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 && !state->error)                    \
  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 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
 119 static 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  */
 132 static 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 /* Fake frame pointer entry -- used as a fallback for generated code */
 140 static struct orc_entry orc_fp_entry = {
 141         .type           = UNWIND_HINT_TYPE_CALL,
 142         .sp_reg         = ORC_REG_BP,
 143         .sp_offset      = 16,
 144         .bp_reg         = ORC_REG_PREV_SP,
 145         .bp_offset      = -16,
 146         .end            = 0,
 147 };
 148
 149 static struct orc_entry *orc_find(unsigned long ip)
 150 {
 151         static struct orc_entry *orc;
 152
 153         if (ip == 0)
 154                 return &null_orc_entry;
 155
 156         /* For non-init vmlinux addresses, use the fast lookup table: */
 157         if (ip >= LOOKUP_START_IP && ip < LOOKUP_STOP_IP) {
 158                 unsigned int idx, start, stop;
 159
 160                 idx = (ip - LOOKUP_START_IP) / LOOKUP_BLOCK_SIZE;
 161
 162                 if (unlikely((idx >= lookup_num_blocks-1))) {
 163                         orc_warn("WARNING: bad lookup idx: idx=%u num=%u ip=%pB\n",
 164                                  idx, lookup_num_blocks, (void *)ip);
 165                         return NULL;
 166                 }
 167
 168                 start = orc_lookup[idx];
 169                 stop = orc_lookup[idx + 1] + 1;
 170
 171                 if (unlikely((__start_orc_unwind + start >= __stop_orc_unwind) ||
 172                              (__start_orc_unwind + stop > __stop_orc_unwind))) {
 173                         orc_warn("WARNING: bad lookup value: idx=%u num=%u start=%u stop=%u ip=%pB\n",
 174                                  idx, lookup_num_blocks, start, stop, (void *)ip);
 175                         return NULL;
 176                 }
 177
 178                 return __orc_find(__start_orc_unwind_ip + start,
 179                                   __start_orc_unwind + start, stop - start, ip);
 180         }
 181
 182         /* vmlinux .init slow lookup: */
 183         if (is_kernel_inittext(ip))
 184                 return __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 185                                   __stop_orc_unwind_ip - __start_orc_unwind_ip, ip);
 186
 187         /* Module lookup: */
 188         orc = orc_module_find(ip);
 189         if (orc)
 190                 return orc;
 191
 192         return orc_ftrace_find(ip);
 193 }
 194
 195 #ifdef CONFIG_MODULES
 196
 197 static DEFINE_MUTEX(sort_mutex);
 198 static int *cur_orc_ip_table = __start_orc_unwind_ip;
 199 static struct orc_entry *cur_orc_table = __start_orc_unwind;
 200
 201 static void orc_sort_swap(void *_a, void *_b, int size)
 202 {
 203         struct orc_entry *orc_a, *orc_b;
 204         struct orc_entry orc_tmp;
 205         int *a = _a, *b = _b, tmp;
 206         int delta = _b - _a;
 207
 208         /* Swap the .orc_unwind_ip entries: */
 209         tmp = *a;
 210         *a = *b + delta;
 211         *b = tmp - delta;
 212
 213         /* Swap the corresponding .orc_unwind entries: */
 214         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 215         orc_b = cur_orc_table + (b - cur_orc_ip_table);
 216         orc_tmp = *orc_a;
 217         *orc_a = *orc_b;
 218         *orc_b = orc_tmp;
 219 }
 220
 221 static int orc_sort_cmp(const void *_a, const void *_b)
 222 {
 223         struct orc_entry *orc_a;
 224         const int *a = _a, *b = _b;
 225         unsigned long a_val = orc_ip(a);
 226         unsigned long b_val = orc_ip(b);
 227
 228         if (a_val > b_val)
 229                 return 1;
 230         if (a_val < b_val)
 231                 return -1;
 232
 233         /*
 234          * The "weak" section terminator entries need to always be on the left
 235          * to ensure the lookup code skips them in favor of real entries.
 236          * These terminator entries exist to handle any gaps created by
 237          * whitelisted .o files which didn't get objtool generation.
 238          */
 239         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 240         return orc_a->sp_reg == ORC_REG_UNDEFINED && !orc_a->end ? -1 : 1;
 241 }
 242
 243 void unwind_module_init(struct module *mod, void *_orc_ip, size_t orc_ip_size,
 244                         void *_orc, size_t orc_size)
 245 {
 246         int *orc_ip = _orc_ip;
 247         struct orc_entry *orc = _orc;
 248         unsigned int num_entries = orc_ip_size / sizeof(int);
 249
 250         WARN_ON_ONCE(orc_ip_size % sizeof(int) != 0 ||
 251                      orc_size % sizeof(*orc) != 0 ||
 252                      num_entries != orc_size / sizeof(*orc));
 253
 254         /*
 255          * The 'cur_orc_*' globals allow the orc_sort_swap() callback to
 256          * associate an .orc_unwind_ip table entry with its corresponding
 257          * .orc_unwind entry so they can both be swapped.
 258          */
 259         mutex_lock(&sort_mutex);
 260         cur_orc_ip_table = orc_ip;
 261         cur_orc_table = orc;
 262         sort(orc_ip, num_entries, sizeof(int), orc_sort_cmp, orc_sort_swap);
 263         mutex_unlock(&sort_mutex);
 264
 265         mod->arch.orc_unwind_ip = orc_ip;
 266         mod->arch.orc_unwind = orc;
 267         mod->arch.num_orcs = num_entries;
 268 }
 269 #endif
 270
 271 void __init unwind_init(void)
 272 {
 273         size_t orc_ip_size = (void *)__stop_orc_unwind_ip - (void *)__start_orc_unwind_ip;
 274         size_t orc_size = (void *)__stop_orc_unwind - (void *)__start_orc_unwind;
 275         size_t num_entries = orc_ip_size / sizeof(int);
 276         struct orc_entry *orc;
 277         int i;
 278
 279         if (!num_entries || orc_ip_size % sizeof(int) != 0 ||
 280             orc_size % sizeof(struct orc_entry) != 0 ||
 281             num_entries != orc_size / sizeof(struct orc_entry)) {
 282                 orc_warn("WARNING: Bad or missing .orc_unwind table.  Disabling unwinder.\n");
 283                 return;
 284         }
 285
 286         /*
 287          * Note, the orc_unwind and orc_unwind_ip tables were already
 288          * sorted at build time via the 'sorttable' tool.
 289          * It's ready for binary search straight away, no need to sort it.
 290          */
 291
 292         /* Initialize the fast lookup table: */
 293         lookup_num_blocks = orc_lookup_end - orc_lookup;
 294         for (i = 0; i < lookup_num_blocks-1; i++) {
 295                 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 296                                  num_entries,
 297                                  LOOKUP_START_IP + (LOOKUP_BLOCK_SIZE * i));
 298                 if (!orc) {
 299                         orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 300                         return;
 301                 }
 302
 303                 orc_lookup[i] = orc - __start_orc_unwind;
 304         }
 305
 306         /* Initialize the ending block: */
 307         orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind, num_entries,
 308                          LOOKUP_STOP_IP);
 309         if (!orc) {
 310                 orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 311                 return;
 312         }
 313         orc_lookup[lookup_num_blocks-1] = orc - __start_orc_unwind;
 314
 315         orc_init = true;
 316 }
 317
 318 unsigned long unwind_get_return_address(struct unwind_state *state)
 319 {
 320         if (unwind_done(state))
 321                 return 0;
 322
 323         return __kernel_text_address(state->ip) ? state->ip : 0;
 324 }
 325 EXPORT_SYMBOL_GPL(unwind_get_return_address);
 326
 327 unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
 328 {
 329         if (unwind_done(state))
 330                 return NULL;
 331
 332         if (state->regs)
 333                 return &state->regs->ip;
 334
 335         if (state->sp)
 336                 return (unsigned long *)state->sp - 1;
 337
 338         return NULL;
 339 }
 340
 341 static bool stack_access_ok(struct unwind_state *state, unsigned long _addr,
 342                             size_t len)
 343 {
 344         struct stack_info *info = &state->stack_info;
 345         void *addr = (void *)_addr;
 346
 347         if (on_stack(info, addr, len))
 348                 return true;
 349
 350         return !get_stack_info(addr, state->task, info, &state->stack_mask) &&
 351                 on_stack(info, addr, len);
 352 }
 353
 354 static bool deref_stack_reg(struct unwind_state *state, unsigned long addr,
 355                             unsigned long *val)
 356 {
 357         if (!stack_access_ok(state, addr, sizeof(long)))
 358                 return false;
 359
 360         *val = READ_ONCE_NOCHECK(*(unsigned long *)addr);
 361         return true;
 362 }
 363
 364 static bool deref_stack_regs(struct unwind_state *state, unsigned long addr,
 365                              unsigned long *ip, unsigned long *sp)
 366 {
 367         struct pt_regs *regs = (struct pt_regs *)addr;
 368
 369         /* x86-32 support will be more complicated due to the &regs->sp hack */
 370         BUILD_BUG_ON(IS_ENABLED(CONFIG_X86_32));
 371
 372         if (!stack_access_ok(state, addr, sizeof(struct pt_regs)))
 373                 return false;
 374
 375         *ip = READ_ONCE_NOCHECK(regs->ip);
 376         *sp = READ_ONCE_NOCHECK(regs->sp);
 377         return true;
 378 }
 379
 380 static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr,
 381                                   unsigned long *ip, unsigned long *sp)
 382 {
 383         struct pt_regs *regs = (void *)addr - IRET_FRAME_OFFSET;
 384
 385         if (!stack_access_ok(state, addr, IRET_FRAME_SIZE))
 386                 return false;
 387
 388         *ip = READ_ONCE_NOCHECK(regs->ip);
 389         *sp = READ_ONCE_NOCHECK(regs->sp);
 390         return true;
 391 }
 392
 393 /*
 394  * If state->regs is non-NULL, and points to a full pt_regs, just get the reg
 395  * value from state->regs.
 396  *
 397  * Otherwise, if state->regs just points to IRET regs, and the previous frame
 398  * had full regs, it's safe to get the value from the previous regs.  This can
 399  * happen when early/late IRQ entry code gets interrupted by an NMI.
 400  */
 401 static bool get_reg(struct unwind_state *state, unsigned int reg_off,
 402                     unsigned long *val)
 403 {
 404         unsigned int reg = reg_off/8;
 405
 406         if (!state->regs)
 407                 return false;
 408
 409         if (state->full_regs) {
 410                 *val = READ_ONCE_NOCHECK(((unsigned long *)state->regs)[reg]);
 411                 return true;
 412         }
 413
 414         if (state->prev_regs) {
 415                 *val = READ_ONCE_NOCHECK(((unsigned long *)state->prev_regs)[reg]);
 416                 return true;
 417         }
 418
 419         return false;
 420 }
 421
 422 bool unwind_next_frame(struct unwind_state *state)
 423 {
 424         unsigned long ip_p, sp, tmp, orig_ip = state->ip, prev_sp = state->sp;
 425         enum stack_type prev_type = state->stack_info.type;
 426         struct orc_entry *orc;
 427         bool indirect = false;
 428
 429         if (unwind_done(state))
 430                 return false;
 431
 432         /* Don't let modules unload while we're reading their ORC data. */
 433         preempt_disable();
 434
 435         /* End-of-stack check for user tasks: */
 436         if (state->regs && user_mode(state->regs))
 437                 goto the_end;
 438
 439         /*
 440          * Find the orc_entry associated with the text address.
 441          *
 442          * For a call frame (as opposed to a signal frame), state->ip points to
 443          * the instruction after the call.  That instruction's stack layout
 444          * could be different from the call instruction's layout, for example
 445          * if the call was to a noreturn function.  So get the ORC data for the
 446          * call instruction itself.
 447          */
 448         orc = orc_find(state->signal ? state->ip : state->ip - 1);
 449         if (!orc) {
 450                 /*
 451                  * As a fallback, try to assume this code uses a frame pointer.
 452                  * This is useful for generated code, like BPF, which ORC
 453                  * doesn't know about.  This is just a guess, so the rest of
 454                  * the unwind is no longer considered reliable.
 455                  */
 456                 orc = &orc_fp_entry;
 457                 state->error = true;
 458         }
 459
 460         /* End-of-stack check for kernel threads: */
 461         if (orc->sp_reg == ORC_REG_UNDEFINED) {
 462                 if (!orc->end)
 463                         goto err;
 464
 465                 goto the_end;
 466         }
 467
 468         /* Find the previous frame's stack: */
 469         switch (orc->sp_reg) {
 470         case ORC_REG_SP:
 471                 sp = state->sp + orc->sp_offset;
 472                 break;
 473
 474         case ORC_REG_BP:
 475                 sp = state->bp + orc->sp_offset;
 476                 break;
 477
 478         case ORC_REG_SP_INDIRECT:
 479                 sp = state->sp;
 480                 indirect = true;
 481                 break;
 482
 483         case ORC_REG_BP_INDIRECT:
 484                 sp = state->bp + orc->sp_offset;
 485                 indirect = true;
 486                 break;
 487
 488         case ORC_REG_R10:
 489                 if (!get_reg(state, offsetof(struct pt_regs, r10), &sp)) {
 490                         orc_warn_current("missing R10 value at %pB\n",
 491                                          (void *)state->ip);
 492                         goto err;
 493                 }
 494                 break;
 495
 496         case ORC_REG_R13:
 497                 if (!get_reg(state, offsetof(struct pt_regs, r13), &sp)) {
 498                         orc_warn_current("missing R13 value at %pB\n",
 499                                          (void *)state->ip);
 500                         goto err;
 501                 }
 502                 break;
 503
 504         case ORC_REG_DI:
 505                 if (!get_reg(state, offsetof(struct pt_regs, di), &sp)) {
 506                         orc_warn_current("missing RDI value at %pB\n",
 507                                          (void *)state->ip);
 508                         goto err;
 509                 }
 510                 break;
 511
 512         case ORC_REG_DX:
 513                 if (!get_reg(state, offsetof(struct pt_regs, dx), &sp)) {
 514                         orc_warn_current("missing DX value at %pB\n",
 515                                          (void *)state->ip);
 516                         goto err;
 517                 }
 518                 break;
 519
 520         default:
 521                 orc_warn("unknown SP base reg %d at %pB\n",
 522                          orc->sp_reg, (void *)state->ip);
 523                 goto err;
 524         }
 525
 526         if (indirect) {
 527                 if (!deref_stack_reg(state, sp, &sp))
 528                         goto err;
 529
 530                 if (orc->sp_reg == ORC_REG_SP_INDIRECT)
 531                         sp += orc->sp_offset;
 532         }
 533
 534         /* Find IP, SP and possibly regs: */
 535         switch (orc->type) {
 536         case UNWIND_HINT_TYPE_CALL:
 537                 ip_p = sp - sizeof(long);
 538
 539                 if (!deref_stack_reg(state, ip_p, &state->ip))
 540                         goto err;
 541
 542                 state->ip = unwind_recover_ret_addr(state, state->ip,
 543                                                     (unsigned long *)ip_p);
 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                  * There is a small chance to interrupt at the entry of
 558                  * arch_rethook_trampoline() where the ORC info doesn't exist.
 559                  * That point is right after the RET to arch_rethook_trampoline()
 560                  * which was modified return address.
 561                  * At that point, the @addr_p of the unwind_recover_rethook()
 562                  * (this has to point the address of the stack entry storing
 563                  * the modified return address) must be "SP - (a stack entry)"
 564                  * because SP is incremented by the RET.
 565                  */
 566                 state->ip = unwind_recover_rethook(state, state->ip,
 567                                 (unsigned long *)(state->sp - sizeof(long)));
 568                 state->regs = (struct pt_regs *)sp;
 569                 state->prev_regs = NULL;
 570                 state->full_regs = true;
 571                 state->signal = true;
 572                 break;
 573
 574         case UNWIND_HINT_TYPE_REGS_PARTIAL:
 575                 if (!deref_stack_iret_regs(state, sp, &state->ip, &state->sp)) {
 576                         orc_warn_current("can't access iret registers at %pB\n",
 577                                          (void *)orig_ip);
 578                         goto err;
 579                 }
 580                 /* See UNWIND_HINT_TYPE_REGS case comment. */
 581                 state->ip = unwind_recover_rethook(state, state->ip,
 582                                 (unsigned long *)(state->sp - sizeof(long)));
 583
 584                 if (state->full_regs)
 585                         state->prev_regs = state->regs;
 586                 state->regs = (void *)sp - IRET_FRAME_OFFSET;
 587                 state->full_regs = false;
 588                 state->signal = true;
 589                 break;
 590
 591         default:
 592                 orc_warn("unknown .orc_unwind entry type %d at %pB\n",
 593                          orc->type, (void *)orig_ip);
 594                 goto err;
 595         }
 596
 597         /* Find BP: */
 598         switch (orc->bp_reg) {
 599         case ORC_REG_UNDEFINED:
 600                 if (get_reg(state, offsetof(struct pt_regs, bp), &tmp))
 601                         state->bp = tmp;
 602                 break;
 603
 604         case ORC_REG_PREV_SP:
 605                 if (!deref_stack_reg(state, sp + orc->bp_offset, &state->bp))
 606                         goto err;
 607                 break;
 608
 609         case ORC_REG_BP:
 610                 if (!deref_stack_reg(state, state->bp + orc->bp_offset, &state->bp))
 611                         goto err;
 612                 break;
 613
 614         default:
 615                 orc_warn("unknown BP base reg %d for ip %pB\n",
 616                          orc->bp_reg, (void *)orig_ip);
 617                 goto err;
 618         }
 619
 620         /* Prevent a recursive loop due to bad ORC data: */
 621         if (state->stack_info.type == prev_type &&
 622             on_stack(&state->stack_info, (void *)state->sp, sizeof(long)) &&
 623             state->sp <= prev_sp) {
 624                 orc_warn_current("stack going in the wrong direction? at %pB\n",
 625                                  (void *)orig_ip);
 626                 goto err;
 627         }
 628
 629         preempt_enable();
 630         return true;
 631
 632 err:
 633         state->error = true;
 634
 635 the_end:
 636         preempt_enable();
 637         state->stack_info.type = STACK_TYPE_UNKNOWN;
 638         return false;
 639 }
 640 EXPORT_SYMBOL_GPL(unwind_next_frame);
 641
 642 void __unwind_start(struct unwind_state *state, struct task_struct *task,
 643                     struct pt_regs *regs, unsigned long *first_frame)
 644 {
 645         memset(state, 0, sizeof(*state));
 646         state->task = task;
 647
 648         if (!orc_init)
 649                 goto err;
 650
 651         /*
 652          * Refuse to unwind the stack of a task while it's executing on another
 653          * CPU.  This check is racy, but that's ok: the unwinder has other
 654          * checks to prevent it from going off the rails.
 655          */
 656         if (task_on_another_cpu(task))
 657                 goto err;
 658
 659         if (regs) {
 660                 if (user_mode(regs))
 661                         goto the_end;
 662
 663                 state->ip = regs->ip;
 664                 state->sp = regs->sp;
 665                 state->bp = regs->bp;
 666                 state->regs = regs;
 667                 state->full_regs = true;
 668                 state->signal = true;
 669
 670         } else if (task == current) {
 671                 asm volatile("lea (%%rip), %0\n\t"
 672                              "mov %%rsp, %1\n\t"
 673                              "mov %%rbp, %2\n\t"
 674                              : "=r" (state->ip), "=r" (state->sp),
 675                                "=r" (state->bp));
 676
 677         } else {
 678                 struct inactive_task_frame *frame = (void *)task->thread.sp;
 679
 680                 state->sp = task->thread.sp + sizeof(*frame);
 681                 state->bp = READ_ONCE_NOCHECK(frame->bp);
 682                 state->ip = READ_ONCE_NOCHECK(frame->ret_addr);
 683                 state->signal = (void *)state->ip == ret_from_fork;
 684         }
 685
 686         if (get_stack_info((unsigned long *)state->sp, state->task,
 687                            &state->stack_info, &state->stack_mask)) {
 688                 /*
 689                  * We weren't on a valid stack.  It's possible that
 690                  * we overflowed a valid stack into a guard page.
 691                  * See if the next page up is valid so that we can
 692                  * generate some kind of backtrace if this happens.
 693                  */
 694                 void *next_page = (void *)PAGE_ALIGN((unsigned long)state->sp);
 695                 state->error = true;
 696                 if (get_stack_info(next_page, state->task, &state->stack_info,
 697                                    &state->stack_mask))
 698                         return;
 699         }
 700
 701         /*
 702          * The caller can provide the address of the first frame directly
 703          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
 704          * to start unwinding at.  Skip ahead until we reach it.
 705          */
 706
 707         /* When starting from regs, skip the regs frame: */
 708         if (regs) {
 709                 unwind_next_frame(state);
 710                 return;
 711         }
 712
 713         /* Otherwise, skip ahead to the user-specified starting frame: */
 714         while (!unwind_done(state) &&
 715                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
 716                         state->sp <= (unsigned long)first_frame))
 717                 unwind_next_frame(state);
 718
 719         return;
 720
 721 err:
 722         state->error = true;
 723 the_end:
 724         state->stack_info.type = STACK_TYPE_UNKNOWN;
 725 }
 726 EXPORT_SYMBOL_GPL(__unwind_start);