Merge tag 'hyperv-next-signed' of git://git.kernel.org/pub/scm/linux/kernel/git/hyper...
[sfrench/cifs-2.6.git] / arch / x86 / kernel / unwind_orc.c
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)                                     \
17                 orc_warn(args);                                         \
18 })
19
20 extern int __start_orc_unwind_ip[];
21 extern int __stop_orc_unwind_ip[];
22 extern struct orc_entry __start_orc_unwind[];
23 extern struct orc_entry __stop_orc_unwind[];
24
25 static bool orc_init __ro_after_init;
26 static unsigned int lookup_num_blocks __ro_after_init;
27
28 static inline unsigned long orc_ip(const int *ip)
29 {
30         return (unsigned long)ip + *ip;
31 }
32
33 static 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
63 static 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
74 static struct orc_entry *orc_module_find(unsigned long ip)
75 {
76         return NULL;
77 }
78 #endif
79
80 #ifdef CONFIG_DYNAMIC_FTRACE
81 static 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  */
93 static struct orc_entry *orc_ftrace_find(unsigned long ip)
94 {
95         struct ftrace_ops *ops;
96         unsigned long caller;
97
98         ops = ftrace_ops_trampoline(ip);
99         if (!ops)
100                 return NULL;
101
102         if (ops->flags & FTRACE_OPS_FL_SAVE_REGS)
103                 caller = (unsigned long)ftrace_regs_call;
104         else
105                 caller = (unsigned long)ftrace_call;
106
107         /* Prevent unlikely recursion */
108         if (ip == caller)
109                 return NULL;
110
111         return orc_find(caller);
112 }
113 #else
114 static struct orc_entry *orc_ftrace_find(unsigned long ip)
115 {
116         return NULL;
117 }
118 #endif
119
120 /*
121  * If we crash with IP==0, the last successfully executed instruction
122  * was probably an indirect function call with a NULL function pointer,
123  * and we don't have unwind information for NULL.
124  * This hardcoded ORC entry for IP==0 allows us to unwind from a NULL function
125  * pointer into its parent and then continue normally from there.
126  */
127 static struct orc_entry null_orc_entry = {
128         .sp_offset = sizeof(long),
129         .sp_reg = ORC_REG_SP,
130         .bp_reg = ORC_REG_UNDEFINED,
131         .type = UNWIND_HINT_TYPE_CALL
132 };
133
134 /* Fake frame pointer entry -- used as a fallback for generated code */
135 static struct orc_entry orc_fp_entry = {
136         .type           = UNWIND_HINT_TYPE_CALL,
137         .sp_reg         = ORC_REG_BP,
138         .sp_offset      = 16,
139         .bp_reg         = ORC_REG_PREV_SP,
140         .bp_offset      = -16,
141         .end            = 0,
142 };
143
144 static struct orc_entry *orc_find(unsigned long ip)
145 {
146         static struct orc_entry *orc;
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
190 #ifdef CONFIG_MODULES
191
192 static DEFINE_MUTEX(sort_mutex);
193 static int *cur_orc_ip_table = __start_orc_unwind_ip;
194 static struct orc_entry *cur_orc_table = __start_orc_unwind;
195
196 static void orc_sort_swap(void *_a, void *_b, int size)
197 {
198         struct orc_entry *orc_a, *orc_b;
199         struct orc_entry orc_tmp;
200         int *a = _a, *b = _b, tmp;
201         int delta = _b - _a;
202
203         /* Swap the .orc_unwind_ip entries: */
204         tmp = *a;
205         *a = *b + delta;
206         *b = tmp - delta;
207
208         /* Swap the corresponding .orc_unwind entries: */
209         orc_a = cur_orc_table + (a - cur_orc_ip_table);
210         orc_b = cur_orc_table + (b - cur_orc_ip_table);
211         orc_tmp = *orc_a;
212         *orc_a = *orc_b;
213         *orc_b = orc_tmp;
214 }
215
216 static int orc_sort_cmp(const void *_a, const void *_b)
217 {
218         struct orc_entry *orc_a;
219         const int *a = _a, *b = _b;
220         unsigned long a_val = orc_ip(a);
221         unsigned long b_val = orc_ip(b);
222
223         if (a_val > b_val)
224                 return 1;
225         if (a_val < b_val)
226                 return -1;
227
228         /*
229          * The "weak" section terminator entries need to always be on the left
230          * to ensure the lookup code skips them in favor of real entries.
231          * These terminator entries exist to handle any gaps created by
232          * whitelisted .o files which didn't get objtool generation.
233          */
234         orc_a = cur_orc_table + (a - cur_orc_ip_table);
235         return orc_a->sp_reg == ORC_REG_UNDEFINED && !orc_a->end ? -1 : 1;
236 }
237
238 void unwind_module_init(struct module *mod, void *_orc_ip, size_t orc_ip_size,
239                         void *_orc, size_t orc_size)
240 {
241         int *orc_ip = _orc_ip;
242         struct orc_entry *orc = _orc;
243         unsigned int num_entries = orc_ip_size / sizeof(int);
244
245         WARN_ON_ONCE(orc_ip_size % sizeof(int) != 0 ||
246                      orc_size % sizeof(*orc) != 0 ||
247                      num_entries != orc_size / sizeof(*orc));
248
249         /*
250          * The 'cur_orc_*' globals allow the orc_sort_swap() callback to
251          * associate an .orc_unwind_ip table entry with its corresponding
252          * .orc_unwind entry so they can both be swapped.
253          */
254         mutex_lock(&sort_mutex);
255         cur_orc_ip_table = orc_ip;
256         cur_orc_table = orc;
257         sort(orc_ip, num_entries, sizeof(int), orc_sort_cmp, orc_sort_swap);
258         mutex_unlock(&sort_mutex);
259
260         mod->arch.orc_unwind_ip = orc_ip;
261         mod->arch.orc_unwind = orc;
262         mod->arch.num_orcs = num_entries;
263 }
264 #endif
265
266 void __init unwind_init(void)
267 {
268         size_t orc_ip_size = (void *)__stop_orc_unwind_ip - (void *)__start_orc_unwind_ip;
269         size_t orc_size = (void *)__stop_orc_unwind - (void *)__start_orc_unwind;
270         size_t num_entries = orc_ip_size / sizeof(int);
271         struct orc_entry *orc;
272         int i;
273
274         if (!num_entries || orc_ip_size % sizeof(int) != 0 ||
275             orc_size % sizeof(struct orc_entry) != 0 ||
276             num_entries != orc_size / sizeof(struct orc_entry)) {
277                 orc_warn("WARNING: Bad or missing .orc_unwind table.  Disabling unwinder.\n");
278                 return;
279         }
280
281         /*
282          * Note, the orc_unwind and orc_unwind_ip tables were already
283          * sorted at build time via the 'sorttable' tool.
284          * It's ready for binary search straight away, no need to sort it.
285          */
286
287         /* Initialize the fast lookup table: */
288         lookup_num_blocks = orc_lookup_end - orc_lookup;
289         for (i = 0; i < lookup_num_blocks-1; i++) {
290                 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
291                                  num_entries,
292                                  LOOKUP_START_IP + (LOOKUP_BLOCK_SIZE * i));
293                 if (!orc) {
294                         orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
295                         return;
296                 }
297
298                 orc_lookup[i] = orc - __start_orc_unwind;
299         }
300
301         /* Initialize the ending block: */
302         orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind, num_entries,
303                          LOOKUP_STOP_IP);
304         if (!orc) {
305                 orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
306                 return;
307         }
308         orc_lookup[lookup_num_blocks-1] = orc - __start_orc_unwind;
309
310         orc_init = true;
311 }
312
313 unsigned long unwind_get_return_address(struct unwind_state *state)
314 {
315         if (unwind_done(state))
316                 return 0;
317
318         return __kernel_text_address(state->ip) ? state->ip : 0;
319 }
320 EXPORT_SYMBOL_GPL(unwind_get_return_address);
321
322 unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
323 {
324         struct task_struct *task = state->task;
325
326         if (unwind_done(state))
327                 return NULL;
328
329         if (state->regs)
330                 return &state->regs->ip;
331
332         if (task != current && state->sp == task->thread.sp) {
333                 struct inactive_task_frame *frame = (void *)task->thread.sp;
334                 return &frame->ret_addr;
335         }
336
337         if (state->sp)
338                 return (unsigned long *)state->sp - 1;
339
340         return NULL;
341 }
342
343 static bool stack_access_ok(struct unwind_state *state, unsigned long _addr,
344                             size_t len)
345 {
346         struct stack_info *info = &state->stack_info;
347         void *addr = (void *)_addr;
348
349         if (!on_stack(info, addr, len) &&
350             (get_stack_info(addr, state->task, info, &state->stack_mask)))
351                 return false;
352
353         return true;
354 }
355
356 static bool deref_stack_reg(struct unwind_state *state, unsigned long addr,
357                             unsigned long *val)
358 {
359         if (!stack_access_ok(state, addr, sizeof(long)))
360                 return false;
361
362         *val = READ_ONCE_NOCHECK(*(unsigned long *)addr);
363         return true;
364 }
365
366 static bool deref_stack_regs(struct unwind_state *state, unsigned long addr,
367                              unsigned long *ip, unsigned long *sp)
368 {
369         struct pt_regs *regs = (struct pt_regs *)addr;
370
371         /* x86-32 support will be more complicated due to the &regs->sp hack */
372         BUILD_BUG_ON(IS_ENABLED(CONFIG_X86_32));
373
374         if (!stack_access_ok(state, addr, sizeof(struct pt_regs)))
375                 return false;
376
377         *ip = regs->ip;
378         *sp = regs->sp;
379         return true;
380 }
381
382 static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr,
383                                   unsigned long *ip, unsigned long *sp)
384 {
385         struct pt_regs *regs = (void *)addr - IRET_FRAME_OFFSET;
386
387         if (!stack_access_ok(state, addr, IRET_FRAME_SIZE))
388                 return false;
389
390         *ip = regs->ip;
391         *sp = regs->sp;
392         return true;
393 }
394
395 /*
396  * If state->regs is non-NULL, and points to a full pt_regs, just get the reg
397  * value from state->regs.
398  *
399  * Otherwise, if state->regs just points to IRET regs, and the previous frame
400  * had full regs, it's safe to get the value from the previous regs.  This can
401  * happen when early/late IRQ entry code gets interrupted by an NMI.
402  */
403 static bool get_reg(struct unwind_state *state, unsigned int reg_off,
404                     unsigned long *val)
405 {
406         unsigned int reg = reg_off/8;
407
408         if (!state->regs)
409                 return false;
410
411         if (state->full_regs) {
412                 *val = ((unsigned long *)state->regs)[reg];
413                 return true;
414         }
415
416         if (state->prev_regs) {
417                 *val = ((unsigned long *)state->prev_regs)[reg];
418                 return true;
419         }
420
421         return false;
422 }
423
424 bool unwind_next_frame(struct unwind_state *state)
425 {
426         unsigned long ip_p, sp, tmp, orig_ip = state->ip, prev_sp = state->sp;
427         enum stack_type prev_type = state->stack_info.type;
428         struct orc_entry *orc;
429         bool indirect = false;
430
431         if (unwind_done(state))
432                 return false;
433
434         /* Don't let modules unload while we're reading their ORC data. */
435         preempt_disable();
436
437         /* End-of-stack check for user tasks: */
438         if (state->regs && user_mode(state->regs))
439                 goto the_end;
440
441         /*
442          * Find the orc_entry associated with the text address.
443          *
444          * For a call frame (as opposed to a signal frame), state->ip points to
445          * the instruction after the call.  That instruction's stack layout
446          * could be different from the call instruction's layout, for example
447          * if the call was to a noreturn function.  So get the ORC data for the
448          * call instruction itself.
449          */
450         orc = orc_find(state->signal ? state->ip : state->ip - 1);
451         if (!orc) {
452                 /*
453                  * As a fallback, try to assume this code uses a frame pointer.
454                  * This is useful for generated code, like BPF, which ORC
455                  * doesn't know about.  This is just a guess, so the rest of
456                  * the unwind is no longer considered reliable.
457                  */
458                 orc = &orc_fp_entry;
459                 state->error = true;
460         }
461
462         /* End-of-stack check for kernel threads: */
463         if (orc->sp_reg == ORC_REG_UNDEFINED) {
464                 if (!orc->end)
465                         goto err;
466
467                 goto the_end;
468         }
469
470         /* Find the previous frame's stack: */
471         switch (orc->sp_reg) {
472         case ORC_REG_SP:
473                 sp = state->sp + orc->sp_offset;
474                 break;
475
476         case ORC_REG_BP:
477                 sp = state->bp + orc->sp_offset;
478                 break;
479
480         case ORC_REG_SP_INDIRECT:
481                 sp = state->sp + orc->sp_offset;
482                 indirect = true;
483                 break;
484
485         case ORC_REG_BP_INDIRECT:
486                 sp = state->bp + orc->sp_offset;
487                 indirect = true;
488                 break;
489
490         case ORC_REG_R10:
491                 if (!get_reg(state, offsetof(struct pt_regs, r10), &sp)) {
492                         orc_warn_current("missing R10 value at %pB\n",
493                                          (void *)state->ip);
494                         goto err;
495                 }
496                 break;
497
498         case ORC_REG_R13:
499                 if (!get_reg(state, offsetof(struct pt_regs, r13), &sp)) {
500                         orc_warn_current("missing R13 value at %pB\n",
501                                          (void *)state->ip);
502                         goto err;
503                 }
504                 break;
505
506         case ORC_REG_DI:
507                 if (!get_reg(state, offsetof(struct pt_regs, di), &sp)) {
508                         orc_warn_current("missing RDI value at %pB\n",
509                                          (void *)state->ip);
510                         goto err;
511                 }
512                 break;
513
514         case ORC_REG_DX:
515                 if (!get_reg(state, offsetof(struct pt_regs, dx), &sp)) {
516                         orc_warn_current("missing DX value at %pB\n",
517                                          (void *)state->ip);
518                         goto err;
519                 }
520                 break;
521
522         default:
523                 orc_warn("unknown SP base reg %d at %pB\n",
524                          orc->sp_reg, (void *)state->ip);
525                 goto err;
526         }
527
528         if (indirect) {
529                 if (!deref_stack_reg(state, sp, &sp))
530                         goto err;
531         }
532
533         /* Find IP, SP and possibly regs: */
534         switch (orc->type) {
535         case UNWIND_HINT_TYPE_CALL:
536                 ip_p = sp - sizeof(long);
537
538                 if (!deref_stack_reg(state, ip_p, &state->ip))
539                         goto err;
540
541                 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
542                                                   state->ip, (void *)ip_p);
543
544                 state->sp = sp;
545                 state->regs = NULL;
546                 state->prev_regs = NULL;
547                 state->signal = false;
548                 break;
549
550         case UNWIND_HINT_TYPE_REGS:
551                 if (!deref_stack_regs(state, sp, &state->ip, &state->sp)) {
552                         orc_warn_current("can't access registers at %pB\n",
553                                          (void *)orig_ip);
554                         goto err;
555                 }
556
557                 state->regs = (struct pt_regs *)sp;
558                 state->prev_regs = NULL;
559                 state->full_regs = true;
560                 state->signal = true;
561                 break;
562
563         case UNWIND_HINT_TYPE_REGS_PARTIAL:
564                 if (!deref_stack_iret_regs(state, sp, &state->ip, &state->sp)) {
565                         orc_warn_current("can't access iret registers at %pB\n",
566                                          (void *)orig_ip);
567                         goto err;
568                 }
569
570                 if (state->full_regs)
571                         state->prev_regs = state->regs;
572                 state->regs = (void *)sp - IRET_FRAME_OFFSET;
573                 state->full_regs = false;
574                 state->signal = true;
575                 break;
576
577         default:
578                 orc_warn("unknown .orc_unwind entry type %d at %pB\n",
579                          orc->type, (void *)orig_ip);
580                 goto err;
581         }
582
583         /* Find BP: */
584         switch (orc->bp_reg) {
585         case ORC_REG_UNDEFINED:
586                 if (get_reg(state, offsetof(struct pt_regs, bp), &tmp))
587                         state->bp = tmp;
588                 break;
589
590         case ORC_REG_PREV_SP:
591                 if (!deref_stack_reg(state, sp + orc->bp_offset, &state->bp))
592                         goto err;
593                 break;
594
595         case ORC_REG_BP:
596                 if (!deref_stack_reg(state, state->bp + orc->bp_offset, &state->bp))
597                         goto err;
598                 break;
599
600         default:
601                 orc_warn("unknown BP base reg %d for ip %pB\n",
602                          orc->bp_reg, (void *)orig_ip);
603                 goto err;
604         }
605
606         /* Prevent a recursive loop due to bad ORC data: */
607         if (state->stack_info.type == prev_type &&
608             on_stack(&state->stack_info, (void *)state->sp, sizeof(long)) &&
609             state->sp <= prev_sp) {
610                 orc_warn_current("stack going in the wrong direction? at %pB\n",
611                                  (void *)orig_ip);
612                 goto err;
613         }
614
615         preempt_enable();
616         return true;
617
618 err:
619         state->error = true;
620
621 the_end:
622         preempt_enable();
623         state->stack_info.type = STACK_TYPE_UNKNOWN;
624         return false;
625 }
626 EXPORT_SYMBOL_GPL(unwind_next_frame);
627
628 void __unwind_start(struct unwind_state *state, struct task_struct *task,
629                     struct pt_regs *regs, unsigned long *first_frame)
630 {
631         memset(state, 0, sizeof(*state));
632         state->task = task;
633
634         if (!orc_init)
635                 goto err;
636
637         /*
638          * Refuse to unwind the stack of a task while it's executing on another
639          * CPU.  This check is racy, but that's ok: the unwinder has other
640          * checks to prevent it from going off the rails.
641          */
642         if (task_on_another_cpu(task))
643                 goto err;
644
645         if (regs) {
646                 if (user_mode(regs))
647                         goto the_end;
648
649                 state->ip = regs->ip;
650                 state->sp = regs->sp;
651                 state->bp = regs->bp;
652                 state->regs = regs;
653                 state->full_regs = true;
654                 state->signal = true;
655
656         } else if (task == current) {
657                 asm volatile("lea (%%rip), %0\n\t"
658                              "mov %%rsp, %1\n\t"
659                              "mov %%rbp, %2\n\t"
660                              : "=r" (state->ip), "=r" (state->sp),
661                                "=r" (state->bp));
662
663         } else {
664                 struct inactive_task_frame *frame = (void *)task->thread.sp;
665
666                 state->sp = task->thread.sp;
667                 state->bp = READ_ONCE_NOCHECK(frame->bp);
668                 state->ip = READ_ONCE_NOCHECK(frame->ret_addr);
669                 state->signal = (void *)state->ip == ret_from_fork;
670         }
671
672         if (get_stack_info((unsigned long *)state->sp, state->task,
673                            &state->stack_info, &state->stack_mask)) {
674                 /*
675                  * We weren't on a valid stack.  It's possible that
676                  * we overflowed a valid stack into a guard page.
677                  * See if the next page up is valid so that we can
678                  * generate some kind of backtrace if this happens.
679                  */
680                 void *next_page = (void *)PAGE_ALIGN((unsigned long)state->sp);
681                 state->error = true;
682                 if (get_stack_info(next_page, state->task, &state->stack_info,
683                                    &state->stack_mask))
684                         return;
685         }
686
687         /*
688          * The caller can provide the address of the first frame directly
689          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
690          * to start unwinding at.  Skip ahead until we reach it.
691          */
692
693         /* When starting from regs, skip the regs frame: */
694         if (regs) {
695                 unwind_next_frame(state);
696                 return;
697         }
698
699         /* Otherwise, skip ahead to the user-specified starting frame: */
700         while (!unwind_done(state) &&
701                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
702                         state->sp < (unsigned long)first_frame))
703                 unwind_next_frame(state);
704
705         return;
706
707 err:
708         state->error = true;
709 the_end:
710         state->stack_info.type = STACK_TYPE_UNKNOWN;
711 }
712 EXPORT_SYMBOL_GPL(__unwind_start);