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 undwinder 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 entrie, 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 static struct orc_entry *orc_find(unsigned long ip)
 132 {
 133         static struct orc_entry *orc;
 134
 135         if (!orc_init)
 136                 return NULL;
 137
 138         if (ip == 0)
 139                 return &null_orc_entry;
 140
 141         /* For non-init vmlinux addresses, use the fast lookup table: */
 142         if (ip >= LOOKUP_START_IP && ip < LOOKUP_STOP_IP) {
 143                 unsigned int idx, start, stop;
 144
 145                 idx = (ip - LOOKUP_START_IP) / LOOKUP_BLOCK_SIZE;
 146
 147                 if (unlikely((idx >= lookup_num_blocks-1))) {
 148                         orc_warn("WARNING: bad lookup idx: idx=%u num=%u ip=%pB\n",
 149                                  idx, lookup_num_blocks, (void *)ip);
 150                         return NULL;
 151                 }
 152
 153                 start = orc_lookup[idx];
 154                 stop = orc_lookup[idx + 1] + 1;
 155
 156                 if (unlikely((__start_orc_unwind + start >= __stop_orc_unwind) ||
 157                              (__start_orc_unwind + stop > __stop_orc_unwind))) {
 158                         orc_warn("WARNING: bad lookup value: idx=%u num=%u start=%u stop=%u ip=%pB\n",
 159                                  idx, lookup_num_blocks, start, stop, (void *)ip);
 160                         return NULL;
 161                 }
 162
 163                 return __orc_find(__start_orc_unwind_ip + start,
 164                                   __start_orc_unwind + start, stop - start, ip);
 165         }
 166
 167         /* vmlinux .init slow lookup: */
 168         if (init_kernel_text(ip))
 169                 return __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 170                                   __stop_orc_unwind_ip - __start_orc_unwind_ip, ip);
 171
 172         /* Module lookup: */
 173         orc = orc_module_find(ip);
 174         if (orc)
 175                 return orc;
 176
 177         return orc_ftrace_find(ip);
 178 }
 179
 180 static void orc_sort_swap(void *_a, void *_b, int size)
 181 {
 182         struct orc_entry *orc_a, *orc_b;
 183         struct orc_entry orc_tmp;
 184         int *a = _a, *b = _b, tmp;
 185         int delta = _b - _a;
 186
 187         /* Swap the .orc_unwind_ip entries: */
 188         tmp = *a;
 189         *a = *b + delta;
 190         *b = tmp - delta;
 191
 192         /* Swap the corresponding .orc_unwind entries: */
 193         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 194         orc_b = cur_orc_table + (b - cur_orc_ip_table);
 195         orc_tmp = *orc_a;
 196         *orc_a = *orc_b;
 197         *orc_b = orc_tmp;
 198 }
 199
 200 static int orc_sort_cmp(const void *_a, const void *_b)
 201 {
 202         struct orc_entry *orc_a;
 203         const int *a = _a, *b = _b;
 204         unsigned long a_val = orc_ip(a);
 205         unsigned long b_val = orc_ip(b);
 206
 207         if (a_val > b_val)
 208                 return 1;
 209         if (a_val < b_val)
 210                 return -1;
 211
 212         /*
 213          * The "weak" section terminator entries need to always be on the left
 214          * to ensure the lookup code skips them in favor of real entries.
 215          * These terminator entries exist to handle any gaps created by
 216          * whitelisted .o files which didn't get objtool generation.
 217          */
 218         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 219         return orc_a->sp_reg == ORC_REG_UNDEFINED && !orc_a->end ? -1 : 1;
 220 }
 221
 222 #ifdef CONFIG_MODULES
 223 void unwind_module_init(struct module *mod, void *_orc_ip, size_t orc_ip_size,
 224                         void *_orc, size_t orc_size)
 225 {
 226         int *orc_ip = _orc_ip;
 227         struct orc_entry *orc = _orc;
 228         unsigned int num_entries = orc_ip_size / sizeof(int);
 229
 230         WARN_ON_ONCE(orc_ip_size % sizeof(int) != 0 ||
 231                      orc_size % sizeof(*orc) != 0 ||
 232                      num_entries != orc_size / sizeof(*orc));
 233
 234         /*
 235          * The 'cur_orc_*' globals allow the orc_sort_swap() callback to
 236          * associate an .orc_unwind_ip table entry with its corresponding
 237          * .orc_unwind entry so they can both be swapped.
 238          */
 239         mutex_lock(&sort_mutex);
 240         cur_orc_ip_table = orc_ip;
 241         cur_orc_table = orc;
 242         sort(orc_ip, num_entries, sizeof(int), orc_sort_cmp, orc_sort_swap);
 243         mutex_unlock(&sort_mutex);
 244
 245         mod->arch.orc_unwind_ip = orc_ip;
 246         mod->arch.orc_unwind = orc;
 247         mod->arch.num_orcs = num_entries;
 248 }
 249 #endif
 250
 251 void __init unwind_init(void)
 252 {
 253         size_t orc_ip_size = (void *)__stop_orc_unwind_ip - (void *)__start_orc_unwind_ip;
 254         size_t orc_size = (void *)__stop_orc_unwind - (void *)__start_orc_unwind;
 255         size_t num_entries = orc_ip_size / sizeof(int);
 256         struct orc_entry *orc;
 257         int i;
 258
 259         if (!num_entries || orc_ip_size % sizeof(int) != 0 ||
 260             orc_size % sizeof(struct orc_entry) != 0 ||
 261             num_entries != orc_size / sizeof(struct orc_entry)) {
 262                 orc_warn("WARNING: Bad or missing .orc_unwind table.  Disabling unwinder.\n");
 263                 return;
 264         }
 265
 266         /* Sort the .orc_unwind and .orc_unwind_ip tables: */
 267         sort(__start_orc_unwind_ip, num_entries, sizeof(int), orc_sort_cmp,
 268              orc_sort_swap);
 269
 270         /* Initialize the fast lookup table: */
 271         lookup_num_blocks = orc_lookup_end - orc_lookup;
 272         for (i = 0; i < lookup_num_blocks-1; i++) {
 273                 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 274                                  num_entries,
 275                                  LOOKUP_START_IP + (LOOKUP_BLOCK_SIZE * i));
 276                 if (!orc) {
 277                         orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 278                         return;
 279                 }
 280
 281                 orc_lookup[i] = orc - __start_orc_unwind;
 282         }
 283
 284         /* Initialize the ending block: */
 285         orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind, num_entries,
 286                          LOOKUP_STOP_IP);
 287         if (!orc) {
 288                 orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 289                 return;
 290         }
 291         orc_lookup[lookup_num_blocks-1] = orc - __start_orc_unwind;
 292
 293         orc_init = true;
 294 }
 295
 296 unsigned long unwind_get_return_address(struct unwind_state *state)
 297 {
 298         if (unwind_done(state))
 299                 return 0;
 300
 301         return __kernel_text_address(state->ip) ? state->ip : 0;
 302 }
 303 EXPORT_SYMBOL_GPL(unwind_get_return_address);
 304
 305 unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
 306 {
 307         if (unwind_done(state))
 308                 return NULL;
 309
 310         if (state->regs)
 311                 return &state->regs->ip;
 312
 313         if (state->sp)
 314                 return (unsigned long *)state->sp - 1;
 315
 316         return NULL;
 317 }
 318
 319 static bool stack_access_ok(struct unwind_state *state, unsigned long _addr,
 320                             size_t len)
 321 {
 322         struct stack_info *info = &state->stack_info;
 323         void *addr = (void *)_addr;
 324
 325         if (!on_stack(info, addr, len) &&
 326             (get_stack_info(addr, state->task, info, &state->stack_mask)))
 327                 return false;
 328
 329         return true;
 330 }
 331
 332 static bool deref_stack_reg(struct unwind_state *state, unsigned long addr,
 333                             unsigned long *val)
 334 {
 335         if (!stack_access_ok(state, addr, sizeof(long)))
 336                 return false;
 337
 338         *val = READ_ONCE_NOCHECK(*(unsigned long *)addr);
 339         return true;
 340 }
 341
 342 static bool deref_stack_regs(struct unwind_state *state, unsigned long addr,
 343                              unsigned long *ip, unsigned long *sp)
 344 {
 345         struct pt_regs *regs = (struct pt_regs *)addr;
 346
 347         /* x86-32 support will be more complicated due to the &regs->sp hack */
 348         BUILD_BUG_ON(IS_ENABLED(CONFIG_X86_32));
 349
 350         if (!stack_access_ok(state, addr, sizeof(struct pt_regs)))
 351                 return false;
 352
 353         *ip = regs->ip;
 354         *sp = regs->sp;
 355         return true;
 356 }
 357
 358 static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr,
 359                                   unsigned long *ip, unsigned long *sp)
 360 {
 361         struct pt_regs *regs = (void *)addr - IRET_FRAME_OFFSET;
 362
 363         if (!stack_access_ok(state, addr, IRET_FRAME_SIZE))
 364                 return false;
 365
 366         *ip = regs->ip;
 367         *sp = regs->sp;
 368         return true;
 369 }
 370
 371 bool unwind_next_frame(struct unwind_state *state)
 372 {
 373         unsigned long ip_p, sp, orig_ip = state->ip, prev_sp = state->sp;
 374         enum stack_type prev_type = state->stack_info.type;
 375         struct orc_entry *orc;
 376         bool indirect = false;
 377
 378         if (unwind_done(state))
 379                 return false;
 380
 381         /* Don't let modules unload while we're reading their ORC data. */
 382         preempt_disable();
 383
 384         /* End-of-stack check for user tasks: */
 385         if (state->regs && user_mode(state->regs))
 386                 goto the_end;
 387
 388         /*
 389          * Find the orc_entry associated with the text address.
 390          *
 391          * Decrement call return addresses by one so they work for sibling
 392          * calls and calls to noreturn functions.
 393          */
 394         orc = orc_find(state->signal ? state->ip : state->ip - 1);
 395         if (!orc)
 396                 goto err;
 397
 398         /* End-of-stack check for kernel threads: */
 399         if (orc->sp_reg == ORC_REG_UNDEFINED) {
 400                 if (!orc->end)
 401                         goto err;
 402
 403                 goto the_end;
 404         }
 405
 406         /* Find the previous frame's stack: */
 407         switch (orc->sp_reg) {
 408         case ORC_REG_SP:
 409                 sp = state->sp + orc->sp_offset;
 410                 break;
 411
 412         case ORC_REG_BP:
 413                 sp = state->bp + orc->sp_offset;
 414                 break;
 415
 416         case ORC_REG_SP_INDIRECT:
 417                 sp = state->sp + orc->sp_offset;
 418                 indirect = true;
 419                 break;
 420
 421         case ORC_REG_BP_INDIRECT:
 422                 sp = state->bp + orc->sp_offset;
 423                 indirect = true;
 424                 break;
 425
 426         case ORC_REG_R10:
 427                 if (!state->regs || !state->full_regs) {
 428                         orc_warn("missing regs for base reg R10 at ip %pB\n",
 429                                  (void *)state->ip);
 430                         goto err;
 431                 }
 432                 sp = state->regs->r10;
 433                 break;
 434
 435         case ORC_REG_R13:
 436                 if (!state->regs || !state->full_regs) {
 437                         orc_warn("missing regs for base reg R13 at ip %pB\n",
 438                                  (void *)state->ip);
 439                         goto err;
 440                 }
 441                 sp = state->regs->r13;
 442                 break;
 443
 444         case ORC_REG_DI:
 445                 if (!state->regs || !state->full_regs) {
 446                         orc_warn("missing regs for base reg DI at ip %pB\n",
 447                                  (void *)state->ip);
 448                         goto err;
 449                 }
 450                 sp = state->regs->di;
 451                 break;
 452
 453         case ORC_REG_DX:
 454                 if (!state->regs || !state->full_regs) {
 455                         orc_warn("missing regs for base reg DX at ip %pB\n",
 456                                  (void *)state->ip);
 457                         goto err;
 458                 }
 459                 sp = state->regs->dx;
 460                 break;
 461
 462         default:
 463                 orc_warn("unknown SP base reg %d for ip %pB\n",
 464                          orc->sp_reg, (void *)state->ip);
 465                 goto err;
 466         }
 467
 468         if (indirect) {
 469                 if (!deref_stack_reg(state, sp, &sp))
 470                         goto err;
 471         }
 472
 473         /* Find IP, SP and possibly regs: */
 474         switch (orc->type) {
 475         case ORC_TYPE_CALL:
 476                 ip_p = sp - sizeof(long);
 477
 478                 if (!deref_stack_reg(state, ip_p, &state->ip))
 479                         goto err;
 480
 481                 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
 482                                                   state->ip, (void *)ip_p);
 483
 484                 state->sp = sp;
 485                 state->regs = NULL;
 486                 state->signal = false;
 487                 break;
 488
 489         case ORC_TYPE_REGS:
 490                 if (!deref_stack_regs(state, sp, &state->ip, &state->sp)) {
 491                         orc_warn("can't dereference registers at %p for ip %pB\n",
 492                                  (void *)sp, (void *)orig_ip);
 493                         goto err;
 494                 }
 495
 496                 state->regs = (struct pt_regs *)sp;
 497                 state->full_regs = true;
 498                 state->signal = true;
 499                 break;
 500
 501         case ORC_TYPE_REGS_IRET:
 502                 if (!deref_stack_iret_regs(state, sp, &state->ip, &state->sp)) {
 503                         orc_warn("can't dereference iret registers at %p for ip %pB\n",
 504                                  (void *)sp, (void *)orig_ip);
 505                         goto err;
 506                 }
 507
 508                 state->regs = (void *)sp - IRET_FRAME_OFFSET;
 509                 state->full_regs = false;
 510                 state->signal = true;
 511                 break;
 512
 513         default:
 514                 orc_warn("unknown .orc_unwind entry type %d for ip %pB\n",
 515                          orc->type, (void *)orig_ip);
 516                 break;
 517         }
 518
 519         /* Find BP: */
 520         switch (orc->bp_reg) {
 521         case ORC_REG_UNDEFINED:
 522                 if (state->regs && state->full_regs)
 523                         state->bp = state->regs->bp;
 524                 break;
 525
 526         case ORC_REG_PREV_SP:
 527                 if (!deref_stack_reg(state, sp + orc->bp_offset, &state->bp))
 528                         goto err;
 529                 break;
 530
 531         case ORC_REG_BP:
 532                 if (!deref_stack_reg(state, state->bp + orc->bp_offset, &state->bp))
 533                         goto err;
 534                 break;
 535
 536         default:
 537                 orc_warn("unknown BP base reg %d for ip %pB\n",
 538                          orc->bp_reg, (void *)orig_ip);
 539                 goto err;
 540         }
 541
 542         /* Prevent a recursive loop due to bad ORC data: */
 543         if (state->stack_info.type == prev_type &&
 544             on_stack(&state->stack_info, (void *)state->sp, sizeof(long)) &&
 545             state->sp <= prev_sp) {
 546                 orc_warn("stack going in the wrong direction? ip=%pB\n",
 547                          (void *)orig_ip);
 548                 goto err;
 549         }
 550
 551         preempt_enable();
 552         return true;
 553
 554 err:
 555         state->error = true;
 556
 557 the_end:
 558         preempt_enable();
 559         state->stack_info.type = STACK_TYPE_UNKNOWN;
 560         return false;
 561 }
 562 EXPORT_SYMBOL_GPL(unwind_next_frame);
 563
 564 void __unwind_start(struct unwind_state *state, struct task_struct *task,
 565                     struct pt_regs *regs, unsigned long *first_frame)
 566 {
 567         memset(state, 0, sizeof(*state));
 568         state->task = task;
 569
 570         /*
 571          * Refuse to unwind the stack of a task while it's executing on another
 572          * CPU.  This check is racy, but that's ok: the unwinder has other
 573          * checks to prevent it from going off the rails.
 574          */
 575         if (task_on_another_cpu(task))
 576                 goto done;
 577
 578         if (regs) {
 579                 if (user_mode(regs))
 580                         goto done;
 581
 582                 state->ip = regs->ip;
 583                 state->sp = kernel_stack_pointer(regs);
 584                 state->bp = regs->bp;
 585                 state->regs = regs;
 586                 state->full_regs = true;
 587                 state->signal = true;
 588
 589         } else if (task == current) {
 590                 asm volatile("lea (%%rip), %0\n\t"
 591                              "mov %%rsp, %1\n\t"
 592                              "mov %%rbp, %2\n\t"
 593                              : "=r" (state->ip), "=r" (state->sp),
 594                                "=r" (state->bp));
 595
 596         } else {
 597                 struct inactive_task_frame *frame = (void *)task->thread.sp;
 598
 599                 state->sp = task->thread.sp;
 600                 state->bp = READ_ONCE_NOCHECK(frame->bp);
 601                 state->ip = READ_ONCE_NOCHECK(frame->ret_addr);
 602         }
 603
 604         if (get_stack_info((unsigned long *)state->sp, state->task,
 605                            &state->stack_info, &state->stack_mask)) {
 606                 /*
 607                  * We weren't on a valid stack.  It's possible that
 608                  * we overflowed a valid stack into a guard page.
 609                  * See if the next page up is valid so that we can
 610                  * generate some kind of backtrace if this happens.
 611                  */
 612                 void *next_page = (void *)PAGE_ALIGN((unsigned long)state->sp);
 613                 if (get_stack_info(next_page, state->task, &state->stack_info,
 614                                    &state->stack_mask))
 615                         return;
 616         }
 617
 618         /*
 619          * The caller can provide the address of the first frame directly
 620          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
 621          * to start unwinding at.  Skip ahead until we reach it.
 622          */
 623
 624         /* When starting from regs, skip the regs frame: */
 625         if (regs) {
 626                 unwind_next_frame(state);
 627                 return;
 628         }
 629
 630         /* Otherwise, skip ahead to the user-specified starting frame: */
 631         while (!unwind_done(state) &&
 632                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
 633                         state->sp <= (unsigned long)first_frame))
 634                 unwind_next_frame(state);
 635
 636         return;
 637
 638 done:
 639         state->stack_info.type = STACK_TYPE_UNKNOWN;
 640         return;
 641 }
 642 EXPORT_SYMBOL_GPL(__unwind_start);