arch/sh/kernel/dwarf.c

   1 /*
   2  * Copyright (C) 2009 Matt Fleming <matt@console-pimps.org>
   3  *
   4  * This file is subject to the terms and conditions of the GNU General Public
   5  * License.  See the file "COPYING" in the main directory of this archive
   6  * for more details.
   7  *
   8  * This is an implementation of a DWARF unwinder. Its main purpose is
   9  * for generating stacktrace information. Based on the DWARF 3
  10  * specification from http://www.dwarfstd.org.
  11  *
  12  * TODO:
  13  *      - DWARF64 doesn't work.
  14  *      - Registers with DWARF_VAL_OFFSET rules aren't handled properly.
  15  */
  16
  17 /* #define DEBUG */
  18 #include <linux/kernel.h>
  19 #include <linux/io.h>
  20 #include <linux/list.h>
  21 #include <linux/mempool.h>
  22 #include <linux/mm.h>
  23 #include <linux/elf.h>
  24 #include <linux/ftrace.h>
  25 #include <asm/dwarf.h>
  26 #include <asm/unwinder.h>
  27 #include <asm/sections.h>
  28 #include <asm/unaligned.h>
  29 #include <asm/stacktrace.h>
  30
  31 /* Reserve enough memory for two stack frames */
  32 #define DWARF_FRAME_MIN_REQ     2
  33 /* ... with 4 registers per frame. */
  34 #define DWARF_REG_MIN_REQ       (DWARF_FRAME_MIN_REQ * 4)
  35
  36 static struct kmem_cache *dwarf_frame_cachep;
  37 static mempool_t *dwarf_frame_pool;
  38
  39 static struct kmem_cache *dwarf_reg_cachep;
  40 static mempool_t *dwarf_reg_pool;
  41
  42 static LIST_HEAD(dwarf_cie_list);
  43 static DEFINE_SPINLOCK(dwarf_cie_lock);
  44
  45 static LIST_HEAD(dwarf_fde_list);
  46 static DEFINE_SPINLOCK(dwarf_fde_lock);
  47
  48 static struct dwarf_cie *cached_cie;
  49
  50 /**
  51  *      dwarf_frame_alloc_reg - allocate memory for a DWARF register
  52  *      @frame: the DWARF frame whose list of registers we insert on
  53  *      @reg_num: the register number
  54  *
  55  *      Allocate space for, and initialise, a dwarf reg from
  56  *      dwarf_reg_pool and insert it onto the (unsorted) linked-list of
  57  *      dwarf registers for @frame.
  58  *
  59  *      Return the initialised DWARF reg.
  60  */
  61 static struct dwarf_reg *dwarf_frame_alloc_reg(struct dwarf_frame *frame,
  62                                                unsigned int reg_num)
  63 {
  64         struct dwarf_reg *reg;
  65
  66         reg = mempool_alloc(dwarf_reg_pool, GFP_ATOMIC);
  67         if (!reg) {
  68                 printk(KERN_WARNING "Unable to allocate a DWARF register\n");
  69                 /*
  70                  * Let's just bomb hard here, we have no way to
  71                  * gracefully recover.
  72                  */
  73                 UNWINDER_BUG();
  74         }
  75
  76         reg->number = reg_num;
  77         reg->addr = 0;
  78         reg->flags = 0;
  79
  80         list_add(&reg->link, &frame->reg_list);
  81
  82         return reg;
  83 }
  84
  85 static void dwarf_frame_free_regs(struct dwarf_frame *frame)
  86 {
  87         struct dwarf_reg *reg, *n;
  88
  89         list_for_each_entry_safe(reg, n, &frame->reg_list, link) {
  90                 list_del(&reg->link);
  91                 mempool_free(reg, dwarf_reg_pool);
  92         }
  93 }
  94
  95 /**
  96  *      dwarf_frame_reg - return a DWARF register
  97  *      @frame: the DWARF frame to search in for @reg_num
  98  *      @reg_num: the register number to search for
  99  *
 100  *      Lookup and return the dwarf reg @reg_num for this frame. Return
 101  *      NULL if @reg_num is an register invalid number.
 102  */
 103 static struct dwarf_reg *dwarf_frame_reg(struct dwarf_frame *frame,
 104                                          unsigned int reg_num)
 105 {
 106         struct dwarf_reg *reg;
 107
 108         list_for_each_entry(reg, &frame->reg_list, link) {
 109                 if (reg->number == reg_num)
 110                         return reg;
 111         }
 112
 113         return NULL;
 114 }
 115
 116 /**
 117  *      dwarf_read_addr - read dwarf data
 118  *      @src: source address of data
 119  *      @dst: destination address to store the data to
 120  *
 121  *      Read 'n' bytes from @src, where 'n' is the size of an address on
 122  *      the native machine. We return the number of bytes read, which
 123  *      should always be 'n'. We also have to be careful when reading
 124  *      from @src and writing to @dst, because they can be arbitrarily
 125  *      aligned. Return 'n' - the number of bytes read.
 126  */
 127 static inline int dwarf_read_addr(unsigned long *src, unsigned long *dst)
 128 {
 129         u32 val = get_unaligned(src);
 130         put_unaligned(val, dst);
 131         return sizeof(unsigned long *);
 132 }
 133
 134 /**
 135  *      dwarf_read_uleb128 - read unsigned LEB128 data
 136  *      @addr: the address where the ULEB128 data is stored
 137  *      @ret: address to store the result
 138  *
 139  *      Decode an unsigned LEB128 encoded datum. The algorithm is taken
 140  *      from Appendix C of the DWARF 3 spec. For information on the
 141  *      encodings refer to section "7.6 - Variable Length Data". Return
 142  *      the number of bytes read.
 143  */
 144 static inline unsigned long dwarf_read_uleb128(char *addr, unsigned int *ret)
 145 {
 146         unsigned int result;
 147         unsigned char byte;
 148         int shift, count;
 149
 150         result = 0;
 151         shift = 0;
 152         count = 0;
 153
 154         while (1) {
 155                 byte = __raw_readb(addr);
 156                 addr++;
 157                 count++;
 158
 159                 result |= (byte & 0x7f) << shift;
 160                 shift += 7;
 161
 162                 if (!(byte & 0x80))
 163                         break;
 164         }
 165
 166         *ret = result;
 167
 168         return count;
 169 }
 170
 171 /**
 172  *      dwarf_read_leb128 - read signed LEB128 data
 173  *      @addr: the address of the LEB128 encoded data
 174  *      @ret: address to store the result
 175  *
 176  *      Decode signed LEB128 data. The algorithm is taken from Appendix
 177  *      C of the DWARF 3 spec. Return the number of bytes read.
 178  */
 179 static inline unsigned long dwarf_read_leb128(char *addr, int *ret)
 180 {
 181         unsigned char byte;
 182         int result, shift;
 183         int num_bits;
 184         int count;
 185
 186         result = 0;
 187         shift = 0;
 188         count = 0;
 189
 190         while (1) {
 191                 byte = __raw_readb(addr);
 192                 addr++;
 193                 result |= (byte & 0x7f) << shift;
 194                 shift += 7;
 195                 count++;
 196
 197                 if (!(byte & 0x80))
 198                         break;
 199         }
 200
 201         /* The number of bits in a signed integer. */
 202         num_bits = 8 * sizeof(result);
 203
 204         if ((shift < num_bits) && (byte & 0x40))
 205                 result |= (-1 << shift);
 206
 207         *ret = result;
 208
 209         return count;
 210 }
 211
 212 /**
 213  *      dwarf_read_encoded_value - return the decoded value at @addr
 214  *      @addr: the address of the encoded value
 215  *      @val: where to write the decoded value
 216  *      @encoding: the encoding with which we can decode @addr
 217  *
 218  *      GCC emits encoded address in the .eh_frame FDE entries. Decode
 219  *      the value at @addr using @encoding. The decoded value is written
 220  *      to @val and the number of bytes read is returned.
 221  */
 222 static int dwarf_read_encoded_value(char *addr, unsigned long *val,
 223                                     char encoding)
 224 {
 225         unsigned long decoded_addr = 0;
 226         int count = 0;
 227
 228         switch (encoding & 0x70) {
 229         case DW_EH_PE_absptr:
 230                 break;
 231         case DW_EH_PE_pcrel:
 232                 decoded_addr = (unsigned long)addr;
 233                 break;
 234         default:
 235                 pr_debug("encoding=0x%x\n", (encoding & 0x70));
 236                 UNWINDER_BUG();
 237         }
 238
 239         if ((encoding & 0x07) == 0x00)
 240                 encoding |= DW_EH_PE_udata4;
 241
 242         switch (encoding & 0x0f) {
 243         case DW_EH_PE_sdata4:
 244         case DW_EH_PE_udata4:
 245                 count += 4;
 246                 decoded_addr += get_unaligned((u32 *)addr);
 247                 __raw_writel(decoded_addr, val);
 248                 break;
 249         default:
 250                 pr_debug("encoding=0x%x\n", encoding);
 251                 UNWINDER_BUG();
 252         }
 253
 254         return count;
 255 }
 256
 257 /**
 258  *      dwarf_entry_len - return the length of an FDE or CIE
 259  *      @addr: the address of the entry
 260  *      @len: the length of the entry
 261  *
 262  *      Read the initial_length field of the entry and store the size of
 263  *      the entry in @len. We return the number of bytes read. Return a
 264  *      count of 0 on error.
 265  */
 266 static inline int dwarf_entry_len(char *addr, unsigned long *len)
 267 {
 268         u32 initial_len;
 269         int count;
 270
 271         initial_len = get_unaligned((u32 *)addr);
 272         count = 4;
 273
 274         /*
 275          * An initial length field value in the range DW_LEN_EXT_LO -
 276          * DW_LEN_EXT_HI indicates an extension, and should not be
 277          * interpreted as a length. The only extension that we currently
 278          * understand is the use of DWARF64 addresses.
 279          */
 280         if (initial_len >= DW_EXT_LO && initial_len <= DW_EXT_HI) {
 281                 /*
 282                  * The 64-bit length field immediately follows the
 283                  * compulsory 32-bit length field.
 284                  */
 285                 if (initial_len == DW_EXT_DWARF64) {
 286                         *len = get_unaligned((u64 *)addr + 4);
 287                         count = 12;
 288                 } else {
 289                         printk(KERN_WARNING "Unknown DWARF extension\n");
 290                         count = 0;
 291                 }
 292         } else
 293                 *len = initial_len;
 294
 295         return count;
 296 }
 297
 298 /**
 299  *      dwarf_lookup_cie - locate the cie
 300  *      @cie_ptr: pointer to help with lookup
 301  */
 302 static struct dwarf_cie *dwarf_lookup_cie(unsigned long cie_ptr)
 303 {
 304         struct dwarf_cie *cie;
 305         unsigned long flags;
 306
 307         spin_lock_irqsave(&dwarf_cie_lock, flags);
 308
 309         /*
 310          * We've cached the last CIE we looked up because chances are
 311          * that the FDE wants this CIE.
 312          */
 313         if (cached_cie && cached_cie->cie_pointer == cie_ptr) {
 314                 cie = cached_cie;
 315                 goto out;
 316         }
 317
 318         list_for_each_entry(cie, &dwarf_cie_list, link) {
 319                 if (cie->cie_pointer == cie_ptr) {
 320                         cached_cie = cie;
 321                         break;
 322                 }
 323         }
 324
 325         /* Couldn't find the entry in the list. */
 326         if (&cie->link == &dwarf_cie_list)
 327                 cie = NULL;
 328 out:
 329         spin_unlock_irqrestore(&dwarf_cie_lock, flags);
 330         return cie;
 331 }
 332
 333 /**
 334  *      dwarf_lookup_fde - locate the FDE that covers pc
 335  *      @pc: the program counter
 336  */
 337 struct dwarf_fde *dwarf_lookup_fde(unsigned long pc)
 338 {
 339         struct dwarf_fde *fde;
 340         unsigned long flags;
 341
 342         spin_lock_irqsave(&dwarf_fde_lock, flags);
 343
 344         list_for_each_entry(fde, &dwarf_fde_list, link) {
 345                 unsigned long start, end;
 346
 347                 start = fde->initial_location;
 348                 end = fde->initial_location + fde->address_range;
 349
 350                 if (pc >= start && pc < end)
 351                         break;
 352         }
 353
 354         /* Couldn't find the entry in the list. */
 355         if (&fde->link == &dwarf_fde_list)
 356                 fde = NULL;
 357
 358         spin_unlock_irqrestore(&dwarf_fde_lock, flags);
 359
 360         return fde;
 361 }
 362
 363 /**
 364  *      dwarf_cfa_execute_insns - execute instructions to calculate a CFA
 365  *      @insn_start: address of the first instruction
 366  *      @insn_end: address of the last instruction
 367  *      @cie: the CIE for this function
 368  *      @fde: the FDE for this function
 369  *      @frame: the instructions calculate the CFA for this frame
 370  *      @pc: the program counter of the address we're interested in
 371  *
 372  *      Execute the Call Frame instruction sequence starting at
 373  *      @insn_start and ending at @insn_end. The instructions describe
 374  *      how to calculate the Canonical Frame Address of a stackframe.
 375  *      Store the results in @frame.
 376  */
 377 static int dwarf_cfa_execute_insns(unsigned char *insn_start,
 378                                    unsigned char *insn_end,
 379                                    struct dwarf_cie *cie,
 380                                    struct dwarf_fde *fde,
 381                                    struct dwarf_frame *frame,
 382                                    unsigned long pc)
 383 {
 384         unsigned char insn;
 385         unsigned char *current_insn;
 386         unsigned int count, delta, reg, expr_len, offset;
 387         struct dwarf_reg *regp;
 388
 389         current_insn = insn_start;
 390
 391         while (current_insn < insn_end && frame->pc <= pc) {
 392                 insn = __raw_readb(current_insn++);
 393
 394                 /*
 395                  * Firstly, handle the opcodes that embed their operands
 396                  * in the instructions.
 397                  */
 398                 switch (DW_CFA_opcode(insn)) {
 399                 case DW_CFA_advance_loc:
 400                         delta = DW_CFA_operand(insn);
 401                         delta *= cie->code_alignment_factor;
 402                         frame->pc += delta;
 403                         continue;
 404                         /* NOTREACHED */
 405                 case DW_CFA_offset:
 406                         reg = DW_CFA_operand(insn);
 407                         count = dwarf_read_uleb128(current_insn, &offset);
 408                         current_insn += count;
 409                         offset *= cie->data_alignment_factor;
 410                         regp = dwarf_frame_alloc_reg(frame, reg);
 411                         regp->addr = offset;
 412                         regp->flags |= DWARF_REG_OFFSET;
 413                         continue;
 414                         /* NOTREACHED */
 415                 case DW_CFA_restore:
 416                         reg = DW_CFA_operand(insn);
 417                         continue;
 418                         /* NOTREACHED */
 419                 }
 420
 421                 /*
 422                  * Secondly, handle the opcodes that don't embed their
 423                  * operands in the instruction.
 424                  */
 425                 switch (insn) {
 426                 case DW_CFA_nop:
 427                         continue;
 428                 case DW_CFA_advance_loc1:
 429                         delta = *current_insn++;
 430                         frame->pc += delta * cie->code_alignment_factor;
 431                         break;
 432                 case DW_CFA_advance_loc2:
 433                         delta = get_unaligned((u16 *)current_insn);
 434                         current_insn += 2;
 435                         frame->pc += delta * cie->code_alignment_factor;
 436                         break;
 437                 case DW_CFA_advance_loc4:
 438                         delta = get_unaligned((u32 *)current_insn);
 439                         current_insn += 4;
 440                         frame->pc += delta * cie->code_alignment_factor;
 441                         break;
 442                 case DW_CFA_offset_extended:
 443                         count = dwarf_read_uleb128(current_insn, &reg);
 444                         current_insn += count;
 445                         count = dwarf_read_uleb128(current_insn, &offset);
 446                         current_insn += count;
 447                         offset *= cie->data_alignment_factor;
 448                         break;
 449                 case DW_CFA_restore_extended:
 450                         count = dwarf_read_uleb128(current_insn, &reg);
 451                         current_insn += count;
 452                         break;
 453                 case DW_CFA_undefined:
 454                         count = dwarf_read_uleb128(current_insn, &reg);
 455                         current_insn += count;
 456                         regp = dwarf_frame_alloc_reg(frame, reg);
 457                         regp->flags |= DWARF_UNDEFINED;
 458                         break;
 459                 case DW_CFA_def_cfa:
 460                         count = dwarf_read_uleb128(current_insn,
 461                                                    &frame->cfa_register);
 462                         current_insn += count;
 463                         count = dwarf_read_uleb128(current_insn,
 464                                                    &frame->cfa_offset);
 465                         current_insn += count;
 466
 467                         frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
 468                         break;
 469                 case DW_CFA_def_cfa_register:
 470                         count = dwarf_read_uleb128(current_insn,
 471                                                    &frame->cfa_register);
 472                         current_insn += count;
 473                         frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
 474                         break;
 475                 case DW_CFA_def_cfa_offset:
 476                         count = dwarf_read_uleb128(current_insn, &offset);
 477                         current_insn += count;
 478                         frame->cfa_offset = offset;
 479                         break;
 480                 case DW_CFA_def_cfa_expression:
 481                         count = dwarf_read_uleb128(current_insn, &expr_len);
 482                         current_insn += count;
 483
 484                         frame->cfa_expr = current_insn;
 485                         frame->cfa_expr_len = expr_len;
 486                         current_insn += expr_len;
 487
 488                         frame->flags |= DWARF_FRAME_CFA_REG_EXP;
 489                         break;
 490                 case DW_CFA_offset_extended_sf:
 491                         count = dwarf_read_uleb128(current_insn, &reg);
 492                         current_insn += count;
 493                         count = dwarf_read_leb128(current_insn, &offset);
 494                         current_insn += count;
 495                         offset *= cie->data_alignment_factor;
 496                         regp = dwarf_frame_alloc_reg(frame, reg);
 497                         regp->flags |= DWARF_REG_OFFSET;
 498                         regp->addr = offset;
 499                         break;
 500                 case DW_CFA_val_offset:
 501                         count = dwarf_read_uleb128(current_insn, &reg);
 502                         current_insn += count;
 503                         count = dwarf_read_leb128(current_insn, &offset);
 504                         offset *= cie->data_alignment_factor;
 505                         regp = dwarf_frame_alloc_reg(frame, reg);
 506                         regp->flags |= DWARF_VAL_OFFSET;
 507                         regp->addr = offset;
 508                         break;
 509                 case DW_CFA_GNU_args_size:
 510                         count = dwarf_read_uleb128(current_insn, &offset);
 511                         current_insn += count;
 512                         break;
 513                 case DW_CFA_GNU_negative_offset_extended:
 514                         count = dwarf_read_uleb128(current_insn, &reg);
 515                         current_insn += count;
 516                         count = dwarf_read_uleb128(current_insn, &offset);
 517                         offset *= cie->data_alignment_factor;
 518
 519                         regp = dwarf_frame_alloc_reg(frame, reg);
 520                         regp->flags |= DWARF_REG_OFFSET;
 521                         regp->addr = -offset;
 522                         break;
 523                 default:
 524                         pr_debug("unhandled DWARF instruction 0x%x\n", insn);
 525                         UNWINDER_BUG();
 526                         break;
 527                 }
 528         }
 529
 530         return 0;
 531 }
 532
 533 /**
 534  *      dwarf_free_frame - free the memory allocated for @frame
 535  *      @frame: the frame to free
 536  */
 537 void dwarf_free_frame(struct dwarf_frame *frame)
 538 {
 539         dwarf_frame_free_regs(frame);
 540         mempool_free(frame, dwarf_frame_pool);
 541 }
 542
 543 extern void ret_from_irq(void);
 544
 545 /**
 546  *      dwarf_unwind_stack - unwind the stack
 547  *
 548  *      @pc: address of the function to unwind
 549  *      @prev: struct dwarf_frame of the previous stackframe on the callstack
 550  *
 551  *      Return a struct dwarf_frame representing the most recent frame
 552  *      on the callstack. Each of the lower (older) stack frames are
 553  *      linked via the "prev" member.
 554  */
 555 struct dwarf_frame * dwarf_unwind_stack(unsigned long pc,
 556                                         struct dwarf_frame *prev)
 557 {
 558         struct dwarf_frame *frame;
 559         struct dwarf_cie *cie;
 560         struct dwarf_fde *fde;
 561         struct dwarf_reg *reg;
 562         unsigned long addr;
 563
 564         /*
 565          * If we're starting at the top of the stack we need get the
 566          * contents of a physical register to get the CFA in order to
 567          * begin the virtual unwinding of the stack.
 568          *
 569          * NOTE: the return address is guaranteed to be setup by the
 570          * time this function makes its first function call.
 571          */
 572         if (!pc || !prev)
 573                 pc = (unsigned long)current_text_addr();
 574
 575 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
 576         /*
 577          * If our stack has been patched by the function graph tracer
 578          * then we might see the address of return_to_handler() where we
 579          * expected to find the real return address.
 580          */
 581         if (pc == (unsigned long)&return_to_handler) {
 582                 int index = current->curr_ret_stack;
 583
 584                 /*
 585                  * We currently have no way of tracking how many
 586                  * return_to_handler()'s we've seen. If there is more
 587                  * than one patched return address on our stack,
 588                  * complain loudly.
 589                  */
 590                 WARN_ON(index > 0);
 591
 592                 pc = current->ret_stack[index].ret;
 593         }
 594 #endif
 595
 596         frame = mempool_alloc(dwarf_frame_pool, GFP_ATOMIC);
 597         if (!frame) {
 598                 printk(KERN_ERR "Unable to allocate a dwarf frame\n");
 599                 UNWINDER_BUG();
 600         }
 601
 602         INIT_LIST_HEAD(&frame->reg_list);
 603         frame->flags = 0;
 604         frame->prev = prev;
 605         frame->return_addr = 0;
 606
 607         fde = dwarf_lookup_fde(pc);
 608         if (!fde) {
 609                 /*
 610                  * This is our normal exit path. There are two reasons
 611                  * why we might exit here,
 612                  *
 613                  *      a) pc has no asscociated DWARF frame info and so
 614                  *      we don't know how to unwind this frame. This is
 615                  *      usually the case when we're trying to unwind a
 616                  *      frame that was called from some assembly code
 617                  *      that has no DWARF info, e.g. syscalls.
 618                  *
 619                  *      b) the DEBUG info for pc is bogus. There's
 620                  *      really no way to distinguish this case from the
 621                  *      case above, which sucks because we could print a
 622                  *      warning here.
 623                  */
 624                 goto bail;
 625         }
 626
 627         cie = dwarf_lookup_cie(fde->cie_pointer);
 628
 629         frame->pc = fde->initial_location;
 630
 631         /* CIE initial instructions */
 632         dwarf_cfa_execute_insns(cie->initial_instructions,
 633                                 cie->instructions_end, cie, fde,
 634                                 frame, pc);
 635
 636         /* FDE instructions */
 637         dwarf_cfa_execute_insns(fde->instructions, fde->end, cie,
 638                                 fde, frame, pc);
 639
 640         /* Calculate the CFA */
 641         switch (frame->flags) {
 642         case DWARF_FRAME_CFA_REG_OFFSET:
 643                 if (prev) {
 644                         reg = dwarf_frame_reg(prev, frame->cfa_register);
 645                         UNWINDER_BUG_ON(!reg);
 646                         UNWINDER_BUG_ON(reg->flags != DWARF_REG_OFFSET);
 647
 648                         addr = prev->cfa + reg->addr;
 649                         frame->cfa = __raw_readl(addr);
 650
 651                 } else {
 652                         /*
 653                          * Again, we're starting from the top of the
 654                          * stack. We need to physically read
 655                          * the contents of a register in order to get
 656                          * the Canonical Frame Address for this
 657                          * function.
 658                          */
 659                         frame->cfa = dwarf_read_arch_reg(frame->cfa_register);
 660                 }
 661
 662                 frame->cfa += frame->cfa_offset;
 663                 break;
 664         default:
 665                 UNWINDER_BUG();
 666         }
 667
 668         reg = dwarf_frame_reg(frame, DWARF_ARCH_RA_REG);
 669
 670         /*
 671          * If we haven't seen the return address register or the return
 672          * address column is undefined then we must assume that this is
 673          * the end of the callstack.
 674          */
 675         if (!reg || reg->flags == DWARF_UNDEFINED)
 676                 goto bail;
 677
 678         UNWINDER_BUG_ON(reg->flags != DWARF_REG_OFFSET);
 679
 680         addr = frame->cfa + reg->addr;
 681         frame->return_addr = __raw_readl(addr);
 682
 683         /*
 684          * Ah, the joys of unwinding through interrupts.
 685          *
 686          * Interrupts are tricky - the DWARF info needs to be _really_
 687          * accurate and unfortunately I'm seeing a lot of bogus DWARF
 688          * info. For example, I've seen interrupts occur in epilogues
 689          * just after the frame pointer (r14) had been restored. The
 690          * problem was that the DWARF info claimed that the CFA could be
 691          * reached by using the value of the frame pointer before it was
 692          * restored.
 693          *
 694          * So until the compiler can be trusted to produce reliable
 695          * DWARF info when it really matters, let's stop unwinding once
 696          * we've calculated the function that was interrupted.
 697          */
 698         if (prev && prev->pc == (unsigned long)ret_from_irq)
 699                 frame->return_addr = 0;
 700
 701         return frame;
 702
 703 bail:
 704         dwarf_free_frame(frame);
 705         return NULL;
 706 }
 707
 708 static int dwarf_parse_cie(void *entry, void *p, unsigned long len,
 709                            unsigned char *end, struct module *mod)
 710 {
 711         struct dwarf_cie *cie;
 712         unsigned long flags;
 713         int count;
 714
 715         cie = kzalloc(sizeof(*cie), GFP_KERNEL);
 716         if (!cie)
 717                 return -ENOMEM;
 718
 719         cie->length = len;
 720
 721         /*
 722          * Record the offset into the .eh_frame section
 723          * for this CIE. It allows this CIE to be
 724          * quickly and easily looked up from the
 725          * corresponding FDE.
 726          */
 727         cie->cie_pointer = (unsigned long)entry;
 728
 729         cie->version = *(char *)p++;
 730         UNWINDER_BUG_ON(cie->version != 1);
 731
 732         cie->augmentation = p;
 733         p += strlen(cie->augmentation) + 1;
 734
 735         count = dwarf_read_uleb128(p, &cie->code_alignment_factor);
 736         p += count;
 737
 738         count = dwarf_read_leb128(p, &cie->data_alignment_factor);
 739         p += count;
 740
 741         /*
 742          * Which column in the rule table contains the
 743          * return address?
 744          */
 745         if (cie->version == 1) {
 746                 cie->return_address_reg = __raw_readb(p);
 747                 p++;
 748         } else {
 749                 count = dwarf_read_uleb128(p, &cie->return_address_reg);
 750                 p += count;
 751         }
 752
 753         if (cie->augmentation[0] == 'z') {
 754                 unsigned int length, count;
 755                 cie->flags |= DWARF_CIE_Z_AUGMENTATION;
 756
 757                 count = dwarf_read_uleb128(p, &length);
 758                 p += count;
 759
 760                 UNWINDER_BUG_ON((unsigned char *)p > end);
 761
 762                 cie->initial_instructions = p + length;
 763                 cie->augmentation++;
 764         }
 765
 766         while (*cie->augmentation) {
 767                 /*
 768                  * "L" indicates a byte showing how the
 769                  * LSDA pointer is encoded. Skip it.
 770                  */
 771                 if (*cie->augmentation == 'L') {
 772                         p++;
 773                         cie->augmentation++;
 774                 } else if (*cie->augmentation == 'R') {
 775                         /*
 776                          * "R" indicates a byte showing
 777                          * how FDE addresses are
 778                          * encoded.
 779                          */
 780                         cie->encoding = *(char *)p++;
 781                         cie->augmentation++;
 782                 } else if (*cie->augmentation == 'P') {
 783                         /*
 784                          * "R" indicates a personality
 785                          * routine in the CIE
 786                          * augmentation.
 787                          */
 788                         UNWINDER_BUG();
 789                 } else if (*cie->augmentation == 'S') {
 790                         UNWINDER_BUG();
 791                 } else {
 792                         /*
 793                          * Unknown augmentation. Assume
 794                          * 'z' augmentation.
 795                          */
 796                         p = cie->initial_instructions;
 797                         UNWINDER_BUG_ON(!p);
 798                         break;
 799                 }
 800         }
 801
 802         cie->initial_instructions = p;
 803         cie->instructions_end = end;
 804
 805         cie->mod = mod;
 806
 807         /* Add to list */
 808         spin_lock_irqsave(&dwarf_cie_lock, flags);
 809         list_add_tail(&cie->link, &dwarf_cie_list);
 810         spin_unlock_irqrestore(&dwarf_cie_lock, flags);
 811
 812         return 0;
 813 }
 814
 815 static int dwarf_parse_fde(void *entry, u32 entry_type,
 816                            void *start, unsigned long len,
 817                            unsigned char *end, struct module *mod)
 818 {
 819         struct dwarf_fde *fde;
 820         struct dwarf_cie *cie;
 821         unsigned long flags;
 822         int count;
 823         void *p = start;
 824
 825         fde = kzalloc(sizeof(*fde), GFP_KERNEL);
 826         if (!fde)
 827                 return -ENOMEM;
 828
 829         fde->length = len;
 830
 831         /*
 832          * In a .eh_frame section the CIE pointer is the
 833          * delta between the address within the FDE
 834          */
 835         fde->cie_pointer = (unsigned long)(p - entry_type - 4);
 836
 837         cie = dwarf_lookup_cie(fde->cie_pointer);
 838         fde->cie = cie;
 839
 840         if (cie->encoding)
 841                 count = dwarf_read_encoded_value(p, &fde->initial_location,
 842                                                  cie->encoding);
 843         else
 844                 count = dwarf_read_addr(p, &fde->initial_location);
 845
 846         p += count;
 847
 848         if (cie->encoding)
 849                 count = dwarf_read_encoded_value(p, &fde->address_range,
 850                                                  cie->encoding & 0x0f);
 851         else
 852                 count = dwarf_read_addr(p, &fde->address_range);
 853
 854         p += count;
 855
 856         if (fde->cie->flags & DWARF_CIE_Z_AUGMENTATION) {
 857                 unsigned int length;
 858                 count = dwarf_read_uleb128(p, &length);
 859                 p += count + length;
 860         }
 861
 862         /* Call frame instructions. */
 863         fde->instructions = p;
 864         fde->end = end;
 865
 866         fde->mod = mod;
 867
 868         /* Add to list. */
 869         spin_lock_irqsave(&dwarf_fde_lock, flags);
 870         list_add_tail(&fde->link, &dwarf_fde_list);
 871         spin_unlock_irqrestore(&dwarf_fde_lock, flags);
 872
 873         return 0;
 874 }
 875
 876 static void dwarf_unwinder_dump(struct task_struct *task,
 877                                 struct pt_regs *regs,
 878                                 unsigned long *sp,
 879                                 const struct stacktrace_ops *ops,
 880                                 void *data)
 881 {
 882         struct dwarf_frame *frame, *_frame;
 883         unsigned long return_addr;
 884
 885         _frame = NULL;
 886         return_addr = 0;
 887
 888         while (1) {
 889                 frame = dwarf_unwind_stack(return_addr, _frame);
 890
 891                 if (_frame)
 892                         dwarf_free_frame(_frame);
 893
 894                 _frame = frame;
 895
 896                 if (!frame || !frame->return_addr)
 897                         break;
 898
 899                 return_addr = frame->return_addr;
 900                 ops->address(data, return_addr, 1);
 901         }
 902
 903         if (frame)
 904                 dwarf_free_frame(frame);
 905 }
 906
 907 static struct unwinder dwarf_unwinder = {
 908         .name = "dwarf-unwinder",
 909         .dump = dwarf_unwinder_dump,
 910         .rating = 150,
 911 };
 912
 913 static void dwarf_unwinder_cleanup(void)
 914 {
 915         struct dwarf_cie *cie, *cie_tmp;
 916         struct dwarf_fde *fde, *fde_tmp;
 917
 918         /*
 919          * Deallocate all the memory allocated for the DWARF unwinder.
 920          * Traverse all the FDE/CIE lists and remove and free all the
 921          * memory associated with those data structures.
 922          */
 923         list_for_each_entry_safe(cie, cie_tmp, &dwarf_cie_list, link)
 924                 kfree(cie);
 925
 926         list_for_each_entry_safe(fde, fde_tmp, &dwarf_fde_list, link)
 927                 kfree(fde);
 928
 929         kmem_cache_destroy(dwarf_reg_cachep);
 930         kmem_cache_destroy(dwarf_frame_cachep);
 931 }
 932
 933 /**
 934  *      dwarf_parse_section - parse DWARF section
 935  *      @eh_frame_start: start address of the .eh_frame section
 936  *      @eh_frame_end: end address of the .eh_frame section
 937  *      @mod: the kernel module containing the .eh_frame section
 938  *
 939  *      Parse the information in a .eh_frame section.
 940  */
 941 static int dwarf_parse_section(char *eh_frame_start, char *eh_frame_end,
 942                                struct module *mod)
 943 {
 944         u32 entry_type;
 945         void *p, *entry;
 946         int count, err = 0;
 947         unsigned long len = 0;
 948         unsigned int c_entries, f_entries;
 949         unsigned char *end;
 950
 951         c_entries = 0;
 952         f_entries = 0;
 953         entry = eh_frame_start;
 954
 955         while ((char *)entry < eh_frame_end) {
 956                 p = entry;
 957
 958                 count = dwarf_entry_len(p, &len);
 959                 if (count == 0) {
 960                         /*
 961                          * We read a bogus length field value. There is
 962                          * nothing we can do here apart from disabling
 963                          * the DWARF unwinder. We can't even skip this
 964                          * entry and move to the next one because 'len'
 965                          * tells us where our next entry is.
 966                          */
 967                         err = -EINVAL;
 968                         goto out;
 969                 } else
 970                         p += count;
 971
 972                 /* initial length does not include itself */
 973                 end = p + len;
 974
 975                 entry_type = get_unaligned((u32 *)p);
 976                 p += 4;
 977
 978                 if (entry_type == DW_EH_FRAME_CIE) {
 979                         err = dwarf_parse_cie(entry, p, len, end, mod);
 980                         if (err < 0)
 981                                 goto out;
 982                         else
 983                                 c_entries++;
 984                 } else {
 985                         err = dwarf_parse_fde(entry, entry_type, p, len,
 986                                               end, mod);
 987                         if (err < 0)
 988                                 goto out;
 989                         else
 990                                 f_entries++;
 991                 }
 992
 993                 entry = (char *)entry + len + 4;
 994         }
 995
 996         printk(KERN_INFO "DWARF unwinder initialised: read %u CIEs, %u FDEs\n",
 997                c_entries, f_entries);
 998
 999         return 0;
1000
1001 out:
1002         return err;
1003 }
1004
1005 #ifdef CONFIG_MODULES
1006 int module_dwarf_finalize(const Elf_Ehdr *hdr, const Elf_Shdr *sechdrs,
1007                           struct module *me)
1008 {
1009         unsigned int i, err;
1010         unsigned long start, end;
1011         char *secstrings = (void *)hdr + sechdrs[hdr->e_shstrndx].sh_offset;
1012
1013         start = end = 0;
1014
1015         for (i = 1; i < hdr->e_shnum; i++) {
1016                 /* Alloc bit cleared means "ignore it." */
1017                 if ((sechdrs[i].sh_flags & SHF_ALLOC)
1018                     && !strcmp(secstrings+sechdrs[i].sh_name, ".eh_frame")) {
1019                         start = sechdrs[i].sh_addr;
1020                         end = start + sechdrs[i].sh_size;
1021                         break;
1022                 }
1023         }
1024
1025         /* Did we find the .eh_frame section? */
1026         if (i != hdr->e_shnum) {
1027                 err = dwarf_parse_section((char *)start, (char *)end, me);
1028                 if (err) {
1029                         printk(KERN_WARNING "%s: failed to parse DWARF info\n",
1030                                me->name);
1031                         return err;
1032                 }
1033         }
1034
1035         return 0;
1036 }
1037
1038 /**
1039  *      module_dwarf_cleanup - remove FDE/CIEs associated with @mod
1040  *      @mod: the module that is being unloaded
1041  *
1042  *      Remove any FDEs and CIEs from the global lists that came from
1043  *      @mod's .eh_frame section because @mod is being unloaded.
1044  */
1045 void module_dwarf_cleanup(struct module *mod)
1046 {
1047         struct dwarf_fde *fde;
1048         struct dwarf_cie *cie;
1049         unsigned long flags;
1050
1051         spin_lock_irqsave(&dwarf_cie_lock, flags);
1052
1053 again_cie:
1054         list_for_each_entry(cie, &dwarf_cie_list, link) {
1055                 if (cie->mod == mod)
1056                         break;
1057         }
1058
1059         if (&cie->link != &dwarf_cie_list) {
1060                 list_del(&cie->link);
1061                 kfree(cie);
1062                 goto again_cie;
1063         }
1064
1065         spin_unlock_irqrestore(&dwarf_cie_lock, flags);
1066
1067         spin_lock_irqsave(&dwarf_fde_lock, flags);
1068
1069 again_fde:
1070         list_for_each_entry(fde, &dwarf_fde_list, link) {
1071                 if (fde->mod == mod)
1072                         break;
1073         }
1074
1075         if (&fde->link != &dwarf_fde_list) {
1076                 list_del(&fde->link);
1077                 kfree(fde);
1078                 goto again_fde;
1079         }
1080
1081         spin_unlock_irqrestore(&dwarf_fde_lock, flags);
1082 }
1083 #endif /* CONFIG_MODULES */
1084
1085 /**
1086  *      dwarf_unwinder_init - initialise the dwarf unwinder
1087  *
1088  *      Build the data structures describing the .dwarf_frame section to
1089  *      make it easier to lookup CIE and FDE entries. Because the
1090  *      .eh_frame section is packed as tightly as possible it is not
1091  *      easy to lookup the FDE for a given PC, so we build a list of FDE
1092  *      and CIE entries that make it easier.
1093  */
1094 static int __init dwarf_unwinder_init(void)
1095 {
1096         int err;
1097         INIT_LIST_HEAD(&dwarf_cie_list);
1098         INIT_LIST_HEAD(&dwarf_fde_list);
1099
1100         dwarf_frame_cachep = kmem_cache_create("dwarf_frames",
1101                         sizeof(struct dwarf_frame), 0,
1102                         SLAB_PANIC | SLAB_HWCACHE_ALIGN | SLAB_NOTRACK, NULL);
1103
1104         dwarf_reg_cachep = kmem_cache_create("dwarf_regs",
1105                         sizeof(struct dwarf_reg), 0,
1106                         SLAB_PANIC | SLAB_HWCACHE_ALIGN | SLAB_NOTRACK, NULL);
1107
1108         dwarf_frame_pool = mempool_create(DWARF_FRAME_MIN_REQ,
1109                                           mempool_alloc_slab,
1110                                           mempool_free_slab,
1111                                           dwarf_frame_cachep);
1112
1113         dwarf_reg_pool = mempool_create(DWARF_REG_MIN_REQ,
1114                                          mempool_alloc_slab,
1115                                          mempool_free_slab,
1116                                          dwarf_reg_cachep);
1117
1118         err = dwarf_parse_section(__start_eh_frame, __stop_eh_frame, NULL);
1119         if (err)
1120                 goto out;
1121
1122         err = unwinder_register(&dwarf_unwinder);
1123         if (err)
1124                 goto out;
1125
1126         return 0;
1127
1128 out:
1129         printk(KERN_ERR "Failed to initialise DWARF unwinder: %d\n", err);
1130         dwarf_unwinder_cleanup();
1131         return -EINVAL;
1132 }
1133 early_initcall(dwarf_unwinder_init);