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 static struct orc_entry *orc_find(unsigned long ip)
  78 {
  79         if (!orc_init)
  80                 return NULL;
  81
  82         /* For non-init vmlinux addresses, use the fast lookup table: */
  83         if (ip >= LOOKUP_START_IP && ip < LOOKUP_STOP_IP) {
  84                 unsigned int idx, start, stop;
  85
  86                 idx = (ip - LOOKUP_START_IP) / LOOKUP_BLOCK_SIZE;
  87
  88                 if (unlikely((idx >= lookup_num_blocks-1))) {
  89                         orc_warn("WARNING: bad lookup idx: idx=%u num=%u ip=%pB\n",
  90                                  idx, lookup_num_blocks, (void *)ip);
  91                         return NULL;
  92                 }
  93
  94                 start = orc_lookup[idx];
  95                 stop = orc_lookup[idx + 1] + 1;
  96
  97                 if (unlikely((__start_orc_unwind + start >= __stop_orc_unwind) ||
  98                              (__start_orc_unwind + stop > __stop_orc_unwind))) {
  99                         orc_warn("WARNING: bad lookup value: idx=%u num=%u start=%u stop=%u ip=%pB\n",
 100                                  idx, lookup_num_blocks, start, stop, (void *)ip);
 101                         return NULL;
 102                 }
 103
 104                 return __orc_find(__start_orc_unwind_ip + start,
 105                                   __start_orc_unwind + start, stop - start, ip);
 106         }
 107
 108         /* vmlinux .init slow lookup: */
 109         if (ip >= (unsigned long)_sinittext && ip < (unsigned long)_einittext)
 110                 return __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 111                                   __stop_orc_unwind_ip - __start_orc_unwind_ip, ip);
 112
 113         /* Module lookup: */
 114         return orc_module_find(ip);
 115 }
 116
 117 static void orc_sort_swap(void *_a, void *_b, int size)
 118 {
 119         struct orc_entry *orc_a, *orc_b;
 120         struct orc_entry orc_tmp;
 121         int *a = _a, *b = _b, tmp;
 122         int delta = _b - _a;
 123
 124         /* Swap the .orc_unwind_ip entries: */
 125         tmp = *a;
 126         *a = *b + delta;
 127         *b = tmp - delta;
 128
 129         /* Swap the corresponding .orc_unwind entries: */
 130         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 131         orc_b = cur_orc_table + (b - cur_orc_ip_table);
 132         orc_tmp = *orc_a;
 133         *orc_a = *orc_b;
 134         *orc_b = orc_tmp;
 135 }
 136
 137 static int orc_sort_cmp(const void *_a, const void *_b)
 138 {
 139         struct orc_entry *orc_a;
 140         const int *a = _a, *b = _b;
 141         unsigned long a_val = orc_ip(a);
 142         unsigned long b_val = orc_ip(b);
 143
 144         if (a_val > b_val)
 145                 return 1;
 146         if (a_val < b_val)
 147                 return -1;
 148
 149         /*
 150          * The "weak" section terminator entries need to always be on the left
 151          * to ensure the lookup code skips them in favor of real entries.
 152          * These terminator entries exist to handle any gaps created by
 153          * whitelisted .o files which didn't get objtool generation.
 154          */
 155         orc_a = cur_orc_table + (a - cur_orc_ip_table);
 156         return orc_a->sp_reg == ORC_REG_UNDEFINED ? -1 : 1;
 157 }
 158
 159 #ifdef CONFIG_MODULES
 160 void unwind_module_init(struct module *mod, void *_orc_ip, size_t orc_ip_size,
 161                         void *_orc, size_t orc_size)
 162 {
 163         int *orc_ip = _orc_ip;
 164         struct orc_entry *orc = _orc;
 165         unsigned int num_entries = orc_ip_size / sizeof(int);
 166
 167         WARN_ON_ONCE(orc_ip_size % sizeof(int) != 0 ||
 168                      orc_size % sizeof(*orc) != 0 ||
 169                      num_entries != orc_size / sizeof(*orc));
 170
 171         /*
 172          * The 'cur_orc_*' globals allow the orc_sort_swap() callback to
 173          * associate an .orc_unwind_ip table entry with its corresponding
 174          * .orc_unwind entry so they can both be swapped.
 175          */
 176         mutex_lock(&sort_mutex);
 177         cur_orc_ip_table = orc_ip;
 178         cur_orc_table = orc;
 179         sort(orc_ip, num_entries, sizeof(int), orc_sort_cmp, orc_sort_swap);
 180         mutex_unlock(&sort_mutex);
 181
 182         mod->arch.orc_unwind_ip = orc_ip;
 183         mod->arch.orc_unwind = orc;
 184         mod->arch.num_orcs = num_entries;
 185 }
 186 #endif
 187
 188 void __init unwind_init(void)
 189 {
 190         size_t orc_ip_size = (void *)__stop_orc_unwind_ip - (void *)__start_orc_unwind_ip;
 191         size_t orc_size = (void *)__stop_orc_unwind - (void *)__start_orc_unwind;
 192         size_t num_entries = orc_ip_size / sizeof(int);
 193         struct orc_entry *orc;
 194         int i;
 195
 196         if (!num_entries || orc_ip_size % sizeof(int) != 0 ||
 197             orc_size % sizeof(struct orc_entry) != 0 ||
 198             num_entries != orc_size / sizeof(struct orc_entry)) {
 199                 orc_warn("WARNING: Bad or missing .orc_unwind table.  Disabling unwinder.\n");
 200                 return;
 201         }
 202
 203         /* Sort the .orc_unwind and .orc_unwind_ip tables: */
 204         sort(__start_orc_unwind_ip, num_entries, sizeof(int), orc_sort_cmp,
 205              orc_sort_swap);
 206
 207         /* Initialize the fast lookup table: */
 208         lookup_num_blocks = orc_lookup_end - orc_lookup;
 209         for (i = 0; i < lookup_num_blocks-1; i++) {
 210                 orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind,
 211                                  num_entries,
 212                                  LOOKUP_START_IP + (LOOKUP_BLOCK_SIZE * i));
 213                 if (!orc) {
 214                         orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 215                         return;
 216                 }
 217
 218                 orc_lookup[i] = orc - __start_orc_unwind;
 219         }
 220
 221         /* Initialize the ending block: */
 222         orc = __orc_find(__start_orc_unwind_ip, __start_orc_unwind, num_entries,
 223                          LOOKUP_STOP_IP);
 224         if (!orc) {
 225                 orc_warn("WARNING: Corrupt .orc_unwind table.  Disabling unwinder.\n");
 226                 return;
 227         }
 228         orc_lookup[lookup_num_blocks-1] = orc - __start_orc_unwind;
 229
 230         orc_init = true;
 231 }
 232
 233 unsigned long unwind_get_return_address(struct unwind_state *state)
 234 {
 235         if (unwind_done(state))
 236                 return 0;
 237
 238         return __kernel_text_address(state->ip) ? state->ip : 0;
 239 }
 240 EXPORT_SYMBOL_GPL(unwind_get_return_address);
 241
 242 unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
 243 {
 244         if (unwind_done(state))
 245                 return NULL;
 246
 247         if (state->regs)
 248                 return &state->regs->ip;
 249
 250         if (state->sp)
 251                 return (unsigned long *)state->sp - 1;
 252
 253         return NULL;
 254 }
 255
 256 static bool stack_access_ok(struct unwind_state *state, unsigned long _addr,
 257                             size_t len)
 258 {
 259         struct stack_info *info = &state->stack_info;
 260         void *addr = (void *)_addr;
 261
 262         if (!on_stack(info, addr, len) &&
 263             (get_stack_info(addr, state->task, info, &state->stack_mask)))
 264                 return false;
 265
 266         return true;
 267 }
 268
 269 static bool deref_stack_reg(struct unwind_state *state, unsigned long addr,
 270                             unsigned long *val)
 271 {
 272         if (!stack_access_ok(state, addr, sizeof(long)))
 273                 return false;
 274
 275         *val = READ_ONCE_NOCHECK(*(unsigned long *)addr);
 276         return true;
 277 }
 278
 279 static bool deref_stack_regs(struct unwind_state *state, unsigned long addr,
 280                              unsigned long *ip, unsigned long *sp)
 281 {
 282         struct pt_regs *regs = (struct pt_regs *)addr;
 283
 284         /* x86-32 support will be more complicated due to the &regs->sp hack */
 285         BUILD_BUG_ON(IS_ENABLED(CONFIG_X86_32));
 286
 287         if (!stack_access_ok(state, addr, sizeof(struct pt_regs)))
 288                 return false;
 289
 290         *ip = regs->ip;
 291         *sp = regs->sp;
 292         return true;
 293 }
 294
 295 static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr,
 296                                   unsigned long *ip, unsigned long *sp)
 297 {
 298         struct pt_regs *regs = (void *)addr - IRET_FRAME_OFFSET;
 299
 300         if (!stack_access_ok(state, addr, IRET_FRAME_SIZE))
 301                 return false;
 302
 303         *ip = regs->ip;
 304         *sp = regs->sp;
 305         return true;
 306 }
 307
 308 bool unwind_next_frame(struct unwind_state *state)
 309 {
 310         unsigned long ip_p, sp, orig_ip, prev_sp = state->sp;
 311         enum stack_type prev_type = state->stack_info.type;
 312         struct orc_entry *orc;
 313         bool indirect = false;
 314
 315         if (unwind_done(state))
 316                 return false;
 317
 318         /* Don't let modules unload while we're reading their ORC data. */
 319         preempt_disable();
 320
 321         /* Have we reached the end? */
 322         if (state->regs && user_mode(state->regs))
 323                 goto done;
 324
 325         /*
 326          * Find the orc_entry associated with the text address.
 327          *
 328          * Decrement call return addresses by one so they work for sibling
 329          * calls and calls to noreturn functions.
 330          */
 331         orc = orc_find(state->signal ? state->ip : state->ip - 1);
 332         if (!orc || orc->sp_reg == ORC_REG_UNDEFINED)
 333                 goto done;
 334         orig_ip = state->ip;
 335
 336         /* Find the previous frame's stack: */
 337         switch (orc->sp_reg) {
 338         case ORC_REG_SP:
 339                 sp = state->sp + orc->sp_offset;
 340                 break;
 341
 342         case ORC_REG_BP:
 343                 sp = state->bp + orc->sp_offset;
 344                 break;
 345
 346         case ORC_REG_SP_INDIRECT:
 347                 sp = state->sp + orc->sp_offset;
 348                 indirect = true;
 349                 break;
 350
 351         case ORC_REG_BP_INDIRECT:
 352                 sp = state->bp + orc->sp_offset;
 353                 indirect = true;
 354                 break;
 355
 356         case ORC_REG_R10:
 357                 if (!state->regs || !state->full_regs) {
 358                         orc_warn("missing regs for base reg R10 at ip %pB\n",
 359                                  (void *)state->ip);
 360                         goto done;
 361                 }
 362                 sp = state->regs->r10;
 363                 break;
 364
 365         case ORC_REG_R13:
 366                 if (!state->regs || !state->full_regs) {
 367                         orc_warn("missing regs for base reg R13 at ip %pB\n",
 368                                  (void *)state->ip);
 369                         goto done;
 370                 }
 371                 sp = state->regs->r13;
 372                 break;
 373
 374         case ORC_REG_DI:
 375                 if (!state->regs || !state->full_regs) {
 376                         orc_warn("missing regs for base reg DI at ip %pB\n",
 377                                  (void *)state->ip);
 378                         goto done;
 379                 }
 380                 sp = state->regs->di;
 381                 break;
 382
 383         case ORC_REG_DX:
 384                 if (!state->regs || !state->full_regs) {
 385                         orc_warn("missing regs for base reg DX at ip %pB\n",
 386                                  (void *)state->ip);
 387                         goto done;
 388                 }
 389                 sp = state->regs->dx;
 390                 break;
 391
 392         default:
 393                 orc_warn("unknown SP base reg %d for ip %pB\n",
 394                          orc->sp_reg, (void *)state->ip);
 395                 goto done;
 396         }
 397
 398         if (indirect) {
 399                 if (!deref_stack_reg(state, sp, &sp))
 400                         goto done;
 401         }
 402
 403         /* Find IP, SP and possibly regs: */
 404         switch (orc->type) {
 405         case ORC_TYPE_CALL:
 406                 ip_p = sp - sizeof(long);
 407
 408                 if (!deref_stack_reg(state, ip_p, &state->ip))
 409                         goto done;
 410
 411                 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
 412                                                   state->ip, (void *)ip_p);
 413
 414                 state->sp = sp;
 415                 state->regs = NULL;
 416                 state->signal = false;
 417                 break;
 418
 419         case ORC_TYPE_REGS:
 420                 if (!deref_stack_regs(state, sp, &state->ip, &state->sp)) {
 421                         orc_warn("can't dereference registers at %p for ip %pB\n",
 422                                  (void *)sp, (void *)orig_ip);
 423                         goto done;
 424                 }
 425
 426                 state->regs = (struct pt_regs *)sp;
 427                 state->full_regs = true;
 428                 state->signal = true;
 429                 break;
 430
 431         case ORC_TYPE_REGS_IRET:
 432                 if (!deref_stack_iret_regs(state, sp, &state->ip, &state->sp)) {
 433                         orc_warn("can't dereference iret registers at %p for ip %pB\n",
 434                                  (void *)sp, (void *)orig_ip);
 435                         goto done;
 436                 }
 437
 438                 state->regs = (void *)sp - IRET_FRAME_OFFSET;
 439                 state->full_regs = false;
 440                 state->signal = true;
 441                 break;
 442
 443         default:
 444                 orc_warn("unknown .orc_unwind entry type %d for ip %pB\n",
 445                          orc->type, (void *)orig_ip);
 446                 break;
 447         }
 448
 449         /* Find BP: */
 450         switch (orc->bp_reg) {
 451         case ORC_REG_UNDEFINED:
 452                 if (state->regs && state->full_regs)
 453                         state->bp = state->regs->bp;
 454                 break;
 455
 456         case ORC_REG_PREV_SP:
 457                 if (!deref_stack_reg(state, sp + orc->bp_offset, &state->bp))
 458                         goto done;
 459                 break;
 460
 461         case ORC_REG_BP:
 462                 if (!deref_stack_reg(state, state->bp + orc->bp_offset, &state->bp))
 463                         goto done;
 464                 break;
 465
 466         default:
 467                 orc_warn("unknown BP base reg %d for ip %pB\n",
 468                          orc->bp_reg, (void *)orig_ip);
 469                 goto done;
 470         }
 471
 472         /* Prevent a recursive loop due to bad ORC data: */
 473         if (state->stack_info.type == prev_type &&
 474             on_stack(&state->stack_info, (void *)state->sp, sizeof(long)) &&
 475             state->sp <= prev_sp) {
 476                 orc_warn("stack going in the wrong direction? ip=%pB\n",
 477                          (void *)orig_ip);
 478                 goto done;
 479         }
 480
 481         preempt_enable();
 482         return true;
 483
 484 done:
 485         preempt_enable();
 486         state->stack_info.type = STACK_TYPE_UNKNOWN;
 487         return false;
 488 }
 489 EXPORT_SYMBOL_GPL(unwind_next_frame);
 490
 491 void __unwind_start(struct unwind_state *state, struct task_struct *task,
 492                     struct pt_regs *regs, unsigned long *first_frame)
 493 {
 494         memset(state, 0, sizeof(*state));
 495         state->task = task;
 496
 497         /*
 498          * Refuse to unwind the stack of a task while it's executing on another
 499          * CPU.  This check is racy, but that's ok: the unwinder has other
 500          * checks to prevent it from going off the rails.
 501          */
 502         if (task_on_another_cpu(task))
 503                 goto done;
 504
 505         if (regs) {
 506                 if (user_mode(regs))
 507                         goto done;
 508
 509                 state->ip = regs->ip;
 510                 state->sp = kernel_stack_pointer(regs);
 511                 state->bp = regs->bp;
 512                 state->regs = regs;
 513                 state->full_regs = true;
 514                 state->signal = true;
 515
 516         } else if (task == current) {
 517                 asm volatile("lea (%%rip), %0\n\t"
 518                              "mov %%rsp, %1\n\t"
 519                              "mov %%rbp, %2\n\t"
 520                              : "=r" (state->ip), "=r" (state->sp),
 521                                "=r" (state->bp));
 522
 523         } else {
 524                 struct inactive_task_frame *frame = (void *)task->thread.sp;
 525
 526                 state->sp = task->thread.sp;
 527                 state->bp = READ_ONCE_NOCHECK(frame->bp);
 528                 state->ip = READ_ONCE_NOCHECK(frame->ret_addr);
 529         }
 530
 531         if (get_stack_info((unsigned long *)state->sp, state->task,
 532                            &state->stack_info, &state->stack_mask)) {
 533                 /*
 534                  * We weren't on a valid stack.  It's possible that
 535                  * we overflowed a valid stack into a guard page.
 536                  * See if the next page up is valid so that we can
 537                  * generate some kind of backtrace if this happens.
 538                  */
 539                 void *next_page = (void *)PAGE_ALIGN((unsigned long)state->sp);
 540                 if (get_stack_info(next_page, state->task, &state->stack_info,
 541                                    &state->stack_mask))
 542                         return;
 543         }
 544
 545         /*
 546          * The caller can provide the address of the first frame directly
 547          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
 548          * to start unwinding at.  Skip ahead until we reach it.
 549          */
 550
 551         /* When starting from regs, skip the regs frame: */
 552         if (regs) {
 553                 unwind_next_frame(state);
 554                 return;
 555         }
 556
 557         /* Otherwise, skip ahead to the user-specified starting frame: */
 558         while (!unwind_done(state) &&
 559                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
 560                         state->sp <= (unsigned long)first_frame))
 561                 unwind_next_frame(state);
 562
 563         return;
 564
 565 done:
 566         state->stack_info.type = STACK_TYPE_UNKNOWN;
 567         return;
 568 }
 569 EXPORT_SYMBOL_GPL(__unwind_start);