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 #include <asm/sections.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 static struct orc_entry *orc_find(unsigned long ip)
 118 {
 119         static struct orc_entry *orc;
 120
 121         if (!orc_init)
 122                 return NULL;
 123
 124         /* For non-init vmlinux addresses, use the fast lookup table: */
 125         if (ip >= LOOKUP_START_IP && ip < LOOKUP_STOP_IP) {
 126                 unsigned int idx, start, stop;
 127
 128                 idx = (ip - LOOKUP_START_IP) / LOOKUP_BLOCK_SIZE;
 129
 130                 if (unlikely((idx >= lookup_num_blocks-1))) {
 131                         orc_warn("WARNING: bad lookup idx: idx=%u num=%u ip=%pB\n",
 132                                  idx, lookup_num_blocks, (void *)ip);
 133                         return NULL;
 134                 }
 135
 136                 start = orc_lookup[idx];
 137                 stop = orc_lookup[idx + 1] + 1;
 138
 139                 if (unlikely((__start_orc_unwind + start >= __stop_orc_unwind) ||
 140                              (__start_orc_unwind + stop > __stop_orc_unwind))) {
 141                         orc_warn("WARNING: bad lookup value: idx=%u num=%u start=%u stop=%u ip=%pB\n",
 142                                  idx, lookup_num_blocks, start, stop, (void *)ip);
 143                         return NULL;
 144                 }
 145
 146                 return __orc_find(__start_orc_unwind_ip + start,
 147                                   __start_orc_unwind + start, stop - start, ip);
 148         }
 149
 150         /* vmlinux .init slow lookup: */
 151         if (ip >= (unsigned long)_sinittext && ip < (unsigned long)_einittext)
 152                 return __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 153                                   __stop_orc_unwind_ip - __start_orc_unwind_ip, ip);
 154
 155         /* Module lookup: */
 156         orc = orc_module_find(ip);
 157         if (orc)
 158                 return orc;
 159
 160         return orc_ftrace_find(ip);
 161 }
 162
 163 static void orc_sort_swap(void *_a, void *_b, int size)
 164 {
 165         struct orc_entry *orc_a, *orc_b;
 166         struct orc_entry orc_tmp;
 167         int *a = _a, *b = _b, tmp;
 168         int delta = _b - _a;
 169
 170         /* Swap the .orc_unwind_ip entries: */
 171         tmp = *a;
 172         *a = *b + delta;
 173         *b = tmp - delta;
 174
 175         /* Swap the corresponding .orc_unwind entries: */
 176         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 177         orc_b = cur_orc_table + (b - cur_orc_ip_table);
 178         orc_tmp = *orc_a;
 179         *orc_a = *orc_b;
 180         *orc_b = orc_tmp;
 181 }
 182
 183 static int orc_sort_cmp(const void *_a, const void *_b)
 184 {
 185         struct orc_entry *orc_a;
 186         const int *a = _a, *b = _b;
 187         unsigned long a_val = orc_ip(a);
 188         unsigned long b_val = orc_ip(b);
 189
 190         if (a_val > b_val)
 191                 return 1;
 192         if (a_val < b_val)
 193                 return -1;
 194
 195         /*
 196          * The "weak" section terminator entries need to always be on the left
 197          * to ensure the lookup code skips them in favor of real entries.
 198          * These terminator entries exist to handle any gaps created by
 199          * whitelisted .o files which didn't get objtool generation.
 200          */
 201         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 202         return orc_a->sp_reg == ORC_REG_UNDEFINED ? -1 : 1;
 203 }
 204
 205 #ifdef CONFIG_MODULES
 206 void unwind_module_init(struct module *mod, void *_orc_ip, size_t orc_ip_size,
 207                         void *_orc, size_t orc_size)
 208 {
 209         int *orc_ip = _orc_ip;
 210         struct orc_entry *orc = _orc;
 211         unsigned int num_entries = orc_ip_size / sizeof(int);
 212
 213         WARN_ON_ONCE(orc_ip_size % sizeof(int) != 0 ||
 214                      orc_size % sizeof(*orc) != 0 ||
 215                      num_entries != orc_size / sizeof(*orc));
 216
 217         /*
 218          * The 'cur_orc_*' globals allow the orc_sort_swap() callback to
 219          * associate an .orc_unwind_ip table entry with its corresponding
 220          * .orc_unwind entry so they can both be swapped.
 221          */
 222         mutex_lock(&sort_mutex);
 223         cur_orc_ip_table = orc_ip;
 224         cur_orc_table = orc;
 225         sort(orc_ip, num_entries, sizeof(int), orc_sort_cmp, orc_sort_swap);
 226         mutex_unlock(&sort_mutex);
 227
 228         mod->arch.orc_unwind_ip = orc_ip;
 229         mod->arch.orc_unwind = orc;
 230         mod->arch.num_orcs = num_entries;
 231 }
 232 #endif
 233
 234 void __init unwind_init(void)
 235 {
 236         size_t orc_ip_size = (void *)__stop_orc_unwind_ip - (void *)__start_orc_unwind_ip;
 237         size_t orc_size = (void *)__stop_orc_unwind - (void *)__start_orc_unwind;
 238         size_t num_entries = orc_ip_size / sizeof(int);
 239         struct orc_entry *orc;
 240         int i;
 241
 242         if (!num_entries || orc_ip_size % sizeof(int) != 0 ||
 243             orc_size % sizeof(struct orc_entry) != 0 ||
 244             num_entries != orc_size / sizeof(struct orc_entry)) {
 245                 orc_warn("WARNING: Bad or missing .orc_unwind table.  Disabling unwinder.\n");
 246                 return;
 247         }
 248
 249         /* Sort the .orc_unwind and .orc_unwind_ip tables: */
 250         sort(__start_orc_unwind_ip, num_entries, sizeof(int), orc_sort_cmp,
 251              orc_sort_swap);
 252
 253         /* Initialize the fast lookup table: */
 254         lookup_num_blocks = orc_lookup_end - orc_lookup;
 255         for (i = 0; i < lookup_num_blocks-1; i++) {
 256                 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 257                                  num_entries,
 258                                  LOOKUP_START_IP + (LOOKUP_BLOCK_SIZE * i));
 259                 if (!orc) {
 260                         orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 261                         return;
 262                 }
 263
 264                 orc_lookup[i] = orc - __start_orc_unwind;
 265         }
 266
 267         /* Initialize the ending block: */
 268         orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind, num_entries,
 269                          LOOKUP_STOP_IP);
 270         if (!orc) {
 271                 orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 272                 return;
 273         }
 274         orc_lookup[lookup_num_blocks-1] = orc - __start_orc_unwind;
 275
 276         orc_init = true;
 277 }
 278
 279 unsigned long unwind_get_return_address(struct unwind_state *state)
 280 {
 281         if (unwind_done(state))
 282                 return 0;
 283
 284         return __kernel_text_address(state->ip) ? state->ip : 0;
 285 }
 286 EXPORT_SYMBOL_GPL(unwind_get_return_address);
 287
 288 unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
 289 {
 290         if (unwind_done(state))
 291                 return NULL;
 292
 293         if (state->regs)
 294                 return &state->regs->ip;
 295
 296         if (state->sp)
 297                 return (unsigned long *)state->sp - 1;
 298
 299         return NULL;
 300 }
 301
 302 static bool stack_access_ok(struct unwind_state *state, unsigned long _addr,
 303                             size_t len)
 304 {
 305         struct stack_info *info = &state->stack_info;
 306         void *addr = (void *)_addr;
 307
 308         if (!on_stack(info, addr, len) &&
 309             (get_stack_info(addr, state->task, info, &state->stack_mask)))
 310                 return false;
 311
 312         return true;
 313 }
 314
 315 static bool deref_stack_reg(struct unwind_state *state, unsigned long addr,
 316                             unsigned long *val)
 317 {
 318         if (!stack_access_ok(state, addr, sizeof(long)))
 319                 return false;
 320
 321         *val = READ_ONCE_NOCHECK(*(unsigned long *)addr);
 322         return true;
 323 }
 324
 325 static bool deref_stack_regs(struct unwind_state *state, unsigned long addr,
 326                              unsigned long *ip, unsigned long *sp)
 327 {
 328         struct pt_regs *regs = (struct pt_regs *)addr;
 329
 330         /* x86-32 support will be more complicated due to the &regs->sp hack */
 331         BUILD_BUG_ON(IS_ENABLED(CONFIG_X86_32));
 332
 333         if (!stack_access_ok(state, addr, sizeof(struct pt_regs)))
 334                 return false;
 335
 336         *ip = regs->ip;
 337         *sp = regs->sp;
 338         return true;
 339 }
 340
 341 static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr,
 342                                   unsigned long *ip, unsigned long *sp)
 343 {
 344         struct pt_regs *regs = (void *)addr - IRET_FRAME_OFFSET;
 345
 346         if (!stack_access_ok(state, addr, IRET_FRAME_SIZE))
 347                 return false;
 348
 349         *ip = regs->ip;
 350         *sp = regs->sp;
 351         return true;
 352 }
 353
 354 bool unwind_next_frame(struct unwind_state *state)
 355 {
 356         unsigned long ip_p, sp, orig_ip, prev_sp = state->sp;
 357         enum stack_type prev_type = state->stack_info.type;
 358         struct orc_entry *orc;
 359         bool indirect = false;
 360
 361         if (unwind_done(state))
 362                 return false;
 363
 364         /* Don't let modules unload while we're reading their ORC data. */
 365         preempt_disable();
 366
 367         /* Have we reached the end? */
 368         if (state->regs && user_mode(state->regs))
 369                 goto done;
 370
 371         /*
 372          * Find the orc_entry associated with the text address.
 373          *
 374          * Decrement call return addresses by one so they work for sibling
 375          * calls and calls to noreturn functions.
 376          */
 377         orc = orc_find(state->signal ? state->ip : state->ip - 1);
 378         if (!orc || orc->sp_reg == ORC_REG_UNDEFINED)
 379                 goto done;
 380         orig_ip = state->ip;
 381
 382         /* Find the previous frame's stack: */
 383         switch (orc->sp_reg) {
 384         case ORC_REG_SP:
 385                 sp = state->sp + orc->sp_offset;
 386                 break;
 387
 388         case ORC_REG_BP:
 389                 sp = state->bp + orc->sp_offset;
 390                 break;
 391
 392         case ORC_REG_SP_INDIRECT:
 393                 sp = state->sp + orc->sp_offset;
 394                 indirect = true;
 395                 break;
 396
 397         case ORC_REG_BP_INDIRECT:
 398                 sp = state->bp + orc->sp_offset;
 399                 indirect = true;
 400                 break;
 401
 402         case ORC_REG_R10:
 403                 if (!state->regs || !state->full_regs) {
 404                         orc_warn("missing regs for base reg R10 at ip %pB\n",
 405                                  (void *)state->ip);
 406                         goto done;
 407                 }
 408                 sp = state->regs->r10;
 409                 break;
 410
 411         case ORC_REG_R13:
 412                 if (!state->regs || !state->full_regs) {
 413                         orc_warn("missing regs for base reg R13 at ip %pB\n",
 414                                  (void *)state->ip);
 415                         goto done;
 416                 }
 417                 sp = state->regs->r13;
 418                 break;
 419
 420         case ORC_REG_DI:
 421                 if (!state->regs || !state->full_regs) {
 422                         orc_warn("missing regs for base reg DI at ip %pB\n",
 423                                  (void *)state->ip);
 424                         goto done;
 425                 }
 426                 sp = state->regs->di;
 427                 break;
 428
 429         case ORC_REG_DX:
 430                 if (!state->regs || !state->full_regs) {
 431                         orc_warn("missing regs for base reg DX at ip %pB\n",
 432                                  (void *)state->ip);
 433                         goto done;
 434                 }
 435                 sp = state->regs->dx;
 436                 break;
 437
 438         default:
 439                 orc_warn("unknown SP base reg %d for ip %pB\n",
 440                          orc->sp_reg, (void *)state->ip);
 441                 goto done;
 442         }
 443
 444         if (indirect) {
 445                 if (!deref_stack_reg(state, sp, &sp))
 446                         goto done;
 447         }
 448
 449         /* Find IP, SP and possibly regs: */
 450         switch (orc->type) {
 451         case ORC_TYPE_CALL:
 452                 ip_p = sp - sizeof(long);
 453
 454                 if (!deref_stack_reg(state, ip_p, &state->ip))
 455                         goto done;
 456
 457                 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
 458                                                   state->ip, (void *)ip_p);
 459
 460                 state->sp = sp;
 461                 state->regs = NULL;
 462                 state->signal = false;
 463                 break;
 464
 465         case ORC_TYPE_REGS:
 466                 if (!deref_stack_regs(state, sp, &state->ip, &state->sp)) {
 467                         orc_warn("can't dereference registers at %p for ip %pB\n",
 468                                  (void *)sp, (void *)orig_ip);
 469                         goto done;
 470                 }
 471
 472                 state->regs = (struct pt_regs *)sp;
 473                 state->full_regs = true;
 474                 state->signal = true;
 475                 break;
 476
 477         case ORC_TYPE_REGS_IRET:
 478                 if (!deref_stack_iret_regs(state, sp, &state->ip, &state->sp)) {
 479                         orc_warn("can't dereference iret registers at %p for ip %pB\n",
 480                                  (void *)sp, (void *)orig_ip);
 481                         goto done;
 482                 }
 483
 484                 state->regs = (void *)sp - IRET_FRAME_OFFSET;
 485                 state->full_regs = false;
 486                 state->signal = true;
 487                 break;
 488
 489         default:
 490                 orc_warn("unknown .orc_unwind entry type %d for ip %pB\n",
 491                          orc->type, (void *)orig_ip);
 492                 break;
 493         }
 494
 495         /* Find BP: */
 496         switch (orc->bp_reg) {
 497         case ORC_REG_UNDEFINED:
 498                 if (state->regs && state->full_regs)
 499                         state->bp = state->regs->bp;
 500                 break;
 501
 502         case ORC_REG_PREV_SP:
 503                 if (!deref_stack_reg(state, sp + orc->bp_offset, &state->bp))
 504                         goto done;
 505                 break;
 506
 507         case ORC_REG_BP:
 508                 if (!deref_stack_reg(state, state->bp + orc->bp_offset, &state->bp))
 509                         goto done;
 510                 break;
 511
 512         default:
 513                 orc_warn("unknown BP base reg %d for ip %pB\n",
 514                          orc->bp_reg, (void *)orig_ip);
 515                 goto done;
 516         }
 517
 518         /* Prevent a recursive loop due to bad ORC data: */
 519         if (state->stack_info.type == prev_type &&
 520             on_stack(&state->stack_info, (void *)state->sp, sizeof(long)) &&
 521             state->sp <= prev_sp) {
 522                 orc_warn("stack going in the wrong direction? ip=%pB\n",
 523                          (void *)orig_ip);
 524                 goto done;
 525         }
 526
 527         preempt_enable();
 528         return true;
 529
 530 done:
 531         preempt_enable();
 532         state->stack_info.type = STACK_TYPE_UNKNOWN;
 533         return false;
 534 }
 535 EXPORT_SYMBOL_GPL(unwind_next_frame);
 536
 537 void __unwind_start(struct unwind_state *state, struct task_struct *task,
 538                     struct pt_regs *regs, unsigned long *first_frame)
 539 {
 540         memset(state, 0, sizeof(*state));
 541         state->task = task;
 542
 543         /*
 544          * Refuse to unwind the stack of a task while it's executing on another
 545          * CPU.  This check is racy, but that's ok: the unwinder has other
 546          * checks to prevent it from going off the rails.
 547          */
 548         if (task_on_another_cpu(task))
 549                 goto done;
 550
 551         if (regs) {
 552                 if (user_mode(regs))
 553                         goto done;
 554
 555                 state->ip = regs->ip;
 556                 state->sp = kernel_stack_pointer(regs);
 557                 state->bp = regs->bp;
 558                 state->regs = regs;
 559                 state->full_regs = true;
 560                 state->signal = true;
 561
 562         } else if (task == current) {
 563                 asm volatile("lea (%%rip), %0\n\t"
 564                              "mov %%rsp, %1\n\t"
 565                              "mov %%rbp, %2\n\t"
 566                              : "=r" (state->ip), "=r" (state->sp),
 567                                "=r" (state->bp));
 568
 569         } else {
 570                 struct inactive_task_frame *frame = (void *)task->thread.sp;
 571
 572                 state->sp = task->thread.sp;
 573                 state->bp = READ_ONCE_NOCHECK(frame->bp);
 574                 state->ip = READ_ONCE_NOCHECK(frame->ret_addr);
 575         }
 576
 577         if (get_stack_info((unsigned long *)state->sp, state->task,
 578                            &state->stack_info, &state->stack_mask)) {
 579                 /*
 580                  * We weren't on a valid stack.  It's possible that
 581                  * we overflowed a valid stack into a guard page.
 582                  * See if the next page up is valid so that we can
 583                  * generate some kind of backtrace if this happens.
 584                  */
 585                 void *next_page = (void *)PAGE_ALIGN((unsigned long)state->sp);
 586                 if (get_stack_info(next_page, state->task, &state->stack_info,
 587                                    &state->stack_mask))
 588                         return;
 589         }
 590
 591         /*
 592          * The caller can provide the address of the first frame directly
 593          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
 594          * to start unwinding at.  Skip ahead until we reach it.
 595          */
 596
 597         /* When starting from regs, skip the regs frame: */
 598         if (regs) {
 599                 unwind_next_frame(state);
 600                 return;
 601         }
 602
 603         /* Otherwise, skip ahead to the user-specified starting frame: */
 604         while (!unwind_done(state) &&
 605                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
 606                         state->sp <= (unsigned long)first_frame))
 607                 unwind_next_frame(state);
 608
 609         return;
 610
 611 done:
 612         state->stack_info.type = STACK_TYPE_UNKNOWN;
 613         return;
 614 }
 615 EXPORT_SYMBOL_GPL(__unwind_start);