util.c

   1 /*
   2  * mdadm - manage Linux "md" devices aka RAID arrays.
   3  *
   4  * Copyright (C) 2001-2009 Neil Brown <neilb@suse.de>
   5  *
   6  *
   7  *    This program is free software; you can redistribute it and/or modify
   8  *    it under the terms of the GNU General Public License as published by
   9  *    the Free Software Foundation; either version 2 of the License, or
  10  *    (at your option) any later version.
  11  *
  12  *    This program is distributed in the hope that it will be useful,
  13  *    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *    GNU General Public License for more details.
  16  *
  17  *    You should have received a copy of the GNU General Public License
  18  *    along with this program; if not, write to the Free Software
  19  *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  20  *
  21  *    Author: Neil Brown
  22  *    Email: <neilb@suse.de>
  23  */
  24
  25 #include        "mdadm.h"
  26 #include        "md_p.h"
  27 #include        <sys/socket.h>
  28 #include        <sys/utsname.h>
  29 #include        <sys/wait.h>
  30 #include        <sys/un.h>
  31 #include        <ctype.h>
  32 #include        <dirent.h>
  33 #include        <signal.h>
  34
  35 /*
  36  * following taken from linux/blkpg.h because they aren't
  37  * anywhere else and it isn't safe to #include linux/ * stuff.
  38  */
  39
  40 #define BLKPG      _IO(0x12,105)
  41
  42 /* The argument structure */
  43 struct blkpg_ioctl_arg {
  44         int op;
  45         int flags;
  46         int datalen;
  47         void *data;
  48 };
  49
  50 /* The subfunctions (for the op field) */
  51 #define BLKPG_ADD_PARTITION     1
  52 #define BLKPG_DEL_PARTITION     2
  53
  54 /* Sizes of name fields. Unused at present. */
  55 #define BLKPG_DEVNAMELTH        64
  56 #define BLKPG_VOLNAMELTH        64
  57
  58 /* The data structure for ADD_PARTITION and DEL_PARTITION */
  59 struct blkpg_partition {
  60         long long start;                /* starting offset in bytes */
  61         long long length;               /* length in bytes */
  62         int pno;                        /* partition number */
  63         char devname[BLKPG_DEVNAMELTH]; /* partition name, like sda5 or c0d1p2,
  64                                            to be used in kernel messages */
  65         char volname[BLKPG_VOLNAMELTH]; /* volume label */
  66 };
  67
  68 /*
  69  * Parse a 128 bit uuid in 4 integers
  70  * format is 32 hexx nibbles with options :.<space> separator
  71  * If not exactly 32 hex digits are found, return 0
  72  * else return 1
  73  */
  74 int parse_uuid(char *str, int uuid[4])
  75 {
  76         int hit = 0; /* number of Hex digIT */
  77         int i;
  78         char c;
  79         for (i=0; i<4; i++) uuid[i]=0;
  80
  81         while ((c= *str++)) {
  82                 int n;
  83                 if (c>='0' && c<='9')
  84                         n = c-'0';
  85                 else if (c>='a' && c <= 'f')
  86                         n = 10 + c - 'a';
  87                 else if (c>='A' && c <= 'F')
  88                         n = 10 + c - 'A';
  89                 else if (strchr(":. -", c))
  90                         continue;
  91                 else return 0;
  92
  93                 if (hit<32) {
  94                         uuid[hit/8] <<= 4;
  95                         uuid[hit/8] += n;
  96                 }
  97                 hit++;
  98         }
  99         if (hit == 32)
 100                 return 1;
 101         return 0;
 102 }
 103
 104
 105 /*
 106  * Get the md version number.
 107  * We use the RAID_VERSION ioctl if it is supported
 108  * If not, but we have a block device with major '9', we assume
 109  * 0.36.0
 110  *
 111  * Return version number as 24 but number - assume version parts
 112  * always < 255
 113  */
 114
 115 int md_get_version(int fd)
 116 {
 117     struct stat stb;
 118     mdu_version_t vers;
 119
 120     if (fstat(fd, &stb)<0)
 121         return -1;
 122     if ((S_IFMT&stb.st_mode) != S_IFBLK)
 123         return -1;
 124
 125     if (ioctl(fd, RAID_VERSION, &vers) == 0)
 126         return  (vers.major*10000) + (vers.minor*100) + vers.patchlevel;
 127     if (errno == EACCES)
 128             return -1;
 129     if (major(stb.st_rdev) == MD_MAJOR)
 130         return (3600);
 131     return -1;
 132 }
 133
 134 int get_linux_version()
 135 {
 136         struct utsname name;
 137         char *cp;
 138         int a,b,c;
 139         if (uname(&name) <0)
 140                 return -1;
 141
 142         cp = name.release;
 143         a = strtoul(cp, &cp, 10);
 144         if (*cp != '.') return -1;
 145         b = strtoul(cp+1, &cp, 10);
 146         if (*cp != '.') return -1;
 147         c = strtoul(cp+1, NULL, 10);
 148
 149         return (a*1000000)+(b*1000)+c;
 150 }
 151
 152 long long parse_size(char *size)
 153 {
 154         /* parse 'size' which should be a number optionally
 155          * followed by 'K', 'M', or 'G'.
 156          * Without a suffix, K is assumed.
 157          * Number returned is in sectors (half-K)
 158          */
 159         char *c;
 160         long long s = strtoll(size, &c, 10);
 161         if (s > 0) {
 162                 switch (*c) {
 163                 case 'K':
 164                         c++;
 165                 default:
 166                         s *= 2;
 167                         break;
 168                 case 'M':
 169                         c++;
 170                         s *= 1024 * 2;
 171                         break;
 172                 case 'G':
 173                         c++;
 174                         s *= 1024 * 1024 * 2;
 175                         break;
 176                 }
 177         }
 178         if (*c)
 179                 s = 0;
 180         return s;
 181 }
 182
 183 int parse_layout_10(char *layout)
 184 {
 185         int copies, rv;
 186         char *cp;
 187         /* Parse the layout string for raid10 */
 188         /* 'f', 'o' or 'n' followed by a number <= raid_disks */
 189         if ((layout[0] !=  'n' && layout[0] != 'f' && layout[0] != 'o') ||
 190             (copies = strtoul(layout+1, &cp, 10)) < 1 ||
 191             copies > 200 ||
 192             *cp)
 193                 return -1;
 194         if (layout[0] == 'n')
 195                 rv = 256 + copies;
 196         else if (layout[0] == 'o')
 197                 rv = 0x10000 + (copies<<8) + 1;
 198         else
 199                 rv = 1 + (copies<<8);
 200         return rv;
 201 }
 202
 203 int parse_layout_faulty(char *layout)
 204 {
 205         /* Parse the layout string for 'faulty' */
 206         int ln = strcspn(layout, "0123456789");
 207         char *m = strdup(layout);
 208         int mode;
 209         m[ln] = 0;
 210         mode = map_name(faultylayout, m);
 211         if (mode == UnSet)
 212                 return -1;
 213
 214         return mode | (atoi(layout+ln)<< ModeShift);
 215 }
 216
 217 void remove_partitions(int fd)
 218 {
 219         /* remove partitions from this block devices.
 220          * This is used for components added to an array
 221          */
 222 #ifdef BLKPG_DEL_PARTITION
 223         struct blkpg_ioctl_arg a;
 224         struct blkpg_partition p;
 225
 226         a.op = BLKPG_DEL_PARTITION;
 227         a.data = (void*)&p;
 228         a.datalen = sizeof(p);
 229         a.flags = 0;
 230         memset(a.data, 0, a.datalen);
 231         for (p.pno=0; p.pno < 16; p.pno++)
 232                 ioctl(fd, BLKPG, &a);
 233 #endif
 234 }
 235
 236 int enough(int level, int raid_disks, int layout, int clean,
 237            char *avail, int avail_disks)
 238 {
 239         int copies, first;
 240         switch (level) {
 241         case 10:
 242                 /* This is the tricky one - we need to check
 243                  * which actual disks are present.
 244                  */
 245                 copies = (layout&255)* ((layout>>8) & 255);
 246                 first=0;
 247                 do {
 248                         /* there must be one of the 'copies' form 'first' */
 249                         int n = copies;
 250                         int cnt=0;
 251                         while (n--) {
 252                                 if (avail[first])
 253                                         cnt++;
 254                                 first = (first+1) % raid_disks;
 255                         }
 256                         if (cnt == 0)
 257                                 return 0;
 258
 259                 } while (first != 0);
 260                 return 1;
 261
 262         case -4:
 263                 return avail_disks>= 1;
 264         case -1:
 265         case 0:
 266                 return avail_disks == raid_disks;
 267         case 1:
 268                 return avail_disks >= 1;
 269         case 4:
 270         case 5:
 271                 if (clean)
 272                         return avail_disks >= raid_disks-1;
 273                 else
 274                         return avail_disks >= raid_disks;
 275         case 6:
 276                 if (clean)
 277                         return avail_disks >= raid_disks-2;
 278                 else
 279                         return avail_disks >= raid_disks;
 280         default:
 281                 return 0;
 282         }
 283 }
 284
 285 const int uuid_match_any[4] = { ~0, ~0, ~0, ~0 };
 286 int same_uuid(int a[4], int b[4], int swapuuid)
 287 {
 288         if (memcmp(a, uuid_match_any, sizeof(int[4])) == 0 ||
 289             memcmp(b, uuid_match_any, sizeof(int[4])) == 0)
 290                 return 1;
 291
 292         if (swapuuid) {
 293                 /* parse uuids are hostendian.
 294                  * uuid's from some superblocks are big-ending
 295                  * if there is a difference, we need to swap..
 296                  */
 297                 unsigned char *ac = (unsigned char *)a;
 298                 unsigned char *bc = (unsigned char *)b;
 299                 int i;
 300                 for (i=0; i<16; i+= 4) {
 301                         if (ac[i+0] != bc[i+3] ||
 302                             ac[i+1] != bc[i+2] ||
 303                             ac[i+2] != bc[i+1] ||
 304                             ac[i+3] != bc[i+0])
 305                                 return 0;
 306                 }
 307                 return 1;
 308         } else {
 309                 if (a[0]==b[0] &&
 310                     a[1]==b[1] &&
 311                     a[2]==b[2] &&
 312                     a[3]==b[3])
 313                         return 1;
 314                 return 0;
 315         }
 316 }
 317 void copy_uuid(void *a, int b[4], int swapuuid)
 318 {
 319         if (swapuuid) {
 320                 /* parse uuids are hostendian.
 321                  * uuid's from some superblocks are big-ending
 322                  * if there is a difference, we need to swap..
 323                  */
 324                 unsigned char *ac = (unsigned char *)a;
 325                 unsigned char *bc = (unsigned char *)b;
 326                 int i;
 327                 for (i=0; i<16; i+= 4) {
 328                         ac[i+0] = bc[i+3];
 329                         ac[i+1] = bc[i+2];
 330                         ac[i+2] = bc[i+1];
 331                         ac[i+3] = bc[i+0];
 332                 }
 333         } else
 334                 memcpy(a, b, 16);
 335 }
 336
 337 char *fname_from_uuid(struct supertype *st, struct mdinfo *info, char *buf, char sep)
 338 {
 339         int i, j;
 340         int id;
 341         char uuid[16];
 342         char *c = buf;
 343         strcpy(c, "UUID-");
 344         c += strlen(c);
 345         copy_uuid(uuid, info->uuid, st->ss->swapuuid);
 346         for (i = 0; i < 4; i++) {
 347                 id = uuid[i];
 348                 if (i)
 349                         *c++ = sep;
 350                 for (j = 3; j >= 0; j--) {
 351                         sprintf(c,"%02x", (unsigned char) uuid[j+4*i]);
 352                         c+= 2;
 353                 }
 354         }
 355         return buf;
 356 }
 357
 358 #ifndef MDASSEMBLE
 359 int check_ext2(int fd, char *name)
 360 {
 361         /*
 362          * Check for an ext2fs file system.
 363          * Superblock is always 1K at 1K offset
 364          *
 365          * s_magic is le16 at 56 == 0xEF53
 366          * report mtime - le32 at 44
 367          * blocks - le32 at 4
 368          * logblksize - le32 at 24
 369          */
 370         unsigned char sb[1024];
 371         time_t mtime;
 372         int size, bsize;
 373         if (lseek(fd, 1024,0)!= 1024)
 374                 return 0;
 375         if (read(fd, sb, 1024)!= 1024)
 376                 return 0;
 377         if (sb[56] != 0x53 || sb[57] != 0xef)
 378                 return 0;
 379
 380         mtime = sb[44]|(sb[45]|(sb[46]|sb[47]<<8)<<8)<<8;
 381         bsize = sb[24]|(sb[25]|(sb[26]|sb[27]<<8)<<8)<<8;
 382         size = sb[4]|(sb[5]|(sb[6]|sb[7]<<8)<<8)<<8;
 383         fprintf(stderr, Name ": %s appears to contain an ext2fs file system\n",
 384                 name);
 385         fprintf(stderr,"    size=%dK  mtime=%s",
 386                 size*(1<<bsize), ctime(&mtime));
 387         return 1;
 388 }
 389
 390 int check_reiser(int fd, char *name)
 391 {
 392         /*
 393          * superblock is at 64K
 394          * size is 1024;
 395          * Magic string "ReIsErFs" or "ReIsEr2Fs" at 52
 396          *
 397          */
 398         unsigned char sb[1024];
 399         unsigned long size;
 400         if (lseek(fd, 64*1024, 0) != 64*1024)
 401                 return 0;
 402         if (read(fd, sb, 1024) != 1024)
 403                 return 0;
 404         if (strncmp((char*)sb+52, "ReIsErFs",8)!=0 &&
 405             strncmp((char*)sb+52, "ReIsEr2Fs",9)!=0)
 406                 return 0;
 407         fprintf(stderr, Name ": %s appears to contain a reiserfs file system\n",name);
 408         size = sb[0]|(sb[1]|(sb[2]|sb[3]<<8)<<8)<<8;
 409         fprintf(stderr, "    size = %luK\n", size*4);
 410
 411         return 1;
 412 }
 413
 414 int check_raid(int fd, char *name)
 415 {
 416         struct mdinfo info;
 417         time_t crtime;
 418         char *level;
 419         struct supertype *st = guess_super(fd);
 420
 421         if (!st) return 0;
 422         st->ss->load_super(st, fd, name);
 423         /* Looks like a raid array .. */
 424         fprintf(stderr, Name ": %s appears to be part of a raid array:\n",
 425                 name);
 426         st->ss->getinfo_super(st, &info);
 427         st->ss->free_super(st);
 428         crtime = info.array.ctime;
 429         level = map_num(pers, info.array.level);
 430         if (!level) level = "-unknown-";
 431         fprintf(stderr, "    level=%s devices=%d ctime=%s",
 432                 level, info.array.raid_disks, ctime(&crtime));
 433         return 1;
 434 }
 435
 436 int ask(char *mesg)
 437 {
 438         char *add = "";
 439         int i;
 440         for (i=0; i<5; i++) {
 441                 char buf[100];
 442                 fprintf(stderr, "%s%s", mesg, add);
 443                 fflush(stderr);
 444                 if (fgets(buf, 100, stdin)==NULL)
 445                         return 0;
 446                 if (buf[0]=='y' || buf[0]=='Y')
 447                         return 1;
 448                 if (buf[0]=='n' || buf[0]=='N')
 449                         return 0;
 450                 add = "(y/n) ";
 451         }
 452         fprintf(stderr, Name ": assuming 'no'\n");
 453         return 0;
 454 }
 455 #endif /* MDASSEMBLE */
 456
 457 char *map_num(mapping_t *map, int num)
 458 {
 459         while (map->name) {
 460                 if (map->num == num)
 461                         return map->name;
 462                 map++;
 463         }
 464         return NULL;
 465 }
 466
 467 int map_name(mapping_t *map, char *name)
 468 {
 469         while (map->name) {
 470                 if (strcmp(map->name, name)==0)
 471                         return map->num;
 472                 map++;
 473         }
 474         return UnSet;
 475 }
 476
 477
 478 int is_standard(char *dev, int *nump)
 479 {
 480         /* tests if dev is a "standard" md dev name.
 481          * i.e if the last component is "/dNN" or "/mdNN",
 482          * where NN is a string of digits
 483          * Returns 1 if a partitionable standard,
 484          *   -1 if non-partitonable,
 485          *   0 if not a standard name.
 486          */
 487         char *d = strrchr(dev, '/');
 488         int type=0;
 489         int num;
 490         if (!d)
 491                 return 0;
 492         if (strncmp(d, "/d",2)==0)
 493                 d += 2, type=1; /* /dev/md/dN{pM} */
 494         else if (strncmp(d, "/md_d", 5)==0)
 495                 d += 5, type=1; /* /dev/md_dN{pM} */
 496         else if (strncmp(d, "/md", 3)==0)
 497                 d += 3, type=-1; /* /dev/mdN */
 498         else if (d-dev > 3 && strncmp(d-2, "md/", 3)==0)
 499                 d += 1, type=-1; /* /dev/md/N */
 500         else
 501                 return 0;
 502         if (!*d)
 503                 return 0;
 504         num = atoi(d);
 505         while (isdigit(*d))
 506                 d++;
 507         if (*d)
 508                 return 0;
 509         if (nump) *nump = num;
 510
 511         return type;
 512 }
 513
 514
 515 /*
 516  * convert a major/minor pair for a block device into a name in /dev, if possible.
 517  * On the first call, walk /dev collecting name.
 518  * Put them in a simple linked listfor now.
 519  */
 520 struct devmap {
 521     int major, minor;
 522     char *name;
 523     struct devmap *next;
 524 } *devlist = NULL;
 525 int devlist_ready = 0;
 526
 527 int add_dev(const char *name, const struct stat *stb, int flag, struct FTW *s)
 528 {
 529         struct stat st;
 530
 531         if (S_ISLNK(stb->st_mode)) {
 532                 if (stat(name, &st) != 0)
 533                         return 0;
 534                 stb = &st;
 535         }
 536
 537         if ((stb->st_mode&S_IFMT)== S_IFBLK) {
 538                 char *n = strdup(name);
 539                 struct devmap *dm = malloc(sizeof(*dm));
 540                 if (strncmp(n, "/dev/./", 7)==0)
 541                         strcpy(n+4, name+6);
 542                 if (dm) {
 543                         dm->major = major(stb->st_rdev);
 544                         dm->minor = minor(stb->st_rdev);
 545                         dm->name = n;
 546                         dm->next = devlist;
 547                         devlist = dm;
 548                 }
 549         }
 550         return 0;
 551 }
 552
 553 #ifndef HAVE_NFTW
 554 #ifdef HAVE_FTW
 555 int add_dev_1(const char *name, const struct stat *stb, int flag)
 556 {
 557         return add_dev(name, stb, flag, NULL);
 558 }
 559 int nftw(const char *path, int (*han)(const char *name, const struct stat *stb, int flag, struct FTW *s), int nopenfd, int flags)
 560 {
 561         return ftw(path, add_dev_1, nopenfd);
 562 }
 563 #else
 564 int nftw(const char *path, int (*han)(const char *name, const struct stat *stb, int flag, struct FTW *s), int nopenfd, int flags)
 565 {
 566         return 0;
 567 }
 568 #endif /* HAVE_FTW */
 569 #endif /* HAVE_NFTW */
 570
 571 /*
 572  * Find a block device with the right major/minor number.
 573  * If we find multiple names, choose the shortest.
 574  * If we find a name in /dev/md/, we prefer that.
 575  * This applies only to names for MD devices.
 576  */
 577 char *map_dev(int major, int minor, int create)
 578 {
 579         struct devmap *p;
 580         char *regular = NULL, *preferred=NULL;
 581         int did_check = 0;
 582
 583         if (major == 0 && minor == 0)
 584                         return NULL;
 585
 586  retry:
 587         if (!devlist_ready) {
 588                 char *dev = "/dev";
 589                 struct stat stb;
 590                 while(devlist) {
 591                         struct devmap *d = devlist;
 592                         devlist = d->next;
 593                         free(d->name);
 594                         free(d);
 595                 }
 596                 if (lstat(dev, &stb)==0 &&
 597                     S_ISLNK(stb.st_mode))
 598                         dev = "/dev/.";
 599                 nftw(dev, add_dev, 10, FTW_PHYS);
 600                 devlist_ready=1;
 601                 did_check = 1;
 602         }
 603
 604         for (p=devlist; p; p=p->next)
 605                 if (p->major == major &&
 606                     p->minor == minor) {
 607                         if (strncmp(p->name, "/dev/md/",8) == 0) {
 608                                 if (preferred == NULL ||
 609                                     strlen(p->name) < strlen(preferred))
 610                                         preferred = p->name;
 611                         } else {
 612                                 if (regular == NULL ||
 613                                     strlen(p->name) < strlen(regular))
 614                                         regular = p->name;
 615                         }
 616                 }
 617         if (!regular && !preferred && !did_check) {
 618                 devlist_ready = 0;
 619                 goto retry;
 620         }
 621         if (create && !regular && !preferred) {
 622                 static char buf[30];
 623                 snprintf(buf, sizeof(buf), "%d:%d", major, minor);
 624                 regular = buf;
 625         }
 626
 627         return preferred ? preferred : regular;
 628 }
 629
 630 unsigned long calc_csum(void *super, int bytes)
 631 {
 632         unsigned long long newcsum = 0;
 633         int i;
 634         unsigned int csum;
 635         unsigned int *superc = (unsigned int*) super;
 636
 637         for(i=0; i<bytes/4; i++)
 638                 newcsum+= superc[i];
 639         csum = (newcsum& 0xffffffff) + (newcsum>>32);
 640 #ifdef __alpha__
 641 /* The in-kernel checksum calculation is always 16bit on
 642  * the alpha, though it is 32 bit on i386...
 643  * I wonder what it is elsewhere... (it uses and API in
 644  * a way that it shouldn't).
 645  */
 646         csum = (csum & 0xffff) + (csum >> 16);
 647         csum = (csum & 0xffff) + (csum >> 16);
 648 #endif
 649         return csum;
 650 }
 651
 652 #ifndef MDASSEMBLE
 653 char *human_size(long long bytes)
 654 {
 655         static char buf[30];
 656
 657         /* We convert bytes to either centi-M{ega,ibi}bytes or
 658          * centi-G{igi,ibi}bytes, with appropriate rounding,
 659          * and then print 1/100th of those as a decimal.
 660          * We allow upto 2048Megabytes before converting to
 661          * gigabytes, as that shows more precision and isn't
 662          * too large a number.
 663          * Terrabytes are not yet handled.
 664          */
 665
 666         if (bytes < 5000*1024)
 667                 buf[0]=0;
 668         else if (bytes < 2*1024LL*1024LL*1024LL) {
 669                 long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;
 670                 long cMB  = (bytes / ( 1000000LL / 200LL ) +1) /2;
 671                 snprintf(buf, sizeof(buf), " (%ld.%02ld MiB %ld.%02ld MB)",
 672                         cMiB/100 , cMiB % 100,
 673                         cMB/100, cMB % 100);
 674         } else {
 675                 long cGiB = (bytes / ( (1LL<<30) / 200LL ) +1) /2;
 676                 long cGB  = (bytes / (1000000000LL/200LL ) +1) /2;
 677                 snprintf(buf, sizeof(buf), " (%ld.%02ld GiB %ld.%02ld GB)",
 678                         cGiB/100 , cGiB % 100,
 679                         cGB/100, cGB % 100);
 680         }
 681         return buf;
 682 }
 683
 684 char *human_size_brief(long long bytes)
 685 {
 686         static char buf[30];
 687
 688         if (bytes < 5000*1024)
 689                 snprintf(buf, sizeof(buf), "%ld.%02ldKiB",
 690                         (long)(bytes>>10), (long)(((bytes&1023)*100+512)/1024)
 691                         );
 692         else if (bytes < 2*1024LL*1024LL*1024LL)
 693                 snprintf(buf, sizeof(buf), "%ld.%02ldMiB",
 694                         (long)(bytes>>20),
 695                         (long)((bytes&0xfffff)+0x100000/200)/(0x100000/100)
 696                         );
 697         else
 698                 snprintf(buf, sizeof(buf), "%ld.%02ldGiB",
 699                         (long)(bytes>>30),
 700                         (long)(((bytes>>10)&0xfffff)+0x100000/200)/(0x100000/100)
 701                         );
 702         return buf;
 703 }
 704
 705 void print_r10_layout(int layout)
 706 {
 707         int near = layout & 255;
 708         int far = (layout >> 8) & 255;
 709         int offset = (layout&0x10000);
 710         char *sep = "";
 711
 712         if (near != 1) {
 713                 printf("%s near=%d", sep, near);
 714                 sep = ",";
 715         }
 716         if (far != 1)
 717                 printf("%s %s=%d", sep, offset?"offset":"far", far);
 718         if (near*far == 1)
 719                 printf("NO REDUNDANCY");
 720 }
 721 #endif
 722
 723 unsigned long long calc_array_size(int level, int raid_disks, int layout,
 724                                    int chunksize, unsigned long long devsize)
 725 {
 726         int data_disks = 0;
 727         switch (level) {
 728         case 0: data_disks = raid_disks; break;
 729         case 1: data_disks = 1; break;
 730         case 4:
 731         case 5: data_disks = raid_disks - 1; break;
 732         case 6: data_disks = raid_disks - 2; break;
 733         case 10: data_disks = raid_disks / (layout & 255) / ((layout>>8)&255);
 734                 break;
 735         }
 736         devsize &= ~(unsigned long long)((chunksize>>9)-1);
 737         return data_disks * devsize;
 738 }
 739
 740 int get_mdp_major(void)
 741 {
 742 static int mdp_major = -1;
 743         FILE *fl;
 744         char *w;
 745         int have_block = 0;
 746         int have_devices = 0;
 747         int last_num = -1;
 748
 749         if (mdp_major != -1)
 750                 return mdp_major;
 751         fl = fopen("/proc/devices", "r");
 752         if (!fl)
 753                 return -1;
 754         while ((w = conf_word(fl, 1))) {
 755                 if (have_block && strcmp(w, "devices:")==0)
 756                         have_devices = 1;
 757                 have_block =  (strcmp(w, "Block")==0);
 758                 if (isdigit(w[0]))
 759                         last_num = atoi(w);
 760                 if (have_devices && strcmp(w, "mdp")==0)
 761                         mdp_major = last_num;
 762                 free(w);
 763         }
 764         fclose(fl);
 765         return mdp_major;
 766 }
 767
 768 #if !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO)
 769 char *get_md_name(int dev)
 770 {
 771         /* find /dev/md%d or /dev/md/%d or make a device /dev/.tmp.md%d */
 772         /* if dev < 0, want /dev/md/d%d or find mdp in /proc/devices ... */
 773         static char devname[50];
 774         struct stat stb;
 775         dev_t rdev;
 776         char *dn;
 777
 778         if (dev < 0) {
 779                 int mdp =  get_mdp_major();
 780                 if (mdp < 0) return NULL;
 781                 rdev = makedev(mdp, (-1-dev)<<6);
 782                 snprintf(devname, sizeof(devname), "/dev/md/d%d", -1-dev);
 783                 if (stat(devname, &stb) == 0
 784                     && (S_IFMT&stb.st_mode) == S_IFBLK
 785                     && (stb.st_rdev == rdev))
 786                         return devname;
 787         } else {
 788                 rdev = makedev(MD_MAJOR, dev);
 789                 snprintf(devname, sizeof(devname), "/dev/md%d", dev);
 790                 if (stat(devname, &stb) == 0
 791                     && (S_IFMT&stb.st_mode) == S_IFBLK
 792                     && (stb.st_rdev == rdev))
 793                         return devname;
 794
 795                 snprintf(devname, sizeof(devname), "/dev/md/%d", dev);
 796                 if (stat(devname, &stb) == 0
 797                     && (S_IFMT&stb.st_mode) == S_IFBLK
 798                     && (stb.st_rdev == rdev))
 799                         return devname;
 800         }
 801         dn = map_dev(major(rdev), minor(rdev), 0);
 802         if (dn)
 803                 return dn;
 804         snprintf(devname, sizeof(devname), "/dev/.tmp.md%d", dev);
 805         if (mknod(devname, S_IFBLK | 0600, rdev) == -1)
 806                 if (errno != EEXIST)
 807                         return NULL;
 808
 809         if (stat(devname, &stb) == 0
 810             && (S_IFMT&stb.st_mode) == S_IFBLK
 811             && (stb.st_rdev == rdev))
 812                 return devname;
 813         unlink(devname);
 814         return NULL;
 815 }
 816
 817 void put_md_name(char *name)
 818 {
 819         if (strncmp(name, "/dev/.tmp.md", 12)==0)
 820                 unlink(name);
 821 }
 822
 823 int find_free_devnum(int use_partitions)
 824 {
 825         int devnum;
 826         for (devnum = 127; devnum != 128;
 827              devnum = devnum ? devnum-1 : (1<<20)-1) {
 828                 char *dn;
 829                 int _devnum;
 830
 831                 _devnum = use_partitions ? (-1-devnum) : devnum;
 832                 if (mddev_busy(_devnum))
 833                         continue;
 834                 /* make sure it is new to /dev too, at least as a
 835                  * non-standard */
 836                 dn = map_dev(dev2major(_devnum), dev2minor(_devnum), 0);
 837                 if (dn && ! is_standard(dn, NULL))
 838                         continue;
 839                 break;
 840         }
 841         if (devnum == 128)
 842                 return NoMdDev;
 843         return use_partitions ? (-1-devnum) : devnum;
 844 }
 845 #endif /* !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO) */
 846
 847 int dev_open(char *dev, int flags)
 848 {
 849         /* like 'open', but if 'dev' matches %d:%d, create a temp
 850          * block device and open that
 851          */
 852         char *e;
 853         int fd = -1;
 854         char devname[32];
 855         int major;
 856         int minor;
 857
 858         if (!dev) return -1;
 859
 860         major = strtoul(dev, &e, 0);
 861         if (e > dev && *e == ':' && e[1] &&
 862             (minor = strtoul(e+1, &e, 0)) >= 0 &&
 863             *e == 0) {
 864                 snprintf(devname, sizeof(devname), "/dev/.tmp.md.%d:%d:%d",
 865                          (int)getpid(), major, minor);
 866                 if (mknod(devname, S_IFBLK|0600, makedev(major, minor))==0) {
 867                         fd = open(devname, flags|O_DIRECT);
 868                         unlink(devname);
 869                 }
 870         } else
 871                 fd = open(dev, flags|O_DIRECT);
 872         return fd;
 873 }
 874
 875 int open_dev(int devnum)
 876 {
 877         char buf[20];
 878
 879         sprintf(buf, "%d:%d", dev2major(devnum), dev2minor(devnum));
 880         return dev_open(buf, O_RDWR);
 881 }
 882
 883 int open_dev_excl(int devnum)
 884 {
 885         char buf[20];
 886         int i;
 887
 888         sprintf(buf, "%d:%d", dev2major(devnum), dev2minor(devnum));
 889         for (i=0 ; i<25 ; i++) {
 890                 int fd = dev_open(buf, O_RDWR|O_EXCL);
 891                 if (fd >= 0)
 892                         return fd;
 893                 if (errno != EBUSY)
 894                         return fd;
 895                 usleep(200000);
 896         }
 897         return -1;
 898 }
 899
 900 int same_dev(char *one, char *two)
 901 {
 902         struct stat st1, st2;
 903         if (stat(one, &st1) != 0)
 904                 return 0;
 905         if (stat(two, &st2) != 0)
 906                 return 0;
 907         if ((st1.st_mode & S_IFMT) != S_IFBLK)
 908                 return 0;
 909         if ((st2.st_mode & S_IFMT) != S_IFBLK)
 910                 return 0;
 911         return st1.st_rdev == st2.st_rdev;
 912 }
 913
 914 void wait_for(char *dev, int fd)
 915 {
 916         int i;
 917         struct stat stb_want;
 918
 919         if (fstat(fd, &stb_want) != 0 ||
 920             (stb_want.st_mode & S_IFMT) != S_IFBLK)
 921                 return;
 922
 923         for (i=0 ; i<25 ; i++) {
 924                 struct stat stb;
 925                 if (stat(dev, &stb) == 0 &&
 926                     (stb.st_mode & S_IFMT) == S_IFBLK &&
 927                     (stb.st_rdev == stb_want.st_rdev))
 928                         return;
 929                 usleep(200000);
 930         }
 931         if (i == 25)
 932                 dprintf("%s: timeout waiting for %s\n", __func__, dev);
 933 }
 934
 935 struct superswitch *superlist[] = { &super0, &super1, &super_ddf, &super_imsm, NULL };
 936
 937 #if !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO)
 938
 939 struct supertype *super_by_fd(int fd)
 940 {
 941         mdu_array_info_t array;
 942         int vers;
 943         int minor;
 944         struct supertype *st = NULL;
 945         struct mdinfo *sra;
 946         char *verstr;
 947         char version[20];
 948         int i;
 949         char *subarray = NULL;
 950
 951         sra = sysfs_read(fd, 0, GET_VERSION);
 952
 953         if (sra) {
 954                 vers = sra->array.major_version;
 955                 minor = sra->array.minor_version;
 956                 verstr = sra->text_version;
 957         } else {
 958                 if (ioctl(fd, GET_ARRAY_INFO, &array))
 959                         array.major_version = array.minor_version = 0;
 960                 vers = array.major_version;
 961                 minor = array.minor_version;
 962                 verstr = "";
 963         }
 964
 965         if (vers != -1) {
 966                 sprintf(version, "%d.%d", vers, minor);
 967                 verstr = version;
 968         }
 969         if (minor == -2 && is_subarray(verstr)) {
 970                 char *dev = verstr+1;
 971                 subarray = strchr(dev, '/');
 972                 int devnum;
 973                 if (subarray)
 974                         *subarray++ = '\0';
 975                 devnum = devname2devnum(dev);
 976                 subarray = strdup(subarray);
 977                 if (sra)
 978                         sysfs_free(sra);
 979                 sra = sysfs_read(-1, devnum, GET_VERSION);
 980                 if (sra && sra->text_version[0])
 981                         verstr = sra->text_version;
 982                 else
 983                         verstr = "-no-metadata-";
 984         }
 985
 986         for (i = 0; st == NULL && superlist[i] ; i++)
 987                 st = superlist[i]->match_metadata_desc(verstr);
 988
 989         if (sra)
 990                 sysfs_free(sra);
 991         if (st) {
 992                 st->sb = NULL;
 993                 if (subarray) {
 994                         strncpy(st->subarray, subarray, 32);
 995                         st->subarray[31] = 0;
 996                         free(subarray);
 997                 } else
 998                         st->subarray[0] = 0;
 999         }
1000         return st;
1001 }
1002 #endif /* !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO) */
1003
1004
1005 struct supertype *dup_super(struct supertype *orig)
1006 {
1007         struct supertype *st;
1008
1009         if (!orig)
1010                 return orig;
1011         st = malloc(sizeof(*st));
1012         if (!st)
1013                 return st;
1014         memset(st, 0, sizeof(*st));
1015         st->ss = orig->ss;
1016         st->max_devs = orig->max_devs;
1017         st->minor_version = orig->minor_version;
1018         strcpy(st->subarray, orig->subarray);
1019         st->sb = NULL;
1020         st->info = NULL;
1021         return st;
1022 }
1023
1024 struct supertype *guess_super(int fd)
1025 {
1026         /* try each load_super to find the best match,
1027          * and return the best superswitch
1028          */
1029         struct superswitch  *ss;
1030         struct supertype *st;
1031         unsigned long besttime = 0;
1032         int bestsuper = -1;
1033         int i;
1034
1035         st = malloc(sizeof(*st));
1036         for (i=0 ; superlist[i]; i++) {
1037                 int rv;
1038                 ss = superlist[i];
1039                 memset(st, 0, sizeof(*st));
1040                 rv = ss->load_super(st, fd, NULL);
1041                 if (rv == 0) {
1042                         struct mdinfo info;
1043                         st->ss->getinfo_super(st, &info);
1044                         if (bestsuper == -1 ||
1045                             besttime < info.array.ctime) {
1046                                 bestsuper = i;
1047                                 besttime = info.array.ctime;
1048                         }
1049                         ss->free_super(st);
1050                 }
1051         }
1052         if (bestsuper != -1) {
1053                 int rv;
1054                 memset(st, 0, sizeof(*st));
1055                 rv = superlist[bestsuper]->load_super(st, fd, NULL);
1056                 if (rv == 0) {
1057                         superlist[bestsuper]->free_super(st);
1058                         return st;
1059                 }
1060         }
1061         free(st);
1062         return NULL;
1063 }
1064
1065 /* Return size of device in bytes */
1066 int get_dev_size(int fd, char *dname, unsigned long long *sizep)
1067 {
1068         unsigned long long ldsize;
1069         struct stat st;
1070
1071         if (fstat(fd, &st) != -1 && S_ISREG(st.st_mode))
1072                 ldsize = (unsigned long long)st.st_size;
1073         else
1074 #ifdef BLKGETSIZE64
1075         if (ioctl(fd, BLKGETSIZE64, &ldsize) != 0)
1076 #endif
1077         {
1078                 unsigned long dsize;
1079                 if (ioctl(fd, BLKGETSIZE, &dsize) == 0) {
1080                         ldsize = dsize;
1081                         ldsize <<= 9;
1082                 } else {
1083                         if (dname)
1084                                 fprintf(stderr, Name ": Cannot get size of %s: %s\b",
1085                                         dname, strerror(errno));
1086                         return 0;
1087                 }
1088         }
1089         *sizep = ldsize;
1090         return 1;
1091 }
1092
1093 void get_one_disk(int mdfd, mdu_array_info_t *ainf, mdu_disk_info_t *disk)
1094 {
1095         int d;
1096         ioctl(mdfd, GET_ARRAY_INFO, ainf);
1097         for (d = 0 ; d < ainf->raid_disks + ainf->nr_disks ; d++)
1098                 if (ioctl(mdfd, GET_DISK_INFO, disk) == 0)
1099                         return;
1100 }
1101
1102 int open_container(int fd)
1103 {
1104         /* 'fd' is a block device.  Find out if it is in use
1105          * by a container, and return an open fd on that container.
1106          */
1107         char path[256];
1108         char *e;
1109         DIR *dir;
1110         struct dirent *de;
1111         int dfd, n;
1112         char buf[200];
1113         int major, minor;
1114         struct stat st;
1115
1116         if (fstat(fd, &st) != 0)
1117                 return -1;
1118         sprintf(path, "/sys/dev/block/%d:%d/holders",
1119                 (int)major(st.st_rdev), (int)minor(st.st_rdev));
1120         e = path + strlen(path);
1121
1122         dir = opendir(path);
1123         if (!dir)
1124                 return -1;
1125         while ((de = readdir(dir))) {
1126                 if (de->d_ino == 0)
1127                         continue;
1128                 if (de->d_name[0] == '.')
1129                         continue;
1130                 sprintf(e, "/%s/dev", de->d_name);
1131                 dfd = open(path, O_RDONLY);
1132                 if (dfd < 0)
1133                         continue;
1134                 n = read(dfd, buf, sizeof(buf));
1135                 close(dfd);
1136                 if (n <= 0 || n >= sizeof(buf))
1137                         continue;
1138                 buf[n] = 0;
1139                 if (sscanf(buf, "%d:%d", &major, &minor) != 2)
1140                         continue;
1141                 sprintf(buf, "%d:%d", major, minor);
1142                 dfd = dev_open(buf, O_RDONLY);
1143                 if (dfd >= 0) {
1144                         closedir(dir);
1145                         return dfd;
1146                 }
1147         }
1148         closedir(dir);
1149         return -1;
1150 }
1151
1152 int add_disk(int mdfd, struct supertype *st,
1153              struct mdinfo *sra, struct mdinfo *info)
1154 {
1155         /* Add a device to an array, in one of 2 ways. */
1156         int rv;
1157 #ifndef MDASSEMBLE
1158         if (st->ss->external) {
1159                 rv = sysfs_add_disk(sra, info,
1160                                     info->disk.state & (1<<MD_DISK_SYNC));
1161                 if (! rv) {
1162                         struct mdinfo *sd2;
1163                         for (sd2 = sra->devs; sd2; sd2=sd2->next)
1164                                 if (sd2 == info)
1165                                         break;
1166                         if (sd2 == NULL) {
1167                                 sd2 = malloc(sizeof(*sd2));
1168                                 *sd2 = *info;
1169                                 sd2->next = sra->devs;
1170                                 sra->devs = sd2;
1171                         }
1172                 }
1173         } else
1174 #endif
1175                 rv = ioctl(mdfd, ADD_NEW_DISK, &info->disk);
1176         return rv;
1177 }
1178
1179 int set_array_info(int mdfd, struct supertype *st, struct mdinfo *info)
1180 {
1181         /* Initialise kernel's knowledge of array.
1182          * This varies between externally managed arrays
1183          * and older kernels
1184          */
1185         int vers = md_get_version(mdfd);
1186         int rv;
1187
1188 #ifndef MDASSEMBLE
1189         if (st->ss->external)
1190                 rv = sysfs_set_array(info, vers);
1191         else
1192 #endif
1193                 if ((vers % 100) >= 1) { /* can use different versions */
1194                 mdu_array_info_t inf;
1195                 memset(&inf, 0, sizeof(inf));
1196                 inf.major_version = info->array.major_version;
1197                 inf.minor_version = info->array.minor_version;
1198                 rv = ioctl(mdfd, SET_ARRAY_INFO, &inf);
1199         } else
1200                 rv = ioctl(mdfd, SET_ARRAY_INFO, NULL);
1201         return rv;
1202 }
1203
1204 char *devnum2devname(int num)
1205 {
1206         char name[100];
1207         if (num > 0)
1208                 sprintf(name, "md%d", num);
1209         else
1210                 sprintf(name, "md_d%d", -1-num);
1211         return strdup(name);
1212 }
1213
1214 int devname2devnum(char *name)
1215 {
1216         char *ep;
1217         int num;
1218         if (strncmp(name, "md_d", 4)==0)
1219                 num = -1-strtoul(name+4, &ep, 10);
1220         else
1221                 num = strtoul(name+2, &ep, 10);
1222         return num;
1223 }
1224
1225 int stat2devnum(struct stat *st)
1226 {
1227         char path[30];
1228         char link[200];
1229         char *cp;
1230         int n;
1231
1232         if ((S_IFMT & st->st_mode) == S_IFBLK) {
1233                 if (major(st->st_rdev) == MD_MAJOR)
1234                         return minor(st->st_rdev);
1235                 else if (major(st->st_rdev) == get_mdp_major())
1236                         return -1- (minor(st->st_rdev)>>MdpMinorShift);
1237
1238                 /* must be an extended-minor partition. Look at the
1239                  * /sys/dev/block/%d:%d link which must look like
1240                  * ../../block/mdXXX/mdXXXpYY
1241                  */
1242                 sprintf(path, "/sys/dev/block/%d:%d", major(st->st_rdev),
1243                         minor(st->st_rdev));
1244                 n = readlink(path, link, sizeof(link)-1);
1245                 if (n <= 0)
1246                         return NoMdDev;
1247                 link[n] = 0;
1248                 cp = strrchr(link, '/');
1249                 if (cp) *cp = 0;
1250                 cp = strchr(link, '/');
1251                 if (cp && strncmp(cp, "/md", 3) == 0)
1252                         return devname2devnum(cp+1);
1253         }
1254         return NoMdDev;
1255
1256 }
1257
1258 int fd2devnum(int fd)
1259 {
1260         struct stat stb;
1261         if (fstat(fd, &stb) == 0)
1262                 return stat2devnum(&stb);
1263         return NoMdDev;
1264 }
1265
1266 int mdmon_running(int devnum)
1267 {
1268         char path[100];
1269         char pid[10];
1270         int fd;
1271         int n;
1272         sprintf(path, "/var/run/mdadm/%s.pid", devnum2devname(devnum));
1273         fd = open(path, O_RDONLY, 0);
1274
1275         if (fd < 0)
1276                 return 0;
1277         n = read(fd, pid, 9);
1278         close(fd);
1279         if (n <= 0)
1280                 return 0;
1281         if (kill(atoi(pid), 0) == 0)
1282                 return 1;
1283         return 0;
1284 }
1285
1286 int signal_mdmon(int devnum)
1287 {
1288         char path[100];
1289         char pid[10];
1290         int fd;
1291         int n;
1292         sprintf(path, "/var/run/mdadm/%s.pid", devnum2devname(devnum));
1293         fd = open(path, O_RDONLY, 0);
1294
1295         if (fd < 0)
1296                 return 0;
1297         n = read(fd, pid, 9);
1298         close(fd);
1299         if (n <= 0)
1300                 return 0;
1301         if (kill(atoi(pid), SIGUSR1) == 0)
1302                 return 1;
1303         return 0;
1304 }
1305
1306 int start_mdmon(int devnum)
1307 {
1308         int i;
1309         int len;
1310         pid_t pid;
1311         int status;
1312         char pathbuf[1024];
1313         char *paths[4] = {
1314                 pathbuf,
1315                 "/sbin/mdmon",
1316                 "mdmon",
1317                 NULL
1318         };
1319
1320         if (check_env("MDADM_NO_MDMON"))
1321                 return 0;
1322
1323         len = readlink("/proc/self/exe", pathbuf, sizeof(pathbuf));
1324         if (len > 0) {
1325                 char *sl;
1326                 pathbuf[len] = 0;
1327                 sl = strrchr(pathbuf, '/');
1328                 if (sl)
1329                         sl++;
1330                 else
1331                         sl = pathbuf;
1332                 strcpy(sl, "mdmon");
1333         } else
1334                 pathbuf[0] = '\0';
1335
1336         switch(fork()) {
1337         case 0:
1338                 /* FIXME yuk. CLOSE_EXEC?? */
1339                 for (i=3; i < 100; i++)
1340                         close(i);
1341                 for (i=0; paths[i]; i++)
1342                         if (paths[i][0])
1343                                 execl(paths[i], "mdmon",
1344                                       devnum2devname(devnum),
1345                                       NULL);
1346                 exit(1);
1347         case -1: fprintf(stderr, Name ": cannot run mdmon. "
1348                          "Array remains readonly\n");
1349                 return -1;
1350         default: /* parent - good */
1351                 pid = wait(&status);
1352                 if (pid < 0 || status != 0)
1353                         return -1;
1354         }
1355         return 0;
1356 }
1357
1358 int check_env(char *name)
1359 {
1360         char *val = getenv(name);
1361
1362         if (val && atoi(val) == 1)
1363                 return 1;
1364
1365         return 0;
1366 }
1367
1368 __u32 random32(void)
1369 {
1370         __u32 rv;
1371         int rfd = open("/dev/urandom", O_RDONLY);
1372         if (rfd < 0 || read(rfd, &rv, 4) != 4)
1373                 rv = random();
1374         if (rfd >= 0)
1375                 close(rfd);
1376         return rv;
1377 }
1378
1379 #ifndef MDASSEMBLE
1380 int flush_metadata_updates(struct supertype *st)
1381 {
1382         int sfd;
1383         if (!st->updates) {
1384                 st->update_tail = NULL;
1385                 return -1;
1386         }
1387
1388         sfd = connect_monitor(devnum2devname(st->container_dev));
1389         if (sfd < 0)
1390                 return -1;
1391
1392         while (st->updates) {
1393                 struct metadata_update *mu = st->updates;
1394                 st->updates = mu->next;
1395
1396                 send_message(sfd, mu, 0);
1397                 wait_reply(sfd, 0);
1398                 free(mu->buf);
1399                 free(mu);
1400         }
1401         ack(sfd, 0);
1402         wait_reply(sfd, 0);
1403         close(sfd);
1404         st->update_tail = NULL;
1405         return 0;
1406 }
1407
1408 void append_metadata_update(struct supertype *st, void *buf, int len)
1409 {
1410
1411         struct metadata_update *mu = malloc(sizeof(*mu));
1412
1413         mu->buf = buf;
1414         mu->len = len;
1415         mu->space = NULL;
1416         mu->next = NULL;
1417         *st->update_tail = mu;
1418         st->update_tail = &mu->next;
1419 }
1420 #endif /* MDASSEMBLE */
1421
1422 #ifdef __TINYC__
1423 /* tinyc doesn't optimize this check in ioctl.h out ... */
1424 unsigned int __invalid_size_argument_for_IOC = 0;
1425 #endif
1426