sys/geom/part/g_part.c

   1 /*-
   2  * Copyright (c) 2002, 2005-2009 Marcel Moolenaar
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  *
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include <sys/param.h>
  31 #include <sys/bio.h>
  32 #include <sys/diskmbr.h>
  33 #include <sys/endian.h>
  34 #include <sys/kernel.h>
  35 #include <sys/kobj.h>
  36 #include <sys/limits.h>
  37 #include <sys/lock.h>
  38 #include <sys/malloc.h>
  39 #include <sys/mutex.h>
  40 #include <sys/queue.h>
  41 #include <sys/sbuf.h>
  42 #include <sys/sysctl.h>
  43 #include <sys/systm.h>
  44 #include <sys/uuid.h>
  45 #include <geom/geom.h>
  46 #include <geom/geom_ctl.h>
  47 #include <geom/geom_int.h>
  48 #include <geom/part/g_part.h>
  49
  50 #include "g_part_if.h"
  51
  52 #ifndef _PATH_DEV
  53 #define _PATH_DEV "/dev/"
  54 #endif
  55
  56 static kobj_method_t g_part_null_methods[] = {
  57         { 0, 0 }
  58 };
  59
  60 static struct g_part_scheme g_part_null_scheme = {
  61         "(none)",
  62         g_part_null_methods,
  63         sizeof(struct g_part_table),
  64 };
  65
  66 TAILQ_HEAD(, g_part_scheme) g_part_schemes =
  67     TAILQ_HEAD_INITIALIZER(g_part_schemes);
  68
  69 struct g_part_alias_list {
  70         const char *lexeme;
  71         enum g_part_alias alias;
  72 } g_part_alias_list[G_PART_ALIAS_COUNT] = {
  73         { "apple-boot", G_PART_ALIAS_APPLE_BOOT },
  74         { "apple-hfs", G_PART_ALIAS_APPLE_HFS },
  75         { "apple-label", G_PART_ALIAS_APPLE_LABEL },
  76         { "apple-raid", G_PART_ALIAS_APPLE_RAID },
  77         { "apple-raid-offline", G_PART_ALIAS_APPLE_RAID_OFFLINE },
  78         { "apple-tv-recovery", G_PART_ALIAS_APPLE_TV_RECOVERY },
  79         { "apple-ufs", G_PART_ALIAS_APPLE_UFS },
  80         { "bios-boot", G_PART_ALIAS_BIOS_BOOT },
  81         { "ebr", G_PART_ALIAS_EBR },
  82         { "efi", G_PART_ALIAS_EFI },
  83         { "fat32", G_PART_ALIAS_MS_FAT32 },
  84         { "freebsd", G_PART_ALIAS_FREEBSD },
  85         { "freebsd-boot", G_PART_ALIAS_FREEBSD_BOOT },
  86         { "freebsd-swap", G_PART_ALIAS_FREEBSD_SWAP },
  87         { "freebsd-ufs", G_PART_ALIAS_FREEBSD_UFS },
  88         { "freebsd-vinum", G_PART_ALIAS_FREEBSD_VINUM },
  89         { "freebsd-zfs", G_PART_ALIAS_FREEBSD_ZFS },
  90         { "linux-data", G_PART_ALIAS_LINUX_DATA },
  91         { "linux-lvm", G_PART_ALIAS_LINUX_LVM },
  92         { "linux-raid", G_PART_ALIAS_LINUX_RAID },
  93         { "linux-swap", G_PART_ALIAS_LINUX_SWAP },
  94         { "mbr", G_PART_ALIAS_MBR },
  95         { "ms-basic-data", G_PART_ALIAS_MS_BASIC_DATA },
  96         { "ms-ldm-data", G_PART_ALIAS_MS_LDM_DATA },
  97         { "ms-ldm-metadata", G_PART_ALIAS_MS_LDM_METADATA },
  98         { "ms-reserved", G_PART_ALIAS_MS_RESERVED },
  99         { "ntfs", G_PART_ALIAS_MS_NTFS },
 100         { "netbsd-ccd", G_PART_ALIAS_NETBSD_CCD },
 101         { "netbsd-cgd", G_PART_ALIAS_NETBSD_CGD },
 102         { "netbsd-ffs", G_PART_ALIAS_NETBSD_FFS },
 103         { "netbsd-lfs", G_PART_ALIAS_NETBSD_LFS },
 104         { "netbsd-raid", G_PART_ALIAS_NETBSD_RAID },
 105         { "netbsd-swap", G_PART_ALIAS_NETBSD_SWAP },
 106 };
 107
 108 SYSCTL_DECL(_kern_geom);
 109 SYSCTL_NODE(_kern_geom, OID_AUTO, part, CTLFLAG_RW, 0,
 110     "GEOM_PART stuff");
 111 static u_int check_integrity = 1;
 112 TUNABLE_INT("kern.geom.part.check_integrity", &check_integrity);
 113 SYSCTL_UINT(_kern_geom_part, OID_AUTO, check_integrity,
 114     CTLFLAG_RW | CTLFLAG_TUN, &check_integrity, 1,
 115     "Enable integrity checking");
 116
 117 /*
 118  * The GEOM partitioning class.
 119  */
 120 static g_ctl_req_t g_part_ctlreq;
 121 static g_ctl_destroy_geom_t g_part_destroy_geom;
 122 static g_fini_t g_part_fini;
 123 static g_init_t g_part_init;
 124 static g_taste_t g_part_taste;
 125
 126 static g_access_t g_part_access;
 127 static g_dumpconf_t g_part_dumpconf;
 128 static g_orphan_t g_part_orphan;
 129 static g_spoiled_t g_part_spoiled;
 130 static g_start_t g_part_start;
 131
 132 static struct g_class g_part_class = {
 133         .name = "PART",
 134         .version = G_VERSION,
 135         /* Class methods. */
 136         .ctlreq = g_part_ctlreq,
 137         .destroy_geom = g_part_destroy_geom,
 138         .fini = g_part_fini,
 139         .init = g_part_init,
 140         .taste = g_part_taste,
 141         /* Geom methods. */
 142         .access = g_part_access,
 143         .dumpconf = g_part_dumpconf,
 144         .orphan = g_part_orphan,
 145         .spoiled = g_part_spoiled,
 146         .start = g_part_start,
 147 };
 148
 149 DECLARE_GEOM_CLASS(g_part_class, g_part);
 150 MODULE_VERSION(g_part, 0);
 151
 152 /*
 153  * Support functions.
 154  */
 155
 156 static void g_part_wither(struct g_geom *, int);
 157
 158 const char *
 159 g_part_alias_name(enum g_part_alias alias)
 160 {
 161         int i;
 162
 163         for (i = 0; i < G_PART_ALIAS_COUNT; i++) {
 164                 if (g_part_alias_list[i].alias != alias)
 165                         continue;
 166                 return (g_part_alias_list[i].lexeme);
 167         }
 168
 169         return (NULL);
 170 }
 171
 172 void
 173 g_part_geometry_heads(off_t blocks, u_int sectors, off_t *bestchs,
 174     u_int *bestheads)
 175 {
 176         static u_int candidate_heads[] = { 1, 2, 16, 32, 64, 128, 255, 0 };
 177         off_t chs, cylinders;
 178         u_int heads;
 179         int idx;
 180
 181         *bestchs = 0;
 182         *bestheads = 0;
 183         for (idx = 0; candidate_heads[idx] != 0; idx++) {
 184                 heads = candidate_heads[idx];
 185                 cylinders = blocks / heads / sectors;
 186                 if (cylinders < heads || cylinders < sectors)
 187                         break;
 188                 if (cylinders > 1023)
 189                         continue;
 190                 chs = cylinders * heads * sectors;
 191                 if (chs > *bestchs || (chs == *bestchs && *bestheads == 1)) {
 192                         *bestchs = chs;
 193                         *bestheads = heads;
 194                 }
 195         }
 196 }
 197
 198 static void
 199 g_part_geometry(struct g_part_table *table, struct g_consumer *cp,
 200     off_t blocks)
 201 {
 202         static u_int candidate_sectors[] = { 1, 9, 17, 33, 63, 0 };
 203         off_t chs, bestchs;
 204         u_int heads, sectors;
 205         int idx;
 206
 207         if (g_getattr("GEOM::fwsectors", cp, &sectors) != 0 || sectors == 0 ||
 208             g_getattr("GEOM::fwheads", cp, &heads) != 0 || heads == 0) {
 209                 table->gpt_fixgeom = 0;
 210                 table->gpt_heads = 0;
 211                 table->gpt_sectors = 0;
 212                 bestchs = 0;
 213                 for (idx = 0; candidate_sectors[idx] != 0; idx++) {
 214                         sectors = candidate_sectors[idx];
 215                         g_part_geometry_heads(blocks, sectors, &chs, &heads);
 216                         if (chs == 0)
 217                                 continue;
 218                         /*
 219                          * Prefer a geometry with sectors > 1, but only if
 220                          * it doesn't bump down the number of heads to 1.
 221                          */
 222                         if (chs > bestchs || (chs == bestchs && heads > 1 &&
 223                             table->gpt_sectors == 1)) {
 224                                 bestchs = chs;
 225                                 table->gpt_heads = heads;
 226                                 table->gpt_sectors = sectors;
 227                         }
 228                 }
 229                 /*
 230                  * If we didn't find a geometry at all, then the disk is
 231                  * too big. This means we can use the maximum number of
 232                  * heads and sectors.
 233                  */
 234                 if (bestchs == 0) {
 235                         table->gpt_heads = 255;
 236                         table->gpt_sectors = 63;
 237                 }
 238         } else {
 239                 table->gpt_fixgeom = 1;
 240                 table->gpt_heads = heads;
 241                 table->gpt_sectors = sectors;
 242         }
 243 }
 244
 245 #define DPRINTF(...)    if (bootverbose) {      \
 246         printf("GEOM_PART: " __VA_ARGS__);      \
 247 }
 248
 249 static int
 250 g_part_check_integrity(struct g_part_table *table, struct g_consumer *cp)
 251 {
 252         struct g_part_entry *e1, *e2;
 253         struct g_provider *pp;
 254         off_t offset;
 255         int failed;
 256
 257         failed = 0;
 258         pp = cp->provider;
 259         if (table->gpt_last < table->gpt_first) {
 260                 DPRINTF("last LBA is below first LBA: %jd < %jd\n",
 261                     (intmax_t)table->gpt_last, (intmax_t)table->gpt_first);
 262                 failed++;
 263         }
 264         if (table->gpt_last > pp->mediasize / pp->sectorsize - 1) {
 265                 DPRINTF("last LBA extends beyond mediasize: "
 266                     "%jd > %jd\n", (intmax_t)table->gpt_last,
 267                     (intmax_t)pp->mediasize / pp->sectorsize - 1);
 268                 failed++;
 269         }
 270         LIST_FOREACH(e1, &table->gpt_entry, gpe_entry) {
 271                 if (e1->gpe_deleted || e1->gpe_internal)
 272                         continue;
 273                 if (e1->gpe_start < table->gpt_first) {
 274                         DPRINTF("partition %d has start offset below first "
 275                             "LBA: %jd < %jd\n", e1->gpe_index,
 276                             (intmax_t)e1->gpe_start,
 277                             (intmax_t)table->gpt_first);
 278                         failed++;
 279                 }
 280                 if (e1->gpe_start > table->gpt_last) {
 281                         DPRINTF("partition %d has start offset beyond last "
 282                             "LBA: %jd > %jd\n", e1->gpe_index,
 283                             (intmax_t)e1->gpe_start,
 284                             (intmax_t)table->gpt_last);
 285                         failed++;
 286                 }
 287                 if (e1->gpe_end < e1->gpe_start) {
 288                         DPRINTF("partition %d has end offset below start "
 289                             "offset: %jd < %jd\n", e1->gpe_index,
 290                             (intmax_t)e1->gpe_end,
 291                             (intmax_t)e1->gpe_start);
 292                         failed++;
 293                 }
 294                 if (e1->gpe_end > table->gpt_last) {
 295                         DPRINTF("partition %d has end offset beyond last "
 296                             "LBA: %jd > %jd\n", e1->gpe_index,
 297                             (intmax_t)e1->gpe_end,
 298                             (intmax_t)table->gpt_last);
 299                         failed++;
 300                 }
 301                 if (pp->stripesize > 0) {
 302                         offset = e1->gpe_start * pp->sectorsize;
 303                         if (e1->gpe_offset > offset)
 304                                 offset = e1->gpe_offset;
 305                         if ((offset + pp->stripeoffset) % pp->stripesize) {
 306                                 DPRINTF("partition %d is not aligned on %u "
 307                                     "bytes\n", e1->gpe_index, pp->stripesize);
 308                                 /* Don't treat this as a critical failure */
 309                         }
 310                 }
 311                 e2 = e1;
 312                 while ((e2 = LIST_NEXT(e2, gpe_entry)) != NULL) {
 313                         if (e2->gpe_deleted || e2->gpe_internal)
 314                                 continue;
 315                         if (e1->gpe_start >= e2->gpe_start &&
 316                             e1->gpe_start <= e2->gpe_end) {
 317                                 DPRINTF("partition %d has start offset inside "
 318                                     "partition %d: start[%d] %jd >= start[%d] "
 319                                     "%jd <= end[%d] %jd\n",
 320                                     e1->gpe_index, e2->gpe_index,
 321                                     e2->gpe_index, (intmax_t)e2->gpe_start,
 322                                     e1->gpe_index, (intmax_t)e1->gpe_start,
 323                                     e2->gpe_index, (intmax_t)e2->gpe_end);
 324                                 failed++;
 325                         }
 326                         if (e1->gpe_end >= e2->gpe_start &&
 327                             e1->gpe_end <= e2->gpe_end) {
 328                                 DPRINTF("partition %d has end offset inside "
 329                                     "partition %d: start[%d] %jd >= end[%d] "
 330                                     "%jd <= end[%d] %jd\n",
 331                                     e1->gpe_index, e2->gpe_index,
 332                                     e2->gpe_index, (intmax_t)e2->gpe_start,
 333                                     e1->gpe_index, (intmax_t)e1->gpe_end,
 334                                     e2->gpe_index, (intmax_t)e2->gpe_end);
 335                                 failed++;
 336                         }
 337                         if (e1->gpe_start < e2->gpe_start &&
 338                             e1->gpe_end > e2->gpe_end) {
 339                                 DPRINTF("partition %d contains partition %d: "
 340                                     "start[%d] %jd > start[%d] %jd, end[%d] "
 341                                     "%jd < end[%d] %jd\n",
 342                                     e1->gpe_index, e2->gpe_index,
 343                                     e1->gpe_index, (intmax_t)e1->gpe_start,
 344                                     e2->gpe_index, (intmax_t)e2->gpe_start,
 345                                     e2->gpe_index, (intmax_t)e2->gpe_end,
 346                                     e1->gpe_index, (intmax_t)e1->gpe_end);
 347                                 failed++;
 348                         }
 349                 }
 350         }
 351         if (failed != 0) {
 352                 printf("GEOM_PART: integrity check failed (%s, %s)\n",
 353                     pp->name, table->gpt_scheme->name);
 354                 if (check_integrity != 0)
 355                         return (EINVAL);
 356                 table->gpt_corrupt = 1;
 357         }
 358         return (0);
 359 }
 360 #undef  DPRINTF
 361
 362 struct g_part_entry *
 363 g_part_new_entry(struct g_part_table *table, int index, quad_t start,
 364     quad_t end)
 365 {
 366         struct g_part_entry *entry, *last;
 367
 368         last = NULL;
 369         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
 370                 if (entry->gpe_index == index)
 371                         break;
 372                 if (entry->gpe_index > index) {
 373                         entry = NULL;
 374                         break;
 375                 }
 376                 last = entry;
 377         }
 378         if (entry == NULL) {
 379                 entry = g_malloc(table->gpt_scheme->gps_entrysz,
 380                     M_WAITOK | M_ZERO);
 381                 entry->gpe_index = index;
 382                 if (last == NULL)
 383                         LIST_INSERT_HEAD(&table->gpt_entry, entry, gpe_entry);
 384                 else
 385                         LIST_INSERT_AFTER(last, entry, gpe_entry);
 386         } else
 387                 entry->gpe_offset = 0;
 388         entry->gpe_start = start;
 389         entry->gpe_end = end;
 390         return (entry);
 391 }
 392
 393 static void
 394 g_part_new_provider(struct g_geom *gp, struct g_part_table *table,
 395     struct g_part_entry *entry)
 396 {
 397         struct g_consumer *cp;
 398         struct g_provider *pp;
 399         struct sbuf *sb;
 400         off_t offset;
 401
 402         cp = LIST_FIRST(&gp->consumer);
 403         pp = cp->provider;
 404
 405         offset = entry->gpe_start * pp->sectorsize;
 406         if (entry->gpe_offset < offset)
 407                 entry->gpe_offset = offset;
 408
 409         if (entry->gpe_pp == NULL) {
 410                 sb = sbuf_new_auto();
 411                 G_PART_FULLNAME(table, entry, sb, gp->name);
 412                 sbuf_finish(sb);
 413                 entry->gpe_pp = g_new_providerf(gp, "%s", sbuf_data(sb));
 414                 sbuf_delete(sb);
 415                 entry->gpe_pp->private = entry;         /* Close the circle. */
 416         }
 417         entry->gpe_pp->index = entry->gpe_index - 1;    /* index is 1-based. */
 418         entry->gpe_pp->mediasize = (entry->gpe_end - entry->gpe_start + 1) *
 419             pp->sectorsize;
 420         entry->gpe_pp->mediasize -= entry->gpe_offset - offset;
 421         entry->gpe_pp->sectorsize = pp->sectorsize;
 422         entry->gpe_pp->flags = pp->flags & G_PF_CANDELETE;
 423         entry->gpe_pp->stripesize = pp->stripesize;
 424         entry->gpe_pp->stripeoffset = pp->stripeoffset + entry->gpe_offset;
 425         if (pp->stripesize > 0)
 426                 entry->gpe_pp->stripeoffset %= pp->stripesize;
 427         g_error_provider(entry->gpe_pp, 0);
 428 }
 429
 430 static struct g_geom*
 431 g_part_find_geom(const char *name)
 432 {
 433         struct g_geom *gp;
 434         LIST_FOREACH(gp, &g_part_class.geom, geom) {
 435                 if (!strcmp(name, gp->name))
 436                         break;
 437         }
 438         return (gp);
 439 }
 440
 441 static int
 442 g_part_parm_geom(struct gctl_req *req, const char *name, struct g_geom **v)
 443 {
 444         struct g_geom *gp;
 445         const char *gname;
 446
 447         gname = gctl_get_asciiparam(req, name);
 448         if (gname == NULL)
 449                 return (ENOATTR);
 450         if (strncmp(gname, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
 451                 gname += sizeof(_PATH_DEV) - 1;
 452         gp = g_part_find_geom(gname);
 453         if (gp == NULL) {
 454                 gctl_error(req, "%d %s '%s'", EINVAL, name, gname);
 455                 return (EINVAL);
 456         }
 457         if ((gp->flags & G_GEOM_WITHER) != 0) {
 458                 gctl_error(req, "%d %s", ENXIO, gname);
 459                 return (ENXIO);
 460         }
 461         *v = gp;
 462         return (0);
 463 }
 464
 465 static int
 466 g_part_parm_provider(struct gctl_req *req, const char *name,
 467     struct g_provider **v)
 468 {
 469         struct g_provider *pp;
 470         const char *pname;
 471
 472         pname = gctl_get_asciiparam(req, name);
 473         if (pname == NULL)
 474                 return (ENOATTR);
 475         if (strncmp(pname, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
 476                 pname += sizeof(_PATH_DEV) - 1;
 477         pp = g_provider_by_name(pname);
 478         if (pp == NULL) {
 479                 gctl_error(req, "%d %s '%s'", EINVAL, name, pname);
 480                 return (EINVAL);
 481         }
 482         *v = pp;
 483         return (0);
 484 }
 485
 486 static int
 487 g_part_parm_quad(struct gctl_req *req, const char *name, quad_t *v)
 488 {
 489         const char *p;
 490         char *x;
 491         quad_t q;
 492
 493         p = gctl_get_asciiparam(req, name);
 494         if (p == NULL)
 495                 return (ENOATTR);
 496         q = strtoq(p, &x, 0);
 497         if (*x != '\0' || q < 0) {
 498                 gctl_error(req, "%d %s '%s'", EINVAL, name, p);
 499                 return (EINVAL);
 500         }
 501         *v = q;
 502         return (0);
 503 }
 504
 505 static int
 506 g_part_parm_scheme(struct gctl_req *req, const char *name,
 507     struct g_part_scheme **v)
 508 {
 509         struct g_part_scheme *s;
 510         const char *p;
 511
 512         p = gctl_get_asciiparam(req, name);
 513         if (p == NULL)
 514                 return (ENOATTR);
 515         TAILQ_FOREACH(s, &g_part_schemes, scheme_list) {
 516                 if (s == &g_part_null_scheme)
 517                         continue;
 518                 if (!strcasecmp(s->name, p))
 519                         break;
 520         }
 521         if (s == NULL) {
 522                 gctl_error(req, "%d %s '%s'", EINVAL, name, p);
 523                 return (EINVAL);
 524         }
 525         *v = s;
 526         return (0);
 527 }
 528
 529 static int
 530 g_part_parm_str(struct gctl_req *req, const char *name, const char **v)
 531 {
 532         const char *p;
 533
 534         p = gctl_get_asciiparam(req, name);
 535         if (p == NULL)
 536                 return (ENOATTR);
 537         /* An empty label is always valid. */
 538         if (strcmp(name, "label") != 0 && p[0] == '\0') {
 539                 gctl_error(req, "%d %s '%s'", EINVAL, name, p);
 540                 return (EINVAL);
 541         }
 542         *v = p;
 543         return (0);
 544 }
 545
 546 static int
 547 g_part_parm_intmax(struct gctl_req *req, const char *name, u_int *v)
 548 {
 549         const intmax_t *p;
 550         int size;
 551
 552         p = gctl_get_param(req, name, &size);
 553         if (p == NULL)
 554                 return (ENOATTR);
 555         if (size != sizeof(*p) || *p < 0 || *p > INT_MAX) {
 556                 gctl_error(req, "%d %s '%jd'", EINVAL, name, *p);
 557                 return (EINVAL);
 558         }
 559         *v = (u_int)*p;
 560         return (0);
 561 }
 562
 563 static int
 564 g_part_parm_uint32(struct gctl_req *req, const char *name, u_int *v)
 565 {
 566         const uint32_t *p;
 567         int size;
 568
 569         p = gctl_get_param(req, name, &size);
 570         if (p == NULL)
 571                 return (ENOATTR);
 572         if (size != sizeof(*p) || *p > INT_MAX) {
 573                 gctl_error(req, "%d %s '%u'", EINVAL, name, (unsigned int)*p);
 574                 return (EINVAL);
 575         }
 576         *v = (u_int)*p;
 577         return (0);
 578 }
 579
 580 static int
 581 g_part_parm_bootcode(struct gctl_req *req, const char *name, const void **v,
 582     unsigned int *s)
 583 {
 584         const void *p;
 585         int size;
 586
 587         p = gctl_get_param(req, name, &size);
 588         if (p == NULL)
 589                 return (ENOATTR);
 590         *v = p;
 591         *s = size;
 592         return (0);
 593 }
 594
 595 static int
 596 g_part_probe(struct g_geom *gp, struct g_consumer *cp, int depth)
 597 {
 598         struct g_part_scheme *iter, *scheme;
 599         struct g_part_table *table;
 600         int pri, probe;
 601
 602         table = gp->softc;
 603         scheme = (table != NULL) ? table->gpt_scheme : NULL;
 604         pri = (scheme != NULL) ? G_PART_PROBE(table, cp) : INT_MIN;
 605         if (pri == 0)
 606                 goto done;
 607         if (pri > 0) {  /* error */
 608                 scheme = NULL;
 609                 pri = INT_MIN;
 610         }
 611
 612         TAILQ_FOREACH(iter, &g_part_schemes, scheme_list) {
 613                 if (iter == &g_part_null_scheme)
 614                         continue;
 615                 table = (void *)kobj_create((kobj_class_t)iter, M_GEOM,
 616                     M_WAITOK);
 617                 table->gpt_gp = gp;
 618                 table->gpt_scheme = iter;
 619                 table->gpt_depth = depth;
 620                 probe = G_PART_PROBE(table, cp);
 621                 if (probe <= 0 && probe > pri) {
 622                         pri = probe;
 623                         scheme = iter;
 624                         if (gp->softc != NULL)
 625                                 kobj_delete((kobj_t)gp->softc, M_GEOM);
 626                         gp->softc = table;
 627                         if (pri == 0)
 628                                 goto done;
 629                 } else
 630                         kobj_delete((kobj_t)table, M_GEOM);
 631         }
 632
 633 done:
 634         return ((scheme == NULL) ? ENXIO : 0);
 635 }
 636
 637 /*
 638  * Control request functions.
 639  */
 640
 641 static int
 642 g_part_ctl_add(struct gctl_req *req, struct g_part_parms *gpp)
 643 {
 644         struct g_geom *gp;
 645         struct g_provider *pp;
 646         struct g_part_entry *delent, *last, *entry;
 647         struct g_part_table *table;
 648         struct sbuf *sb;
 649         quad_t end;
 650         unsigned int index;
 651         int error;
 652
 653         gp = gpp->gpp_geom;
 654         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
 655         g_topology_assert();
 656
 657         pp = LIST_FIRST(&gp->consumer)->provider;
 658         table = gp->softc;
 659         end = gpp->gpp_start + gpp->gpp_size - 1;
 660
 661         if (gpp->gpp_start < table->gpt_first ||
 662             gpp->gpp_start > table->gpt_last) {
 663                 gctl_error(req, "%d start '%jd'", EINVAL,
 664                     (intmax_t)gpp->gpp_start);
 665                 return (EINVAL);
 666         }
 667         if (end < gpp->gpp_start || end > table->gpt_last) {
 668                 gctl_error(req, "%d size '%jd'", EINVAL,
 669                     (intmax_t)gpp->gpp_size);
 670                 return (EINVAL);
 671         }
 672         if (gpp->gpp_index > table->gpt_entries) {
 673                 gctl_error(req, "%d index '%d'", EINVAL, gpp->gpp_index);
 674                 return (EINVAL);
 675         }
 676
 677         delent = last = NULL;
 678         index = (gpp->gpp_index > 0) ? gpp->gpp_index : 1;
 679         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
 680                 if (entry->gpe_deleted) {
 681                         if (entry->gpe_index == index)
 682                                 delent = entry;
 683                         continue;
 684                 }
 685                 if (entry->gpe_index == index)
 686                         index = entry->gpe_index + 1;
 687                 if (entry->gpe_index < index)
 688                         last = entry;
 689                 if (entry->gpe_internal)
 690                         continue;
 691                 if (gpp->gpp_start >= entry->gpe_start &&
 692                     gpp->gpp_start <= entry->gpe_end) {
 693                         gctl_error(req, "%d start '%jd'", ENOSPC,
 694                             (intmax_t)gpp->gpp_start);
 695                         return (ENOSPC);
 696                 }
 697                 if (end >= entry->gpe_start && end <= entry->gpe_end) {
 698                         gctl_error(req, "%d end '%jd'", ENOSPC, (intmax_t)end);
 699                         return (ENOSPC);
 700                 }
 701                 if (gpp->gpp_start < entry->gpe_start && end > entry->gpe_end) {
 702                         gctl_error(req, "%d size '%jd'", ENOSPC,
 703                             (intmax_t)gpp->gpp_size);
 704                         return (ENOSPC);
 705                 }
 706         }
 707         if (gpp->gpp_index > 0 && index != gpp->gpp_index) {
 708                 gctl_error(req, "%d index '%d'", EEXIST, gpp->gpp_index);
 709                 return (EEXIST);
 710         }
 711         if (index > table->gpt_entries) {
 712                 gctl_error(req, "%d index '%d'", ENOSPC, index);
 713                 return (ENOSPC);
 714         }
 715
 716         entry = (delent == NULL) ? g_malloc(table->gpt_scheme->gps_entrysz,
 717             M_WAITOK | M_ZERO) : delent;
 718         entry->gpe_index = index;
 719         entry->gpe_start = gpp->gpp_start;
 720         entry->gpe_end = end;
 721         error = G_PART_ADD(table, entry, gpp);
 722         if (error) {
 723                 gctl_error(req, "%d", error);
 724                 if (delent == NULL)
 725                         g_free(entry);
 726                 return (error);
 727         }
 728         if (delent == NULL) {
 729                 if (last == NULL)
 730                         LIST_INSERT_HEAD(&table->gpt_entry, entry, gpe_entry);
 731                 else
 732                         LIST_INSERT_AFTER(last, entry, gpe_entry);
 733                 entry->gpe_created = 1;
 734         } else {
 735                 entry->gpe_deleted = 0;
 736                 entry->gpe_modified = 1;
 737         }
 738         g_part_new_provider(gp, table, entry);
 739
 740         /* Provide feedback if so requested. */
 741         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
 742                 sb = sbuf_new_auto();
 743                 G_PART_FULLNAME(table, entry, sb, gp->name);
 744                 if (pp->stripesize > 0 && entry->gpe_pp->stripeoffset != 0)
 745                         sbuf_printf(sb, " added, but partition is not "
 746                             "aligned on %u bytes\n", pp->stripesize);
 747                 else
 748                         sbuf_cat(sb, " added\n");
 749                 sbuf_finish(sb);
 750                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
 751                 sbuf_delete(sb);
 752         }
 753         return (0);
 754 }
 755
 756 static int
 757 g_part_ctl_bootcode(struct gctl_req *req, struct g_part_parms *gpp)
 758 {
 759         struct g_geom *gp;
 760         struct g_part_table *table;
 761         struct sbuf *sb;
 762         int error, sz;
 763
 764         gp = gpp->gpp_geom;
 765         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
 766         g_topology_assert();
 767
 768         table = gp->softc;
 769         sz = table->gpt_scheme->gps_bootcodesz;
 770         if (sz == 0) {
 771                 error = ENODEV;
 772                 goto fail;
 773         }
 774         if (gpp->gpp_codesize > sz) {
 775                 error = EFBIG;
 776                 goto fail;
 777         }
 778
 779         error = G_PART_BOOTCODE(table, gpp);
 780         if (error)
 781                 goto fail;
 782
 783         /* Provide feedback if so requested. */
 784         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
 785                 sb = sbuf_new_auto();
 786                 sbuf_printf(sb, "bootcode written to %s\n", gp->name);
 787                 sbuf_finish(sb);
 788                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
 789                 sbuf_delete(sb);
 790         }
 791         return (0);
 792
 793  fail:
 794         gctl_error(req, "%d", error);
 795         return (error);
 796 }
 797
 798 static int
 799 g_part_ctl_commit(struct gctl_req *req, struct g_part_parms *gpp)
 800 {
 801         struct g_consumer *cp;
 802         struct g_geom *gp;
 803         struct g_provider *pp;
 804         struct g_part_entry *entry, *tmp;
 805         struct g_part_table *table;
 806         char *buf;
 807         int error, i;
 808
 809         gp = gpp->gpp_geom;
 810         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
 811         g_topology_assert();
 812
 813         table = gp->softc;
 814         if (!table->gpt_opened) {
 815                 gctl_error(req, "%d", EPERM);
 816                 return (EPERM);
 817         }
 818
 819         g_topology_unlock();
 820
 821         cp = LIST_FIRST(&gp->consumer);
 822         if ((table->gpt_smhead | table->gpt_smtail) != 0) {
 823                 pp = cp->provider;
 824                 buf = g_malloc(pp->sectorsize, M_WAITOK | M_ZERO);
 825                 while (table->gpt_smhead != 0) {
 826                         i = ffs(table->gpt_smhead) - 1;
 827                         error = g_write_data(cp, i * pp->sectorsize, buf,
 828                             pp->sectorsize);
 829                         if (error) {
 830                                 g_free(buf);
 831                                 goto fail;
 832                         }
 833                         table->gpt_smhead &= ~(1 << i);
 834                 }
 835                 while (table->gpt_smtail != 0) {
 836                         i = ffs(table->gpt_smtail) - 1;
 837                         error = g_write_data(cp, pp->mediasize - (i + 1) *
 838                             pp->sectorsize, buf, pp->sectorsize);
 839                         if (error) {
 840                                 g_free(buf);
 841                                 goto fail;
 842                         }
 843                         table->gpt_smtail &= ~(1 << i);
 844                 }
 845                 g_free(buf);
 846         }
 847
 848         if (table->gpt_scheme == &g_part_null_scheme) {
 849                 g_topology_lock();
 850                 g_access(cp, -1, -1, -1);
 851                 g_part_wither(gp, ENXIO);
 852                 return (0);
 853         }
 854
 855         error = G_PART_WRITE(table, cp);
 856         if (error)
 857                 goto fail;
 858
 859         LIST_FOREACH_SAFE(entry, &table->gpt_entry, gpe_entry, tmp) {
 860                 if (!entry->gpe_deleted) {
 861                         entry->gpe_created = 0;
 862                         entry->gpe_modified = 0;
 863                         continue;
 864                 }
 865                 LIST_REMOVE(entry, gpe_entry);
 866                 g_free(entry);
 867         }
 868         table->gpt_created = 0;
 869         table->gpt_opened = 0;
 870
 871         g_topology_lock();
 872         g_access(cp, -1, -1, -1);
 873         return (0);
 874
 875 fail:
 876         g_topology_lock();
 877         gctl_error(req, "%d", error);
 878         return (error);
 879 }
 880
 881 static int
 882 g_part_ctl_create(struct gctl_req *req, struct g_part_parms *gpp)
 883 {
 884         struct g_consumer *cp;
 885         struct g_geom *gp;
 886         struct g_provider *pp;
 887         struct g_part_scheme *scheme;
 888         struct g_part_table *null, *table;
 889         struct sbuf *sb;
 890         int attr, error;
 891
 892         pp = gpp->gpp_provider;
 893         scheme = gpp->gpp_scheme;
 894         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, pp->name));
 895         g_topology_assert();
 896
 897         /* Check that there isn't already a g_part geom on the provider. */
 898         gp = g_part_find_geom(pp->name);
 899         if (gp != NULL) {
 900                 null = gp->softc;
 901                 if (null->gpt_scheme != &g_part_null_scheme) {
 902                         gctl_error(req, "%d geom '%s'", EEXIST, pp->name);
 903                         return (EEXIST);
 904                 }
 905         } else
 906                 null = NULL;
 907
 908         if ((gpp->gpp_parms & G_PART_PARM_ENTRIES) &&
 909             (gpp->gpp_entries < scheme->gps_minent ||
 910              gpp->gpp_entries > scheme->gps_maxent)) {
 911                 gctl_error(req, "%d entries '%d'", EINVAL, gpp->gpp_entries);
 912                 return (EINVAL);
 913         }
 914
 915         if (null == NULL)
 916                 gp = g_new_geomf(&g_part_class, "%s", pp->name);
 917         gp->softc = kobj_create((kobj_class_t)gpp->gpp_scheme, M_GEOM,
 918             M_WAITOK);
 919         table = gp->softc;
 920         table->gpt_gp = gp;
 921         table->gpt_scheme = gpp->gpp_scheme;
 922         table->gpt_entries = (gpp->gpp_parms & G_PART_PARM_ENTRIES) ?
 923             gpp->gpp_entries : scheme->gps_minent;
 924         LIST_INIT(&table->gpt_entry);
 925         if (null == NULL) {
 926                 cp = g_new_consumer(gp);
 927                 error = g_attach(cp, pp);
 928                 if (error == 0)
 929                         error = g_access(cp, 1, 1, 1);
 930                 if (error != 0) {
 931                         g_part_wither(gp, error);
 932                         gctl_error(req, "%d geom '%s'", error, pp->name);
 933                         return (error);
 934                 }
 935                 table->gpt_opened = 1;
 936         } else {
 937                 cp = LIST_FIRST(&gp->consumer);
 938                 table->gpt_opened = null->gpt_opened;
 939                 table->gpt_smhead = null->gpt_smhead;
 940                 table->gpt_smtail = null->gpt_smtail;
 941         }
 942
 943         g_topology_unlock();
 944
 945         /* Make sure the provider has media. */
 946         if (pp->mediasize == 0 || pp->sectorsize == 0) {
 947                 error = ENODEV;
 948                 goto fail;
 949         }
 950
 951         /* Make sure we can nest and if so, determine our depth. */
 952         error = g_getattr("PART::isleaf", cp, &attr);
 953         if (!error && attr) {
 954                 error = ENODEV;
 955                 goto fail;
 956         }
 957         error = g_getattr("PART::depth", cp, &attr);
 958         table->gpt_depth = (!error) ? attr + 1 : 0;
 959
 960         /*
 961          * Synthesize a disk geometry. Some partitioning schemes
 962          * depend on it and since some file systems need it even
 963          * when the partitition scheme doesn't, we do it here in
 964          * scheme-independent code.
 965          */
 966         g_part_geometry(table, cp, pp->mediasize / pp->sectorsize);
 967
 968         error = G_PART_CREATE(table, gpp);
 969         if (error)
 970                 goto fail;
 971
 972         g_topology_lock();
 973
 974         table->gpt_created = 1;
 975         if (null != NULL)
 976                 kobj_delete((kobj_t)null, M_GEOM);
 977
 978         /*
 979          * Support automatic commit by filling in the gpp_geom
 980          * parameter.
 981          */
 982         gpp->gpp_parms |= G_PART_PARM_GEOM;
 983         gpp->gpp_geom = gp;
 984
 985         /* Provide feedback if so requested. */
 986         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
 987                 sb = sbuf_new_auto();
 988                 sbuf_printf(sb, "%s created\n", gp->name);
 989                 sbuf_finish(sb);
 990                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
 991                 sbuf_delete(sb);
 992         }
 993         return (0);
 994
 995 fail:
 996         g_topology_lock();
 997         if (null == NULL) {
 998                 g_access(cp, -1, -1, -1);
 999                 g_part_wither(gp, error);
1000         } else {
1001                 kobj_delete((kobj_t)gp->softc, M_GEOM);
1002                 gp->softc = null;
1003         }
1004         gctl_error(req, "%d provider", error);
1005         return (error);
1006 }
1007
1008 static int
1009 g_part_ctl_delete(struct gctl_req *req, struct g_part_parms *gpp)
1010 {
1011         struct g_geom *gp;
1012         struct g_provider *pp;
1013         struct g_part_entry *entry;
1014         struct g_part_table *table;
1015         struct sbuf *sb;
1016
1017         gp = gpp->gpp_geom;
1018         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1019         g_topology_assert();
1020
1021         table = gp->softc;
1022
1023         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1024                 if (entry->gpe_deleted || entry->gpe_internal)
1025                         continue;
1026                 if (entry->gpe_index == gpp->gpp_index)
1027                         break;
1028         }
1029         if (entry == NULL) {
1030                 gctl_error(req, "%d index '%d'", ENOENT, gpp->gpp_index);
1031                 return (ENOENT);
1032         }
1033
1034         pp = entry->gpe_pp;
1035         if (pp != NULL) {
1036                 if (pp->acr > 0 || pp->acw > 0 || pp->ace > 0) {
1037                         gctl_error(req, "%d", EBUSY);
1038                         return (EBUSY);
1039                 }
1040
1041                 pp->private = NULL;
1042                 entry->gpe_pp = NULL;
1043         }
1044
1045         if (pp != NULL)
1046                 g_wither_provider(pp, ENXIO);
1047
1048         /* Provide feedback if so requested. */
1049         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1050                 sb = sbuf_new_auto();
1051                 G_PART_FULLNAME(table, entry, sb, gp->name);
1052                 sbuf_cat(sb, " deleted\n");
1053                 sbuf_finish(sb);
1054                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1055                 sbuf_delete(sb);
1056         }
1057
1058         if (entry->gpe_created) {
1059                 LIST_REMOVE(entry, gpe_entry);
1060                 g_free(entry);
1061         } else {
1062                 entry->gpe_modified = 0;
1063                 entry->gpe_deleted = 1;
1064         }
1065         return (0);
1066 }
1067
1068 static int
1069 g_part_ctl_destroy(struct gctl_req *req, struct g_part_parms *gpp)
1070 {
1071         struct g_consumer *cp;
1072         struct g_geom *gp;
1073         struct g_provider *pp;
1074         struct g_part_entry *entry, *tmp;
1075         struct g_part_table *null, *table;
1076         struct sbuf *sb;
1077         int error;
1078
1079         gp = gpp->gpp_geom;
1080         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1081         g_topology_assert();
1082
1083         table = gp->softc;
1084         /* Check for busy providers. */
1085         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1086                 if (entry->gpe_deleted || entry->gpe_internal)
1087                         continue;
1088                 if (gpp->gpp_force) {
1089                         pp = entry->gpe_pp;
1090                         if (pp == NULL)
1091                                 continue;
1092                         if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1093                                 continue;
1094                 }
1095                 gctl_error(req, "%d", EBUSY);
1096                 return (EBUSY);
1097         }
1098
1099         if (gpp->gpp_force) {
1100                 /* Destroy all providers. */
1101                 LIST_FOREACH_SAFE(entry, &table->gpt_entry, gpe_entry, tmp) {
1102                         pp = entry->gpe_pp;
1103                         if (pp != NULL) {
1104                                 pp->private = NULL;
1105                                 g_wither_provider(pp, ENXIO);
1106                         }
1107                         LIST_REMOVE(entry, gpe_entry);
1108                         g_free(entry);
1109                 }
1110         }
1111
1112         error = G_PART_DESTROY(table, gpp);
1113         if (error) {
1114                 gctl_error(req, "%d", error);
1115                 return (error);
1116         }
1117
1118         gp->softc = kobj_create((kobj_class_t)&g_part_null_scheme, M_GEOM,
1119             M_WAITOK);
1120         null = gp->softc;
1121         null->gpt_gp = gp;
1122         null->gpt_scheme = &g_part_null_scheme;
1123         LIST_INIT(&null->gpt_entry);
1124
1125         cp = LIST_FIRST(&gp->consumer);
1126         pp = cp->provider;
1127         null->gpt_last = pp->mediasize / pp->sectorsize - 1;
1128
1129         null->gpt_depth = table->gpt_depth;
1130         null->gpt_opened = table->gpt_opened;
1131         null->gpt_smhead = table->gpt_smhead;
1132         null->gpt_smtail = table->gpt_smtail;
1133
1134         while ((entry = LIST_FIRST(&table->gpt_entry)) != NULL) {
1135                 LIST_REMOVE(entry, gpe_entry);
1136                 g_free(entry);
1137         }
1138         kobj_delete((kobj_t)table, M_GEOM);
1139
1140         /* Provide feedback if so requested. */
1141         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1142                 sb = sbuf_new_auto();
1143                 sbuf_printf(sb, "%s destroyed\n", gp->name);
1144                 sbuf_finish(sb);
1145                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1146                 sbuf_delete(sb);
1147         }
1148         return (0);
1149 }
1150
1151 static int
1152 g_part_ctl_modify(struct gctl_req *req, struct g_part_parms *gpp)
1153 {
1154         struct g_geom *gp;
1155         struct g_part_entry *entry;
1156         struct g_part_table *table;
1157         struct sbuf *sb;
1158         int error;
1159
1160         gp = gpp->gpp_geom;
1161         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1162         g_topology_assert();
1163
1164         table = gp->softc;
1165
1166         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1167                 if (entry->gpe_deleted || entry->gpe_internal)
1168                         continue;
1169                 if (entry->gpe_index == gpp->gpp_index)
1170                         break;
1171         }
1172         if (entry == NULL) {
1173                 gctl_error(req, "%d index '%d'", ENOENT, gpp->gpp_index);
1174                 return (ENOENT);
1175         }
1176
1177         error = G_PART_MODIFY(table, entry, gpp);
1178         if (error) {
1179                 gctl_error(req, "%d", error);
1180                 return (error);
1181         }
1182
1183         if (!entry->gpe_created)
1184                 entry->gpe_modified = 1;
1185
1186         /* Provide feedback if so requested. */
1187         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1188                 sb = sbuf_new_auto();
1189                 G_PART_FULLNAME(table, entry, sb, gp->name);
1190                 sbuf_cat(sb, " modified\n");
1191                 sbuf_finish(sb);
1192                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1193                 sbuf_delete(sb);
1194         }
1195         return (0);
1196 }
1197
1198 static int
1199 g_part_ctl_move(struct gctl_req *req, struct g_part_parms *gpp)
1200 {
1201         gctl_error(req, "%d verb 'move'", ENOSYS);
1202         return (ENOSYS);
1203 }
1204
1205 static int
1206 g_part_ctl_recover(struct gctl_req *req, struct g_part_parms *gpp)
1207 {
1208         struct g_part_table *table;
1209         struct g_geom *gp;
1210         struct sbuf *sb;
1211         int error, recovered;
1212
1213         gp = gpp->gpp_geom;
1214         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1215         g_topology_assert();
1216         table = gp->softc;
1217         error = recovered = 0;
1218
1219         if (table->gpt_corrupt) {
1220                 error = G_PART_RECOVER(table);
1221                 if (error == 0)
1222                         error = g_part_check_integrity(table,
1223                             LIST_FIRST(&gp->consumer));
1224                 if (error) {
1225                         gctl_error(req, "%d recovering '%s' failed",
1226                             error, gp->name);
1227                         return (error);
1228                 }
1229                 recovered = 1;
1230         }
1231         /* Provide feedback if so requested. */
1232         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1233                 sb = sbuf_new_auto();
1234                 if (recovered)
1235                         sbuf_printf(sb, "%s recovered\n", gp->name);
1236                 else
1237                         sbuf_printf(sb, "%s recovering is not needed\n",
1238                             gp->name);
1239                 sbuf_finish(sb);
1240                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1241                 sbuf_delete(sb);
1242         }
1243         return (0);
1244 }
1245
1246 static int
1247 g_part_ctl_resize(struct gctl_req *req, struct g_part_parms *gpp)
1248 {
1249         struct g_geom *gp;
1250         struct g_provider *pp;
1251         struct g_part_entry *pe, *entry;
1252         struct g_part_table *table;
1253         struct sbuf *sb;
1254         quad_t end;
1255         int error;
1256
1257         gp = gpp->gpp_geom;
1258         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1259         g_topology_assert();
1260         table = gp->softc;
1261
1262         /* check gpp_index */
1263         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1264                 if (entry->gpe_deleted || entry->gpe_internal)
1265                         continue;
1266                 if (entry->gpe_index == gpp->gpp_index)
1267                         break;
1268         }
1269         if (entry == NULL) {
1270                 gctl_error(req, "%d index '%d'", ENOENT, gpp->gpp_index);
1271                 return (ENOENT);
1272         }
1273
1274         /* check gpp_size */
1275         end = entry->gpe_start + gpp->gpp_size - 1;
1276         if (gpp->gpp_size < 1 || end > table->gpt_last) {
1277                 gctl_error(req, "%d size '%jd'", EINVAL,
1278                     (intmax_t)gpp->gpp_size);
1279                 return (EINVAL);
1280         }
1281
1282         LIST_FOREACH(pe, &table->gpt_entry, gpe_entry) {
1283                 if (pe->gpe_deleted || pe->gpe_internal || pe == entry)
1284                         continue;
1285                 if (end >= pe->gpe_start && end <= pe->gpe_end) {
1286                         gctl_error(req, "%d end '%jd'", ENOSPC,
1287                             (intmax_t)end);
1288                         return (ENOSPC);
1289                 }
1290                 if (entry->gpe_start < pe->gpe_start && end > pe->gpe_end) {
1291                         gctl_error(req, "%d size '%jd'", ENOSPC,
1292                             (intmax_t)gpp->gpp_size);
1293                         return (ENOSPC);
1294                 }
1295         }
1296
1297         pp = entry->gpe_pp;
1298         if ((g_debugflags & 16) == 0 &&
1299             (pp->acr > 0 || pp->acw > 0 || pp->ace > 0)) {
1300                 gctl_error(req, "%d", EBUSY);
1301                 return (EBUSY);
1302         }
1303
1304         error = G_PART_RESIZE(table, entry, gpp);
1305         if (error) {
1306                 gctl_error(req, "%d", error);
1307                 return (error);
1308         }
1309
1310         if (!entry->gpe_created)
1311                 entry->gpe_modified = 1;
1312
1313         /* update mediasize of changed provider */
1314         pp->mediasize = (entry->gpe_end - entry->gpe_start + 1) *
1315                 pp->sectorsize;
1316
1317         /* Provide feedback if so requested. */
1318         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1319                 sb = sbuf_new_auto();
1320                 G_PART_FULLNAME(table, entry, sb, gp->name);
1321                 sbuf_cat(sb, " resized\n");
1322                 sbuf_finish(sb);
1323                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1324                 sbuf_delete(sb);
1325         }
1326         return (0);
1327 }
1328
1329 static int
1330 g_part_ctl_setunset(struct gctl_req *req, struct g_part_parms *gpp,
1331     unsigned int set)
1332 {
1333         struct g_geom *gp;
1334         struct g_part_entry *entry;
1335         struct g_part_table *table;
1336         struct sbuf *sb;
1337         int error;
1338
1339         gp = gpp->gpp_geom;
1340         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1341         g_topology_assert();
1342
1343         table = gp->softc;
1344
1345         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1346                 if (entry->gpe_deleted || entry->gpe_internal)
1347                         continue;
1348                 if (entry->gpe_index == gpp->gpp_index)
1349                         break;
1350         }
1351         if (entry == NULL) {
1352                 gctl_error(req, "%d index '%d'", ENOENT, gpp->gpp_index);
1353                 return (ENOENT);
1354         }
1355
1356         error = G_PART_SETUNSET(table, entry, gpp->gpp_attrib, set);
1357         if (error) {
1358                 gctl_error(req, "%d attrib '%s'", error, gpp->gpp_attrib);
1359                 return (error);
1360         }
1361
1362         /* Provide feedback if so requested. */
1363         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1364                 sb = sbuf_new_auto();
1365                 sbuf_printf(sb, "%s %sset on ", gpp->gpp_attrib,
1366                     (set) ? "" : "un");
1367                 G_PART_FULLNAME(table, entry, sb, gp->name);
1368                 sbuf_printf(sb, "\n");
1369                 sbuf_finish(sb);
1370                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1371                 sbuf_delete(sb);
1372         }
1373         return (0);
1374 }
1375
1376 static int
1377 g_part_ctl_undo(struct gctl_req *req, struct g_part_parms *gpp)
1378 {
1379         struct g_consumer *cp;
1380         struct g_provider *pp;
1381         struct g_geom *gp;
1382         struct g_part_entry *entry, *tmp;
1383         struct g_part_table *table;
1384         int error, reprobe;
1385
1386         gp = gpp->gpp_geom;
1387         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1388         g_topology_assert();
1389
1390         table = gp->softc;
1391         if (!table->gpt_opened) {
1392                 gctl_error(req, "%d", EPERM);
1393                 return (EPERM);
1394         }
1395
1396         cp = LIST_FIRST(&gp->consumer);
1397         LIST_FOREACH_SAFE(entry, &table->gpt_entry, gpe_entry, tmp) {
1398                 entry->gpe_modified = 0;
1399                 if (entry->gpe_created) {
1400                         pp = entry->gpe_pp;
1401                         if (pp != NULL) {
1402                                 pp->private = NULL;
1403                                 entry->gpe_pp = NULL;
1404                                 g_wither_provider(pp, ENXIO);
1405                         }
1406                         entry->gpe_deleted = 1;
1407                 }
1408                 if (entry->gpe_deleted) {
1409                         LIST_REMOVE(entry, gpe_entry);
1410                         g_free(entry);
1411                 }
1412         }
1413
1414         g_topology_unlock();
1415
1416         reprobe = (table->gpt_scheme == &g_part_null_scheme ||
1417             table->gpt_created) ? 1 : 0;
1418
1419         if (reprobe) {
1420                 LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1421                         if (entry->gpe_internal)
1422                                 continue;
1423                         error = EBUSY;
1424                         goto fail;
1425                 }
1426                 while ((entry = LIST_FIRST(&table->gpt_entry)) != NULL) {
1427                         LIST_REMOVE(entry, gpe_entry);
1428                         g_free(entry);
1429                 }
1430                 error = g_part_probe(gp, cp, table->gpt_depth);
1431                 if (error) {
1432                         g_topology_lock();
1433                         g_access(cp, -1, -1, -1);
1434                         g_part_wither(gp, error);
1435                         return (0);
1436                 }
1437                 table = gp->softc;
1438
1439                 /*
1440                  * Synthesize a disk geometry. Some partitioning schemes
1441                  * depend on it and since some file systems need it even
1442                  * when the partitition scheme doesn't, we do it here in
1443                  * scheme-independent code.
1444                  */
1445                 pp = cp->provider;
1446                 g_part_geometry(table, cp, pp->mediasize / pp->sectorsize);
1447         }
1448
1449         error = G_PART_READ(table, cp);
1450         if (error)
1451                 goto fail;
1452         error = g_part_check_integrity(table, cp);
1453         if (error)
1454                 goto fail;
1455
1456         g_topology_lock();
1457         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1458                 if (!entry->gpe_internal)
1459                         g_part_new_provider(gp, table, entry);
1460         }
1461
1462         table->gpt_opened = 0;
1463         g_access(cp, -1, -1, -1);
1464         return (0);
1465
1466 fail:
1467         g_topology_lock();
1468         gctl_error(req, "%d", error);
1469         return (error);
1470 }
1471
1472 static void
1473 g_part_wither(struct g_geom *gp, int error)
1474 {
1475         struct g_part_entry *entry;
1476         struct g_part_table *table;
1477
1478         table = gp->softc;
1479         if (table != NULL) {
1480                 G_PART_DESTROY(table, NULL);
1481                 while ((entry = LIST_FIRST(&table->gpt_entry)) != NULL) {
1482                         LIST_REMOVE(entry, gpe_entry);
1483                         g_free(entry);
1484                 }
1485                 if (gp->softc != NULL) {
1486                         kobj_delete((kobj_t)gp->softc, M_GEOM);
1487                         gp->softc = NULL;
1488                 }
1489         }
1490         g_wither_geom(gp, error);
1491 }
1492
1493 /*
1494  * Class methods.
1495  */
1496
1497 static void
1498 g_part_ctlreq(struct gctl_req *req, struct g_class *mp, const char *verb)
1499 {
1500         struct g_part_parms gpp;
1501         struct g_part_table *table;
1502         struct gctl_req_arg *ap;
1503         enum g_part_ctl ctlreq;
1504         unsigned int i, mparms, oparms, parm;
1505         int auto_commit, close_on_error;
1506         int error, modifies;
1507
1508         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s,%s)", __func__, mp->name, verb));
1509         g_topology_assert();
1510
1511         ctlreq = G_PART_CTL_NONE;
1512         modifies = 1;
1513         mparms = 0;
1514         oparms = G_PART_PARM_FLAGS | G_PART_PARM_OUTPUT | G_PART_PARM_VERSION;
1515         switch (*verb) {
1516         case 'a':
1517                 if (!strcmp(verb, "add")) {
1518                         ctlreq = G_PART_CTL_ADD;
1519                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_SIZE |
1520                             G_PART_PARM_START | G_PART_PARM_TYPE;
1521                         oparms |= G_PART_PARM_INDEX | G_PART_PARM_LABEL;
1522                 }
1523                 break;
1524         case 'b':
1525                 if (!strcmp(verb, "bootcode")) {
1526                         ctlreq = G_PART_CTL_BOOTCODE;
1527                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_BOOTCODE;
1528                 }
1529                 break;
1530         case 'c':
1531                 if (!strcmp(verb, "commit")) {
1532                         ctlreq = G_PART_CTL_COMMIT;
1533                         mparms |= G_PART_PARM_GEOM;
1534                         modifies = 0;
1535                 } else if (!strcmp(verb, "create")) {
1536                         ctlreq = G_PART_CTL_CREATE;
1537                         mparms |= G_PART_PARM_PROVIDER | G_PART_PARM_SCHEME;
1538                         oparms |= G_PART_PARM_ENTRIES;
1539                 }
1540                 break;
1541         case 'd':
1542                 if (!strcmp(verb, "delete")) {
1543                         ctlreq = G_PART_CTL_DELETE;
1544                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_INDEX;
1545                 } else if (!strcmp(verb, "destroy")) {
1546                         ctlreq = G_PART_CTL_DESTROY;
1547                         mparms |= G_PART_PARM_GEOM;
1548                         oparms |= G_PART_PARM_FORCE;
1549                 }
1550                 break;
1551         case 'm':
1552                 if (!strcmp(verb, "modify")) {
1553                         ctlreq = G_PART_CTL_MODIFY;
1554                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_INDEX;
1555                         oparms |= G_PART_PARM_LABEL | G_PART_PARM_TYPE;
1556                 } else if (!strcmp(verb, "move")) {
1557                         ctlreq = G_PART_CTL_MOVE;
1558                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_INDEX;
1559                 }
1560                 break;
1561         case 'r':
1562                 if (!strcmp(verb, "recover")) {
1563                         ctlreq = G_PART_CTL_RECOVER;
1564                         mparms |= G_PART_PARM_GEOM;
1565                 } else if (!strcmp(verb, "resize")) {
1566                         ctlreq = G_PART_CTL_RESIZE;
1567                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_INDEX |
1568                             G_PART_PARM_SIZE;
1569                 }
1570                 break;
1571         case 's':
1572                 if (!strcmp(verb, "set")) {
1573                         ctlreq = G_PART_CTL_SET;
1574                         mparms |= G_PART_PARM_ATTRIB | G_PART_PARM_GEOM |
1575                             G_PART_PARM_INDEX;
1576                 }
1577                 break;
1578         case 'u':
1579                 if (!strcmp(verb, "undo")) {
1580                         ctlreq = G_PART_CTL_UNDO;
1581                         mparms |= G_PART_PARM_GEOM;
1582                         modifies = 0;
1583                 } else if (!strcmp(verb, "unset")) {
1584                         ctlreq = G_PART_CTL_UNSET;
1585                         mparms |= G_PART_PARM_ATTRIB | G_PART_PARM_GEOM |
1586                             G_PART_PARM_INDEX;
1587                 }
1588                 break;
1589         }
1590         if (ctlreq == G_PART_CTL_NONE) {
1591                 gctl_error(req, "%d verb '%s'", EINVAL, verb);
1592                 return;
1593         }
1594
1595         bzero(&gpp, sizeof(gpp));
1596         for (i = 0; i < req->narg; i++) {
1597                 ap = &req->arg[i];
1598                 parm = 0;
1599                 switch (ap->name[0]) {
1600                 case 'a':
1601                         if (!strcmp(ap->name, "arg0")) {
1602                                 parm = mparms &
1603                                     (G_PART_PARM_GEOM | G_PART_PARM_PROVIDER);
1604                         }
1605                         if (!strcmp(ap->name, "attrib"))
1606                                 parm = G_PART_PARM_ATTRIB;
1607                         break;
1608                 case 'b':
1609                         if (!strcmp(ap->name, "bootcode"))
1610                                 parm = G_PART_PARM_BOOTCODE;
1611                         break;
1612                 case 'c':
1613                         if (!strcmp(ap->name, "class"))
1614                                 continue;
1615                         break;
1616                 case 'e':
1617                         if (!strcmp(ap->name, "entries"))
1618                                 parm = G_PART_PARM_ENTRIES;
1619                         break;
1620                 case 'f':
1621                         if (!strcmp(ap->name, "flags"))
1622                                 parm = G_PART_PARM_FLAGS;
1623                         else if (!strcmp(ap->name, "force"))
1624                                 parm = G_PART_PARM_FORCE;
1625                         break;
1626                 case 'i':
1627                         if (!strcmp(ap->name, "index"))
1628                                 parm = G_PART_PARM_INDEX;
1629                         break;
1630                 case 'l':
1631                         if (!strcmp(ap->name, "label"))
1632                                 parm = G_PART_PARM_LABEL;
1633                         break;
1634                 case 'o':
1635                         if (!strcmp(ap->name, "output"))
1636                                 parm = G_PART_PARM_OUTPUT;
1637                         break;
1638                 case 's':
1639                         if (!strcmp(ap->name, "scheme"))
1640                                 parm = G_PART_PARM_SCHEME;
1641                         else if (!strcmp(ap->name, "size"))
1642                                 parm = G_PART_PARM_SIZE;
1643                         else if (!strcmp(ap->name, "start"))
1644                                 parm = G_PART_PARM_START;
1645                         break;
1646                 case 't':
1647                         if (!strcmp(ap->name, "type"))
1648                                 parm = G_PART_PARM_TYPE;
1649                         break;
1650                 case 'v':
1651                         if (!strcmp(ap->name, "verb"))
1652                                 continue;
1653                         else if (!strcmp(ap->name, "version"))
1654                                 parm = G_PART_PARM_VERSION;
1655                         break;
1656                 }
1657                 if ((parm & (mparms | oparms)) == 0) {
1658                         gctl_error(req, "%d param '%s'", EINVAL, ap->name);
1659                         return;
1660                 }
1661                 switch (parm) {
1662                 case G_PART_PARM_ATTRIB:
1663                         error = g_part_parm_str(req, ap->name,
1664                             &gpp.gpp_attrib);
1665                         break;
1666                 case G_PART_PARM_BOOTCODE:
1667                         error = g_part_parm_bootcode(req, ap->name,
1668                             &gpp.gpp_codeptr, &gpp.gpp_codesize);
1669                         break;
1670                 case G_PART_PARM_ENTRIES:
1671                         error = g_part_parm_intmax(req, ap->name,
1672                             &gpp.gpp_entries);
1673                         break;
1674                 case G_PART_PARM_FLAGS:
1675                         error = g_part_parm_str(req, ap->name, &gpp.gpp_flags);
1676                         break;
1677                 case G_PART_PARM_FORCE:
1678                         error = g_part_parm_uint32(req, ap->name,
1679                             &gpp.gpp_force);
1680                         break;
1681                 case G_PART_PARM_GEOM:
1682                         error = g_part_parm_geom(req, ap->name, &gpp.gpp_geom);
1683                         break;
1684                 case G_PART_PARM_INDEX:
1685                         error = g_part_parm_intmax(req, ap->name,
1686                             &gpp.gpp_index);
1687                         break;
1688                 case G_PART_PARM_LABEL:
1689                         error = g_part_parm_str(req, ap->name, &gpp.gpp_label);
1690                         break;
1691                 case G_PART_PARM_OUTPUT:
1692                         error = 0;      /* Write-only parameter */
1693                         break;
1694                 case G_PART_PARM_PROVIDER:
1695                         error = g_part_parm_provider(req, ap->name,
1696                             &gpp.gpp_provider);
1697                         break;
1698                 case G_PART_PARM_SCHEME:
1699                         error = g_part_parm_scheme(req, ap->name,
1700                             &gpp.gpp_scheme);
1701                         break;
1702                 case G_PART_PARM_SIZE:
1703                         error = g_part_parm_quad(req, ap->name, &gpp.gpp_size);
1704                         break;
1705                 case G_PART_PARM_START:
1706                         error = g_part_parm_quad(req, ap->name,
1707                             &gpp.gpp_start);
1708                         break;
1709                 case G_PART_PARM_TYPE:
1710                         error = g_part_parm_str(req, ap->name, &gpp.gpp_type);
1711                         break;
1712                 case G_PART_PARM_VERSION:
1713                         error = g_part_parm_uint32(req, ap->name,
1714                             &gpp.gpp_version);
1715                         break;
1716                 default:
1717                         error = EDOOFUS;
1718                         gctl_error(req, "%d %s", error, ap->name);
1719                         break;
1720                 }
1721                 if (error != 0) {
1722                         if (error == ENOATTR) {
1723                                 gctl_error(req, "%d param '%s'", error,
1724                                     ap->name);
1725                         }
1726                         return;
1727                 }
1728                 gpp.gpp_parms |= parm;
1729         }
1730         if ((gpp.gpp_parms & mparms) != mparms) {
1731                 parm = mparms - (gpp.gpp_parms & mparms);
1732                 gctl_error(req, "%d param '%x'", ENOATTR, parm);
1733                 return;
1734         }
1735
1736         /* Obtain permissions if possible/necessary. */
1737         close_on_error = 0;
1738         table = NULL;
1739         if (modifies && (gpp.gpp_parms & G_PART_PARM_GEOM)) {
1740                 table = gpp.gpp_geom->softc;
1741                 if (table != NULL && table->gpt_corrupt &&
1742                     ctlreq != G_PART_CTL_DESTROY &&
1743                     ctlreq != G_PART_CTL_RECOVER) {
1744                         gctl_error(req, "%d table '%s' is corrupt",
1745                             EPERM, gpp.gpp_geom->name);
1746                         return;
1747                 }
1748                 if (table != NULL && !table->gpt_opened) {
1749                         error = g_access(LIST_FIRST(&gpp.gpp_geom->consumer),
1750                             1, 1, 1);
1751                         if (error) {
1752                                 gctl_error(req, "%d geom '%s'", error,
1753                                     gpp.gpp_geom->name);
1754                                 return;
1755                         }
1756                         table->gpt_opened = 1;
1757                         close_on_error = 1;
1758                 }
1759         }
1760
1761         /* Allow the scheme to check or modify the parameters. */
1762         if (table != NULL) {
1763                 error = G_PART_PRECHECK(table, ctlreq, &gpp);
1764                 if (error) {
1765                         gctl_error(req, "%d pre-check failed", error);
1766                         goto out;
1767                 }
1768         } else
1769                 error = EDOOFUS;        /* Prevent bogus uninit. warning. */
1770
1771         switch (ctlreq) {
1772         case G_PART_CTL_NONE:
1773                 panic("%s", __func__);
1774         case G_PART_CTL_ADD:
1775                 error = g_part_ctl_add(req, &gpp);
1776                 break;
1777         case G_PART_CTL_BOOTCODE:
1778                 error = g_part_ctl_bootcode(req, &gpp);
1779                 break;
1780         case G_PART_CTL_COMMIT:
1781                 error = g_part_ctl_commit(req, &gpp);
1782                 break;
1783         case G_PART_CTL_CREATE:
1784                 error = g_part_ctl_create(req, &gpp);
1785                 break;
1786         case G_PART_CTL_DELETE:
1787                 error = g_part_ctl_delete(req, &gpp);
1788                 break;
1789         case G_PART_CTL_DESTROY:
1790                 error = g_part_ctl_destroy(req, &gpp);
1791                 break;
1792         case G_PART_CTL_MODIFY:
1793                 error = g_part_ctl_modify(req, &gpp);
1794                 break;
1795         case G_PART_CTL_MOVE:
1796                 error = g_part_ctl_move(req, &gpp);
1797                 break;
1798         case G_PART_CTL_RECOVER:
1799                 error = g_part_ctl_recover(req, &gpp);
1800                 break;
1801         case G_PART_CTL_RESIZE:
1802                 error = g_part_ctl_resize(req, &gpp);
1803                 break;
1804         case G_PART_CTL_SET:
1805                 error = g_part_ctl_setunset(req, &gpp, 1);
1806                 break;
1807         case G_PART_CTL_UNDO:
1808                 error = g_part_ctl_undo(req, &gpp);
1809                 break;
1810         case G_PART_CTL_UNSET:
1811                 error = g_part_ctl_setunset(req, &gpp, 0);
1812                 break;
1813         }
1814
1815         /* Implement automatic commit. */
1816         if (!error) {
1817                 auto_commit = (modifies &&
1818                     (gpp.gpp_parms & G_PART_PARM_FLAGS) &&
1819                     strchr(gpp.gpp_flags, 'C') != NULL) ? 1 : 0;
1820                 if (auto_commit) {
1821                         KASSERT(gpp.gpp_parms & G_PART_PARM_GEOM, ("%s",
1822                             __func__));
1823                         error = g_part_ctl_commit(req, &gpp);
1824                 }
1825         }
1826
1827  out:
1828         if (error && close_on_error) {
1829                 g_access(LIST_FIRST(&gpp.gpp_geom->consumer), -1, -1, -1);
1830                 table->gpt_opened = 0;
1831         }
1832 }
1833
1834 static int
1835 g_part_destroy_geom(struct gctl_req *req, struct g_class *mp,
1836     struct g_geom *gp)
1837 {
1838
1839         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s,%s)", __func__, mp->name, gp->name));
1840         g_topology_assert();
1841
1842         g_part_wither(gp, EINVAL);
1843         return (0);
1844 }
1845
1846 static struct g_geom *
1847 g_part_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
1848 {
1849         struct g_consumer *cp;
1850         struct g_geom *gp;
1851         struct g_part_entry *entry;
1852         struct g_part_table *table;
1853         struct root_hold_token *rht;
1854         int attr, depth;
1855         int error;
1856
1857         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s,%s)", __func__, mp->name, pp->name));
1858         g_topology_assert();
1859
1860         /* Skip providers that are already open for writing. */
1861         if (pp->acw > 0)
1862                 return (NULL);
1863
1864         /*
1865          * Create a GEOM with consumer and hook it up to the provider.
1866          * With that we become part of the topology. Optain read access
1867          * to the provider.
1868          */
1869         gp = g_new_geomf(mp, "%s", pp->name);
1870         cp = g_new_consumer(gp);
1871         error = g_attach(cp, pp);
1872         if (error == 0)
1873                 error = g_access(cp, 1, 0, 0);
1874         if (error != 0) {
1875                 g_part_wither(gp, error);
1876                 return (NULL);
1877         }
1878
1879         rht = root_mount_hold(mp->name);
1880         g_topology_unlock();
1881
1882         /*
1883          * Short-circuit the whole probing galore when there's no
1884          * media present.
1885          */
1886         if (pp->mediasize == 0 || pp->sectorsize == 0) {
1887                 error = ENODEV;
1888                 goto fail;
1889         }
1890
1891         /* Make sure we can nest and if so, determine our depth. */
1892         error = g_getattr("PART::isleaf", cp, &attr);
1893         if (!error && attr) {
1894                 error = ENODEV;
1895                 goto fail;
1896         }
1897         error = g_getattr("PART::depth", cp, &attr);
1898         depth = (!error) ? attr + 1 : 0;
1899
1900         error = g_part_probe(gp, cp, depth);
1901         if (error)
1902                 goto fail;
1903
1904         table = gp->softc;
1905
1906         /*
1907          * Synthesize a disk geometry. Some partitioning schemes
1908          * depend on it and since some file systems need it even
1909          * when the partitition scheme doesn't, we do it here in
1910          * scheme-independent code.
1911          */
1912         g_part_geometry(table, cp, pp->mediasize / pp->sectorsize);
1913
1914         error = G_PART_READ(table, cp);
1915         if (error)
1916                 goto fail;
1917         error = g_part_check_integrity(table, cp);
1918         if (error)
1919                 goto fail;
1920
1921         g_topology_lock();
1922         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1923                 if (!entry->gpe_internal)
1924                         g_part_new_provider(gp, table, entry);
1925         }
1926
1927         root_mount_rel(rht);
1928         g_access(cp, -1, 0, 0);
1929         return (gp);
1930
1931  fail:
1932         g_topology_lock();
1933         root_mount_rel(rht);
1934         g_access(cp, -1, 0, 0);
1935         g_part_wither(gp, error);
1936         return (NULL);
1937 }
1938
1939 /*
1940  * Geom methods.
1941  */
1942
1943 static int
1944 g_part_access(struct g_provider *pp, int dr, int dw, int de)
1945 {
1946         struct g_consumer *cp;
1947
1948         G_PART_TRACE((G_T_ACCESS, "%s(%s,%d,%d,%d)", __func__, pp->name, dr,
1949             dw, de));
1950
1951         cp = LIST_FIRST(&pp->geom->consumer);
1952
1953         /* We always gain write-exclusive access. */
1954         return (g_access(cp, dr, dw, dw + de));
1955 }
1956
1957 static void
1958 g_part_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1959     struct g_consumer *cp, struct g_provider *pp)
1960 {
1961         char buf[64];
1962         struct g_part_entry *entry;
1963         struct g_part_table *table;
1964
1965         KASSERT(sb != NULL && gp != NULL, ("%s", __func__));
1966         table = gp->softc;
1967
1968         if (indent == NULL) {
1969                 KASSERT(cp == NULL && pp != NULL, ("%s", __func__));
1970                 entry = pp->private;
1971                 if (entry == NULL)
1972                         return;
1973                 sbuf_printf(sb, " i %u o %ju ty %s", entry->gpe_index,
1974                     (uintmax_t)entry->gpe_offset,
1975                     G_PART_TYPE(table, entry, buf, sizeof(buf)));
1976                 /*
1977                  * libdisk compatibility quirk - the scheme dumps the
1978                  * slicer name and partition type in a way that is
1979                  * compatible with libdisk. When libdisk is not used
1980                  * anymore, this should go away.
1981                  */
1982                 G_PART_DUMPCONF(table, entry, sb, indent);
1983         } else if (cp != NULL) {        /* Consumer configuration. */
1984                 KASSERT(pp == NULL, ("%s", __func__));
1985                 /* none */
1986         } else if (pp != NULL) {        /* Provider configuration. */
1987                 entry = pp->private;
1988                 if (entry == NULL)
1989                         return;
1990                 sbuf_printf(sb, "%s<start>%ju</start>\n", indent,
1991                     (uintmax_t)entry->gpe_start);
1992                 sbuf_printf(sb, "%s<end>%ju</end>\n", indent,
1993                     (uintmax_t)entry->gpe_end);
1994                 sbuf_printf(sb, "%s<index>%u</index>\n", indent,
1995                     entry->gpe_index);
1996                 sbuf_printf(sb, "%s<type>%s</type>\n", indent,
1997                     G_PART_TYPE(table, entry, buf, sizeof(buf)));
1998                 sbuf_printf(sb, "%s<offset>%ju</offset>\n", indent,
1999                     (uintmax_t)entry->gpe_offset);
2000                 sbuf_printf(sb, "%s<length>%ju</length>\n", indent,
2001                     (uintmax_t)pp->mediasize);
2002                 G_PART_DUMPCONF(table, entry, sb, indent);
2003         } else {                        /* Geom configuration. */
2004                 sbuf_printf(sb, "%s<scheme>%s</scheme>\n", indent,
2005                     table->gpt_scheme->name);
2006                 sbuf_printf(sb, "%s<entries>%u</entries>\n", indent,
2007                     table->gpt_entries);
2008                 sbuf_printf(sb, "%s<first>%ju</first>\n", indent,
2009                     (uintmax_t)table->gpt_first);
2010                 sbuf_printf(sb, "%s<last>%ju</last>\n", indent,
2011                     (uintmax_t)table->gpt_last);
2012                 sbuf_printf(sb, "%s<fwsectors>%u</fwsectors>\n", indent,
2013                     table->gpt_sectors);
2014                 sbuf_printf(sb, "%s<fwheads>%u</fwheads>\n", indent,
2015                     table->gpt_heads);
2016                 sbuf_printf(sb, "%s<state>%s</state>\n", indent,
2017                     table->gpt_corrupt ? "CORRUPT": "OK");
2018                 sbuf_printf(sb, "%s<modified>%s</modified>\n", indent,
2019                     table->gpt_opened ? "true": "false");
2020                 G_PART_DUMPCONF(table, NULL, sb, indent);
2021         }
2022 }
2023
2024 static void
2025 g_part_orphan(struct g_consumer *cp)
2026 {
2027         struct g_provider *pp;
2028         struct g_part_table *table;
2029
2030         pp = cp->provider;
2031         KASSERT(pp != NULL, ("%s", __func__));
2032         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, pp->name));
2033         g_topology_assert();
2034
2035         KASSERT(pp->error != 0, ("%s", __func__));
2036         table = cp->geom->softc;
2037         if (table != NULL && table->gpt_opened)
2038                 g_access(cp, -1, -1, -1);
2039         g_part_wither(cp->geom, pp->error);
2040 }
2041
2042 static void
2043 g_part_spoiled(struct g_consumer *cp)
2044 {
2045
2046         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, cp->provider->name));
2047         g_topology_assert();
2048
2049         g_part_wither(cp->geom, ENXIO);
2050 }
2051
2052 static void
2053 g_part_start(struct bio *bp)
2054 {
2055         struct bio *bp2;
2056         struct g_consumer *cp;
2057         struct g_geom *gp;
2058         struct g_part_entry *entry;
2059         struct g_part_table *table;
2060         struct g_kerneldump *gkd;
2061         struct g_provider *pp;
2062         char buf[64];
2063
2064         pp = bp->bio_to;
2065         gp = pp->geom;
2066         table = gp->softc;
2067         cp = LIST_FIRST(&gp->consumer);
2068
2069         G_PART_TRACE((G_T_BIO, "%s: cmd=%d, provider=%s", __func__, bp->bio_cmd,
2070             pp->name));
2071
2072         entry = pp->private;
2073         if (entry == NULL) {
2074                 g_io_deliver(bp, ENXIO);
2075                 return;
2076         }
2077
2078         switch(bp->bio_cmd) {
2079         case BIO_DELETE:
2080         case BIO_READ:
2081         case BIO_WRITE:
2082                 if (bp->bio_offset >= pp->mediasize) {
2083                         g_io_deliver(bp, EIO);
2084                         return;
2085                 }
2086                 bp2 = g_clone_bio(bp);
2087                 if (bp2 == NULL) {
2088                         g_io_deliver(bp, ENOMEM);
2089                         return;
2090                 }
2091                 if (bp2->bio_offset + bp2->bio_length > pp->mediasize)
2092                         bp2->bio_length = pp->mediasize - bp2->bio_offset;
2093                 bp2->bio_done = g_std_done;
2094                 bp2->bio_offset += entry->gpe_offset;
2095                 g_io_request(bp2, cp);
2096                 return;
2097         case BIO_FLUSH:
2098                 break;
2099         case BIO_GETATTR:
2100                 if (g_handleattr_int(bp, "GEOM::fwheads", table->gpt_heads))
2101                         return;
2102                 if (g_handleattr_int(bp, "GEOM::fwsectors", table->gpt_sectors))
2103                         return;
2104                 if (g_handleattr_int(bp, "PART::isleaf", table->gpt_isleaf))
2105                         return;
2106                 if (g_handleattr_int(bp, "PART::depth", table->gpt_depth))
2107                         return;
2108                 if (g_handleattr_str(bp, "PART::scheme",
2109                     table->gpt_scheme->name))
2110                         return;
2111                 if (g_handleattr_str(bp, "PART::type",
2112                     G_PART_TYPE(table, entry, buf, sizeof(buf))))
2113                         return;
2114                 if (!strcmp("GEOM::kerneldump", bp->bio_attribute)) {
2115                         /*
2116                          * Check that the partition is suitable for kernel
2117                          * dumps. Typically only swap partitions should be
2118                          * used. If the request comes from the nested scheme
2119                          * we allow dumping there as well.
2120                          */
2121                         if ((bp->bio_from == NULL ||
2122                             bp->bio_from->geom->class != &g_part_class) &&
2123                             G_PART_DUMPTO(table, entry) == 0) {
2124                                 g_io_deliver(bp, ENODEV);
2125                                 printf("GEOM_PART: Partition '%s' not suitable"
2126                                     " for kernel dumps (wrong type?)\n",
2127                                     pp->name);
2128                                 return;
2129                         }
2130                         gkd = (struct g_kerneldump *)bp->bio_data;
2131                         if (gkd->offset >= pp->mediasize) {
2132                                 g_io_deliver(bp, EIO);
2133                                 return;
2134                         }
2135                         if (gkd->offset + gkd->length > pp->mediasize)
2136                                 gkd->length = pp->mediasize - gkd->offset;
2137                         gkd->offset += entry->gpe_offset;
2138                 }
2139                 break;
2140         default:
2141                 g_io_deliver(bp, EOPNOTSUPP);
2142                 return;
2143         }
2144
2145         bp2 = g_clone_bio(bp);
2146         if (bp2 == NULL) {
2147                 g_io_deliver(bp, ENOMEM);
2148                 return;
2149         }
2150         bp2->bio_done = g_std_done;
2151         g_io_request(bp2, cp);
2152 }
2153
2154 static void
2155 g_part_init(struct g_class *mp)
2156 {
2157
2158         TAILQ_INSERT_HEAD(&g_part_schemes, &g_part_null_scheme, scheme_list);
2159 }
2160
2161 static void
2162 g_part_fini(struct g_class *mp)
2163 {
2164
2165         TAILQ_REMOVE(&g_part_schemes, &g_part_null_scheme, scheme_list);
2166 }
2167
2168 static void
2169 g_part_unload_event(void *arg, int flag)
2170 {
2171         struct g_consumer *cp;
2172         struct g_geom *gp;
2173         struct g_provider *pp;
2174         struct g_part_scheme *scheme;
2175         struct g_part_table *table;
2176         uintptr_t *xchg;
2177         int acc, error;
2178
2179         if (flag == EV_CANCEL)
2180                 return;
2181
2182         xchg = arg;
2183         error = 0;
2184         scheme = (void *)(*xchg);
2185
2186         g_topology_assert();
2187
2188         LIST_FOREACH(gp, &g_part_class.geom, geom) {
2189                 table = gp->softc;
2190                 if (table->gpt_scheme != scheme)
2191                         continue;
2192
2193                 acc = 0;
2194                 LIST_FOREACH(pp, &gp->provider, provider)
2195                         acc += pp->acr + pp->acw + pp->ace;
2196                 LIST_FOREACH(cp, &gp->consumer, consumer)
2197                         acc += cp->acr + cp->acw + cp->ace;
2198
2199                 if (!acc)
2200                         g_part_wither(gp, ENOSYS);
2201                 else
2202                         error = EBUSY;
2203         }
2204
2205         if (!error)
2206                 TAILQ_REMOVE(&g_part_schemes, scheme, scheme_list);
2207
2208         *xchg = error;
2209 }
2210
2211 int
2212 g_part_modevent(module_t mod, int type, struct g_part_scheme *scheme)
2213 {
2214         struct g_part_scheme *iter;
2215         uintptr_t arg;
2216         int error;
2217
2218         error = 0;
2219         switch (type) {
2220         case MOD_LOAD:
2221                 TAILQ_FOREACH(iter, &g_part_schemes, scheme_list) {
2222                         if (scheme == iter) {
2223                                 printf("GEOM_PART: scheme %s is already "
2224                                     "registered!\n", scheme->name);
2225                                 break;
2226                         }
2227                 }
2228                 if (iter == NULL) {
2229                         TAILQ_INSERT_TAIL(&g_part_schemes, scheme,
2230                             scheme_list);
2231                         g_retaste(&g_part_class);
2232                 }
2233                 break;
2234         case MOD_UNLOAD:
2235                 arg = (uintptr_t)scheme;
2236                 error = g_waitfor_event(g_part_unload_event, &arg, M_WAITOK,
2237                     NULL);
2238                 if (error == 0)
2239                         error = arg;
2240                 break;
2241         default:
2242                 error = EOPNOTSUPP;
2243                 break;
2244         }
2245
2246         return (error);
2247 }