arch/x86/kernel/unwind_orc.c

   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
  13 extern int __start_orc_unwind_ip[];
  14 extern int __stop_orc_unwind_ip[];
  15 extern struct orc_entry __start_orc_unwind[];
  16 extern struct orc_entry __stop_orc_unwind[];
  17
  18 static DEFINE_MUTEX(sort_mutex);
  19 int *cur_orc_ip_table = __start_orc_unwind_ip;
  20 struct orc_entry *cur_orc_table = __start_orc_unwind;
  21
  22 unsigned int lookup_num_blocks;
  23 bool orc_init;
  24
  25 static inline unsigned long orc_ip(const int *ip)
  26 {
  27         return (unsigned long)ip + *ip;
  28 }
  29
  30 static 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
  60 static 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
  71 static struct orc_entry *orc_module_find(unsigned long ip)
  72 {
  73         return NULL;
  74 }
  75 #endif
  76
  77 #ifdef CONFIG_DYNAMIC_FTRACE
  78 static 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  */
  90 static 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
 111 static 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  */
 124 static 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 */
 132 static 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
 141 static 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
 190 static 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
 210 static 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
 233 void 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
 261 void __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
 306 unsigned 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 }
 313 EXPORT_SYMBOL_GPL(unwind_get_return_address);
 314
 315 unsigned 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
 329 static 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
 342 static 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
 352 static 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 &regs->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
 368 static 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
 381 bool 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
 572 err:
 573         state->error = true;
 574
 575 the_end:
 576         preempt_enable();
 577         state->stack_info.type = STACK_TYPE_UNKNOWN;
 578         return false;
 579 }
 580 EXPORT_SYMBOL_GPL(unwind_next_frame);
 581
 582 void __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
 656 done:
 657         state->stack_info.type = STACK_TYPE_UNKNOWN;
 658         return;
 659 }
 660 EXPORT_SYMBOL_GPL(__unwind_start);