util.c

   1 /*
   2  * mdadm - manage Linux "md" devices aka RAID arrays.
   3  *
   4  * Copyright (C) 2001-2006 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@cse.unsw.edu.au>
  23  *    Paper: Neil Brown
  24  *           School of Computer Science and Engineering
  25  *           The University of New South Wales
  26  *           Sydney, 2052
  27  *           Australia
  28  */
  29
  30 #include        "mdadm.h"
  31 #include        "md_p.h"
  32 #include        <sys/socket.h>
  33 #include        <sys/utsname.h>
  34 #include        <sys/wait.h>
  35 #include        <sys/un.h>
  36 #include        <ctype.h>
  37 #include        <dirent.h>
  38 #include        <signal.h>
  39
  40 /*
  41  * following taken from linux/blkpg.h because they aren't
  42  * anywhere else and it isn't safe to #include linux/ * stuff.
  43  */
  44
  45 #define BLKPG      _IO(0x12,105)
  46
  47 /* The argument structure */
  48 struct blkpg_ioctl_arg {
  49         int op;
  50         int flags;
  51         int datalen;
  52         void *data;
  53 };
  54
  55 /* The subfunctions (for the op field) */
  56 #define BLKPG_ADD_PARTITION     1
  57 #define BLKPG_DEL_PARTITION     2
  58
  59 /* Sizes of name fields. Unused at present. */
  60 #define BLKPG_DEVNAMELTH        64
  61 #define BLKPG_VOLNAMELTH        64
  62
  63 /* The data structure for ADD_PARTITION and DEL_PARTITION */
  64 struct blkpg_partition {
  65         long long start;                /* starting offset in bytes */
  66         long long length;               /* length in bytes */
  67         int pno;                        /* partition number */
  68         char devname[BLKPG_DEVNAMELTH]; /* partition name, like sda5 or c0d1p2,
  69                                            to be used in kernel messages */
  70         char volname[BLKPG_VOLNAMELTH]; /* volume label */
  71 };
  72
  73 /*
  74  * Parse a 128 bit uuid in 4 integers
  75  * format is 32 hexx nibbles with options :.<space> separator
  76  * If not exactly 32 hex digits are found, return 0
  77  * else return 1
  78  */
  79 int parse_uuid(char *str, int uuid[4])
  80 {
  81         int hit = 0; /* number of Hex digIT */
  82         int i;
  83         char c;
  84         for (i=0; i<4; i++) uuid[i]=0;
  85
  86         while ((c= *str++)) {
  87                 int n;
  88                 if (c>='0' && c<='9')
  89                         n = c-'0';
  90                 else if (c>='a' && c <= 'f')
  91                         n = 10 + c - 'a';
  92                 else if (c>='A' && c <= 'F')
  93                         n = 10 + c - 'A';
  94                 else if (strchr(":. -", c))
  95                         continue;
  96                 else return 0;
  97
  98                 if (hit<32) {
  99                         uuid[hit/8] <<= 4;
 100                         uuid[hit/8] += n;
 101                 }
 102                 hit++;
 103         }
 104         if (hit == 32)
 105                 return 1;
 106         return 0;
 107 }
 108
 109
 110 /*
 111  * Get the md version number.
 112  * We use the RAID_VERSION ioctl if it is supported
 113  * If not, but we have a block device with major '9', we assume
 114  * 0.36.0
 115  *
 116  * Return version number as 24 but number - assume version parts
 117  * always < 255
 118  */
 119
 120 int md_get_version(int fd)
 121 {
 122     struct stat stb;
 123     mdu_version_t vers;
 124
 125     if (fstat(fd, &stb)<0)
 126         return -1;
 127     if ((S_IFMT&stb.st_mode) != S_IFBLK)
 128         return -1;
 129
 130     if (ioctl(fd, RAID_VERSION, &vers) == 0)
 131         return  (vers.major*10000) + (vers.minor*100) + vers.patchlevel;
 132     if (errno == EACCES)
 133             return -1;
 134     if (major(stb.st_rdev) == MD_MAJOR)
 135         return (3600);
 136     return -1;
 137 }
 138
 139 int get_linux_version()
 140 {
 141         struct utsname name;
 142         char *cp;
 143         int a,b,c;
 144         if (uname(&name) <0)
 145                 return -1;
 146
 147         cp = name.release;
 148         a = strtoul(cp, &cp, 10);
 149         if (*cp != '.') return -1;
 150         b = strtoul(cp+1, &cp, 10);
 151         if (*cp != '.') return -1;
 152         c = strtoul(cp+1, NULL, 10);
 153
 154         return (a*1000000)+(b*1000)+c;
 155 }
 156
 157 long long parse_size(char *size)
 158 {
 159         /* parse 'size' which should be a number optionally
 160          * followed by 'K', 'M', or 'G'.
 161          * Without a suffix, K is assumed.
 162          * Number returned is in sectors (half-K)
 163          */
 164         char *c;
 165         long long s = strtoll(size, &c, 10);
 166         if (s > 0) {
 167                 switch (*c) {
 168                 case 'K':
 169                         c++;
 170                 default:
 171                         s *= 2;
 172                         break;
 173                 case 'M':
 174                         c++;
 175                         s *= 1024 * 2;
 176                         break;
 177                 case 'G':
 178                         c++;
 179                         s *= 1024 * 1024 * 2;
 180                         break;
 181                 }
 182         }
 183         if (*c)
 184                 s = 0;
 185         return s;
 186 }
 187
 188 int parse_layout_10(char *layout)
 189 {
 190         int copies, rv;
 191         char *cp;
 192         /* Parse the layout string for raid10 */
 193         /* 'f', 'o' or 'n' followed by a number <= raid_disks */
 194         if ((layout[0] !=  'n' && layout[0] != 'f' && layout[0] != 'o') ||
 195             (copies = strtoul(layout+1, &cp, 10)) < 1 ||
 196             copies > 200 ||
 197             *cp)
 198                 return -1;
 199         if (layout[0] == 'n')
 200                 rv = 256 + copies;
 201         else if (layout[0] == 'o')
 202                 rv = 0x10000 + (copies<<8) + 1;
 203         else
 204                 rv = 1 + (copies<<8);
 205         return rv;
 206 }
 207
 208 int parse_layout_faulty(char *layout)
 209 {
 210         /* Parse the layout string for 'faulty' */
 211         int ln = strcspn(layout, "0123456789");
 212         char *m = strdup(layout);
 213         int mode;
 214         m[ln] = 0;
 215         mode = map_name(faultylayout, m);
 216         if (mode == UnSet)
 217                 return -1;
 218
 219         return mode | (atoi(layout+ln)<< ModeShift);
 220 }
 221
 222 void remove_partitions(int fd)
 223 {
 224         /* remove partitions from this block devices.
 225          * This is used for components added to an array
 226          */
 227 #ifdef BLKPG_DEL_PARTITION
 228         struct blkpg_ioctl_arg a;
 229         struct blkpg_partition p;
 230
 231         a.op = BLKPG_DEL_PARTITION;
 232         a.data = (void*)&p;
 233         a.datalen = sizeof(p);
 234         a.flags = 0;
 235         memset(a.data, 0, a.datalen);
 236         for (p.pno=0; p.pno < 16; p.pno++)
 237                 ioctl(fd, BLKPG, &a);
 238 #endif
 239 }
 240
 241 int enough(int level, int raid_disks, int layout, int clean,
 242            char *avail, int avail_disks)
 243 {
 244         int copies, first;
 245         switch (level) {
 246         case 10:
 247                 /* This is the tricky one - we need to check
 248                  * which actual disks are present.
 249                  */
 250                 copies = (layout&255)* ((layout>>8) & 255);
 251                 first=0;
 252                 do {
 253                         /* there must be one of the 'copies' form 'first' */
 254                         int n = copies;
 255                         int cnt=0;
 256                         while (n--) {
 257                                 if (avail[first])
 258                                         cnt++;
 259                                 first = (first+1) % raid_disks;
 260                         }
 261                         if (cnt == 0)
 262                                 return 0;
 263
 264                 } while (first != 0);
 265                 return 1;
 266
 267         case -4:
 268                 return avail_disks>= 1;
 269         case -1:
 270         case 0:
 271                 return avail_disks == raid_disks;
 272         case 1:
 273                 return avail_disks >= 1;
 274         case 4:
 275         case 5:
 276                 if (clean)
 277                         return avail_disks >= raid_disks-1;
 278                 else
 279                         return avail_disks >= raid_disks;
 280         case 6:
 281                 if (clean)
 282                         return avail_disks >= raid_disks-2;
 283                 else
 284                         return avail_disks >= raid_disks;
 285         default:
 286                 return 0;
 287         }
 288 }
 289
 290 const int uuid_match_any[4] = { ~0, ~0, ~0, ~0 };
 291 int same_uuid(int a[4], int b[4], int swapuuid)
 292 {
 293         if (memcmp(a, uuid_match_any, sizeof(int[4])) == 0 ||
 294             memcmp(b, uuid_match_any, sizeof(int[4])) == 0)
 295                 return 1;
 296
 297         if (swapuuid) {
 298                 /* parse uuids are hostendian.
 299                  * uuid's from some superblocks are big-ending
 300                  * if there is a difference, we need to swap..
 301                  */
 302                 unsigned char *ac = (unsigned char *)a;
 303                 unsigned char *bc = (unsigned char *)b;
 304                 int i;
 305                 for (i=0; i<16; i+= 4) {
 306                         if (ac[i+0] != bc[i+3] ||
 307                             ac[i+1] != bc[i+2] ||
 308                             ac[i+2] != bc[i+1] ||
 309                             ac[i+3] != bc[i+0])
 310                                 return 0;
 311                 }
 312                 return 1;
 313         } else {
 314                 if (a[0]==b[0] &&
 315                     a[1]==b[1] &&
 316                     a[2]==b[2] &&
 317                     a[3]==b[3])
 318                         return 1;
 319                 return 0;
 320         }
 321 }
 322 void copy_uuid(void *a, int b[4], int swapuuid)
 323 {
 324         if (swapuuid) {
 325                 /* parse uuids are hostendian.
 326                  * uuid's from some superblocks are big-ending
 327                  * if there is a difference, we need to swap..
 328                  */
 329                 unsigned char *ac = (unsigned char *)a;
 330                 unsigned char *bc = (unsigned char *)b;
 331                 int i;
 332                 for (i=0; i<16; i+= 4) {
 333                         ac[i+0] = bc[i+3];
 334                         ac[i+1] = bc[i+2];
 335                         ac[i+2] = bc[i+1];
 336                         ac[i+3] = bc[i+0];
 337                 }
 338         } else
 339                 memcpy(a, b, 16);
 340 }
 341
 342 char *fname_from_uuid(struct supertype *st, struct mdinfo *info, char *buf, char sep)
 343 {
 344         int i, j;
 345         int id;
 346         char uuid[16];
 347         char *c = buf;
 348         strcpy(c, "UUID-");
 349         c += strlen(c);
 350         copy_uuid(uuid, info->uuid, st->ss->swapuuid);
 351         for (i = 0; i < 4; i++) {
 352                 id = uuid[i];
 353                 if (i)
 354                         *c++ = sep;
 355                 for (j = 3; j >= 0; j--) {
 356                         sprintf(c,"%02x", (unsigned char) uuid[j+4*i]);
 357                         c+= 2;
 358                 }
 359         }
 360         return buf;
 361 }
 362
 363 #ifndef MDASSEMBLE
 364 int check_ext2(int fd, char *name)
 365 {
 366         /*
 367          * Check for an ext2fs file system.
 368          * Superblock is always 1K at 1K offset
 369          *
 370          * s_magic is le16 at 56 == 0xEF53
 371          * report mtime - le32 at 44
 372          * blocks - le32 at 4
 373          * logblksize - le32 at 24
 374          */
 375         unsigned char sb[1024];
 376         time_t mtime;
 377         int size, bsize;
 378         if (lseek(fd, 1024,0)!= 1024)
 379                 return 0;
 380         if (read(fd, sb, 1024)!= 1024)
 381                 return 0;
 382         if (sb[56] != 0x53 || sb[57] != 0xef)
 383                 return 0;
 384
 385         mtime = sb[44]|(sb[45]|(sb[46]|sb[47]<<8)<<8)<<8;
 386         bsize = sb[24]|(sb[25]|(sb[26]|sb[27]<<8)<<8)<<8;
 387         size = sb[4]|(sb[5]|(sb[6]|sb[7]<<8)<<8)<<8;
 388         fprintf(stderr, Name ": %s appears to contain an ext2fs file system\n",
 389                 name);
 390         fprintf(stderr,"    size=%dK  mtime=%s",
 391                 size*(1<<bsize), ctime(&mtime));
 392         return 1;
 393 }
 394
 395 int check_reiser(int fd, char *name)
 396 {
 397         /*
 398          * superblock is at 64K
 399          * size is 1024;
 400          * Magic string "ReIsErFs" or "ReIsEr2Fs" at 52
 401          *
 402          */
 403         unsigned char sb[1024];
 404         unsigned long size;
 405         if (lseek(fd, 64*1024, 0) != 64*1024)
 406                 return 0;
 407         if (read(fd, sb, 1024) != 1024)
 408                 return 0;
 409         if (strncmp((char*)sb+52, "ReIsErFs",8)!=0 &&
 410             strncmp((char*)sb+52, "ReIsEr2Fs",9)!=0)
 411                 return 0;
 412         fprintf(stderr, Name ": %s appears to contain a reiserfs file system\n",name);
 413         size = sb[0]|(sb[1]|(sb[2]|sb[3]<<8)<<8)<<8;
 414         fprintf(stderr, "    size = %luK\n", size*4);
 415
 416         return 1;
 417 }
 418
 419 int check_raid(int fd, char *name)
 420 {
 421         struct mdinfo info;
 422         time_t crtime;
 423         char *level;
 424         struct supertype *st = guess_super(fd);
 425
 426         if (!st) return 0;
 427         st->ss->load_super(st, fd, name);
 428         /* Looks like a raid array .. */
 429         fprintf(stderr, Name ": %s appears to be part of a raid array:\n",
 430                 name);
 431         st->ss->getinfo_super(st, &info);
 432         st->ss->free_super(st);
 433         crtime = info.array.ctime;
 434         level = map_num(pers, info.array.level);
 435         if (!level) level = "-unknown-";
 436         fprintf(stderr, "    level=%s devices=%d ctime=%s",
 437                 level, info.array.raid_disks, ctime(&crtime));
 438         return 1;
 439 }
 440
 441 int ask(char *mesg)
 442 {
 443         char *add = "";
 444         int i;
 445         for (i=0; i<5; i++) {
 446                 char buf[100];
 447                 fprintf(stderr, "%s%s", mesg, add);
 448                 fflush(stderr);
 449                 if (fgets(buf, 100, stdin)==NULL)
 450                         return 0;
 451                 if (buf[0]=='y' || buf[0]=='Y')
 452                         return 1;
 453                 if (buf[0]=='n' || buf[0]=='N')
 454                         return 0;
 455                 add = "(y/n) ";
 456         }
 457         fprintf(stderr, Name ": assuming 'no'\n");
 458         return 0;
 459 }
 460 #endif /* MDASSEMBLE */
 461
 462 char *map_num(mapping_t *map, int num)
 463 {
 464         while (map->name) {
 465                 if (map->num == num)
 466                         return map->name;
 467                 map++;
 468         }
 469         return NULL;
 470 }
 471
 472 int map_name(mapping_t *map, char *name)
 473 {
 474         while (map->name) {
 475                 if (strcmp(map->name, name)==0)
 476                         return map->num;
 477                 map++;
 478         }
 479         return UnSet;
 480 }
 481
 482
 483 int is_standard(char *dev, int *nump)
 484 {
 485         /* tests if dev is a "standard" md dev name.
 486          * i.e if the last component is "/dNN" or "/mdNN",
 487          * where NN is a string of digits
 488          * Returns 1 if a partitionable standard,
 489          *   -1 if non-partitonable,
 490          *   0 if not a standard name.
 491          */
 492         char *d = strrchr(dev, '/');
 493         int type=0;
 494         int num;
 495         if (!d)
 496                 return 0;
 497         if (strncmp(d, "/d",2)==0)
 498                 d += 2, type=1; /* /dev/md/dN{pM} */
 499         else if (strncmp(d, "/md_d", 5)==0)
 500                 d += 5, type=1; /* /dev/md_dN{pM} */
 501         else if (strncmp(d, "/md", 3)==0)
 502                 d += 3, type=-1; /* /dev/mdN */
 503         else if (d-dev > 3 && strncmp(d-2, "md/", 3)==0)
 504                 d += 1, type=-1; /* /dev/md/N */
 505         else
 506                 return 0;
 507         if (!*d)
 508                 return 0;
 509         num = atoi(d);
 510         while (isdigit(*d))
 511                 d++;
 512         if (*d)
 513                 return 0;
 514         if (nump) *nump = num;
 515
 516         return type;
 517 }
 518
 519
 520 /*
 521  * convert a major/minor pair for a block device into a name in /dev, if possible.
 522  * On the first call, walk /dev collecting name.
 523  * Put them in a simple linked listfor now.
 524  */
 525 struct devmap {
 526     int major, minor;
 527     char *name;
 528     struct devmap *next;
 529 } *devlist = NULL;
 530 int devlist_ready = 0;
 531
 532 int add_dev(const char *name, const struct stat *stb, int flag, struct FTW *s)
 533 {
 534         struct stat st;
 535
 536         if (S_ISLNK(stb->st_mode)) {
 537                 if (stat(name, &st) != 0)
 538                         return 0;
 539                 stb = &st;
 540         }
 541
 542         if ((stb->st_mode&S_IFMT)== S_IFBLK) {
 543                 char *n = strdup(name);
 544                 struct devmap *dm = malloc(sizeof(*dm));
 545                 if (strncmp(n, "/dev/./", 7)==0)
 546                         strcpy(n+4, name+6);
 547                 if (dm) {
 548                         dm->major = major(stb->st_rdev);
 549                         dm->minor = minor(stb->st_rdev);
 550                         dm->name = n;
 551                         dm->next = devlist;
 552                         devlist = dm;
 553                 }
 554         }
 555         return 0;
 556 }
 557
 558 #ifndef HAVE_NFTW
 559 #ifdef HAVE_FTW
 560 int add_dev_1(const char *name, const struct stat *stb, int flag)
 561 {
 562         return add_dev(name, stb, flag, NULL);
 563 }
 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 ftw(path, add_dev_1, nopenfd);
 567 }
 568 #else
 569 int nftw(const char *path, int (*han)(const char *name, const struct stat *stb, int flag, struct FTW *s), int nopenfd, int flags)
 570 {
 571         return 0;
 572 }
 573 #endif /* HAVE_FTW */
 574 #endif /* HAVE_NFTW */
 575
 576 /*
 577  * Find a block device with the right major/minor number.
 578  * If we find multiple names, choose the shortest.
 579  * If we find a name in /dev/md/, we prefer that.
 580  * This applies only to names for MD devices.
 581  */
 582 char *map_dev(int major, int minor, int create)
 583 {
 584         struct devmap *p;
 585         char *regular = NULL, *preferred=NULL;
 586         int did_check = 0;
 587
 588         if (major == 0 && minor == 0)
 589                         return NULL;
 590
 591  retry:
 592         if (!devlist_ready) {
 593                 char *dev = "/dev";
 594                 struct stat stb;
 595                 while(devlist) {
 596                         struct devmap *d = devlist;
 597                         devlist = d->next;
 598                         free(d->name);
 599                         free(d);
 600                 }
 601                 if (lstat(dev, &stb)==0 &&
 602                     S_ISLNK(stb.st_mode))
 603                         dev = "/dev/.";
 604                 nftw(dev, add_dev, 10, FTW_PHYS);
 605                 devlist_ready=1;
 606                 did_check = 1;
 607         }
 608
 609         for (p=devlist; p; p=p->next)
 610                 if (p->major == major &&
 611                     p->minor == minor) {
 612                         if (strncmp(p->name, "/dev/md/",8) == 0) {
 613                                 if (preferred == NULL ||
 614                                     strlen(p->name) < strlen(preferred))
 615                                         preferred = p->name;
 616                         } else {
 617                                 if (regular == NULL ||
 618                                     strlen(p->name) < strlen(regular))
 619                                         regular = p->name;
 620                         }
 621                 }
 622         if (!regular && !preferred && !did_check) {
 623                 devlist_ready = 0;
 624                 goto retry;
 625         }
 626         if (create && !regular && !preferred) {
 627                 static char buf[30];
 628                 snprintf(buf, sizeof(buf), "%d:%d", major, minor);
 629                 regular = buf;
 630         }
 631
 632         return preferred ? preferred : regular;
 633 }
 634
 635 unsigned long calc_csum(void *super, int bytes)
 636 {
 637         unsigned long long newcsum = 0;
 638         int i;
 639         unsigned int csum;
 640         unsigned int *superc = (unsigned int*) super;
 641
 642         for(i=0; i<bytes/4; i++)
 643                 newcsum+= superc[i];
 644         csum = (newcsum& 0xffffffff) + (newcsum>>32);
 645 #ifdef __alpha__
 646 /* The in-kernel checksum calculation is always 16bit on
 647  * the alpha, though it is 32 bit on i386...
 648  * I wonder what it is elsewhere... (it uses and API in
 649  * a way that it shouldn't).
 650  */
 651         csum = (csum & 0xffff) + (csum >> 16);
 652         csum = (csum & 0xffff) + (csum >> 16);
 653 #endif
 654         return csum;
 655 }
 656
 657 #ifndef MDASSEMBLE
 658 char *human_size(long long bytes)
 659 {
 660         static char buf[30];
 661
 662         /* We convert bytes to either centi-M{ega,ibi}bytes or
 663          * centi-G{igi,ibi}bytes, with appropriate rounding,
 664          * and then print 1/100th of those as a decimal.
 665          * We allow upto 2048Megabytes before converting to
 666          * gigabytes, as that shows more precision and isn't
 667          * too large a number.
 668          * Terrabytes are not yet handled.
 669          */
 670
 671         if (bytes < 5000*1024)
 672                 buf[0]=0;
 673         else if (bytes < 2*1024LL*1024LL*1024LL) {
 674                 long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;
 675                 long cMB  = (bytes / ( 1000000LL / 200LL ) +1) /2;
 676                 snprintf(buf, sizeof(buf), " (%ld.%02ld MiB %ld.%02ld MB)",
 677                         cMiB/100 , cMiB % 100,
 678                         cMB/100, cMB % 100);
 679         } else {
 680                 long cGiB = (bytes / ( (1LL<<30) / 200LL ) +1) /2;
 681                 long cGB  = (bytes / (1000000000LL/200LL ) +1) /2;
 682                 snprintf(buf, sizeof(buf), " (%ld.%02ld GiB %ld.%02ld GB)",
 683                         cGiB/100 , cGiB % 100,
 684                         cGB/100, cGB % 100);
 685         }
 686         return buf;
 687 }
 688
 689 char *human_size_brief(long long bytes)
 690 {
 691         static char buf[30];
 692
 693         if (bytes < 5000*1024)
 694                 snprintf(buf, sizeof(buf), "%ld.%02ldKiB",
 695                         (long)(bytes>>10), (long)(((bytes&1023)*100+512)/1024)
 696                         );
 697         else if (bytes < 2*1024LL*1024LL*1024LL)
 698                 snprintf(buf, sizeof(buf), "%ld.%02ldMiB",
 699                         (long)(bytes>>20),
 700                         (long)((bytes&0xfffff)+0x100000/200)/(0x100000/100)
 701                         );
 702         else
 703                 snprintf(buf, sizeof(buf), "%ld.%02ldGiB",
 704                         (long)(bytes>>30),
 705                         (long)(((bytes>>10)&0xfffff)+0x100000/200)/(0x100000/100)
 706                         );
 707         return buf;
 708 }
 709
 710 void print_r10_layout(int layout)
 711 {
 712         int near = layout & 255;
 713         int far = (layout >> 8) & 255;
 714         int offset = (layout&0x10000);
 715         char *sep = "";
 716
 717         if (near != 1) {
 718                 printf("%s near=%d", sep, near);
 719                 sep = ",";
 720         }
 721         if (far != 1)
 722                 printf("%s %s=%d", sep, offset?"offset":"far", far);
 723         if (near*far == 1)
 724                 printf("NO REDUNDANCY");
 725 }
 726 #endif
 727
 728 unsigned long long calc_array_size(int level, int raid_disks, int layout,
 729                                    int chunksize, unsigned long long devsize)
 730 {
 731         int data_disks = 0;
 732         switch (level) {
 733         case 0: data_disks = raid_disks; break;
 734         case 1: data_disks = 1; break;
 735         case 4:
 736         case 5: data_disks = raid_disks - 1; break;
 737         case 6: data_disks = raid_disks - 2; break;
 738         case 10: data_disks = raid_disks / (layout & 255) / ((layout>>8)&255);
 739                 break;
 740         }
 741         devsize &= ~(unsigned long long)((chunksize>>9)-1);
 742         return data_disks * devsize;
 743 }
 744
 745 int get_mdp_major(void)
 746 {
 747 static int mdp_major = -1;
 748         FILE *fl;
 749         char *w;
 750         int have_block = 0;
 751         int have_devices = 0;
 752         int last_num = -1;
 753
 754         if (mdp_major != -1)
 755                 return mdp_major;
 756         fl = fopen("/proc/devices", "r");
 757         if (!fl)
 758                 return -1;
 759         while ((w = conf_word(fl, 1))) {
 760                 if (have_block && strcmp(w, "devices:")==0)
 761                         have_devices = 1;
 762                 have_block =  (strcmp(w, "Block")==0);
 763                 if (isdigit(w[0]))
 764                         last_num = atoi(w);
 765                 if (have_devices && strcmp(w, "mdp")==0)
 766                         mdp_major = last_num;
 767                 free(w);
 768         }
 769         fclose(fl);
 770         return mdp_major;
 771 }
 772
 773 #if !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO)
 774 char *get_md_name(int dev)
 775 {
 776         /* find /dev/md%d or /dev/md/%d or make a device /dev/.tmp.md%d */
 777         /* if dev < 0, want /dev/md/d%d or find mdp in /proc/devices ... */
 778         static char devname[50];
 779         struct stat stb;
 780         dev_t rdev;
 781         char *dn;
 782
 783         if (dev < 0) {
 784                 int mdp =  get_mdp_major();
 785                 if (mdp < 0) return NULL;
 786                 rdev = makedev(mdp, (-1-dev)<<6);
 787                 snprintf(devname, sizeof(devname), "/dev/md/d%d", -1-dev);
 788                 if (stat(devname, &stb) == 0
 789                     && (S_IFMT&stb.st_mode) == S_IFBLK
 790                     && (stb.st_rdev == rdev))
 791                         return devname;
 792         } else {
 793                 rdev = makedev(MD_MAJOR, dev);
 794                 snprintf(devname, sizeof(devname), "/dev/md%d", dev);
 795                 if (stat(devname, &stb) == 0
 796                     && (S_IFMT&stb.st_mode) == S_IFBLK
 797                     && (stb.st_rdev == rdev))
 798                         return devname;
 799
 800                 snprintf(devname, sizeof(devname), "/dev/md/%d", dev);
 801                 if (stat(devname, &stb) == 0
 802                     && (S_IFMT&stb.st_mode) == S_IFBLK
 803                     && (stb.st_rdev == rdev))
 804                         return devname;
 805         }
 806         dn = map_dev(major(rdev), minor(rdev), 0);
 807         if (dn)
 808                 return dn;
 809         snprintf(devname, sizeof(devname), "/dev/.tmp.md%d", dev);
 810         if (mknod(devname, S_IFBLK | 0600, rdev) == -1)
 811                 if (errno != EEXIST)
 812                         return NULL;
 813
 814         if (stat(devname, &stb) == 0
 815             && (S_IFMT&stb.st_mode) == S_IFBLK
 816             && (stb.st_rdev == rdev))
 817                 return devname;
 818         unlink(devname);
 819         return NULL;
 820 }
 821
 822 void put_md_name(char *name)
 823 {
 824         if (strncmp(name, "/dev/.tmp.md", 12)==0)
 825                 unlink(name);
 826 }
 827
 828 int find_free_devnum(int use_partitions)
 829 {
 830         int devnum;
 831         for (devnum = 127; devnum != 128;
 832              devnum = devnum ? devnum-1 : (1<<20)-1) {
 833                 char *dn;
 834                 int _devnum;
 835
 836                 _devnum = use_partitions ? (-1-devnum) : devnum;
 837                 if (mddev_busy(_devnum))
 838                         continue;
 839                 /* make sure it is new to /dev too, at least as a
 840                  * non-standard */
 841                 dn = map_dev(dev2major(_devnum), dev2minor(_devnum), 0);
 842                 if (dn && ! is_standard(dn, NULL))
 843                         continue;
 844                 break;
 845         }
 846         if (devnum == 128)
 847                 return NoMdDev;
 848         return use_partitions ? (-1-devnum) : devnum;
 849 }
 850 #endif /* !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO) */
 851
 852 int dev_open(char *dev, int flags)
 853 {
 854         /* like 'open', but if 'dev' matches %d:%d, create a temp
 855          * block device and open that
 856          */
 857         char *e;
 858         int fd = -1;
 859         char devname[32];
 860         int major;
 861         int minor;
 862
 863         if (!dev) return -1;
 864
 865         major = strtoul(dev, &e, 0);
 866         if (e > dev && *e == ':' && e[1] &&
 867             (minor = strtoul(e+1, &e, 0)) >= 0 &&
 868             *e == 0) {
 869                 snprintf(devname, sizeof(devname), "/dev/.tmp.md.%d:%d:%d",
 870                          (int)getpid(), major, minor);
 871                 if (mknod(devname, S_IFBLK|0600, makedev(major, minor))==0) {
 872                         fd = open(devname, flags|O_DIRECT);
 873                         unlink(devname);
 874                 }
 875         } else
 876                 fd = open(dev, flags|O_DIRECT);
 877         return fd;
 878 }
 879
 880 int open_dev(int devnum)
 881 {
 882         char buf[20];
 883
 884         sprintf(buf, "%d:%d", dev2major(devnum), dev2minor(devnum));
 885         return dev_open(buf, O_RDWR);
 886 }
 887
 888 int open_dev_excl(int devnum)
 889 {
 890         char buf[20];
 891         int i;
 892
 893         sprintf(buf, "%d:%d", dev2major(devnum), dev2minor(devnum));
 894         for (i=0 ; i<25 ; i++) {
 895                 int fd = dev_open(buf, O_RDWR|O_EXCL);
 896                 if (fd >= 0)
 897                         return fd;
 898                 if (errno != EBUSY)
 899                         return fd;
 900                 usleep(200000);
 901         }
 902         return -1;
 903 }
 904
 905 int same_dev(char *one, char *two)
 906 {
 907         struct stat st1, st2;
 908         if (stat(one, &st1) != 0)
 909                 return 0;
 910         if (stat(two, &st2) != 0)
 911                 return 0;
 912         if ((st1.st_mode & S_IFMT) != S_IFBLK)
 913                 return 0;
 914         if ((st2.st_mode & S_IFMT) != S_IFBLK)
 915                 return 0;
 916         return st1.st_rdev == st2.st_rdev;
 917 }
 918
 919 void wait_for(char *dev, int fd)
 920 {
 921         int i;
 922         struct stat stb_want;
 923
 924         if (fstat(fd, &stb_want) != 0 ||
 925             (stb_want.st_mode & S_IFMT) != S_IFBLK)
 926                 return;
 927
 928         for (i=0 ; i<25 ; i++) {
 929                 struct stat stb;
 930                 if (stat(dev, &stb) == 0 &&
 931                     (stb.st_mode & S_IFMT) == S_IFBLK &&
 932                     (stb.st_rdev == stb_want.st_rdev))
 933                         return;
 934                 usleep(200000);
 935         }
 936 }
 937
 938 struct superswitch *superlist[] = { &super0, &super1, &super_ddf, &super_imsm, NULL };
 939
 940 #if !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO)
 941
 942 struct supertype *super_by_fd(int fd)
 943 {
 944         mdu_array_info_t array;
 945         int vers;
 946         int minor;
 947         struct supertype *st = NULL;
 948         struct mdinfo *sra;
 949         char *verstr;
 950         char version[20];
 951         int i;
 952         char *subarray = NULL;
 953
 954         sra = sysfs_read(fd, 0, GET_VERSION);
 955
 956         if (sra) {
 957                 vers = sra->array.major_version;
 958                 minor = sra->array.minor_version;
 959                 verstr = sra->text_version;
 960         } else {
 961                 if (ioctl(fd, GET_ARRAY_INFO, &array))
 962                         array.major_version = array.minor_version = 0;
 963                 vers = array.major_version;
 964                 minor = array.minor_version;
 965                 verstr = "";
 966         }
 967
 968         if (vers != -1) {
 969                 sprintf(version, "%d.%d", vers, minor);
 970                 verstr = version;
 971         }
 972         if (minor == -2 && is_subarray(verstr)) {
 973                 char *dev = verstr+1;
 974                 subarray = strchr(dev, '/');
 975                 int devnum;
 976                 if (subarray)
 977                         *subarray++ = '\0';
 978                 devnum = devname2devnum(dev);
 979                 subarray = strdup(subarray);
 980                 if (sra)
 981                         sysfs_free(sra);
 982                 sra = sysfs_read(-1, devnum, GET_VERSION);
 983                 if (sra && sra->text_version[0])
 984                         verstr = sra->text_version;
 985                 else
 986                         verstr = "-no-metadata-";
 987         }
 988
 989         for (i = 0; st == NULL && superlist[i] ; i++)
 990                 st = superlist[i]->match_metadata_desc(verstr);
 991
 992         if (sra)
 993                 sysfs_free(sra);
 994         if (st) {
 995                 st->sb = NULL;
 996                 if (subarray) {
 997                         strncpy(st->subarray, subarray, 32);
 998                         st->subarray[31] = 0;
 999                         free(subarray);
1000                 } else
1001                         st->subarray[0] = 0;
1002         }
1003         return st;
1004 }
1005 #endif /* !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO) */
1006
1007
1008 struct supertype *dup_super(struct supertype *orig)
1009 {
1010         struct supertype *st;
1011
1012         if (!orig)
1013                 return orig;
1014         st = malloc(sizeof(*st));
1015         if (!st)
1016                 return st;
1017         memset(st, 0, sizeof(*st));
1018         st->ss = orig->ss;
1019         st->max_devs = orig->max_devs;
1020         st->minor_version = orig->minor_version;
1021         strcpy(st->subarray, orig->subarray);
1022         st->sb = NULL;
1023         st->info = NULL;
1024         return st;
1025 }
1026
1027 struct supertype *guess_super(int fd)
1028 {
1029         /* try each load_super to find the best match,
1030          * and return the best superswitch
1031          */
1032         struct superswitch  *ss;
1033         struct supertype *st;
1034         unsigned long besttime = 0;
1035         int bestsuper = -1;
1036         int i;
1037
1038         st = malloc(sizeof(*st));
1039         for (i=0 ; superlist[i]; i++) {
1040                 int rv;
1041                 ss = superlist[i];
1042                 memset(st, 0, sizeof(*st));
1043                 rv = ss->load_super(st, fd, NULL);
1044                 if (rv == 0) {
1045                         struct mdinfo info;
1046                         st->ss->getinfo_super(st, &info);
1047                         if (bestsuper == -1 ||
1048                             besttime < info.array.ctime) {
1049                                 bestsuper = i;
1050                                 besttime = info.array.ctime;
1051                         }
1052                         ss->free_super(st);
1053                 }
1054         }
1055         if (bestsuper != -1) {
1056                 int rv;
1057                 memset(st, 0, sizeof(*st));
1058                 rv = superlist[bestsuper]->load_super(st, fd, NULL);
1059                 if (rv == 0) {
1060                         superlist[bestsuper]->free_super(st);
1061                         return st;
1062                 }
1063         }
1064         free(st);
1065         return NULL;
1066 }
1067
1068 /* Return size of device in bytes */
1069 int get_dev_size(int fd, char *dname, unsigned long long *sizep)
1070 {
1071         unsigned long long ldsize;
1072         struct stat st;
1073
1074         if (fstat(fd, &st) != -1 && S_ISREG(st.st_mode))
1075                 ldsize = (unsigned long long)st.st_size;
1076         else
1077 #ifdef BLKGETSIZE64
1078         if (ioctl(fd, BLKGETSIZE64, &ldsize) != 0)
1079 #endif
1080         {
1081                 unsigned long dsize;
1082                 if (ioctl(fd, BLKGETSIZE, &dsize) == 0) {
1083                         ldsize = dsize;
1084                         ldsize <<= 9;
1085                 } else {
1086                         if (dname)
1087                                 fprintf(stderr, Name ": Cannot get size of %s: %s\b",
1088                                         dname, strerror(errno));
1089                         return 0;
1090                 }
1091         }
1092         *sizep = ldsize;
1093         return 1;
1094 }
1095
1096 void get_one_disk(int mdfd, mdu_array_info_t *ainf, mdu_disk_info_t *disk)
1097 {
1098         int d;
1099         ioctl(mdfd, GET_ARRAY_INFO, ainf);
1100         for (d = 0 ; d < ainf->raid_disks + ainf->nr_disks ; d++)
1101                 if (ioctl(mdfd, GET_DISK_INFO, disk) == 0)
1102                         return;
1103 }
1104
1105 int open_container(int fd)
1106 {
1107         /* 'fd' is a block device.  Find out if it is in use
1108          * by a container, and return an open fd on that container.
1109          */
1110         char path[256];
1111         char *e;
1112         DIR *dir;
1113         struct dirent *de;
1114         int dfd, n;
1115         char buf[200];
1116         int major, minor;
1117         struct stat st;
1118
1119         if (fstat(fd, &st) != 0)
1120                 return -1;
1121         sprintf(path, "/sys/dev/block/%d:%d/holders",
1122                 (int)major(st.st_rdev), (int)minor(st.st_rdev));
1123         e = path + strlen(path);
1124
1125         dir = opendir(path);
1126         if (!dir)
1127                 return -1;
1128         while ((de = readdir(dir))) {
1129                 if (de->d_ino == 0)
1130                         continue;
1131                 if (de->d_name[0] == '.')
1132                         continue;
1133                 sprintf(e, "/%s/dev", de->d_name);
1134                 dfd = open(path, O_RDONLY);
1135                 if (dfd < 0)
1136                         continue;
1137                 n = read(dfd, buf, sizeof(buf));
1138                 close(dfd);
1139                 if (n <= 0 || n >= sizeof(buf))
1140                         continue;
1141                 buf[n] = 0;
1142                 if (sscanf(buf, "%d:%d", &major, &minor) != 2)
1143                         continue;
1144                 sprintf(buf, "%d:%d", major, minor);
1145                 dfd = dev_open(buf, O_RDONLY);
1146                 if (dfd >= 0) {
1147                         closedir(dir);
1148                         return dfd;
1149                 }
1150         }
1151         closedir(dir);
1152         return -1;
1153 }
1154
1155 int add_disk(int mdfd, struct supertype *st,
1156              struct mdinfo *sra, struct mdinfo *info)
1157 {
1158         /* Add a device to an array, in one of 2 ways. */
1159         int rv;
1160 #ifndef MDASSEMBLE
1161         if (st->ss->external) {
1162                 rv = sysfs_add_disk(sra, info,
1163                                     info->disk.state & (1<<MD_DISK_SYNC));
1164                 if (! rv) {
1165                         struct mdinfo *sd2;
1166                         for (sd2 = sra->devs; sd2; sd2=sd2->next)
1167                                 if (sd2 == info)
1168                                         break;
1169                         if (sd2 == NULL) {
1170                                 sd2 = malloc(sizeof(*sd2));
1171                                 *sd2 = *info;
1172                                 sd2->next = sra->devs;
1173                                 sra->devs = sd2;
1174                         }
1175                 }
1176         } else
1177 #endif
1178                 rv = ioctl(mdfd, ADD_NEW_DISK, &info->disk);
1179         return rv;
1180 }
1181
1182 int set_array_info(int mdfd, struct supertype *st, struct mdinfo *info)
1183 {
1184         /* Initialise kernel's knowledge of array.
1185          * This varies between externally managed arrays
1186          * and older kernels
1187          */
1188         int vers = md_get_version(mdfd);
1189         int rv;
1190
1191 #ifndef MDASSEMBLE
1192         if (st->ss->external)
1193                 rv = sysfs_set_array(info, vers);
1194         else
1195 #endif
1196                 if ((vers % 100) >= 1) { /* can use different versions */
1197                 mdu_array_info_t inf;
1198                 memset(&inf, 0, sizeof(inf));
1199                 inf.major_version = info->array.major_version;
1200                 inf.minor_version = info->array.minor_version;
1201                 rv = ioctl(mdfd, SET_ARRAY_INFO, &inf);
1202         } else
1203                 rv = ioctl(mdfd, SET_ARRAY_INFO, NULL);
1204         return rv;
1205 }
1206
1207 char *devnum2devname(int num)
1208 {
1209         char name[100];
1210         if (num > 0)
1211                 sprintf(name, "md%d", num);
1212         else
1213                 sprintf(name, "md_d%d", -1-num);
1214         return strdup(name);
1215 }
1216
1217 int devname2devnum(char *name)
1218 {
1219         char *ep;
1220         int num;
1221         if (strncmp(name, "md_d", 4)==0)
1222                 num = -1-strtoul(name+4, &ep, 10);
1223         else
1224                 num = strtoul(name+2, &ep, 10);
1225         return num;
1226 }
1227
1228 int stat2devnum(struct stat *st)
1229 {
1230         char path[30];
1231         char link[200];
1232         char *cp;
1233         int n;
1234
1235         if ((S_IFMT & st->st_mode) == S_IFBLK) {
1236                 if (major(st->st_rdev) == MD_MAJOR)
1237                         return minor(st->st_rdev);
1238                 else if (major(st->st_rdev) == get_mdp_major())
1239                         return -1- (minor(st->st_rdev)>>MdpMinorShift);
1240
1241                 /* must be an extended-minor partition. Look at the
1242                  * /sys/dev/block/%d:%d link which must look like
1243                  * ../../block/mdXXX/mdXXXpYY
1244                  */
1245                 sprintf(path, "/sys/dev/block/%d:%d", major(st->st_rdev),
1246                         minor(st->st_rdev));
1247                 n = readlink(path, link, sizeof(link)-1);
1248                 if (n <= 0)
1249                         return NoMdDev;
1250                 link[n] = 0;
1251                 cp = strrchr(link, '/');
1252                 if (cp) *cp = 0;
1253                 cp = strchr(link, '/');
1254                 if (cp && strncmp(cp, "/md", 3) == 0)
1255                         return devname2devnum(cp+1);
1256         }
1257         return NoMdDev;
1258
1259 }
1260
1261 int fd2devnum(int fd)
1262 {
1263         struct stat stb;
1264         if (fstat(fd, &stb) == 0)
1265                 return stat2devnum(&stb);
1266         return NoMdDev;
1267 }
1268
1269 int mdmon_running(int devnum)
1270 {
1271         char path[100];
1272         char pid[10];
1273         int fd;
1274         int n;
1275         sprintf(path, "/var/run/mdadm/%s.pid", devnum2devname(devnum));
1276         fd = open(path, O_RDONLY, 0);
1277
1278         if (fd < 0)
1279                 return 0;
1280         n = read(fd, pid, 9);
1281         close(fd);
1282         if (n <= 0)
1283                 return 0;
1284         if (kill(atoi(pid), 0) == 0)
1285                 return 1;
1286         return 0;
1287 }
1288
1289 int signal_mdmon(int devnum)
1290 {
1291         char path[100];
1292         char pid[10];
1293         int fd;
1294         int n;
1295         sprintf(path, "/var/run/mdadm/%s.pid", devnum2devname(devnum));
1296         fd = open(path, O_RDONLY, 0);
1297
1298         if (fd < 0)
1299                 return 0;
1300         n = read(fd, pid, 9);
1301         close(fd);
1302         if (n <= 0)
1303                 return 0;
1304         if (kill(atoi(pid), SIGUSR1) == 0)
1305                 return 1;
1306         return 0;
1307 }
1308
1309 int start_mdmon(int devnum)
1310 {
1311         int i;
1312         int len;
1313         pid_t pid;
1314         int status;
1315         char pathbuf[1024];
1316         char *paths[4] = {
1317                 pathbuf,
1318                 "/sbin/mdmon",
1319                 "mdmon",
1320                 NULL
1321         };
1322
1323         if (check_env("MDADM_NO_MDMON"))
1324                 return 0;
1325
1326         len = readlink("/proc/self/exe", pathbuf, sizeof(pathbuf));
1327         if (len > 0) {
1328                 char *sl;
1329                 pathbuf[len] = 0;
1330                 sl = strrchr(pathbuf, '/');
1331                 if (sl)
1332                         sl++;
1333                 else
1334                         sl = pathbuf;
1335                 strcpy(sl, "mdmon");
1336         } else
1337                 pathbuf[0] = '\0';
1338
1339         switch(fork()) {
1340         case 0:
1341                 /* FIXME yuk. CLOSE_EXEC?? */
1342                 for (i=3; i < 100; i++)
1343                         close(i);
1344                 for (i=0; paths[i]; i++)
1345                         if (paths[i][0])
1346                                 execl(paths[i], "mdmon",
1347                                       devnum2devname(devnum),
1348                                       NULL);
1349                 exit(1);
1350         case -1: fprintf(stderr, Name ": cannot run mdmon. "
1351                          "Array remains readonly\n");
1352                 return -1;
1353         default: /* parent - good */
1354                 pid = wait(&status);
1355                 if (pid < 0 || status != 0)
1356                         return -1;
1357         }
1358         return 0;
1359 }
1360
1361 int check_env(char *name)
1362 {
1363         char *val = getenv(name);
1364
1365         if (val && atoi(val) == 1)
1366                 return 1;
1367
1368         return 0;
1369 }
1370
1371 #ifndef MDASSEMBLE
1372 int flush_metadata_updates(struct supertype *st)
1373 {
1374         int sfd;
1375         if (!st->updates) {
1376                 st->update_tail = NULL;
1377                 return -1;
1378         }
1379
1380         sfd = connect_monitor(devnum2devname(st->container_dev));
1381         if (sfd < 0)
1382                 return -1;
1383
1384         while (st->updates) {
1385                 struct metadata_update *mu = st->updates;
1386                 st->updates = mu->next;
1387
1388                 send_message(sfd, mu, 0);
1389                 wait_reply(sfd, 0);
1390                 free(mu->buf);
1391                 free(mu);
1392         }
1393         ack(sfd, 0);
1394         wait_reply(sfd, 0);
1395         close(sfd);
1396         st->update_tail = NULL;
1397         return 0;
1398 }
1399
1400 void append_metadata_update(struct supertype *st, void *buf, int len)
1401 {
1402
1403         struct metadata_update *mu = malloc(sizeof(*mu));
1404
1405         mu->buf = buf;
1406         mu->len = len;
1407         mu->space = NULL;
1408         mu->next = NULL;
1409         *st->update_tail = mu;
1410         st->update_tail = &mu->next;
1411 }
1412 #endif /* MDASSEMBLE */
1413
1414 #ifdef __TINYC__
1415 /* tinyc doesn't optimize this check in ioctl.h out ... */
1416 unsigned int __invalid_size_argument_for_IOC = 0;
1417 #endif
1418