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