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