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