sp_offset;
491 break;
492
493 case ORC_REG_BP:
494 sp = state->bp + orc->sp_offset;
495 break;
496
497 case ORC_REG_SP_INDIRECT:
498 sp = state->sp;
499 indirect = true;
500 break;
501
502 case ORC_REG_BP_INDIRECT:
503 sp = state->bp + orc->sp_offset;
504 indirect = true;
505 break;
506
507 case ORC_REG_R10:
508 if (!get_reg(state, offsetof(struct pt_regs, r10), &sp)) {
509 orc_warn_current("missing R10 value at %pB\n",
510 (void *)state->ip);
511 goto err;
512 }
513 break;
514
515 case ORC_REG_R13:
516 if (!get_reg(state, offsetof(struct pt_regs, r13), &sp)) {
517 orc_warn_current("missing R13 value at %pB\n",
518 (void *)state->ip);
519 goto err;
520 }
521 break;
522
523 case ORC_REG_DI:
524 if (!get_reg(state, offsetof(struct pt_regs, di), &sp)) {
525 orc_warn_current("missing RDI value at %pB\n",
526 (void *)state->ip);
527 goto err;
528 }
529 break;
530
531 case ORC_REG_DX:
532 if (!get_reg(state, offsetof(struct pt_regs, dx), &sp)) {
533 orc_warn_current("missing DX value at %pB\n",
534 (void *)state->ip);
535 goto err;
536 }
537 break;
538
539 default:
540 orc_warn("unknown SP base reg %d at %pB\n",
541 orc->sp_reg, (void *)state->ip);
542 goto err;
543 }
544
545 if (indirect) {
546 if (!deref_stack_reg(state, sp, &sp))
547 goto err;
548
549 if (orc->sp_reg == ORC_REG_SP_INDIRECT)
550 sp += orc->sp_offset;
551 }
552
553 /* Find IP, SP and possibly regs: */
554 switch (orc->type) {
555 case UNWIND_HINT_TYPE_CALL:
556 ip_p = sp - sizeof(long);
557
558 if (!deref_stack_reg(state, ip_p, &state->ip))
559 goto err;
560
561 state->ip = unwind_recover_ret_addr(state, state->ip,
562 (unsigned long *)ip_p);
563 state->sp = sp;
564 state->regs = NULL;
565 state->prev_regs = NULL;
566 state->signal = false;
567 break;
568
569 case UNWIND_HINT_TYPE_REGS:
570 if (!deref_stack_regs(state, sp, &state->ip, &state->sp)) {
571 orc_warn_current("can't access registers at %pB\n",
572 (void *)orig_ip);
573 goto err;
574 }
575 /*
576 * There is a small chance to interrupt at the entry of
577 * arch_rethook_trampoline() where the ORC info doesn't exist.
578 * That point is right after the RET to arch_rethook_trampoline()
579 * which was modified return address.
580 * At that point, the @addr_p of the unwind_recover_rethook()
581 * (this has to point the address of the stack entry storing
582 * the modified return address) must be "SP - (a stack entry)"
583 * because SP is incremented by the RET.
584 */
585 state->ip = unwind_recover_rethook(state, state->ip,
586 (unsigned long *)(state->sp - sizeof(long)));
587 state->regs = (struct pt_regs *)sp;
588 state->prev_regs = NULL;
589 state->full_regs = true;
590 state->signal = true;
591 break;
592
593 case UNWIND_HINT_TYPE_REGS_PARTIAL:
594 if (!deref_stack_iret_regs(state, sp, &state->ip, &state->sp)) {
595 orc_warn_current("can't access iret registers at %pB\n",
596 (void *)orig_ip);
597 goto err;
598 }
599 /* See UNWIND_HINT_TYPE_REGS case comment. */
600 state->ip = unwind_recover_rethook(state, state->ip,
601 (unsigned long *)(state->sp - sizeof(long)));
602
603 if (state->full_regs)
604 state->prev_regs = state->regs;
605 state->regs = (void *)sp - IRET_FRAME_OFFSET;
606 state->full_regs = false;
607 state->signal = true;
608 break;
609
610 default:
611 orc_warn("unknown .orc_unwind entry type %d at %pB\n",
612 orc->type, (void *)orig_ip);
613 goto err;
614 }
615
616 /* Find BP: */
617 switch (orc->bp_reg) {
618 case ORC_REG_UNDEFINED:
619 if (get_reg(state, offsetof(struct pt_regs, bp), &tmp))
620 state->bp = tmp;
621 break;
622
623 case ORC_REG_PREV_SP:
624 if (!deref_stack_reg(state, sp + orc->bp_offset, &state->bp))
625 goto err;
626 break;
627
628 case ORC_REG_BP:
629 if (!deref_stack_reg(state, state->bp + orc->bp_offset, &state->bp))
630 goto err;
631 break;
632
633 default:
634 orc_warn("unknown BP base reg %d for ip %pB\n",
635 orc->bp_reg, (void *)orig_ip);
636 goto err;
637 }
638
639 /* Prevent a recursive loop due to bad ORC data: */
640 if (state->stack_info.type == prev_type &&
641 on_stack(&state->stack_info, (void *)state->sp, sizeof(long)) &&
642 state->sp <= prev_sp) {
643 orc_warn_current("stack going in the wrong direction? at %pB\n",
644 (void *)orig_ip);
645 goto err;
646 }
647
648 preempt_enable();
649 return true;
650
651err:
652 state->error = true;
653
654the_end:
655 preempt_enable();
656 state->stack_info.type = STACK_TYPE_UNKNOWN;
657 return false;
658}
659EXPORT_SYMBOL_GPL(unwind_next_frame);
660
661void __unwind_start(struct unwind_state *state, struct task_struct *task,
662 struct pt_regs *regs, unsigned long *first_frame)
663{
664 memset(state, 0, sizeof(*state));
665 state->task = task;
666
667 if (!orc_init)
668 goto err;
669
670 /*
671 * Refuse to unwind the stack of a task while it's executing on another
672 * CPU. This check is racy, but that's ok: the unwinder has other
673 * checks to prevent it from going off the rails.
674 */
675 if (task_on_another_cpu(task))
676 goto err;
677
678 if (regs) {
679 if (user_mode(regs))
680 goto the_end;
681
682 state->ip = regs->ip;
683 state->sp = regs->sp;
684 state->bp = regs->bp;
685 state->regs = regs;
686 state->full_regs = true;
687 state->signal = true;
688
689 } else if (task == current) {
690 asm volatile("lea (%%rip), %0\n\t"
691 "mov %%rsp, %1\n\t"
692 "mov %%rbp, %2\n\t"
693 : "=r" (state->ip), "=r" (state->sp),
694 "=r" (state->bp));
695
696 } else {
697 struct inactive_task_frame *frame = (void *)task->thread.sp;
698
699 state->sp = task->thread.sp + sizeof(*frame);
700 state->bp = READ_ONCE_NOCHECK(frame->bp);
701 state->ip = READ_ONCE_NOCHECK(frame->ret_addr);
702 state->signal = (void *)state->ip == ret_from_fork;
703 }
704
705 if (get_stack_info((unsigned long *)state->sp, state->task,
706 &state->stack_info, &state->stack_mask)) {
707 /*
708 * We weren't on a valid stack. It's possible that
709 * we overflowed a valid stack into a guard page.
710 * See if the next page up is valid so that we can
711 * generate some kind of backtrace if this happens.
712 */
713 void *next_page = (void *)PAGE_ALIGN((unsigned long)state->sp);
714 state->error = true;
715 if (get_stack_info(next_page, state->task, &state->stack_info,
716 &state->stack_mask))
717 return;
718 }
719
720 /*
721 * The caller can provide the address of the first frame directly
722 * (first_frame) or indirectly (regs->sp) to indicate which stack frame
723 * to start unwinding at. Skip ahead until we reach it.
724 */
725
726 /* When starting from regs, skip the regs frame: */
727 if (regs) {
728 unwind_next_frame(state);
729 return;
730 }
731
732 /* Otherwise, skip ahead to the user-specified starting frame: */
733 while (!unwind_done(state) &&
734 (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
735 state->sp <= (unsigned long)first_frame))
736 unwind_next_frame(state);
737
738 return;
739
740err:
741 state->error = true;
742the_end:
743 state->stack_info.type = STACK_TYPE_UNKNOWN;
744}
745EXPORT_SYMBOL_GPL(__unwind_start);