arch/x86/kernel/unwind_orc.c

   1 #include <linux/module.h>
   2 #include <linux/sort.h>
   3 #include <asm/ptrace.h>
   4 #include <asm/stacktrace.h>
   5 #include <asm/unwind.h>
   6 #include <asm/orc_types.h>
   7 #include <asm/orc_lookup.h>
   8
   9 #define orc_warn(fmt, ...) \
  10         printk_deferred_once(KERN_WARNING pr_fmt("WARNING: " fmt), ##__VA_ARGS__)
  11
  12 extern int __start_orc_unwind_ip[];
  13 extern int __stop_orc_unwind_ip[];
  14 extern struct orc_entry __start_orc_unwind[];
  15 extern struct orc_entry __stop_orc_unwind[];
  16
  17 static DEFINE_MUTEX(sort_mutex);
  18 int *cur_orc_ip_table = __start_orc_unwind_ip;
  19 struct orc_entry *cur_orc_table = __start_orc_unwind;
  20
  21 unsigned int lookup_num_blocks;
  22 bool orc_init;
  23
  24 static inline unsigned long orc_ip(const int *ip)
  25 {
  26         return (unsigned long)ip + *ip;
  27 }
  28
  29 static struct orc_entry *__orc_find(int *ip_table, struct orc_entry *u_table,
  30                                     unsigned int num_entries, unsigned long ip)
  31 {
  32         int *first = ip_table;
  33         int *last = ip_table + num_entries - 1;
  34         int *mid = first, *found = first;
  35
  36         if (!num_entries)
  37                 return NULL;
  38
  39         /*
  40          * Do a binary range search to find the rightmost duplicate of a given
  41          * starting address.  Some entries are section terminators which are
  42          * "weak" entries for ensuring there are no gaps.  They should be
  43          * ignored when they conflict with a real entry.
  44          */
  45         while (first <= last) {
  46                 mid = first + ((last - first) / 2);
  47
  48                 if (orc_ip(mid) <= ip) {
  49                         found = mid;
  50                         first = mid + 1;
  51                 } else
  52                         last = mid - 1;
  53         }
  54
  55         return u_table + (found - ip_table);
  56 }
  57
  58 #ifdef CONFIG_MODULES
  59 static struct orc_entry *orc_module_find(unsigned long ip)
  60 {
  61         struct module *mod;
  62
  63         mod = __module_address(ip);
  64         if (!mod || !mod->arch.orc_unwind || !mod->arch.orc_unwind_ip)
  65                 return NULL;
  66         return __orc_find(mod->arch.orc_unwind_ip, mod->arch.orc_unwind,
  67                           mod->arch.num_orcs, ip);
  68 }
  69 #else
  70 static struct orc_entry *orc_module_find(unsigned long ip)
  71 {
  72         return NULL;
  73 }
  74 #endif
  75
  76 #ifdef CONFIG_DYNAMIC_FTRACE
  77 static struct orc_entry *orc_find(unsigned long ip);
  78
  79 /*
  80  * Ftrace dynamic trampolines do not have orc entries of their own.
  81  * But they are copies of the ftrace entries that are static and
  82  * defined in ftrace_*.S, which do have orc entries.
  83  *
  84  * If the undwinder comes across a ftrace trampoline, then find the
  85  * ftrace function that was used to create it, and use that ftrace
  86  * function's orc entrie, as the placement of the return code in
  87  * the stack will be identical.
  88  */
  89 static struct orc_entry *orc_ftrace_find(unsigned long ip)
  90 {
  91         struct ftrace_ops *ops;
  92         unsigned long caller;
  93
  94         ops = ftrace_ops_trampoline(ip);
  95         if (!ops)
  96                 return NULL;
  97
  98         if (ops->flags & FTRACE_OPS_FL_SAVE_REGS)
  99                 caller = (unsigned long)ftrace_regs_call;
 100         else
 101                 caller = (unsigned long)ftrace_call;
 102
 103         /* Prevent unlikely recursion */
 104         if (ip == caller)
 105                 return NULL;
 106
 107         return orc_find(caller);
 108 }
 109 #else
 110 static struct orc_entry *orc_ftrace_find(unsigned long ip)
 111 {
 112         return NULL;
 113 }
 114 #endif
 115
 116 /*
 117  * If we crash with IP==0, the last successfully executed instruction
 118  * was probably an indirect function call with a NULL function pointer,
 119  * and we don't have unwind information for NULL.
 120  * This hardcoded ORC entry for IP==0 allows us to unwind from a NULL function
 121  * pointer into its parent and then continue normally from there.
 122  */
 123 static struct orc_entry null_orc_entry = {
 124         .sp_offset = sizeof(long),
 125         .sp_reg = ORC_REG_SP,
 126         .bp_reg = ORC_REG_UNDEFINED,
 127         .type = ORC_TYPE_CALL
 128 };
 129
 130 static struct orc_entry *orc_find(unsigned long ip)
 131 {
 132         static struct orc_entry *orc;
 133
 134         if (!orc_init)
 135                 return NULL;
 136
 137         if (ip == 0)
 138                 return &null_orc_entry;
 139
 140         /* For non-init vmlinux addresses, use the fast lookup table: */
 141         if (ip >= LOOKUP_START_IP && ip < LOOKUP_STOP_IP) {
 142                 unsigned int idx, start, stop;
 143
 144                 idx = (ip - LOOKUP_START_IP) / LOOKUP_BLOCK_SIZE;
 145
 146                 if (unlikely((idx >= lookup_num_blocks-1))) {
 147                         orc_warn("WARNING: bad lookup idx: idx=%u num=%u ip=%pB\n",
 148                                  idx, lookup_num_blocks, (void *)ip);
 149                         return NULL;
 150                 }
 151
 152                 start = orc_lookup[idx];
 153                 stop = orc_lookup[idx + 1] + 1;
 154
 155                 if (unlikely((__start_orc_unwind + start >= __stop_orc_unwind) ||
 156                              (__start_orc_unwind + stop > __stop_orc_unwind))) {
 157                         orc_warn("WARNING: bad lookup value: idx=%u num=%u start=%u stop=%u ip=%pB\n",
 158                                  idx, lookup_num_blocks, start, stop, (void *)ip);
 159                         return NULL;
 160                 }
 161
 162                 return __orc_find(__start_orc_unwind_ip + start,
 163                                   __start_orc_unwind + start, stop - start, ip);
 164         }
 165
 166         /* vmlinux .init slow lookup: */
 167         if (init_kernel_text(ip))
 168                 return __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 169                                   __stop_orc_unwind_ip - __start_orc_unwind_ip, ip);
 170
 171         /* Module lookup: */
 172         orc = orc_module_find(ip);
 173         if (orc)
 174                 return orc;
 175
 176         return orc_ftrace_find(ip);
 177 }
 178
 179 static void orc_sort_swap(void *_a, void *_b, int size)
 180 {
 181         struct orc_entry *orc_a, *orc_b;
 182         struct orc_entry orc_tmp;
 183         int *a = _a, *b = _b, tmp;
 184         int delta = _b - _a;
 185
 186         /* Swap the .orc_unwind_ip entries: */
 187         tmp = *a;
 188         *a = *b + delta;
 189         *b = tmp - delta;
 190
 191         /* Swap the corresponding .orc_unwind entries: */
 192         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 193         orc_b = cur_orc_table + (b - cur_orc_ip_table);
 194         orc_tmp = *orc_a;
 195         *orc_a = *orc_b;
 196         *orc_b = orc_tmp;
 197 }
 198
 199 static int orc_sort_cmp(const void *_a, const void *_b)
 200 {
 201         struct orc_entry *orc_a;
 202         const int *a = _a, *b = _b;
 203         unsigned long a_val = orc_ip(a);
 204         unsigned long b_val = orc_ip(b);
 205
 206         if (a_val > b_val)
 207                 return 1;
 208         if (a_val < b_val)
 209                 return -1;
 210
 211         /*
 212          * The "weak" section terminator entries need to always be on the left
 213          * to ensure the lookup code skips them in favor of real entries.
 214          * These terminator entries exist to handle any gaps created by
 215          * whitelisted .o files which didn't get objtool generation.
 216          */
 217         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 218         return orc_a->sp_reg == ORC_REG_UNDEFINED && !orc_a->end ? -1 : 1;
 219 }
 220
 221 #ifdef CONFIG_MODULES
 222 void unwind_module_init(struct module *mod, void *_orc_ip, size_t orc_ip_size,
 223                         void *_orc, size_t orc_size)
 224 {
 225         int *orc_ip = _orc_ip;
 226         struct orc_entry *orc = _orc;
 227         unsigned int num_entries = orc_ip_size / sizeof(int);
 228
 229         WARN_ON_ONCE(orc_ip_size % sizeof(int) != 0 ||
 230                      orc_size % sizeof(*orc) != 0 ||
 231                      num_entries != orc_size / sizeof(*orc));
 232
 233         /*
 234          * The 'cur_orc_*' globals allow the orc_sort_swap() callback to
 235          * associate an .orc_unwind_ip table entry with its corresponding
 236          * .orc_unwind entry so they can both be swapped.
 237          */
 238         mutex_lock(&sort_mutex);
 239         cur_orc_ip_table = orc_ip;
 240         cur_orc_table = orc;
 241         sort(orc_ip, num_entries, sizeof(int), orc_sort_cmp, orc_sort_swap);
 242         mutex_unlock(&sort_mutex);
 243
 244         mod->arch.orc_unwind_ip = orc_ip;
 245         mod->arch.orc_unwind = orc;
 246         mod->arch.num_orcs = num_entries;
 247 }
 248 #endif
 249
 250 void __init unwind_init(void)
 251 {
 252         size_t orc_ip_size = (void *)__stop_orc_unwind_ip - (void *)__start_orc_unwind_ip;
 253         size_t orc_size = (void *)__stop_orc_unwind - (void *)__start_orc_unwind;
 254         size_t num_entries = orc_ip_size / sizeof(int);
 255         struct orc_entry *orc;
 256         int i;
 257
 258         if (!num_entries || orc_ip_size % sizeof(int) != 0 ||
 259             orc_size % sizeof(struct orc_entry) != 0 ||
 260             num_entries != orc_size / sizeof(struct orc_entry)) {
 261                 orc_warn("WARNING: Bad or missing .orc_unwind table.  Disabling unwinder.\n");
 262                 return;
 263         }
 264
 265         /* Sort the .orc_unwind and .orc_unwind_ip tables: */
 266         sort(__start_orc_unwind_ip, num_entries, sizeof(int), orc_sort_cmp,
 267              orc_sort_swap);
 268
 269         /* Initialize the fast lookup table: */
 270         lookup_num_blocks = orc_lookup_end - orc_lookup;
 271         for (i = 0; i < lookup_num_blocks-1; i++) {
 272                 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 273                                  num_entries,
 274                                  LOOKUP_START_IP + (LOOKUP_BLOCK_SIZE * i));
 275                 if (!orc) {
 276                         orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 277                         return;
 278                 }
 279
 280                 orc_lookup[i] = orc - __start_orc_unwind;
 281         }
 282
 283         /* Initialize the ending block: */
 284         orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind, num_entries,
 285                          LOOKUP_STOP_IP);
 286         if (!orc) {
 287                 orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 288                 return;
 289         }
 290         orc_lookup[lookup_num_blocks-1] = orc - __start_orc_unwind;
 291
 292         orc_init = true;
 293 }
 294
 295 unsigned long unwind_get_return_address(struct unwind_state *state)
 296 {
 297         if (unwind_done(state))
 298                 return 0;
 299
 300         return __kernel_text_address(state->ip) ? state->ip : 0;
 301 }
 302 EXPORT_SYMBOL_GPL(unwind_get_return_address);
 303
 304 unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
 305 {
 306         if (unwind_done(state))
 307                 return NULL;
 308
 309         if (state->regs)
 310                 return &state->regs->ip;
 311
 312         if (state->sp)
 313                 return (unsigned long *)state->sp - 1;
 314
 315         return NULL;
 316 }
 317
 318 static bool stack_access_ok(struct unwind_state *state, unsigned long _addr,
 319                             size_t len)
 320 {
 321         struct stack_info *info = &state->stack_info;
 322         void *addr = (void *)_addr;
 323
 324         if (!on_stack(info, addr, len) &&
 325             (get_stack_info(addr, state->task, info, &state->stack_mask)))
 326                 return false;
 327
 328         return true;
 329 }
 330
 331 static bool deref_stack_reg(struct unwind_state *state, unsigned long addr,
 332                             unsigned long *val)
 333 {
 334         if (!stack_access_ok(state, addr, sizeof(long)))
 335                 return false;
 336
 337         *val = READ_ONCE_NOCHECK(*(unsigned long *)addr);
 338         return true;
 339 }
 340
 341 static bool deref_stack_regs(struct unwind_state *state, unsigned long addr,
 342                              unsigned long *ip, unsigned long *sp)
 343 {
 344         struct pt_regs *regs = (struct pt_regs *)addr;
 345
 346         /* x86-32 support will be more complicated due to the &regs->sp hack */
 347         BUILD_BUG_ON(IS_ENABLED(CONFIG_X86_32));
 348
 349         if (!stack_access_ok(state, addr, sizeof(struct pt_regs)))
 350                 return false;
 351
 352         *ip = regs->ip;
 353         *sp = regs->sp;
 354         return true;
 355 }
 356
 357 static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr,
 358                                   unsigned long *ip, unsigned long *sp)
 359 {
 360         struct pt_regs *regs = (void *)addr - IRET_FRAME_OFFSET;
 361
 362         if (!stack_access_ok(state, addr, IRET_FRAME_SIZE))
 363                 return false;
 364
 365         *ip = regs->ip;
 366         *sp = regs->sp;
 367         return true;
 368 }
 369
 370 bool unwind_next_frame(struct unwind_state *state)
 371 {
 372         unsigned long ip_p, sp, orig_ip = state->ip, prev_sp = state->sp;
 373         enum stack_type prev_type = state->stack_info.type;
 374         struct orc_entry *orc;
 375         bool indirect = false;
 376
 377         if (unwind_done(state))
 378                 return false;
 379
 380         /* Don't let modules unload while we're reading their ORC data. */
 381         preempt_disable();
 382
 383         /* End-of-stack check for user tasks: */
 384         if (state->regs && user_mode(state->regs))
 385                 goto the_end;
 386
 387         /*
 388          * Find the orc_entry associated with the text address.
 389          *
 390          * Decrement call return addresses by one so they work for sibling
 391          * calls and calls to noreturn functions.
 392          */
 393         orc = orc_find(state->signal ? state->ip : state->ip - 1);
 394         if (!orc)
 395                 goto err;
 396
 397         /* End-of-stack check for kernel threads: */
 398         if (orc->sp_reg == ORC_REG_UNDEFINED) {
 399                 if (!orc->end)
 400                         goto err;
 401
 402                 goto the_end;
 403         }
 404
 405         /* Find the previous frame's stack: */
 406         switch (orc->sp_reg) {
 407         case ORC_REG_SP:
 408                 sp = state->sp + orc->sp_offset;
 409                 break;
 410
 411         case ORC_REG_BP:
 412                 sp = state->bp + orc->sp_offset;
 413                 break;
 414
 415         case ORC_REG_SP_INDIRECT:
 416                 sp = state->sp + orc->sp_offset;
 417                 indirect = true;
 418                 break;
 419
 420         case ORC_REG_BP_INDIRECT:
 421                 sp = state->bp + orc->sp_offset;
 422                 indirect = true;
 423                 break;
 424
 425         case ORC_REG_R10:
 426                 if (!state->regs || !state->full_regs) {
 427                         orc_warn("missing regs for base reg R10 at ip %pB\n",
 428                                  (void *)state->ip);
 429                         goto err;
 430                 }
 431                 sp = state->regs->r10;
 432                 break;
 433
 434         case ORC_REG_R13:
 435                 if (!state->regs || !state->full_regs) {
 436                         orc_warn("missing regs for base reg R13 at ip %pB\n",
 437                                  (void *)state->ip);
 438                         goto err;
 439                 }
 440                 sp = state->regs->r13;
 441                 break;
 442
 443         case ORC_REG_DI:
 444                 if (!state->regs || !state->full_regs) {
 445                         orc_warn("missing regs for base reg DI at ip %pB\n",
 446                                  (void *)state->ip);
 447                         goto err;
 448                 }
 449                 sp = state->regs->di;
 450                 break;
 451
 452         case ORC_REG_DX:
 453                 if (!state->regs || !state->full_regs) {
 454                         orc_warn("missing regs for base reg DX at ip %pB\n",
 455                                  (void *)state->ip);
 456                         goto err;
 457                 }
 458                 sp = state->regs->dx;
 459                 break;
 460
 461         default:
 462                 orc_warn("unknown SP base reg %d for ip %pB\n",
 463                          orc->sp_reg, (void *)state->ip);
 464                 goto err;
 465         }
 466
 467         if (indirect) {
 468                 if (!deref_stack_reg(state, sp, &sp))
 469                         goto err;
 470         }
 471
 472         /* Find IP, SP and possibly regs: */
 473         switch (orc->type) {
 474         case ORC_TYPE_CALL:
 475                 ip_p = sp - sizeof(long);
 476
 477                 if (!deref_stack_reg(state, ip_p, &state->ip))
 478                         goto err;
 479
 480                 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
 481                                                   state->ip, (void *)ip_p);
 482
 483                 state->sp = sp;
 484                 state->regs = NULL;
 485                 state->signal = false;
 486                 break;
 487
 488         case ORC_TYPE_REGS:
 489                 if (!deref_stack_regs(state, sp, &state->ip, &state->sp)) {
 490                         orc_warn("can't dereference registers at %p for ip %pB\n",
 491                                  (void *)sp, (void *)orig_ip);
 492                         goto err;
 493                 }
 494
 495                 state->regs = (struct pt_regs *)sp;
 496                 state->full_regs = true;
 497                 state->signal = true;
 498                 break;
 499
 500         case ORC_TYPE_REGS_IRET:
 501                 if (!deref_stack_iret_regs(state, sp, &state->ip, &state->sp)) {
 502                         orc_warn("can't dereference iret registers at %p for ip %pB\n",
 503                                  (void *)sp, (void *)orig_ip);
 504                         goto err;
 505                 }
 506
 507                 state->regs = (void *)sp - IRET_FRAME_OFFSET;
 508                 state->full_regs = false;
 509                 state->signal = true;
 510                 break;
 511
 512         default:
 513                 orc_warn("unknown .orc_unwind entry type %d for ip %pB\n",
 514                          orc->type, (void *)orig_ip);
 515                 break;
 516         }
 517
 518         /* Find BP: */
 519         switch (orc->bp_reg) {
 520         case ORC_REG_UNDEFINED:
 521                 if (state->regs && state->full_regs)
 522                         state->bp = state->regs->bp;
 523                 break;
 524
 525         case ORC_REG_PREV_SP:
 526                 if (!deref_stack_reg(state, sp + orc->bp_offset, &state->bp))
 527                         goto err;
 528                 break;
 529
 530         case ORC_REG_BP:
 531                 if (!deref_stack_reg(state, state->bp + orc->bp_offset, &state->bp))
 532                         goto err;
 533                 break;
 534
 535         default:
 536                 orc_warn("unknown BP base reg %d for ip %pB\n",
 537                          orc->bp_reg, (void *)orig_ip);
 538                 goto err;
 539         }
 540
 541         /* Prevent a recursive loop due to bad ORC data: */
 542         if (state->stack_info.type == prev_type &&
 543             on_stack(&state->stack_info, (void *)state->sp, sizeof(long)) &&
 544             state->sp <= prev_sp) {
 545                 orc_warn("stack going in the wrong direction? ip=%pB\n",
 546                          (void *)orig_ip);
 547                 goto err;
 548         }
 549
 550         preempt_enable();
 551         return true;
 552
 553 err:
 554         state->error = true;
 555
 556 the_end:
 557         preempt_enable();
 558         state->stack_info.type = STACK_TYPE_UNKNOWN;
 559         return false;
 560 }
 561 EXPORT_SYMBOL_GPL(unwind_next_frame);
 562
 563 void __unwind_start(struct unwind_state *state, struct task_struct *task,
 564                     struct pt_regs *regs, unsigned long *first_frame)
 565 {
 566         memset(state, 0, sizeof(*state));
 567         state->task = task;
 568
 569         /*
 570          * Refuse to unwind the stack of a task while it's executing on another
 571          * CPU.  This check is racy, but that's ok: the unwinder has other
 572          * checks to prevent it from going off the rails.
 573          */
 574         if (task_on_another_cpu(task))
 575                 goto done;
 576
 577         if (regs) {
 578                 if (user_mode(regs))
 579                         goto done;
 580
 581                 state->ip = regs->ip;
 582                 state->sp = kernel_stack_pointer(regs);
 583                 state->bp = regs->bp;
 584                 state->regs = regs;
 585                 state->full_regs = true;
 586                 state->signal = true;
 587
 588         } else if (task == current) {
 589                 asm volatile("lea (%%rip), %0\n\t"
 590                              "mov %%rsp, %1\n\t"
 591                              "mov %%rbp, %2\n\t"
 592                              : "=r" (state->ip), "=r" (state->sp),
 593                                "=r" (state->bp));
 594
 595         } else {
 596                 struct inactive_task_frame *frame = (void *)task->thread.sp;
 597
 598                 state->sp = task->thread.sp;
 599                 state->bp = READ_ONCE_NOCHECK(frame->bp);
 600                 state->ip = READ_ONCE_NOCHECK(frame->ret_addr);
 601         }
 602
 603         if (get_stack_info((unsigned long *)state->sp, state->task,
 604                            &state->stack_info, &state->stack_mask)) {
 605                 /*
 606                  * We weren't on a valid stack.  It's possible that
 607                  * we overflowed a valid stack into a guard page.
 608                  * See if the next page up is valid so that we can
 609                  * generate some kind of backtrace if this happens.
 610                  */
 611                 void *next_page = (void *)PAGE_ALIGN((unsigned long)state->sp);
 612                 if (get_stack_info(next_page, state->task, &state->stack_info,
 613                                    &state->stack_mask))
 614                         return;
 615         }
 616
 617         /*
 618          * The caller can provide the address of the first frame directly
 619          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
 620          * to start unwinding at.  Skip ahead until we reach it.
 621          */
 622
 623         /* When starting from regs, skip the regs frame: */
 624         if (regs) {
 625                 unwind_next_frame(state);
 626                 return;
 627         }
 628
 629         /* Otherwise, skip ahead to the user-specified starting frame: */
 630         while (!unwind_done(state) &&
 631                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
 632                         state->sp <= (unsigned long)first_frame))
 633                 unwind_next_frame(state);
 634
 635         return;
 636
 637 done:
 638         state->stack_info.type = STACK_TYPE_UNKNOWN;
 639         return;
 640 }
 641 EXPORT_SYMBOL_GPL(__unwind_start);