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         if (S_ISLNK(stb->st_mode)) {
 471                 stat(name, &st);
 472                 stb = &st;
 473         }
 474
 475         if ((stb->st_mode&S_IFMT)== S_IFBLK) {
 476                 char *n = strdup(name);
 477                 struct devmap *dm = malloc(sizeof(*dm));
 478                 if (strncmp(n, "/dev/./", 7)==0)
 479                         strcpy(n+4, name+6);
 480                 if (dm) {
 481                         dm->major = major(stb->st_rdev);
 482                         dm->minor = minor(stb->st_rdev);
 483                         dm->name = n;
 484                         dm->next = devlist;
 485                         devlist = dm;
 486                 }
 487         }
 488         return 0;
 489 }
 490
 491 #ifndef HAVE_NFTW
 492 #ifdef HAVE_FTW
 493 int add_dev_1(const char *name, const struct stat *stb, int flag)
 494 {
 495         return add_dev(name, stb, flag, NULL);
 496 }
 497 int nftw(const char *path, int (*han)(const char *name, const struct stat *stb, int flag, struct FTW *s), int nopenfd, int flags)
 498 {
 499         return ftw(path, add_dev_1, nopenfd);
 500 }
 501 #else
 502 int nftw(const char *path, int (*han)(const char *name, const struct stat *stb, int flag, struct FTW *s), int nopenfd, int flags)
 503 {
 504         return 0;
 505 }
 506 #endif /* HAVE_FTW */
 507 #endif /* HAVE_NFTW */
 508
 509 /*
 510  * Find a block device with the right major/minor number.
 511  * If we find multiple names, choose the shortest.
 512  * If we find a non-standard name, it is probably there
 513  * deliberately so prefer it over a standard name.
 514  * This applies only to names for MD devices.
 515  */
 516 char *map_dev(int major, int minor, int create)
 517 {
 518         struct devmap *p;
 519         char *std = NULL, *nonstd=NULL;
 520         int did_check = 0;
 521
 522         if (major == 0 && minor == 0)
 523                         return NULL;
 524
 525  retry:
 526         if (!devlist_ready) {
 527                 char *dev = "/dev";
 528                 struct stat stb;
 529                 while(devlist) {
 530                         struct devmap *d = devlist;
 531                         devlist = d->next;
 532                         free(d->name);
 533                         free(d);
 534                 }
 535                 if (lstat(dev, &stb)==0 &&
 536                     S_ISLNK(stb.st_mode))
 537                         dev = "/dev/.";
 538                 nftw(dev, add_dev, 10, FTW_PHYS);
 539                 devlist_ready=1;
 540                 did_check = 1;
 541         }
 542
 543         for (p=devlist; p; p=p->next)
 544                 if (p->major == major &&
 545                     p->minor == minor) {
 546                         if (is_standard(p->name, NULL)) {
 547                                 if (std == NULL ||
 548                                     strlen(p->name) < strlen(std))
 549                                         std = p->name;
 550                         } else {
 551                                 if (nonstd == NULL ||
 552                                     strlen(p->name) < strlen(nonstd))
 553                                         nonstd = p->name;
 554                         }
 555                 }
 556         if (!std && !nonstd && !did_check) {
 557                 devlist_ready = 0;
 558                 goto retry;
 559         }
 560         if (create && !std && !nonstd) {
 561                 static char buf[30];
 562                 snprintf(buf, sizeof(buf), "%d:%d", major, minor);
 563                 nonstd = buf;
 564         }
 565
 566         return nonstd ? nonstd : std;
 567 }
 568
 569 unsigned long calc_csum(void *super, int bytes)
 570 {
 571         unsigned long long newcsum = 0;
 572         int i;
 573         unsigned int csum;
 574         unsigned int *superc = (unsigned int*) super;
 575
 576         for(i=0; i<bytes/4; i++)
 577                 newcsum+= superc[i];
 578         csum = (newcsum& 0xffffffff) + (newcsum>>32);
 579 #ifdef __alpha__
 580 /* The in-kernel checksum calculation is always 16bit on
 581  * the alpha, though it is 32 bit on i386...
 582  * I wonder what it is elsewhere... (it uses and API in
 583  * a way that it shouldn't).
 584  */
 585         csum = (csum & 0xffff) + (csum >> 16);
 586         csum = (csum & 0xffff) + (csum >> 16);
 587 #endif
 588         return csum;
 589 }
 590
 591 #ifndef MDASSEMBLE
 592 char *human_size(long long bytes)
 593 {
 594         static char buf[30];
 595
 596         /* We convert bytes to either centi-M{ega,ibi}bytes or
 597          * centi-G{igi,ibi}bytes, with appropriate rounding,
 598          * and then print 1/100th of those as a decimal.
 599          * We allow upto 2048Megabytes before converting to
 600          * gigabytes, as that shows more precision and isn't
 601          * too large a number.
 602          * Terrabytes are not yet handled.
 603          */
 604
 605         if (bytes < 5000*1024)
 606                 buf[0]=0;
 607         else if (bytes < 2*1024LL*1024LL*1024LL) {
 608                 long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;
 609                 long cMB  = (bytes / ( 1000000LL / 200LL ) +1) /2;
 610                 snprintf(buf, sizeof(buf), " (%ld.%02ld MiB %ld.%02ld MB)",
 611                         cMiB/100 , cMiB % 100,
 612                         cMB/100, cMB % 100);
 613         } else {
 614                 long cGiB = (bytes / ( (1LL<<30) / 200LL ) +1) /2;
 615                 long cGB  = (bytes / (1000000000LL/200LL ) +1) /2;
 616                 snprintf(buf, sizeof(buf), " (%ld.%02ld GiB %ld.%02ld GB)",
 617                         cGiB/100 , cGiB % 100,
 618                         cGB/100, cGB % 100);
 619         }
 620         return buf;
 621 }
 622
 623 char *human_size_brief(long long bytes)
 624 {
 625         static char buf[30];
 626
 627         if (bytes < 5000*1024)
 628                 snprintf(buf, sizeof(buf), "%ld.%02ldKiB",
 629                         (long)(bytes>>10), (long)(((bytes&1023)*100+512)/1024)
 630                         );
 631         else if (bytes < 2*1024LL*1024LL*1024LL)
 632                 snprintf(buf, sizeof(buf), "%ld.%02ldMiB",
 633                         (long)(bytes>>20),
 634                         (long)((bytes&0xfffff)+0x100000/200)/(0x100000/100)
 635                         );
 636         else
 637                 snprintf(buf, sizeof(buf), "%ld.%02ldGiB",
 638                         (long)(bytes>>30),
 639                         (long)(((bytes>>10)&0xfffff)+0x100000/200)/(0x100000/100)
 640                         );
 641         return buf;
 642 }
 643
 644 void print_r10_layout(int layout)
 645 {
 646         int near = layout & 255;
 647         int far = (layout >> 8) & 255;
 648         int offset = (layout&0x10000);
 649         char *sep = "";
 650
 651         if (near != 1) {
 652                 printf("%s near=%d", sep, near);
 653                 sep = ",";
 654         }
 655         if (far != 1)
 656                 printf("%s %s=%d", sep, offset?"offset":"far", far);
 657         if (near*far == 1)
 658                 printf("NO REDUNDANCY");
 659 }
 660 #endif
 661
 662 unsigned long long calc_array_size(int level, int raid_disks, int layout,
 663                                    int chunksize, unsigned long long devsize)
 664 {
 665         int data_disks = 0;
 666         switch (level) {
 667         case 0: data_disks = raid_disks; break;
 668         case 1: data_disks = 1; break;
 669         case 4:
 670         case 5: data_disks = raid_disks - 1; break;
 671         case 6: data_disks = raid_disks - 2; break;
 672         case 10: data_disks = raid_disks / (layout & 255) / ((layout>>8)&255);
 673                 break;
 674         }
 675         devsize &= ~(unsigned long long)((chunksize>>9)-1);
 676         return data_disks * devsize;
 677 }
 678
 679 int get_mdp_major(void)
 680 {
 681 static int mdp_major = -1;
 682         FILE *fl;
 683         char *w;
 684         int have_block = 0;
 685         int have_devices = 0;
 686         int last_num = -1;
 687
 688         if (mdp_major != -1)
 689                 return mdp_major;
 690         fl = fopen("/proc/devices", "r");
 691         if (!fl)
 692                 return -1;
 693         while ((w = conf_word(fl, 1))) {
 694                 if (have_block && strcmp(w, "devices:")==0)
 695                         have_devices = 1;
 696                 have_block =  (strcmp(w, "Block")==0);
 697                 if (isdigit(w[0]))
 698                         last_num = atoi(w);
 699                 if (have_devices && strcmp(w, "mdp")==0)
 700                         mdp_major = last_num;
 701                 free(w);
 702         }
 703         fclose(fl);
 704         return mdp_major;
 705 }
 706
 707 #if !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO)
 708 char *get_md_name(int dev)
 709 {
 710         /* find /dev/md%d or /dev/md/%d or make a device /dev/.tmp.md%d */
 711         /* if dev < 0, want /dev/md/d%d or find mdp in /proc/devices ... */
 712         static char devname[50];
 713         struct stat stb;
 714         dev_t rdev;
 715         char *dn;
 716
 717         if (dev < 0) {
 718                 int mdp =  get_mdp_major();
 719                 if (mdp < 0) return NULL;
 720                 rdev = makedev(mdp, (-1-dev)<<6);
 721                 snprintf(devname, sizeof(devname), "/dev/md/d%d", -1-dev);
 722                 if (stat(devname, &stb) == 0
 723                     && (S_IFMT&stb.st_mode) == S_IFBLK
 724                     && (stb.st_rdev == rdev))
 725                         return devname;
 726         } else {
 727                 rdev = makedev(MD_MAJOR, dev);
 728                 snprintf(devname, sizeof(devname), "/dev/md%d", dev);
 729                 if (stat(devname, &stb) == 0
 730                     && (S_IFMT&stb.st_mode) == S_IFBLK
 731                     && (stb.st_rdev == rdev))
 732                         return devname;
 733
 734                 snprintf(devname, sizeof(devname), "/dev/md/%d", dev);
 735                 if (stat(devname, &stb) == 0
 736                     && (S_IFMT&stb.st_mode) == S_IFBLK
 737                     && (stb.st_rdev == rdev))
 738                         return devname;
 739         }
 740         dn = map_dev(major(rdev), minor(rdev), 0);
 741         if (dn)
 742                 return dn;
 743         snprintf(devname, sizeof(devname), "/dev/.tmp.md%d", dev);
 744         if (mknod(devname, S_IFBLK | 0600, rdev) == -1)
 745                 if (errno != EEXIST)
 746                         return NULL;
 747
 748         if (stat(devname, &stb) == 0
 749             && (S_IFMT&stb.st_mode) == S_IFBLK
 750             && (stb.st_rdev == rdev))
 751                 return devname;
 752         unlink(devname);
 753         return NULL;
 754 }
 755
 756 void put_md_name(char *name)
 757 {
 758         if (strncmp(name, "/dev/.tmp.md", 12)==0)
 759                 unlink(name);
 760 }
 761
 762 int find_free_devnum(int use_partitions)
 763 {
 764         int devnum;
 765         for (devnum = 127; devnum != 128;
 766              devnum = devnum ? devnum-1 : (1<<22)-1) {
 767                 char *dn;
 768                 int _devnum;
 769
 770                 _devnum = use_partitions ? (-1-devnum) : devnum;
 771                 if (mddev_busy(_devnum))
 772                         continue;
 773                 /* make sure it is new to /dev too, at least as a
 774                  * non-standard */
 775                 dn = map_dev(dev2major(_devnum), dev2minor(_devnum), 0);
 776                 if (dn && ! is_standard(dn, NULL))
 777                         continue;
 778                 break;
 779         }
 780         if (devnum == 128)
 781                 return NoMdDev;
 782         return use_partitions ? (-1-devnum) : devnum;
 783 }
 784 #endif /* !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO) */
 785
 786 int dev_open(char *dev, int flags)
 787 {
 788         /* like 'open', but if 'dev' matches %d:%d, create a temp
 789          * block device and open that
 790          */
 791         char *e;
 792         int fd = -1;
 793         char devname[32];
 794         int major;
 795         int minor;
 796
 797         if (!dev) return -1;
 798
 799         major = strtoul(dev, &e, 0);
 800         if (e > dev && *e == ':' && e[1] &&
 801             (minor = strtoul(e+1, &e, 0)) >= 0 &&
 802             *e == 0) {
 803                 snprintf(devname, sizeof(devname), "/dev/.tmp.md.%d:%d:%d",
 804                          (int)getpid(), major, minor);
 805                 if (mknod(devname, S_IFBLK|0600, makedev(major, minor))==0) {
 806                         fd = open(devname, flags|O_DIRECT);
 807                         unlink(devname);
 808                 }
 809         } else
 810                 fd = open(dev, flags|O_DIRECT);
 811         return fd;
 812 }
 813
 814 int open_dev_excl(int devnum)
 815 {
 816         char buf[20];
 817         int i;
 818
 819         sprintf(buf, "%d:%d", dev2major(devnum), dev2minor(devnum));
 820         for (i=0 ; i<25 ; i++) {
 821                 int fd = dev_open(buf, O_RDWR|O_EXCL);
 822                 if (fd >= 0)
 823                         return fd;
 824                 if (errno != EBUSY)
 825                         return fd;
 826                 usleep(200000);
 827         }
 828         return -1;
 829 }
 830
 831 struct superswitch *superlist[] = { &super0, &super1, &super_ddf, &super_imsm, NULL };
 832
 833 #if !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO)
 834
 835 struct supertype *super_by_fd(int fd)
 836 {
 837         mdu_array_info_t array;
 838         int vers;
 839         int minor;
 840         struct supertype *st = NULL;
 841         struct mdinfo *sra;
 842         char *verstr;
 843         char version[20];
 844         int i;
 845         char *subarray = NULL;
 846
 847         sra = sysfs_read(fd, 0, GET_VERSION);
 848
 849         if (sra) {
 850                 vers = sra->array.major_version;
 851                 minor = sra->array.minor_version;
 852                 verstr = sra->text_version;
 853         } else {
 854                 if (ioctl(fd, GET_ARRAY_INFO, &array))
 855                         array.major_version = array.minor_version = 0;
 856                 vers = array.major_version;
 857                 minor = array.minor_version;
 858                 verstr = "";
 859         }
 860
 861         if (vers != -1) {
 862                 sprintf(version, "%d.%d", vers, minor);
 863                 verstr = version;
 864         }
 865         if (minor == -2 && is_subarray(verstr)) {
 866                 char *dev = verstr+1;
 867                 subarray = strchr(dev, '/');
 868                 int devnum;
 869                 if (subarray)
 870                         *subarray++ = '\0';
 871                 devnum = devname2devnum(dev);
 872                 subarray = strdup(subarray);
 873                 if (sra)
 874                         sysfs_free(sra);
 875                 sra = sysfs_read(-1, devnum, GET_VERSION);
 876                 verstr = sra->text_version ? : "-no-metadata-";
 877         }
 878
 879         for (i = 0; st == NULL && superlist[i] ; i++)
 880                 st = superlist[i]->match_metadata_desc(verstr);
 881
 882         if (sra)
 883                 sysfs_free(sra);
 884         if (st) {
 885                 st->sb = NULL;
 886                 if (subarray) {
 887                         strncpy(st->subarray, subarray, 32);
 888                         st->subarray[31] = 0;
 889                         free(subarray);
 890                 } else
 891                         st->subarray[0] = 0;
 892         }
 893         return st;
 894 }
 895 #endif /* !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO) */
 896
 897
 898 struct supertype *dup_super(struct supertype *orig)
 899 {
 900         struct supertype *st;
 901
 902         if (!orig)
 903                 return orig;
 904         st = malloc(sizeof(*st));
 905         if (!st)
 906                 return st;
 907         memset(st, 0, sizeof(*st));
 908         st->ss = orig->ss;
 909         st->max_devs = orig->max_devs;
 910         st->minor_version = orig->minor_version;
 911         strcpy(st->subarray, orig->subarray);
 912         st->sb = NULL;
 913         st->info = NULL;
 914         return st;
 915 }
 916
 917 struct supertype *guess_super(int fd)
 918 {
 919         /* try each load_super to find the best match,
 920          * and return the best superswitch
 921          */
 922         struct superswitch  *ss;
 923         struct supertype *st;
 924         unsigned long besttime = 0;
 925         int bestsuper = -1;
 926         int i;
 927
 928         st = malloc(sizeof(*st));
 929         for (i=0 ; superlist[i]; i++) {
 930                 int rv;
 931                 ss = superlist[i];
 932                 memset(st, 0, sizeof(*st));
 933                 rv = ss->load_super(st, fd, NULL);
 934                 if (rv == 0) {
 935                         struct mdinfo info;
 936                         st->ss->getinfo_super(st, &info);
 937                         if (bestsuper == -1 ||
 938                             besttime < info.array.ctime) {
 939                                 bestsuper = i;
 940                                 besttime = info.array.ctime;
 941                         }
 942                         ss->free_super(st);
 943                 }
 944         }
 945         if (bestsuper != -1) {
 946                 int rv;
 947                 memset(st, 0, sizeof(*st));
 948                 rv = superlist[bestsuper]->load_super(st, fd, NULL);
 949                 if (rv == 0) {
 950                         superlist[bestsuper]->free_super(st);
 951                         return st;
 952                 }
 953         }
 954         free(st);
 955         return NULL;
 956 }
 957
 958 /* Return size of device in bytes */
 959 int get_dev_size(int fd, char *dname, unsigned long long *sizep)
 960 {
 961         unsigned long long ldsize;
 962         struct stat st;
 963
 964         if (fstat(fd, &st) != -1 && S_ISREG(st.st_mode))
 965                 ldsize = (unsigned long long)st.st_size;
 966         else
 967 #ifdef BLKGETSIZE64
 968         if (ioctl(fd, BLKGETSIZE64, &ldsize) != 0)
 969 #endif
 970         {
 971                 unsigned long dsize;
 972                 if (ioctl(fd, BLKGETSIZE, &dsize) == 0) {
 973                         ldsize = dsize;
 974                         ldsize <<= 9;
 975                 } else {
 976                         if (dname)
 977                                 fprintf(stderr, Name ": Cannot get size of %s: %s\b",
 978                                         dname, strerror(errno));
 979                         return 0;
 980                 }
 981         }
 982         *sizep = ldsize;
 983         return 1;
 984 }
 985
 986 void get_one_disk(int mdfd, mdu_array_info_t *ainf, mdu_disk_info_t *disk)
 987 {
 988         int d;
 989         ioctl(mdfd, GET_ARRAY_INFO, ainf);
 990         for (d = 0 ; d < ainf->raid_disks + ainf->nr_disks ; d++)
 991                 if (ioctl(mdfd, GET_DISK_INFO, disk) == 0)
 992                         return;
 993 }
 994
 995 int open_container(int fd)
 996 {
 997         /* 'fd' is a block device.  Find out if it is in use
 998          * by a container, and return an open fd on that container.
 999          */
1000         char path[256];
1001         char *e;
1002         DIR *dir;
1003         struct dirent *de;
1004         int dfd, n;
1005         char buf[200];
1006         int major, minor;
1007         struct stat st;
1008
1009         if (fstat(fd, &st) != 0)
1010                 return -1;
1011         sprintf(path, "/sys/dev/block/%d:%d/holders",
1012                 (int)major(st.st_rdev), (int)minor(st.st_rdev));
1013         e = path + strlen(path);
1014
1015         dir = opendir(path);
1016         if (!dir)
1017                 return -1;
1018         while ((de = readdir(dir))) {
1019                 if (de->d_ino == 0)
1020                         continue;
1021                 if (de->d_name[0] == '.')
1022                         continue;
1023                 sprintf(e, "/%s/dev", de->d_name);
1024                 dfd = open(path, O_RDONLY);
1025                 if (dfd < 0)
1026                         continue;
1027                 n = read(dfd, buf, sizeof(buf));
1028                 close(dfd);
1029                 if (n <= 0 || n >= sizeof(buf))
1030                         continue;
1031                 buf[n] = 0;
1032                 if (sscanf(buf, "%d:%d", &major, &minor) != 2)
1033                         continue;
1034                 sprintf(buf, "%d:%d", major, minor);
1035                 dfd = dev_open(buf, O_RDONLY);
1036                 if (dfd >= 0) {
1037                         closedir(dir);
1038                         return dfd;
1039                 }
1040         }
1041         closedir(dir);
1042         return -1;
1043 }
1044
1045 int add_disk(int mdfd, struct supertype *st,
1046              struct mdinfo *sra, struct mdinfo *info)
1047 {
1048         /* Add a device to an array, in one of 2 ways. */
1049         int rv;
1050 #ifndef MDASSEMBLE
1051         if (st->ss->external) {
1052                 rv = sysfs_add_disk(sra, info);
1053                 if (! rv) {
1054                         struct mdinfo *sd2;
1055                         for (sd2 = sra->devs; sd2; sd2=sd2->next)
1056                                 if (sd2 == info)
1057                                         break;
1058                         if (sd2 == NULL) {
1059                                 sd2 = malloc(sizeof(*sd2));
1060                                 *sd2 = *info;
1061                                 sd2->next = sra->devs;
1062                                 sra->devs = sd2;
1063                         }
1064                 }
1065         } else
1066 #endif
1067                 rv = ioctl(mdfd, ADD_NEW_DISK, &info->disk);
1068         return rv;
1069 }
1070
1071 int set_array_info(int mdfd, struct supertype *st, struct mdinfo *info)
1072 {
1073         /* Initialise kernel's knowledge of array.
1074          * This varies between externally managed arrays
1075          * and older kernels
1076          */
1077         int vers = md_get_version(mdfd);
1078         int rv;
1079
1080 #ifndef MDASSEMBLE
1081         if (st->ss->external)
1082                 rv = sysfs_set_array(info, vers);
1083         else
1084 #endif
1085                 if ((vers % 100) >= 1) { /* can use different versions */
1086                 mdu_array_info_t inf;
1087                 memset(&inf, 0, sizeof(inf));
1088                 inf.major_version = info->array.major_version;
1089                 inf.minor_version = info->array.minor_version;
1090                 rv = ioctl(mdfd, SET_ARRAY_INFO, &inf);
1091         } else
1092                 rv = ioctl(mdfd, SET_ARRAY_INFO, NULL);
1093         return rv;
1094 }
1095
1096 char *devnum2devname(int num)
1097 {
1098         char name[100];
1099         if (num > 0)
1100                 sprintf(name, "md%d", num);
1101         else
1102                 sprintf(name, "md_d%d", -1-num);
1103         return strdup(name);
1104 }
1105
1106 int devname2devnum(char *name)
1107 {
1108         char *ep;
1109         int num;
1110         if (strncmp(name, "md_d", 4)==0)
1111                 num = -1-strtoul(name+4, &ep, 10);
1112         else
1113                 num = strtoul(name+2, &ep, 10);
1114         return num;
1115 }
1116
1117 int stat2devnum(struct stat *st)
1118 {
1119         char path[30];
1120         char link[200];
1121         char *cp;
1122         int n;
1123
1124         if ((S_IFMT & st->st_mode) == S_IFBLK) {
1125                 if (major(st->st_rdev) == MD_MAJOR)
1126                         return minor(st->st_rdev);
1127                 else if (major(st->st_rdev) == get_mdp_major())
1128                         return -1- (minor(st->st_rdev)>>MdpMinorShift);
1129
1130                 /* must be an extended-minor partition. Look at the
1131                  * /sys/dev/block/%d:%d link which must look like
1132                  * ../../block/mdXXX/mdXXXpYY
1133                  */
1134                 sprintf(path, "/sys/dev/block/%d:%d", major(st->st_rdev),
1135                         minor(st->st_rdev));
1136                 n = readlink(path, link, sizeof(link)-1);
1137                 if (n <= 0)
1138                         return NoMdDev;
1139                 link[n] = 0;
1140                 cp = strrchr(link, '/');
1141                 if (cp) *cp = 0;
1142                 cp = strchr(link, '/');
1143                 if (cp && strncmp(cp, "/md", 3) == 0)
1144                         return devname2devnum(cp+1);
1145         }
1146         return NoMdDev;
1147
1148 }
1149
1150 int fd2devnum(int fd)
1151 {
1152         struct stat stb;
1153         if (fstat(fd, &stb) == 0)
1154                 return stat2devnum(&stb);
1155         return NoMdDev;
1156 }
1157
1158 int mdmon_running(int devnum)
1159 {
1160         char path[100];
1161         char pid[10];
1162         int fd;
1163         int n;
1164         sprintf(path, "/var/run/mdadm/%s.pid", devnum2devname(devnum));
1165         fd = open(path, O_RDONLY, 0);
1166
1167         if (fd < 0)
1168                 return 0;
1169         n = read(fd, pid, 9);
1170         close(fd);
1171         if (n <= 0)
1172                 return 0;
1173         if (kill(atoi(pid), 0) == 0)
1174                 return 1;
1175         return 0;
1176 }
1177
1178 int signal_mdmon(int devnum)
1179 {
1180         char path[100];
1181         char pid[10];
1182         int fd;
1183         int n;
1184         sprintf(path, "/var/run/mdadm/%s.pid", devnum2devname(devnum));
1185         fd = open(path, O_RDONLY, 0);
1186
1187         if (fd < 0)
1188                 return 0;
1189         n = read(fd, pid, 9);
1190         close(fd);
1191         if (n <= 0)
1192                 return 0;
1193         if (kill(atoi(pid), SIGUSR1) == 0)
1194                 return 1;
1195         return 0;
1196 }
1197
1198 int start_mdmon(int devnum)
1199 {
1200         int i;
1201         int len;
1202         pid_t pid;
1203         int status;
1204         char pathbuf[1024];
1205         char *paths[4] = {
1206                 pathbuf,
1207                 "/sbin/mdmon",
1208                 "mdmon",
1209                 NULL
1210         };
1211
1212         if (check_env("MDADM_NO_MDMON"))
1213                 return 0;
1214
1215         len = readlink("/proc/self/exe", pathbuf, sizeof(pathbuf));
1216         if (len > 0) {
1217                 char *sl;
1218                 pathbuf[len] = 0;
1219                 sl = strrchr(pathbuf, '/');
1220                 if (sl)
1221                         sl++;
1222                 else
1223                         sl = pathbuf;
1224                 strcpy(sl, "mdmon");
1225         } else
1226                 pathbuf[0] = '\0';
1227
1228         switch(fork()) {
1229         case 0:
1230                 /* FIXME yuk. CLOSE_EXEC?? */
1231                 for (i=3; i < 100; i++)
1232                         close(i);
1233                 for (i=0; paths[i]; i++)
1234                         if (paths[i][0])
1235                                 execl(paths[i], "mdmon",
1236                                       map_dev(dev2major(devnum),
1237                                               dev2minor(devnum),
1238                                               1), NULL);
1239                 exit(1);
1240         case -1: fprintf(stderr, Name ": cannot run mdmon. "
1241                          "Array remains readonly\n");
1242                 return -1;
1243         default: /* parent - good */
1244                 pid = wait(&status);
1245                 if (pid < 0 || status != 0)
1246                         return -1;
1247         }
1248         return 0;
1249 }
1250
1251 int check_env(char *name)
1252 {
1253         char *val = getenv(name);
1254
1255         if (val && atoi(val) == 1)
1256                 return 1;
1257
1258         return 0;
1259 }
1260
1261 #ifndef MDASSEMBLE
1262 int flush_metadata_updates(struct supertype *st)
1263 {
1264         int sfd;
1265         if (!st->updates) {
1266                 st->update_tail = NULL;
1267                 return -1;
1268         }
1269
1270         sfd = connect_monitor(devnum2devname(st->container_dev));
1271         if (sfd < 0)
1272                 return -1;
1273
1274         while (st->updates) {
1275                 struct metadata_update *mu = st->updates;
1276                 st->updates = mu->next;
1277
1278                 send_message(sfd, mu, 0);
1279                 wait_reply(sfd, 0);
1280                 free(mu->buf);
1281                 free(mu);
1282         }
1283         ack(sfd, 0);
1284         wait_reply(sfd, 0);
1285         close(sfd);
1286         st->update_tail = NULL;
1287         return 0;
1288 }
1289
1290 void append_metadata_update(struct supertype *st, void *buf, int len)
1291 {
1292
1293         struct metadata_update *mu = malloc(sizeof(*mu));
1294
1295         mu->buf = buf;
1296         mu->len = len;
1297         mu->space = NULL;
1298         mu->next = NULL;
1299         *st->update_tail = mu;
1300         st->update_tail = &mu->next;
1301 }
1302 #endif /* MDASSEMBLE */
1303
1304 #ifdef __TINYC__
1305 /* tinyc doesn't optimize this check in ioctl.h out ... */
1306 unsigned int __invalid_size_argument_for_IOC = 0;
1307 #endif
1308