src/runtime/coreparse.c

   1 /*
   2  * A saved SBCL system is a .core file; the code here helps us accept
   3  * such a file as input.
   4  */
   5
   6 /*
   7  * This software is part of the SBCL system. See the README file for
   8  * more information.
   9  *
  10  * This software is derived from the CMU CL system, which was
  11  * written at Carnegie Mellon University and released into the
  12  * public domain. The software is in the public domain and is
  13  * provided with absolutely no warranty. See the COPYING and CREDITS
  14  * files for more information.
  15  */
  16
  17 #include <stdio.h>
  18 #include <stdlib.h>
  19 #include <string.h>
  20 #include <sys/file.h>
  21 #include <sys/types.h>
  22 #include <sys/stat.h>
  23 #include <fcntl.h>
  24 #include <unistd.h>
  25
  26 #include "sbcl.h"
  27 #include "os.h"
  28 #include "runtime.h"
  29 #include "globals.h"
  30 #include "core.h"
  31 #include "arch.h"
  32 #include "interr.h"
  33 #include "thread.h"
  34
  35 #include "validate.h"
  36 #include "gc-internal.h"
  37
  38 /* lutex stuff */
  39 #if defined(LISP_FEATURE_SB_THREAD) && defined(LISP_FEATURE_SB_LUTEX)
  40 #include "genesis/sap.h"
  41 #include "pthread-lutex.h"
  42 #endif
  43
  44
  45 unsigned char build_id[] =
  46 #include "../../output/build-id.tmp"
  47 ;
  48
  49 int
  50 open_binary(char *filename, int mode)
  51 {
  52 #ifdef LISP_FEATURE_WIN32
  53     mode |= O_BINARY;
  54 #endif
  55
  56     return open(filename, mode);
  57 }
  58
  59
  60 static struct runtime_options *
  61 read_runtime_options(int fd)
  62 {
  63     size_t optarray[RUNTIME_OPTIONS_WORDS];
  64     struct runtime_options *options = NULL;
  65
  66     if (read(fd, optarray, RUNTIME_OPTIONS_WORDS * sizeof(size_t)) !=
  67         RUNTIME_OPTIONS_WORDS * sizeof(size_t)) {
  68         return NULL;
  69     }
  70
  71     if ((RUNTIME_OPTIONS_MAGIC != optarray[0]) || (0 == optarray[1])) {
  72         return NULL;
  73     }
  74
  75     options = successful_malloc(sizeof(struct runtime_options));
  76
  77     options->dynamic_space_size = optarray[2];
  78     options->thread_control_stack_size = optarray[3];
  79
  80     return options;
  81 }
  82
  83 void
  84 maybe_initialize_runtime_options(int fd)
  85 {
  86     struct runtime_options *new_runtime_options;
  87     off_t end_offset = sizeof(lispobj) +
  88         sizeof(os_vm_offset_t) +
  89         (RUNTIME_OPTIONS_WORDS * sizeof(size_t));
  90
  91     lseek(fd, -end_offset, SEEK_END);
  92
  93     if (new_runtime_options = read_runtime_options(fd)) {
  94         runtime_options = new_runtime_options;
  95     }
  96 }
  97
  98 /* Search 'filename' for an embedded core.  An SBCL core has, at the
  99  * end of the file, a trailer containing optional saved runtime
 100  * options, the start of the core (an os_vm_offset_t), and a final
 101  * signature word (the lispobj CORE_MAGIC).  If this trailer is found
 102  * at the end of the file, the start of the core can be determined
 103  * from the core size.
 104  *
 105  * If an embedded core is present, this returns the offset into the
 106  * file to load the core from, or -1 if no core is present. */
 107 os_vm_offset_t
 108 search_for_embedded_core(char *filename)
 109 {
 110     lispobj header;
 111     os_vm_offset_t lispobj_size = sizeof(lispobj);
 112     os_vm_offset_t trailer_size = lispobj_size + sizeof(os_vm_offset_t);
 113     os_vm_offset_t core_start, pos;
 114     int fd = -1;
 115
 116     if ((fd = open_binary(filename, O_RDONLY)) < 0)
 117         goto lose;
 118
 119     if (read(fd, &header, (size_t)lispobj_size) < lispobj_size)
 120         goto lose;
 121     if (header == CORE_MAGIC) {
 122         /* This file is a real core, not an embedded core.  Return 0 to
 123          * indicate where the core starts, and do not look for runtime
 124          * options in this case. */
 125         return 0;
 126     }
 127
 128     if (lseek(fd, -lispobj_size, SEEK_END) < 0)
 129         goto lose;
 130     if (read(fd, &header, (size_t)lispobj_size) < lispobj_size)
 131         goto lose;
 132
 133     if (header == CORE_MAGIC) {
 134         if (lseek(fd, -trailer_size, SEEK_END) < 0)
 135             goto lose;
 136         if (read(fd, &core_start, sizeof(os_vm_offset_t)) < 0)
 137             goto lose;
 138
 139         if (lseek(fd, core_start, SEEK_SET) < 0)
 140             goto lose;
 141         pos = lseek(fd, 0, SEEK_CUR);
 142
 143         if (read(fd, &header, (size_t)lispobj_size) < lispobj_size)
 144             goto lose;
 145
 146         if (header != CORE_MAGIC)
 147             goto lose;
 148
 149         maybe_initialize_runtime_options(fd);
 150
 151         close(fd);
 152         return pos;
 153     }
 154
 155 lose:
 156     if (fd != -1)
 157         close(fd);
 158
 159     return -1;
 160 }
 161
 162 /* If more platforms doesn't support overlapping mmap rename this
 163  * def to something like ifdef nommapoverlap */
 164 /* currently hpux only */
 165 #ifdef LISP_FEATURE_HPUX
 166 os_vm_address_t copy_core_bytes(int fd, os_vm_offset_t offset,
 167                                 os_vm_address_t addr, int len)
 168 {
 169   unsigned char buf[4096];
 170   int c,x;
 171   int old_fd = lseek(fd, 0, SEEK_CUR);
 172
 173   if(len & (4096-1)){
 174     fprintf(stderr, "cant copy a slice of core because slice-length is not of page size(4096)\n");
 175     exit(-1);
 176   }
 177   if(old_fd < 0){
 178     fprintf(stderr, "cant perform lseek() on corefile\n");
 179   }
 180   lseek(fd, offset, SEEK_SET);
 181   if(fd < 0){
 182     fprintf(stderr, "cant perform lseek(%u,%lu,SEEK_SET) on corefile\n", fd, offset);
 183   }
 184   for(x = 0; x < len; x += 4096){
 185     c = read(fd, buf, 4096);
 186     if(c != 4096){
 187       fprintf(stderr, "cant read memory area from corefile at position %lu, got %d\n", offset + x, c);
 188       exit(-1);
 189     }
 190     memcpy(addr+x, buf, 4096);
 191   }
 192   os_flush_icache(addr, len);
 193   return addr;
 194 }
 195 #endif
 196
 197 static void
 198 process_directory(int fd, lispobj *ptr, int count, os_vm_offset_t file_offset)
 199 {
 200     struct ndir_entry *entry;
 201
 202     FSHOW((stderr, "/process_directory(..), count=%d\n", count));
 203
 204     for (entry = (struct ndir_entry *) ptr; --count>= 0; ++entry) {
 205
 206         long id = entry->identifier;
 207         long offset = os_vm_page_size * (1 + entry->data_page);
 208         os_vm_address_t addr =
 209             (os_vm_address_t) (os_vm_page_size * entry->address);
 210         lispobj *free_pointer = (lispobj *) addr + entry->nwords;
 211         unsigned long len = os_vm_page_size * entry->page_count;
 212         if (len != 0) {
 213             os_vm_address_t real_addr;
 214             FSHOW((stderr, "/mapping %ld(0x%lx) bytes at 0x%lx\n",
 215                    (long)len, (long)len, (unsigned long)addr));
 216 #ifdef LISP_FEATURE_HPUX
 217             real_addr = copy_core_bytes(fd, offset + file_offset, addr, len);
 218 #else
 219             real_addr = os_map(fd, offset + file_offset, addr, len);
 220 #endif
 221             if (real_addr != addr) {
 222                 lose("file mapped in wrong place! "
 223                      "(0x%08x != 0x%08lx)\n",
 224                      real_addr,
 225                      addr);
 226             }
 227         }
 228
 229         FSHOW((stderr, "/space id = %ld, free pointer = 0x%lx\n",
 230                id, (unsigned long)free_pointer));
 231
 232         switch (id) {
 233         case DYNAMIC_CORE_SPACE_ID:
 234             if (len > dynamic_space_size) {
 235                 fprintf(stderr,
 236                         "dynamic space too small for core: %ldKiB required, %ldKiB available.\n",
 237                         len >> 10,
 238                         (long)dynamic_space_size >> 10);
 239                 exit(1);
 240             }
 241 #ifdef LISP_FEATURE_GENCGC
 242             if (addr != (os_vm_address_t)DYNAMIC_SPACE_START) {
 243                 fprintf(stderr, "in core: 0x%lx; in runtime: 0x%lx \n",
 244                         (long)addr, (long)DYNAMIC_SPACE_START);
 245                 lose("core/runtime address mismatch: DYNAMIC_SPACE_START\n");
 246             }
 247 #else
 248             if ((addr != (os_vm_address_t)DYNAMIC_0_SPACE_START) &&
 249                 (addr != (os_vm_address_t)DYNAMIC_1_SPACE_START)) {
 250                 fprintf(stderr, "in core: 0x%lx; in runtime: 0x%lx or 0x%lx\n",
 251                         (long)addr,
 252                         (long)DYNAMIC_0_SPACE_START,
 253                         (long)DYNAMIC_1_SPACE_START);
 254                 lose("warning: core/runtime address mismatch: DYNAMIC_SPACE_START\n");
 255             }
 256 #endif
 257 #if defined(ALLOCATION_POINTER)
 258             SetSymbolValue(ALLOCATION_POINTER, (lispobj)free_pointer,0);
 259 #else
 260             dynamic_space_free_pointer = free_pointer;
 261 #endif
 262             /* For stop-and-copy GC, this will be whatever the GC was
 263              * using at the time. With GENCGC, this will always be
 264              * space 0. (We checked above that for GENCGC,
 265              * addr==DYNAMIC_SPACE_START.) */
 266             current_dynamic_space = (lispobj *)addr;
 267             break;
 268         case STATIC_CORE_SPACE_ID:
 269             if (addr != (os_vm_address_t)STATIC_SPACE_START) {
 270                 fprintf(stderr, "in core: 0x%lx - in runtime: 0x%lx\n",
 271                         (long)addr, (long)STATIC_SPACE_START);
 272                 lose("core/runtime address mismatch: STATIC_SPACE_START\n");
 273             }
 274             break;
 275         case READ_ONLY_CORE_SPACE_ID:
 276             if (addr != (os_vm_address_t)READ_ONLY_SPACE_START) {
 277                 fprintf(stderr, "in core: 0x%lx - in runtime: 0x%lx\n",
 278                         (long)addr, (long)READ_ONLY_SPACE_START);
 279                 lose("core/runtime address mismatch: READ_ONLY_SPACE_START\n");
 280             }
 281             break;
 282         default:
 283             lose("unknown space ID %ld addr 0x%lx\n", id, (long)addr);
 284         }
 285     }
 286 }
 287
 288 lispobj
 289 load_core_file(char *file, os_vm_offset_t file_offset)
 290 {
 291     lispobj *header, val, len, *ptr, remaining_len;
 292     int fd = open_binary(file, O_RDONLY);
 293     unsigned int count;
 294
 295     lispobj initial_function = NIL;
 296     FSHOW((stderr, "/entering load_core_file(%s)\n", file));
 297     if (fd < 0) {
 298         fprintf(stderr, "could not open file \"%s\"\n", file);
 299         perror("open");
 300         exit(1);
 301     }
 302
 303     lseek(fd, file_offset, SEEK_SET);
 304     header = calloc(os_vm_page_size / sizeof(u32), sizeof(u32));
 305
 306     count = read(fd, header, os_vm_page_size);
 307     if (count < os_vm_page_size) {
 308         lose("premature end of core file\n");
 309     }
 310     SHOW("successfully read first page of core");
 311
 312     ptr = header;
 313     val = *ptr++;
 314
 315     if (val != CORE_MAGIC) {
 316         lose("invalid magic number in core: 0x%lx should have been 0x%x.\n",
 317              val,
 318              CORE_MAGIC);
 319     }
 320     SHOW("found CORE_MAGIC");
 321
 322     while (val != END_CORE_ENTRY_TYPE_CODE) {
 323         val = *ptr++;
 324         len = *ptr++;
 325         remaining_len = len - 2; /* (-2 to cancel the two ++ operations) */
 326         FSHOW((stderr, "/val=0x%ld, remaining_len=0x%ld\n",
 327                (long)val, (long)remaining_len));
 328
 329         switch (val) {
 330
 331         case END_CORE_ENTRY_TYPE_CODE:
 332             SHOW("END_CORE_ENTRY_TYPE_CODE case");
 333             break;
 334
 335         case VERSION_CORE_ENTRY_TYPE_CODE:
 336             SHOW("VERSION_CORE_ENTRY_TYPE_CODE case");
 337             if (*ptr != SBCL_CORE_VERSION_INTEGER) {
 338                 lose("core file version (%d) != runtime library version (%d)\n",
 339                      *ptr,
 340                      SBCL_CORE_VERSION_INTEGER);
 341             }
 342             break;
 343
 344         case BUILD_ID_CORE_ENTRY_TYPE_CODE:
 345             SHOW("BUILD_ID_CORE_ENTRY_TYPE_CODE case");
 346             {
 347                 unsigned int i;
 348
 349                 FSHOW((stderr, "build_id[]=\"%s\"\n", build_id));
 350                 FSHOW((stderr, "remaining_len = %d\n", remaining_len));
 351                 if (remaining_len != strlen((const char *)build_id))
 352                     goto losing_build_id;
 353                 for (i = 0; i < remaining_len; ++i) {
 354                     FSHOW((stderr, "ptr[%d] = char = %d, expected=%d\n",
 355                            i, ptr[i], build_id[i]));
 356                     if (ptr[i] != build_id[i])
 357                         goto losing_build_id;
 358                 }
 359                 break;
 360             losing_build_id:
 361                 /* .core files are not binary-compatible between
 362                  * builds because we can't easily detect whether the
 363                  * sources were patched between the time the
 364                  * dumping-the-.core runtime was built and the time
 365                  * that the loading-the-.core runtime was built.
 366                  *
 367                  * (We could easily detect whether version.lisp-expr
 368                  * was changed, but people experimenting with patches
 369                  * don't necessarily update version.lisp-expr.) */
 370
 371                 lose("can't load .core for different runtime, sorry\n");
 372             }
 373
 374         case NEW_DIRECTORY_CORE_ENTRY_TYPE_CODE:
 375             SHOW("NEW_DIRECTORY_CORE_ENTRY_TYPE_CODE case");
 376             process_directory(fd,
 377                               ptr,
 378 #ifndef LISP_FEATURE_ALPHA
 379                               remaining_len / (sizeof(struct ndir_entry) /
 380                                                sizeof(long)),
 381 #else
 382                               remaining_len / (sizeof(struct ndir_entry) /
 383                                                sizeof(u32)),
 384 #endif
 385                               file_offset);
 386             break;
 387
 388         case INITIAL_FUN_CORE_ENTRY_TYPE_CODE:
 389             SHOW("INITIAL_FUN_CORE_ENTRY_TYPE_CODE case");
 390             initial_function = (lispobj)*ptr;
 391             break;
 392
 393 #if defined(LISP_FEATURE_SB_THREAD) && defined(LISP_FEATURE_SB_LUTEX)
 394         case LUTEX_TABLE_CORE_ENTRY_TYPE_CODE:
 395             SHOW("LUTEX_TABLE_CORE_ENTRY_TYPE_CODE case");
 396             {
 397                 size_t n_lutexes = *ptr;
 398                 size_t fdoffset = (*(ptr + 1) + 1) * (os_vm_page_size);
 399                 size_t data_length = n_lutexes * sizeof(struct sap *);
 400                 struct lutex **lutexes_to_resurrect = malloc(data_length);
 401                 long bytes_read;
 402
 403                 lseek(fd, fdoffset + file_offset, SEEK_SET);
 404
 405                 FSHOW((stderr, "attempting to read %ld lutexes from core\n", n_lutexes));
 406                 bytes_read = read(fd, lutexes_to_resurrect, data_length);
 407
 408                 /* XXX */
 409                 if (bytes_read != data_length) {
 410                     lose("Could not read the lutex table");
 411                 }
 412                 else {
 413                     int i;
 414
 415                     for (i=0; i<n_lutexes; ++i) {
 416                         struct lutex *lutex = lutexes_to_resurrect[i];
 417
 418                         FSHOW((stderr, "re-init'ing lutex @ %p\n", lutex));
 419                         lutex_init((tagged_lutex_t) lutex);
 420                     }
 421
 422                     free(lutexes_to_resurrect);
 423                 }
 424                 break;
 425             }
 426 #endif
 427
 428 #ifdef LISP_FEATURE_GENCGC
 429         case PAGE_TABLE_CORE_ENTRY_TYPE_CODE:
 430         {
 431             size_t size = *ptr;
 432             size_t fdoffset = (*(ptr+1) + 1) * (os_vm_page_size);
 433             size_t offset = 0;
 434             long bytes_read;
 435             unsigned long data[4096];
 436             unsigned long word;
 437             lseek(fd, fdoffset + file_offset, SEEK_SET);
 438             while ((bytes_read = read(fd, data, (size < 4096 ? size : 4096 )))
 439                     > 0)
 440             {
 441                 int i = 0;
 442                 size -= bytes_read;
 443                 while (bytes_read) {
 444                     bytes_read -= sizeof(long);
 445                     /* Ignore all zeroes. The size of the page table
 446                      * core entry was rounded up to os_vm_page_size
 447                      * during the save, and might now have more
 448                      * elements than the page table.
 449                      *
 450                      * The low bits of each word are allocation flags.
 451                      */
 452                     if (word=data[i]) {
 453                         page_table[offset].region_start_offset = word & ~0x03;
 454                         page_table[offset].allocated = word & 0x03;
 455                     }
 456                     i++;
 457                     offset++;
 458                 }
 459             }
 460
 461             gencgc_partial_pickup = 1;
 462             break;
 463         }
 464 #endif
 465         default:
 466             lose("unknown core file entry: %ld\n", (long)val);
 467         }
 468
 469         ptr += remaining_len;
 470         FSHOW((stderr, "/new ptr=%lx\n", (unsigned long)ptr));
 471     }
 472     SHOW("about to free(header)");
 473     free(header);
 474     SHOW("returning from load_core_file(..)");
 475     return initial_function;
 476 }
 477