sys/geom/part/g_part.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2002, 2005-2009 Marcel Moolenaar
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/bio.h>
  34 #include <sys/endian.h>
  35 #include <sys/kernel.h>
  36 #include <sys/kobj.h>
  37 #include <sys/limits.h>
  38 #include <sys/lock.h>
  39 #include <sys/malloc.h>
  40 #include <sys/mutex.h>
  41 #include <sys/queue.h>
  42 #include <sys/sbuf.h>
  43 #include <sys/sysctl.h>
  44 #include <sys/systm.h>
  45 #include <sys/uuid.h>
  46 #include <geom/geom.h>
  47 #include <geom/geom_ctl.h>
  48 #include <geom/geom_int.h>
  49 #include <geom/part/g_part.h>
  50
  51 #include "g_part_if.h"
  52
  53 static kobj_method_t g_part_null_methods[] = {
  54         { 0, 0 }
  55 };
  56
  57 static struct g_part_scheme g_part_null_scheme = {
  58         "(none)",
  59         g_part_null_methods,
  60         sizeof(struct g_part_table),
  61 };
  62
  63 TAILQ_HEAD(, g_part_scheme) g_part_schemes =
  64     TAILQ_HEAD_INITIALIZER(g_part_schemes);
  65
  66 struct g_part_alias_list {
  67         const char *lexeme;
  68         enum g_part_alias alias;
  69 } g_part_alias_list[G_PART_ALIAS_COUNT] = {
  70         { "apple-apfs", G_PART_ALIAS_APPLE_APFS },
  71         { "apple-boot", G_PART_ALIAS_APPLE_BOOT },
  72         { "apple-core-storage", G_PART_ALIAS_APPLE_CORE_STORAGE },
  73         { "apple-hfs", G_PART_ALIAS_APPLE_HFS },
  74         { "apple-label", G_PART_ALIAS_APPLE_LABEL },
  75         { "apple-raid", G_PART_ALIAS_APPLE_RAID },
  76         { "apple-raid-offline", G_PART_ALIAS_APPLE_RAID_OFFLINE },
  77         { "apple-tv-recovery", G_PART_ALIAS_APPLE_TV_RECOVERY },
  78         { "apple-ufs", G_PART_ALIAS_APPLE_UFS },
  79         { "apple-zfs", G_PART_ALIAS_APPLE_ZFS },
  80         { "bios-boot", G_PART_ALIAS_BIOS_BOOT },
  81         { "chromeos-firmware", G_PART_ALIAS_CHROMEOS_FIRMWARE },
  82         { "chromeos-kernel", G_PART_ALIAS_CHROMEOS_KERNEL },
  83         { "chromeos-reserved", G_PART_ALIAS_CHROMEOS_RESERVED },
  84         { "chromeos-root", G_PART_ALIAS_CHROMEOS_ROOT },
  85         { "dragonfly-ccd", G_PART_ALIAS_DFBSD_CCD },
  86         { "dragonfly-hammer", G_PART_ALIAS_DFBSD_HAMMER },
  87         { "dragonfly-hammer2", G_PART_ALIAS_DFBSD_HAMMER2 },
  88         { "dragonfly-label32", G_PART_ALIAS_DFBSD },
  89         { "dragonfly-label64", G_PART_ALIAS_DFBSD64 },
  90         { "dragonfly-legacy", G_PART_ALIAS_DFBSD_LEGACY },
  91         { "dragonfly-swap", G_PART_ALIAS_DFBSD_SWAP },
  92         { "dragonfly-ufs", G_PART_ALIAS_DFBSD_UFS },
  93         { "dragonfly-vinum", G_PART_ALIAS_DFBSD_VINUM },
  94         { "ebr", G_PART_ALIAS_EBR },
  95         { "efi", G_PART_ALIAS_EFI },
  96         { "fat16", G_PART_ALIAS_MS_FAT16 },
  97         { "fat32", G_PART_ALIAS_MS_FAT32 },
  98         { "fat32lba", G_PART_ALIAS_MS_FAT32LBA },
  99         { "freebsd", G_PART_ALIAS_FREEBSD },
 100         { "freebsd-boot", G_PART_ALIAS_FREEBSD_BOOT },
 101         { "freebsd-nandfs", G_PART_ALIAS_FREEBSD_NANDFS },
 102         { "freebsd-swap", G_PART_ALIAS_FREEBSD_SWAP },
 103         { "freebsd-ufs", G_PART_ALIAS_FREEBSD_UFS },
 104         { "freebsd-vinum", G_PART_ALIAS_FREEBSD_VINUM },
 105         { "freebsd-zfs", G_PART_ALIAS_FREEBSD_ZFS },
 106         { "linux-data", G_PART_ALIAS_LINUX_DATA },
 107         { "linux-lvm", G_PART_ALIAS_LINUX_LVM },
 108         { "linux-raid", G_PART_ALIAS_LINUX_RAID },
 109         { "linux-swap", G_PART_ALIAS_LINUX_SWAP },
 110         { "mbr", G_PART_ALIAS_MBR },
 111         { "ms-basic-data", G_PART_ALIAS_MS_BASIC_DATA },
 112         { "ms-ldm-data", G_PART_ALIAS_MS_LDM_DATA },
 113         { "ms-ldm-metadata", G_PART_ALIAS_MS_LDM_METADATA },
 114         { "ms-recovery", G_PART_ALIAS_MS_RECOVERY },
 115         { "ms-reserved", G_PART_ALIAS_MS_RESERVED },
 116         { "ms-spaces", G_PART_ALIAS_MS_SPACES },
 117         { "netbsd-ccd", G_PART_ALIAS_NETBSD_CCD },
 118         { "netbsd-cgd", G_PART_ALIAS_NETBSD_CGD },
 119         { "netbsd-ffs", G_PART_ALIAS_NETBSD_FFS },
 120         { "netbsd-lfs", G_PART_ALIAS_NETBSD_LFS },
 121         { "netbsd-raid", G_PART_ALIAS_NETBSD_RAID },
 122         { "netbsd-swap", G_PART_ALIAS_NETBSD_SWAP },
 123         { "ntfs", G_PART_ALIAS_MS_NTFS },
 124         { "openbsd-data", G_PART_ALIAS_OPENBSD_DATA },
 125         { "prep-boot", G_PART_ALIAS_PREP_BOOT },
 126         { "solaris-boot", G_PART_ALIAS_SOLARIS_BOOT },
 127         { "solaris-root", G_PART_ALIAS_SOLARIS_ROOT },
 128         { "solaris-swap", G_PART_ALIAS_SOLARIS_SWAP },
 129         { "solaris-backup", G_PART_ALIAS_SOLARIS_BACKUP },
 130         { "solaris-var", G_PART_ALIAS_SOLARIS_VAR },
 131         { "solaris-home", G_PART_ALIAS_SOLARIS_HOME },
 132         { "solaris-altsec", G_PART_ALIAS_SOLARIS_ALTSEC },
 133         { "solaris-reserved", G_PART_ALIAS_SOLARIS_RESERVED },
 134         { "vmware-reserved", G_PART_ALIAS_VMRESERVED },
 135         { "vmware-vmfs", G_PART_ALIAS_VMFS },
 136         { "vmware-vmkdiag", G_PART_ALIAS_VMKDIAG },
 137         { "vmware-vsanhdr", G_PART_ALIAS_VMVSANHDR },
 138 };
 139
 140 SYSCTL_DECL(_kern_geom);
 141 SYSCTL_NODE(_kern_geom, OID_AUTO, part, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 142     "GEOM_PART stuff");
 143 u_int geom_part_check_integrity = 1;
 144 SYSCTL_UINT(_kern_geom_part, OID_AUTO, check_integrity,
 145     CTLFLAG_RWTUN, &geom_part_check_integrity, 1,
 146     "Enable integrity checking");
 147 static u_int auto_resize = 1;
 148 SYSCTL_UINT(_kern_geom_part, OID_AUTO, auto_resize,
 149     CTLFLAG_RWTUN, &auto_resize, 1,
 150     "Enable auto resize");
 151 static u_int allow_nesting = 0;
 152 SYSCTL_UINT(_kern_geom_part, OID_AUTO, allow_nesting,
 153     CTLFLAG_RWTUN, &allow_nesting, 0,
 154     "Allow additional levels of nesting");
 155 char g_part_separator[MAXPATHLEN] = "";
 156 SYSCTL_STRING(_kern_geom_part, OID_AUTO, separator,
 157     CTLFLAG_RDTUN, &g_part_separator, sizeof(g_part_separator),
 158     "Partition name separator");
 159
 160 /*
 161  * The GEOM partitioning class.
 162  */
 163 static g_ctl_req_t g_part_ctlreq;
 164 static g_ctl_destroy_geom_t g_part_destroy_geom;
 165 static g_fini_t g_part_fini;
 166 static g_init_t g_part_init;
 167 static g_taste_t g_part_taste;
 168
 169 static g_access_t g_part_access;
 170 static g_dumpconf_t g_part_dumpconf;
 171 static g_orphan_t g_part_orphan;
 172 static g_spoiled_t g_part_spoiled;
 173 static g_start_t g_part_start;
 174 static g_resize_t g_part_resize;
 175 static g_ioctl_t g_part_ioctl;
 176
 177 static struct g_class g_part_class = {
 178         .name = "PART",
 179         .version = G_VERSION,
 180         /* Class methods. */
 181         .ctlreq = g_part_ctlreq,
 182         .destroy_geom = g_part_destroy_geom,
 183         .fini = g_part_fini,
 184         .init = g_part_init,
 185         .taste = g_part_taste,
 186         /* Geom methods. */
 187         .access = g_part_access,
 188         .dumpconf = g_part_dumpconf,
 189         .orphan = g_part_orphan,
 190         .spoiled = g_part_spoiled,
 191         .start = g_part_start,
 192         .resize = g_part_resize,
 193         .ioctl = g_part_ioctl,
 194 };
 195
 196 DECLARE_GEOM_CLASS(g_part_class, g_part);
 197 MODULE_VERSION(g_part, 0);
 198
 199 /*
 200  * Support functions.
 201  */
 202
 203 static void g_part_wither(struct g_geom *, int);
 204
 205 const char *
 206 g_part_alias_name(enum g_part_alias alias)
 207 {
 208         int i;
 209
 210         for (i = 0; i < G_PART_ALIAS_COUNT; i++) {
 211                 if (g_part_alias_list[i].alias != alias)
 212                         continue;
 213                 return (g_part_alias_list[i].lexeme);
 214         }
 215
 216         return (NULL);
 217 }
 218
 219 void
 220 g_part_geometry_heads(off_t blocks, u_int sectors, off_t *bestchs,
 221     u_int *bestheads)
 222 {
 223         static u_int candidate_heads[] = { 1, 2, 16, 32, 64, 128, 255, 0 };
 224         off_t chs, cylinders;
 225         u_int heads;
 226         int idx;
 227
 228         *bestchs = 0;
 229         *bestheads = 0;
 230         for (idx = 0; candidate_heads[idx] != 0; idx++) {
 231                 heads = candidate_heads[idx];
 232                 cylinders = blocks / heads / sectors;
 233                 if (cylinders < heads || cylinders < sectors)
 234                         break;
 235                 if (cylinders > 1023)
 236                         continue;
 237                 chs = cylinders * heads * sectors;
 238                 if (chs > *bestchs || (chs == *bestchs && *bestheads == 1)) {
 239                         *bestchs = chs;
 240                         *bestheads = heads;
 241                 }
 242         }
 243 }
 244
 245 static void
 246 g_part_geometry(struct g_part_table *table, struct g_consumer *cp,
 247     off_t blocks)
 248 {
 249         static u_int candidate_sectors[] = { 1, 9, 17, 33, 63, 0 };
 250         off_t chs, bestchs;
 251         u_int heads, sectors;
 252         int idx;
 253
 254         if (g_getattr("GEOM::fwsectors", cp, &sectors) != 0 || sectors == 0 ||
 255             g_getattr("GEOM::fwheads", cp, &heads) != 0 || heads == 0) {
 256                 table->gpt_fixgeom = 0;
 257                 table->gpt_heads = 0;
 258                 table->gpt_sectors = 0;
 259                 bestchs = 0;
 260                 for (idx = 0; candidate_sectors[idx] != 0; idx++) {
 261                         sectors = candidate_sectors[idx];
 262                         g_part_geometry_heads(blocks, sectors, &chs, &heads);
 263                         if (chs == 0)
 264                                 continue;
 265                         /*
 266                          * Prefer a geometry with sectors > 1, but only if
 267                          * it doesn't bump down the number of heads to 1.
 268                          */
 269                         if (chs > bestchs || (chs == bestchs && heads > 1 &&
 270                             table->gpt_sectors == 1)) {
 271                                 bestchs = chs;
 272                                 table->gpt_heads = heads;
 273                                 table->gpt_sectors = sectors;
 274                         }
 275                 }
 276                 /*
 277                  * If we didn't find a geometry at all, then the disk is
 278                  * too big. This means we can use the maximum number of
 279                  * heads and sectors.
 280                  */
 281                 if (bestchs == 0) {
 282                         table->gpt_heads = 255;
 283                         table->gpt_sectors = 63;
 284                 }
 285         } else {
 286                 table->gpt_fixgeom = 1;
 287                 table->gpt_heads = heads;
 288                 table->gpt_sectors = sectors;
 289         }
 290 }
 291
 292 static void
 293 g_part_get_physpath_done(struct bio *bp)
 294 {
 295         struct g_geom *gp;
 296         struct g_part_entry *entry;
 297         struct g_part_table *table;
 298         struct g_provider *pp;
 299         struct bio *pbp;
 300
 301         pbp = bp->bio_parent;
 302         pp = pbp->bio_to;
 303         gp = pp->geom;
 304         table = gp->softc;
 305         entry = pp->private;
 306
 307         if (bp->bio_error == 0) {
 308                 char *end;
 309                 size_t len, remainder;
 310                 len = strlcat(bp->bio_data, "/", bp->bio_length);
 311                 if (len < bp->bio_length) {
 312                         end = bp->bio_data + len;
 313                         remainder = bp->bio_length - len;
 314                         G_PART_NAME(table, entry, end, remainder);
 315                 }
 316         }
 317         g_std_done(bp);
 318 }
 319
 320 #define DPRINTF(...)    if (bootverbose) {      \
 321         printf("GEOM_PART: " __VA_ARGS__);      \
 322 }
 323
 324 static int
 325 g_part_check_integrity(struct g_part_table *table, struct g_consumer *cp)
 326 {
 327         struct g_part_entry *e1, *e2;
 328         struct g_provider *pp;
 329         off_t offset;
 330         int failed;
 331
 332         failed = 0;
 333         pp = cp->provider;
 334         if (table->gpt_last < table->gpt_first) {
 335                 DPRINTF("last LBA is below first LBA: %jd < %jd\n",
 336                     (intmax_t)table->gpt_last, (intmax_t)table->gpt_first);
 337                 failed++;
 338         }
 339         if (table->gpt_last > pp->mediasize / pp->sectorsize - 1) {
 340                 DPRINTF("last LBA extends beyond mediasize: "
 341                     "%jd > %jd\n", (intmax_t)table->gpt_last,
 342                     (intmax_t)pp->mediasize / pp->sectorsize - 1);
 343                 failed++;
 344         }
 345         LIST_FOREACH(e1, &table->gpt_entry, gpe_entry) {
 346                 if (e1->gpe_deleted || e1->gpe_internal)
 347                         continue;
 348                 if (e1->gpe_start < table->gpt_first) {
 349                         DPRINTF("partition %d has start offset below first "
 350                             "LBA: %jd < %jd\n", e1->gpe_index,
 351                             (intmax_t)e1->gpe_start,
 352                             (intmax_t)table->gpt_first);
 353                         failed++;
 354                 }
 355                 if (e1->gpe_start > table->gpt_last) {
 356                         DPRINTF("partition %d has start offset beyond last "
 357                             "LBA: %jd > %jd\n", e1->gpe_index,
 358                             (intmax_t)e1->gpe_start,
 359                             (intmax_t)table->gpt_last);
 360                         failed++;
 361                 }
 362                 if (e1->gpe_end < e1->gpe_start) {
 363                         DPRINTF("partition %d has end offset below start "
 364                             "offset: %jd < %jd\n", e1->gpe_index,
 365                             (intmax_t)e1->gpe_end,
 366                             (intmax_t)e1->gpe_start);
 367                         failed++;
 368                 }
 369                 if (e1->gpe_end > table->gpt_last) {
 370                         DPRINTF("partition %d has end offset beyond last "
 371                             "LBA: %jd > %jd\n", e1->gpe_index,
 372                             (intmax_t)e1->gpe_end,
 373                             (intmax_t)table->gpt_last);
 374                         failed++;
 375                 }
 376                 if (pp->stripesize > 0) {
 377                         offset = e1->gpe_start * pp->sectorsize;
 378                         if (e1->gpe_offset > offset)
 379                                 offset = e1->gpe_offset;
 380                         if ((offset + pp->stripeoffset) % pp->stripesize) {
 381                                 DPRINTF("partition %d on (%s, %s) is not "
 382                                     "aligned on %ju bytes\n", e1->gpe_index,
 383                                     pp->name, table->gpt_scheme->name,
 384                                     (uintmax_t)pp->stripesize);
 385                                 /* Don't treat this as a critical failure */
 386                         }
 387                 }
 388                 e2 = e1;
 389                 while ((e2 = LIST_NEXT(e2, gpe_entry)) != NULL) {
 390                         if (e2->gpe_deleted || e2->gpe_internal)
 391                                 continue;
 392                         if (e1->gpe_start >= e2->gpe_start &&
 393                             e1->gpe_start <= e2->gpe_end) {
 394                                 DPRINTF("partition %d has start offset inside "
 395                                     "partition %d: start[%d] %jd >= start[%d] "
 396                                     "%jd <= end[%d] %jd\n",
 397                                     e1->gpe_index, e2->gpe_index,
 398                                     e2->gpe_index, (intmax_t)e2->gpe_start,
 399                                     e1->gpe_index, (intmax_t)e1->gpe_start,
 400                                     e2->gpe_index, (intmax_t)e2->gpe_end);
 401                                 failed++;
 402                         }
 403                         if (e1->gpe_end >= e2->gpe_start &&
 404                             e1->gpe_end <= e2->gpe_end) {
 405                                 DPRINTF("partition %d has end offset inside "
 406                                     "partition %d: start[%d] %jd >= end[%d] "
 407                                     "%jd <= end[%d] %jd\n",
 408                                     e1->gpe_index, e2->gpe_index,
 409                                     e2->gpe_index, (intmax_t)e2->gpe_start,
 410                                     e1->gpe_index, (intmax_t)e1->gpe_end,
 411                                     e2->gpe_index, (intmax_t)e2->gpe_end);
 412                                 failed++;
 413                         }
 414                         if (e1->gpe_start < e2->gpe_start &&
 415                             e1->gpe_end > e2->gpe_end) {
 416                                 DPRINTF("partition %d contains partition %d: "
 417                                     "start[%d] %jd > start[%d] %jd, end[%d] "
 418                                     "%jd < end[%d] %jd\n",
 419                                     e1->gpe_index, e2->gpe_index,
 420                                     e1->gpe_index, (intmax_t)e1->gpe_start,
 421                                     e2->gpe_index, (intmax_t)e2->gpe_start,
 422                                     e2->gpe_index, (intmax_t)e2->gpe_end,
 423                                     e1->gpe_index, (intmax_t)e1->gpe_end);
 424                                 failed++;
 425                         }
 426                 }
 427         }
 428         if (failed != 0) {
 429                 printf("GEOM_PART: integrity check failed (%s, %s)\n",
 430                     pp->name, table->gpt_scheme->name);
 431                 if (geom_part_check_integrity != 0)
 432                         return (EINVAL);
 433                 table->gpt_corrupt = 1;
 434         }
 435         return (0);
 436 }
 437 #undef  DPRINTF
 438
 439 struct g_part_entry *
 440 g_part_new_entry(struct g_part_table *table, int index, quad_t start,
 441     quad_t end)
 442 {
 443         struct g_part_entry *entry, *last;
 444
 445         last = NULL;
 446         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
 447                 if (entry->gpe_index == index)
 448                         break;
 449                 if (entry->gpe_index > index) {
 450                         entry = NULL;
 451                         break;
 452                 }
 453                 last = entry;
 454         }
 455         if (entry == NULL) {
 456                 entry = g_malloc(table->gpt_scheme->gps_entrysz,
 457                     M_WAITOK | M_ZERO);
 458                 entry->gpe_index = index;
 459                 if (last == NULL)
 460                         LIST_INSERT_HEAD(&table->gpt_entry, entry, gpe_entry);
 461                 else
 462                         LIST_INSERT_AFTER(last, entry, gpe_entry);
 463         } else
 464                 entry->gpe_offset = 0;
 465         entry->gpe_start = start;
 466         entry->gpe_end = end;
 467         return (entry);
 468 }
 469
 470 static void
 471 g_part_new_provider(struct g_geom *gp, struct g_part_table *table,
 472     struct g_part_entry *entry)
 473 {
 474         struct g_consumer *cp;
 475         struct g_provider *pp;
 476         struct g_geom_alias *gap;
 477         off_t offset;
 478
 479         cp = LIST_FIRST(&gp->consumer);
 480         pp = cp->provider;
 481
 482         offset = entry->gpe_start * pp->sectorsize;
 483         if (entry->gpe_offset < offset)
 484                 entry->gpe_offset = offset;
 485
 486         if (entry->gpe_pp == NULL) {
 487                 entry->gpe_pp = G_PART_NEW_PROVIDER(table, gp, entry, gp->name);
 488                 /*
 489                  * If our parent provider had any aliases, then copy them to our
 490                  * provider so when geom DEV tastes things later, they will be
 491                  * there for it to create the aliases with those name used in
 492                  * place of the geom's name we use to create the provider. The
 493                  * kobj interface that generates names makes this awkward.
 494                  */
 495                 LIST_FOREACH(gap, &pp->aliases, ga_next)
 496                         G_PART_ADD_ALIAS(table, entry->gpe_pp, entry, gap->ga_alias);
 497                 entry->gpe_pp->flags |= G_PF_DIRECT_SEND | G_PF_DIRECT_RECEIVE;
 498                 entry->gpe_pp->private = entry;         /* Close the circle. */
 499         }
 500         entry->gpe_pp->index = entry->gpe_index - 1;    /* index is 1-based. */
 501         entry->gpe_pp->mediasize = (entry->gpe_end - entry->gpe_start + 1) *
 502             pp->sectorsize;
 503         entry->gpe_pp->mediasize -= entry->gpe_offset - offset;
 504         entry->gpe_pp->sectorsize = pp->sectorsize;
 505         entry->gpe_pp->stripesize = pp->stripesize;
 506         entry->gpe_pp->stripeoffset = pp->stripeoffset + entry->gpe_offset;
 507         if (pp->stripesize > 0)
 508                 entry->gpe_pp->stripeoffset %= pp->stripesize;
 509         entry->gpe_pp->flags |= pp->flags & G_PF_ACCEPT_UNMAPPED;
 510         g_error_provider(entry->gpe_pp, 0);
 511 }
 512
 513 static struct g_geom*
 514 g_part_find_geom(const char *name)
 515 {
 516         struct g_geom *gp;
 517         LIST_FOREACH(gp, &g_part_class.geom, geom) {
 518                 if ((gp->flags & G_GEOM_WITHER) == 0 &&
 519                     strcmp(name, gp->name) == 0)
 520                         break;
 521         }
 522         return (gp);
 523 }
 524
 525 static int
 526 g_part_parm_geom(struct gctl_req *req, const char *name, struct g_geom **v)
 527 {
 528         struct g_geom *gp;
 529         const char *gname;
 530
 531         gname = gctl_get_asciiparam(req, name);
 532         if (gname == NULL)
 533                 return (ENOATTR);
 534         if (strncmp(gname, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
 535                 gname += sizeof(_PATH_DEV) - 1;
 536         gp = g_part_find_geom(gname);
 537         if (gp == NULL) {
 538                 gctl_error(req, "%d %s '%s'", EINVAL, name, gname);
 539                 return (EINVAL);
 540         }
 541         *v = gp;
 542         return (0);
 543 }
 544
 545 static int
 546 g_part_parm_provider(struct gctl_req *req, const char *name,
 547     struct g_provider **v)
 548 {
 549         struct g_provider *pp;
 550         const char *pname;
 551
 552         pname = gctl_get_asciiparam(req, name);
 553         if (pname == NULL)
 554                 return (ENOATTR);
 555         if (strncmp(pname, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
 556                 pname += sizeof(_PATH_DEV) - 1;
 557         pp = g_provider_by_name(pname);
 558         if (pp == NULL) {
 559                 gctl_error(req, "%d %s '%s'", EINVAL, name, pname);
 560                 return (EINVAL);
 561         }
 562         *v = pp;
 563         return (0);
 564 }
 565
 566 static int
 567 g_part_parm_quad(struct gctl_req *req, const char *name, quad_t *v)
 568 {
 569         const char *p;
 570         char *x;
 571         quad_t q;
 572
 573         p = gctl_get_asciiparam(req, name);
 574         if (p == NULL)
 575                 return (ENOATTR);
 576         q = strtoq(p, &x, 0);
 577         if (*x != '\0' || q < 0) {
 578                 gctl_error(req, "%d %s '%s'", EINVAL, name, p);
 579                 return (EINVAL);
 580         }
 581         *v = q;
 582         return (0);
 583 }
 584
 585 static int
 586 g_part_parm_scheme(struct gctl_req *req, const char *name,
 587     struct g_part_scheme **v)
 588 {
 589         struct g_part_scheme *s;
 590         const char *p;
 591
 592         p = gctl_get_asciiparam(req, name);
 593         if (p == NULL)
 594                 return (ENOATTR);
 595         TAILQ_FOREACH(s, &g_part_schemes, scheme_list) {
 596                 if (s == &g_part_null_scheme)
 597                         continue;
 598                 if (!strcasecmp(s->name, p))
 599                         break;
 600         }
 601         if (s == NULL) {
 602                 gctl_error(req, "%d %s '%s'", EINVAL, name, p);
 603                 return (EINVAL);
 604         }
 605         *v = s;
 606         return (0);
 607 }
 608
 609 static int
 610 g_part_parm_str(struct gctl_req *req, const char *name, const char **v)
 611 {
 612         const char *p;
 613
 614         p = gctl_get_asciiparam(req, name);
 615         if (p == NULL)
 616                 return (ENOATTR);
 617         /* An empty label is always valid. */
 618         if (strcmp(name, "label") != 0 && p[0] == '\0') {
 619                 gctl_error(req, "%d %s '%s'", EINVAL, name, p);
 620                 return (EINVAL);
 621         }
 622         *v = p;
 623         return (0);
 624 }
 625
 626 static int
 627 g_part_parm_intmax(struct gctl_req *req, const char *name, u_int *v)
 628 {
 629         const intmax_t *p;
 630         int size;
 631
 632         p = gctl_get_param(req, name, &size);
 633         if (p == NULL)
 634                 return (ENOATTR);
 635         if (size != sizeof(*p) || *p < 0 || *p > INT_MAX) {
 636                 gctl_error(req, "%d %s '%jd'", EINVAL, name, *p);
 637                 return (EINVAL);
 638         }
 639         *v = (u_int)*p;
 640         return (0);
 641 }
 642
 643 static int
 644 g_part_parm_uint32(struct gctl_req *req, const char *name, u_int *v)
 645 {
 646         const uint32_t *p;
 647         int size;
 648
 649         p = gctl_get_param(req, name, &size);
 650         if (p == NULL)
 651                 return (ENOATTR);
 652         if (size != sizeof(*p) || *p > INT_MAX) {
 653                 gctl_error(req, "%d %s '%u'", EINVAL, name, (unsigned int)*p);
 654                 return (EINVAL);
 655         }
 656         *v = (u_int)*p;
 657         return (0);
 658 }
 659
 660 static int
 661 g_part_parm_bootcode(struct gctl_req *req, const char *name, const void **v,
 662     unsigned int *s)
 663 {
 664         const void *p;
 665         int size;
 666
 667         p = gctl_get_param(req, name, &size);
 668         if (p == NULL)
 669                 return (ENOATTR);
 670         *v = p;
 671         *s = size;
 672         return (0);
 673 }
 674
 675 static int
 676 g_part_probe(struct g_geom *gp, struct g_consumer *cp, int depth)
 677 {
 678         struct g_part_scheme *iter, *scheme;
 679         struct g_part_table *table;
 680         int pri, probe;
 681
 682         table = gp->softc;
 683         scheme = (table != NULL) ? table->gpt_scheme : NULL;
 684         pri = (scheme != NULL) ? G_PART_PROBE(table, cp) : INT_MIN;
 685         if (pri == 0)
 686                 goto done;
 687         if (pri > 0) {  /* error */
 688                 scheme = NULL;
 689                 pri = INT_MIN;
 690         }
 691
 692         TAILQ_FOREACH(iter, &g_part_schemes, scheme_list) {
 693                 if (iter == &g_part_null_scheme)
 694                         continue;
 695                 table = (void *)kobj_create((kobj_class_t)iter, M_GEOM,
 696                     M_WAITOK);
 697                 table->gpt_gp = gp;
 698                 table->gpt_scheme = iter;
 699                 table->gpt_depth = depth;
 700                 probe = G_PART_PROBE(table, cp);
 701                 if (probe <= 0 && probe > pri) {
 702                         pri = probe;
 703                         scheme = iter;
 704                         if (gp->softc != NULL)
 705                                 kobj_delete((kobj_t)gp->softc, M_GEOM);
 706                         gp->softc = table;
 707                         if (pri == 0)
 708                                 goto done;
 709                 } else
 710                         kobj_delete((kobj_t)table, M_GEOM);
 711         }
 712
 713 done:
 714         return ((scheme == NULL) ? ENXIO : 0);
 715 }
 716
 717 /*
 718  * Control request functions.
 719  */
 720
 721 static int
 722 g_part_ctl_add(struct gctl_req *req, struct g_part_parms *gpp)
 723 {
 724         struct g_geom *gp;
 725         struct g_provider *pp;
 726         struct g_part_entry *delent, *last, *entry;
 727         struct g_part_table *table;
 728         struct sbuf *sb;
 729         quad_t end;
 730         unsigned int index;
 731         int error;
 732
 733         gp = gpp->gpp_geom;
 734         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
 735         g_topology_assert();
 736
 737         pp = LIST_FIRST(&gp->consumer)->provider;
 738         table = gp->softc;
 739         end = gpp->gpp_start + gpp->gpp_size - 1;
 740
 741         if (gpp->gpp_start < table->gpt_first ||
 742             gpp->gpp_start > table->gpt_last) {
 743                 gctl_error(req, "%d start '%jd'", EINVAL,
 744                     (intmax_t)gpp->gpp_start);
 745                 return (EINVAL);
 746         }
 747         if (end < gpp->gpp_start || end > table->gpt_last) {
 748                 gctl_error(req, "%d size '%jd'", EINVAL,
 749                     (intmax_t)gpp->gpp_size);
 750                 return (EINVAL);
 751         }
 752         if (gpp->gpp_index > table->gpt_entries) {
 753                 gctl_error(req, "%d index '%d'", EINVAL, gpp->gpp_index);
 754                 return (EINVAL);
 755         }
 756
 757         delent = last = NULL;
 758         index = (gpp->gpp_index > 0) ? gpp->gpp_index : 1;
 759         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
 760                 if (entry->gpe_deleted) {
 761                         if (entry->gpe_index == index)
 762                                 delent = entry;
 763                         continue;
 764                 }
 765                 if (entry->gpe_index == index)
 766                         index = entry->gpe_index + 1;
 767                 if (entry->gpe_index < index)
 768                         last = entry;
 769                 if (entry->gpe_internal)
 770                         continue;
 771                 if (gpp->gpp_start >= entry->gpe_start &&
 772                     gpp->gpp_start <= entry->gpe_end) {
 773                         gctl_error(req, "%d start '%jd'", ENOSPC,
 774                             (intmax_t)gpp->gpp_start);
 775                         return (ENOSPC);
 776                 }
 777                 if (end >= entry->gpe_start && end <= entry->gpe_end) {
 778                         gctl_error(req, "%d end '%jd'", ENOSPC, (intmax_t)end);
 779                         return (ENOSPC);
 780                 }
 781                 if (gpp->gpp_start < entry->gpe_start && end > entry->gpe_end) {
 782                         gctl_error(req, "%d size '%jd'", ENOSPC,
 783                             (intmax_t)gpp->gpp_size);
 784                         return (ENOSPC);
 785                 }
 786         }
 787         if (gpp->gpp_index > 0 && index != gpp->gpp_index) {
 788                 gctl_error(req, "%d index '%d'", EEXIST, gpp->gpp_index);
 789                 return (EEXIST);
 790         }
 791         if (index > table->gpt_entries) {
 792                 gctl_error(req, "%d index '%d'", ENOSPC, index);
 793                 return (ENOSPC);
 794         }
 795
 796         entry = (delent == NULL) ? g_malloc(table->gpt_scheme->gps_entrysz,
 797             M_WAITOK | M_ZERO) : delent;
 798         entry->gpe_index = index;
 799         entry->gpe_start = gpp->gpp_start;
 800         entry->gpe_end = end;
 801         error = G_PART_ADD(table, entry, gpp);
 802         if (error) {
 803                 gctl_error(req, "%d", error);
 804                 if (delent == NULL)
 805                         g_free(entry);
 806                 return (error);
 807         }
 808         if (delent == NULL) {
 809                 if (last == NULL)
 810                         LIST_INSERT_HEAD(&table->gpt_entry, entry, gpe_entry);
 811                 else
 812                         LIST_INSERT_AFTER(last, entry, gpe_entry);
 813                 entry->gpe_created = 1;
 814         } else {
 815                 entry->gpe_deleted = 0;
 816                 entry->gpe_modified = 1;
 817         }
 818         g_part_new_provider(gp, table, entry);
 819
 820         /* Provide feedback if so requested. */
 821         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
 822                 sb = sbuf_new_auto();
 823                 G_PART_FULLNAME(table, entry, sb, gp->name);
 824                 if (pp->stripesize > 0 && entry->gpe_pp->stripeoffset != 0)
 825                         sbuf_printf(sb, " added, but partition is not "
 826                             "aligned on %ju bytes\n", (uintmax_t)pp->stripesize);
 827                 else
 828                         sbuf_cat(sb, " added\n");
 829                 sbuf_finish(sb);
 830                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
 831                 sbuf_delete(sb);
 832         }
 833         return (0);
 834 }
 835
 836 static int
 837 g_part_ctl_bootcode(struct gctl_req *req, struct g_part_parms *gpp)
 838 {
 839         struct g_geom *gp;
 840         struct g_part_table *table;
 841         struct sbuf *sb;
 842         int error, sz;
 843
 844         gp = gpp->gpp_geom;
 845         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
 846         g_topology_assert();
 847
 848         table = gp->softc;
 849         sz = table->gpt_scheme->gps_bootcodesz;
 850         if (sz == 0) {
 851                 error = ENODEV;
 852                 goto fail;
 853         }
 854         if (gpp->gpp_codesize > sz) {
 855                 error = EFBIG;
 856                 goto fail;
 857         }
 858
 859         error = G_PART_BOOTCODE(table, gpp);
 860         if (error)
 861                 goto fail;
 862
 863         /* Provide feedback if so requested. */
 864         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
 865                 sb = sbuf_new_auto();
 866                 sbuf_printf(sb, "bootcode written to %s\n", gp->name);
 867                 sbuf_finish(sb);
 868                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
 869                 sbuf_delete(sb);
 870         }
 871         return (0);
 872
 873  fail:
 874         gctl_error(req, "%d", error);
 875         return (error);
 876 }
 877
 878 static int
 879 g_part_ctl_commit(struct gctl_req *req, struct g_part_parms *gpp)
 880 {
 881         struct g_consumer *cp;
 882         struct g_geom *gp;
 883         struct g_provider *pp;
 884         struct g_part_entry *entry, *tmp;
 885         struct g_part_table *table;
 886         char *buf;
 887         int error, i;
 888
 889         gp = gpp->gpp_geom;
 890         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
 891         g_topology_assert();
 892
 893         table = gp->softc;
 894         if (!table->gpt_opened) {
 895                 gctl_error(req, "%d", EPERM);
 896                 return (EPERM);
 897         }
 898
 899         g_topology_unlock();
 900
 901         cp = LIST_FIRST(&gp->consumer);
 902         if ((table->gpt_smhead | table->gpt_smtail) != 0) {
 903                 pp = cp->provider;
 904                 buf = g_malloc(pp->sectorsize, M_WAITOK | M_ZERO);
 905                 while (table->gpt_smhead != 0) {
 906                         i = ffs(table->gpt_smhead) - 1;
 907                         error = g_write_data(cp, i * pp->sectorsize, buf,
 908                             pp->sectorsize);
 909                         if (error) {
 910                                 g_free(buf);
 911                                 goto fail;
 912                         }
 913                         table->gpt_smhead &= ~(1 << i);
 914                 }
 915                 while (table->gpt_smtail != 0) {
 916                         i = ffs(table->gpt_smtail) - 1;
 917                         error = g_write_data(cp, pp->mediasize - (i + 1) *
 918                             pp->sectorsize, buf, pp->sectorsize);
 919                         if (error) {
 920                                 g_free(buf);
 921                                 goto fail;
 922                         }
 923                         table->gpt_smtail &= ~(1 << i);
 924                 }
 925                 g_free(buf);
 926         }
 927
 928         if (table->gpt_scheme == &g_part_null_scheme) {
 929                 g_topology_lock();
 930                 g_access(cp, -1, -1, -1);
 931                 g_part_wither(gp, ENXIO);
 932                 return (0);
 933         }
 934
 935         error = G_PART_WRITE(table, cp);
 936         if (error)
 937                 goto fail;
 938
 939         LIST_FOREACH_SAFE(entry, &table->gpt_entry, gpe_entry, tmp) {
 940                 if (!entry->gpe_deleted) {
 941                         /* Notify consumers that provider might be changed. */
 942                         if (entry->gpe_modified && (
 943                             entry->gpe_pp->acw + entry->gpe_pp->ace +
 944                             entry->gpe_pp->acr) == 0)
 945                                 g_media_changed(entry->gpe_pp, M_NOWAIT);
 946                         entry->gpe_created = 0;
 947                         entry->gpe_modified = 0;
 948                         continue;
 949                 }
 950                 LIST_REMOVE(entry, gpe_entry);
 951                 g_free(entry);
 952         }
 953         table->gpt_created = 0;
 954         table->gpt_opened = 0;
 955
 956         g_topology_lock();
 957         g_access(cp, -1, -1, -1);
 958         return (0);
 959
 960 fail:
 961         g_topology_lock();
 962         gctl_error(req, "%d", error);
 963         return (error);
 964 }
 965
 966 static int
 967 g_part_ctl_create(struct gctl_req *req, struct g_part_parms *gpp)
 968 {
 969         struct g_consumer *cp;
 970         struct g_geom *gp;
 971         struct g_provider *pp;
 972         struct g_part_scheme *scheme;
 973         struct g_part_table *null, *table;
 974         struct sbuf *sb;
 975         int attr, error;
 976
 977         pp = gpp->gpp_provider;
 978         scheme = gpp->gpp_scheme;
 979         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, pp->name));
 980         g_topology_assert();
 981
 982         /* Check that there isn't already a g_part geom on the provider. */
 983         gp = g_part_find_geom(pp->name);
 984         if (gp != NULL) {
 985                 null = gp->softc;
 986                 if (null->gpt_scheme != &g_part_null_scheme) {
 987                         gctl_error(req, "%d geom '%s'", EEXIST, pp->name);
 988                         return (EEXIST);
 989                 }
 990         } else
 991                 null = NULL;
 992
 993         if ((gpp->gpp_parms & G_PART_PARM_ENTRIES) &&
 994             (gpp->gpp_entries < scheme->gps_minent ||
 995              gpp->gpp_entries > scheme->gps_maxent)) {
 996                 gctl_error(req, "%d entries '%d'", EINVAL, gpp->gpp_entries);
 997                 return (EINVAL);
 998         }
 999
1000         if (null == NULL)
1001                 gp = g_new_geomf(&g_part_class, "%s", pp->name);
1002         gp->softc = kobj_create((kobj_class_t)gpp->gpp_scheme, M_GEOM,
1003             M_WAITOK);
1004         table = gp->softc;
1005         table->gpt_gp = gp;
1006         table->gpt_scheme = gpp->gpp_scheme;
1007         table->gpt_entries = (gpp->gpp_parms & G_PART_PARM_ENTRIES) ?
1008             gpp->gpp_entries : scheme->gps_minent;
1009         LIST_INIT(&table->gpt_entry);
1010         if (null == NULL) {
1011                 cp = g_new_consumer(gp);
1012                 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
1013                 error = g_attach(cp, pp);
1014                 if (error == 0)
1015                         error = g_access(cp, 1, 1, 1);
1016                 if (error != 0) {
1017                         g_part_wither(gp, error);
1018                         gctl_error(req, "%d geom '%s'", error, pp->name);
1019                         return (error);
1020                 }
1021                 table->gpt_opened = 1;
1022         } else {
1023                 cp = LIST_FIRST(&gp->consumer);
1024                 table->gpt_opened = null->gpt_opened;
1025                 table->gpt_smhead = null->gpt_smhead;
1026                 table->gpt_smtail = null->gpt_smtail;
1027         }
1028
1029         g_topology_unlock();
1030
1031         /* Make sure the provider has media. */
1032         if (pp->mediasize == 0 || pp->sectorsize == 0) {
1033                 error = ENODEV;
1034                 goto fail;
1035         }
1036
1037         /* Make sure we can nest and if so, determine our depth. */
1038         error = g_getattr("PART::isleaf", cp, &attr);
1039         if (!error && attr) {
1040                 error = ENODEV;
1041                 goto fail;
1042         }
1043         error = g_getattr("PART::depth", cp, &attr);
1044         table->gpt_depth = (!error) ? attr + 1 : 0;
1045
1046         /*
1047          * Synthesize a disk geometry. Some partitioning schemes
1048          * depend on it and since some file systems need it even
1049          * when the partitition scheme doesn't, we do it here in
1050          * scheme-independent code.
1051          */
1052         g_part_geometry(table, cp, pp->mediasize / pp->sectorsize);
1053
1054         error = G_PART_CREATE(table, gpp);
1055         if (error)
1056                 goto fail;
1057
1058         g_topology_lock();
1059
1060         table->gpt_created = 1;
1061         if (null != NULL)
1062                 kobj_delete((kobj_t)null, M_GEOM);
1063
1064         /*
1065          * Support automatic commit by filling in the gpp_geom
1066          * parameter.
1067          */
1068         gpp->gpp_parms |= G_PART_PARM_GEOM;
1069         gpp->gpp_geom = gp;
1070
1071         /* Provide feedback if so requested. */
1072         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1073                 sb = sbuf_new_auto();
1074                 sbuf_printf(sb, "%s created\n", gp->name);
1075                 sbuf_finish(sb);
1076                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1077                 sbuf_delete(sb);
1078         }
1079         return (0);
1080
1081 fail:
1082         g_topology_lock();
1083         if (null == NULL) {
1084                 g_access(cp, -1, -1, -1);
1085                 g_part_wither(gp, error);
1086         } else {
1087                 kobj_delete((kobj_t)gp->softc, M_GEOM);
1088                 gp->softc = null;
1089         }
1090         gctl_error(req, "%d provider", error);
1091         return (error);
1092 }
1093
1094 static int
1095 g_part_ctl_delete(struct gctl_req *req, struct g_part_parms *gpp)
1096 {
1097         struct g_geom *gp;
1098         struct g_provider *pp;
1099         struct g_part_entry *entry;
1100         struct g_part_table *table;
1101         struct sbuf *sb;
1102
1103         gp = gpp->gpp_geom;
1104         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1105         g_topology_assert();
1106
1107         table = gp->softc;
1108
1109         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1110                 if (entry->gpe_deleted || entry->gpe_internal)
1111                         continue;
1112                 if (entry->gpe_index == gpp->gpp_index)
1113                         break;
1114         }
1115         if (entry == NULL) {
1116                 gctl_error(req, "%d index '%d'", ENOENT, gpp->gpp_index);
1117                 return (ENOENT);
1118         }
1119
1120         pp = entry->gpe_pp;
1121         if (pp != NULL) {
1122                 if (pp->acr > 0 || pp->acw > 0 || pp->ace > 0) {
1123                         gctl_error(req, "%d", EBUSY);
1124                         return (EBUSY);
1125                 }
1126
1127                 pp->private = NULL;
1128                 entry->gpe_pp = NULL;
1129         }
1130
1131         if (pp != NULL)
1132                 g_wither_provider(pp, ENXIO);
1133
1134         /* Provide feedback if so requested. */
1135         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1136                 sb = sbuf_new_auto();
1137                 G_PART_FULLNAME(table, entry, sb, gp->name);
1138                 sbuf_cat(sb, " deleted\n");
1139                 sbuf_finish(sb);
1140                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1141                 sbuf_delete(sb);
1142         }
1143
1144         if (entry->gpe_created) {
1145                 LIST_REMOVE(entry, gpe_entry);
1146                 g_free(entry);
1147         } else {
1148                 entry->gpe_modified = 0;
1149                 entry->gpe_deleted = 1;
1150         }
1151         return (0);
1152 }
1153
1154 static int
1155 g_part_ctl_destroy(struct gctl_req *req, struct g_part_parms *gpp)
1156 {
1157         struct g_consumer *cp;
1158         struct g_geom *gp;
1159         struct g_provider *pp;
1160         struct g_part_entry *entry, *tmp;
1161         struct g_part_table *null, *table;
1162         struct sbuf *sb;
1163         int error;
1164
1165         gp = gpp->gpp_geom;
1166         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1167         g_topology_assert();
1168
1169         table = gp->softc;
1170         /* Check for busy providers. */
1171         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1172                 if (entry->gpe_deleted || entry->gpe_internal)
1173                         continue;
1174                 if (gpp->gpp_force) {
1175                         pp = entry->gpe_pp;
1176                         if (pp == NULL)
1177                                 continue;
1178                         if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1179                                 continue;
1180                 }
1181                 gctl_error(req, "%d", EBUSY);
1182                 return (EBUSY);
1183         }
1184
1185         if (gpp->gpp_force) {
1186                 /* Destroy all providers. */
1187                 LIST_FOREACH_SAFE(entry, &table->gpt_entry, gpe_entry, tmp) {
1188                         pp = entry->gpe_pp;
1189                         if (pp != NULL) {
1190                                 pp->private = NULL;
1191                                 g_wither_provider(pp, ENXIO);
1192                         }
1193                         LIST_REMOVE(entry, gpe_entry);
1194                         g_free(entry);
1195                 }
1196         }
1197
1198         error = G_PART_DESTROY(table, gpp);
1199         if (error) {
1200                 gctl_error(req, "%d", error);
1201                 return (error);
1202         }
1203
1204         gp->softc = kobj_create((kobj_class_t)&g_part_null_scheme, M_GEOM,
1205             M_WAITOK);
1206         null = gp->softc;
1207         null->gpt_gp = gp;
1208         null->gpt_scheme = &g_part_null_scheme;
1209         LIST_INIT(&null->gpt_entry);
1210
1211         cp = LIST_FIRST(&gp->consumer);
1212         pp = cp->provider;
1213         null->gpt_last = pp->mediasize / pp->sectorsize - 1;
1214
1215         null->gpt_depth = table->gpt_depth;
1216         null->gpt_opened = table->gpt_opened;
1217         null->gpt_smhead = table->gpt_smhead;
1218         null->gpt_smtail = table->gpt_smtail;
1219
1220         while ((entry = LIST_FIRST(&table->gpt_entry)) != NULL) {
1221                 LIST_REMOVE(entry, gpe_entry);
1222                 g_free(entry);
1223         }
1224         kobj_delete((kobj_t)table, M_GEOM);
1225
1226         /* Provide feedback if so requested. */
1227         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1228                 sb = sbuf_new_auto();
1229                 sbuf_printf(sb, "%s destroyed\n", gp->name);
1230                 sbuf_finish(sb);
1231                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1232                 sbuf_delete(sb);
1233         }
1234         return (0);
1235 }
1236
1237 static int
1238 g_part_ctl_modify(struct gctl_req *req, struct g_part_parms *gpp)
1239 {
1240         struct g_geom *gp;
1241         struct g_part_entry *entry;
1242         struct g_part_table *table;
1243         struct sbuf *sb;
1244         int error;
1245
1246         gp = gpp->gpp_geom;
1247         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1248         g_topology_assert();
1249
1250         table = gp->softc;
1251
1252         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1253                 if (entry->gpe_deleted || entry->gpe_internal)
1254                         continue;
1255                 if (entry->gpe_index == gpp->gpp_index)
1256                         break;
1257         }
1258         if (entry == NULL) {
1259                 gctl_error(req, "%d index '%d'", ENOENT, gpp->gpp_index);
1260                 return (ENOENT);
1261         }
1262
1263         error = G_PART_MODIFY(table, entry, gpp);
1264         if (error) {
1265                 gctl_error(req, "%d", error);
1266                 return (error);
1267         }
1268
1269         if (!entry->gpe_created)
1270                 entry->gpe_modified = 1;
1271
1272         /* Provide feedback if so requested. */
1273         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1274                 sb = sbuf_new_auto();
1275                 G_PART_FULLNAME(table, entry, sb, gp->name);
1276                 sbuf_cat(sb, " modified\n");
1277                 sbuf_finish(sb);
1278                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1279                 sbuf_delete(sb);
1280         }
1281         return (0);
1282 }
1283
1284 static int
1285 g_part_ctl_move(struct gctl_req *req, struct g_part_parms *gpp)
1286 {
1287         gctl_error(req, "%d verb 'move'", ENOSYS);
1288         return (ENOSYS);
1289 }
1290
1291 static int
1292 g_part_ctl_recover(struct gctl_req *req, struct g_part_parms *gpp)
1293 {
1294         struct g_part_table *table;
1295         struct g_geom *gp;
1296         struct sbuf *sb;
1297         int error, recovered;
1298
1299         gp = gpp->gpp_geom;
1300         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1301         g_topology_assert();
1302         table = gp->softc;
1303         error = recovered = 0;
1304
1305         if (table->gpt_corrupt) {
1306                 error = G_PART_RECOVER(table);
1307                 if (error == 0)
1308                         error = g_part_check_integrity(table,
1309                             LIST_FIRST(&gp->consumer));
1310                 if (error) {
1311                         gctl_error(req, "%d recovering '%s' failed",
1312                             error, gp->name);
1313                         return (error);
1314                 }
1315                 recovered = 1;
1316         }
1317         /* Provide feedback if so requested. */
1318         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1319                 sb = sbuf_new_auto();
1320                 if (recovered)
1321                         sbuf_printf(sb, "%s recovered\n", gp->name);
1322                 else
1323                         sbuf_printf(sb, "%s recovering is not needed\n",
1324                             gp->name);
1325                 sbuf_finish(sb);
1326                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1327                 sbuf_delete(sb);
1328         }
1329         return (0);
1330 }
1331
1332 static int
1333 g_part_ctl_resize(struct gctl_req *req, struct g_part_parms *gpp)
1334 {
1335         struct g_geom *gp;
1336         struct g_provider *pp;
1337         struct g_part_entry *pe, *entry;
1338         struct g_part_table *table;
1339         struct sbuf *sb;
1340         quad_t end;
1341         int error;
1342         off_t mediasize;
1343
1344         gp = gpp->gpp_geom;
1345         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1346         g_topology_assert();
1347         table = gp->softc;
1348
1349         /* check gpp_index */
1350         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1351                 if (entry->gpe_deleted || entry->gpe_internal)
1352                         continue;
1353                 if (entry->gpe_index == gpp->gpp_index)
1354                         break;
1355         }
1356         if (entry == NULL) {
1357                 gctl_error(req, "%d index '%d'", ENOENT, gpp->gpp_index);
1358                 return (ENOENT);
1359         }
1360
1361         /* check gpp_size */
1362         end = entry->gpe_start + gpp->gpp_size - 1;
1363         if (gpp->gpp_size < 1 || end > table->gpt_last) {
1364                 gctl_error(req, "%d size '%jd'", EINVAL,
1365                     (intmax_t)gpp->gpp_size);
1366                 return (EINVAL);
1367         }
1368
1369         LIST_FOREACH(pe, &table->gpt_entry, gpe_entry) {
1370                 if (pe->gpe_deleted || pe->gpe_internal || pe == entry)
1371                         continue;
1372                 if (end >= pe->gpe_start && end <= pe->gpe_end) {
1373                         gctl_error(req, "%d end '%jd'", ENOSPC,
1374                             (intmax_t)end);
1375                         return (ENOSPC);
1376                 }
1377                 if (entry->gpe_start < pe->gpe_start && end > pe->gpe_end) {
1378                         gctl_error(req, "%d size '%jd'", ENOSPC,
1379                             (intmax_t)gpp->gpp_size);
1380                         return (ENOSPC);
1381                 }
1382         }
1383
1384         pp = entry->gpe_pp;
1385         if ((g_debugflags & G_F_FOOTSHOOTING) == 0 &&
1386             (pp->acr > 0 || pp->acw > 0 || pp->ace > 0)) {
1387                 if (entry->gpe_end - entry->gpe_start + 1 > gpp->gpp_size) {
1388                         /* Deny shrinking of an opened partition. */
1389                         gctl_error(req, "%d", EBUSY);
1390                         return (EBUSY);
1391                 }
1392         }
1393
1394         error = G_PART_RESIZE(table, entry, gpp);
1395         if (error) {
1396                 gctl_error(req, "%d%s", error, error != EBUSY ? "":
1397                     " resizing will lead to unexpected shrinking"
1398                     " due to alignment");
1399                 return (error);
1400         }
1401
1402         if (!entry->gpe_created)
1403                 entry->gpe_modified = 1;
1404
1405         /* update mediasize of changed provider */
1406         mediasize = (entry->gpe_end - entry->gpe_start + 1) *
1407                 pp->sectorsize;
1408         g_resize_provider(pp, mediasize);
1409
1410         /* Provide feedback if so requested. */
1411         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1412                 sb = sbuf_new_auto();
1413                 G_PART_FULLNAME(table, entry, sb, gp->name);
1414                 sbuf_cat(sb, " resized\n");
1415                 sbuf_finish(sb);
1416                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1417                 sbuf_delete(sb);
1418         }
1419         return (0);
1420 }
1421
1422 static int
1423 g_part_ctl_setunset(struct gctl_req *req, struct g_part_parms *gpp,
1424     unsigned int set)
1425 {
1426         struct g_geom *gp;
1427         struct g_part_entry *entry;
1428         struct g_part_table *table;
1429         struct sbuf *sb;
1430         int error;
1431
1432         gp = gpp->gpp_geom;
1433         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1434         g_topology_assert();
1435
1436         table = gp->softc;
1437
1438         if (gpp->gpp_parms & G_PART_PARM_INDEX) {
1439                 LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1440                         if (entry->gpe_deleted || entry->gpe_internal)
1441                                 continue;
1442                         if (entry->gpe_index == gpp->gpp_index)
1443                                 break;
1444                 }
1445                 if (entry == NULL) {
1446                         gctl_error(req, "%d index '%d'", ENOENT,
1447                             gpp->gpp_index);
1448                         return (ENOENT);
1449                 }
1450         } else
1451                 entry = NULL;
1452
1453         error = G_PART_SETUNSET(table, entry, gpp->gpp_attrib, set);
1454         if (error) {
1455                 gctl_error(req, "%d attrib '%s'", error, gpp->gpp_attrib);
1456                 return (error);
1457         }
1458
1459         /* Provide feedback if so requested. */
1460         if (gpp->gpp_parms & G_PART_PARM_OUTPUT) {
1461                 sb = sbuf_new_auto();
1462                 sbuf_printf(sb, "%s %sset on ", gpp->gpp_attrib,
1463                     (set) ? "" : "un");
1464                 if (entry)
1465                         G_PART_FULLNAME(table, entry, sb, gp->name);
1466                 else
1467                         sbuf_cat(sb, gp->name);
1468                 sbuf_cat(sb, "\n");
1469                 sbuf_finish(sb);
1470                 gctl_set_param(req, "output", sbuf_data(sb), sbuf_len(sb) + 1);
1471                 sbuf_delete(sb);
1472         }
1473         return (0);
1474 }
1475
1476 static int
1477 g_part_ctl_undo(struct gctl_req *req, struct g_part_parms *gpp)
1478 {
1479         struct g_consumer *cp;
1480         struct g_provider *pp;
1481         struct g_geom *gp;
1482         struct g_part_entry *entry, *tmp;
1483         struct g_part_table *table;
1484         int error, reprobe;
1485
1486         gp = gpp->gpp_geom;
1487         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, gp->name));
1488         g_topology_assert();
1489
1490         table = gp->softc;
1491         if (!table->gpt_opened) {
1492                 gctl_error(req, "%d", EPERM);
1493                 return (EPERM);
1494         }
1495
1496         cp = LIST_FIRST(&gp->consumer);
1497         LIST_FOREACH_SAFE(entry, &table->gpt_entry, gpe_entry, tmp) {
1498                 entry->gpe_modified = 0;
1499                 if (entry->gpe_created) {
1500                         pp = entry->gpe_pp;
1501                         if (pp != NULL) {
1502                                 pp->private = NULL;
1503                                 entry->gpe_pp = NULL;
1504                                 g_wither_provider(pp, ENXIO);
1505                         }
1506                         entry->gpe_deleted = 1;
1507                 }
1508                 if (entry->gpe_deleted) {
1509                         LIST_REMOVE(entry, gpe_entry);
1510                         g_free(entry);
1511                 }
1512         }
1513
1514         g_topology_unlock();
1515
1516         reprobe = (table->gpt_scheme == &g_part_null_scheme ||
1517             table->gpt_created) ? 1 : 0;
1518
1519         if (reprobe) {
1520                 LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1521                         if (entry->gpe_internal)
1522                                 continue;
1523                         error = EBUSY;
1524                         goto fail;
1525                 }
1526                 while ((entry = LIST_FIRST(&table->gpt_entry)) != NULL) {
1527                         LIST_REMOVE(entry, gpe_entry);
1528                         g_free(entry);
1529                 }
1530                 error = g_part_probe(gp, cp, table->gpt_depth);
1531                 if (error) {
1532                         g_topology_lock();
1533                         g_access(cp, -1, -1, -1);
1534                         g_part_wither(gp, error);
1535                         return (0);
1536                 }
1537                 table = gp->softc;
1538
1539                 /*
1540                  * Synthesize a disk geometry. Some partitioning schemes
1541                  * depend on it and since some file systems need it even
1542                  * when the partitition scheme doesn't, we do it here in
1543                  * scheme-independent code.
1544                  */
1545                 pp = cp->provider;
1546                 g_part_geometry(table, cp, pp->mediasize / pp->sectorsize);
1547         }
1548
1549         error = G_PART_READ(table, cp);
1550         if (error)
1551                 goto fail;
1552         error = g_part_check_integrity(table, cp);
1553         if (error)
1554                 goto fail;
1555
1556         g_topology_lock();
1557         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
1558                 if (!entry->gpe_internal)
1559                         g_part_new_provider(gp, table, entry);
1560         }
1561
1562         table->gpt_opened = 0;
1563         g_access(cp, -1, -1, -1);
1564         return (0);
1565
1566 fail:
1567         g_topology_lock();
1568         gctl_error(req, "%d", error);
1569         return (error);
1570 }
1571
1572 static void
1573 g_part_wither(struct g_geom *gp, int error)
1574 {
1575         struct g_part_entry *entry;
1576         struct g_part_table *table;
1577         struct g_provider *pp;
1578
1579         table = gp->softc;
1580         if (table != NULL) {
1581                 gp->softc = NULL;
1582                 while ((entry = LIST_FIRST(&table->gpt_entry)) != NULL) {
1583                         LIST_REMOVE(entry, gpe_entry);
1584                         pp = entry->gpe_pp;
1585                         entry->gpe_pp = NULL;
1586                         if (pp != NULL) {
1587                                 pp->private = NULL;
1588                                 g_wither_provider(pp, error);
1589                         }
1590                         g_free(entry);
1591                 }
1592                 G_PART_DESTROY(table, NULL);
1593                 kobj_delete((kobj_t)table, M_GEOM);
1594         }
1595         g_wither_geom(gp, error);
1596 }
1597
1598 /*
1599  * Class methods.
1600  */
1601
1602 static void
1603 g_part_ctlreq(struct gctl_req *req, struct g_class *mp, const char *verb)
1604 {
1605         struct g_part_parms gpp;
1606         struct g_part_table *table;
1607         struct gctl_req_arg *ap;
1608         enum g_part_ctl ctlreq;
1609         unsigned int i, mparms, oparms, parm;
1610         int auto_commit, close_on_error;
1611         int error, modifies;
1612
1613         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s,%s)", __func__, mp->name, verb));
1614         g_topology_assert();
1615
1616         ctlreq = G_PART_CTL_NONE;
1617         modifies = 1;
1618         mparms = 0;
1619         oparms = G_PART_PARM_FLAGS | G_PART_PARM_OUTPUT | G_PART_PARM_VERSION;
1620         switch (*verb) {
1621         case 'a':
1622                 if (!strcmp(verb, "add")) {
1623                         ctlreq = G_PART_CTL_ADD;
1624                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_SIZE |
1625                             G_PART_PARM_START | G_PART_PARM_TYPE;
1626                         oparms |= G_PART_PARM_INDEX | G_PART_PARM_LABEL;
1627                 }
1628                 break;
1629         case 'b':
1630                 if (!strcmp(verb, "bootcode")) {
1631                         ctlreq = G_PART_CTL_BOOTCODE;
1632                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_BOOTCODE;
1633                         oparms |= G_PART_PARM_SKIP_DSN;
1634                 }
1635                 break;
1636         case 'c':
1637                 if (!strcmp(verb, "commit")) {
1638                         ctlreq = G_PART_CTL_COMMIT;
1639                         mparms |= G_PART_PARM_GEOM;
1640                         modifies = 0;
1641                 } else if (!strcmp(verb, "create")) {
1642                         ctlreq = G_PART_CTL_CREATE;
1643                         mparms |= G_PART_PARM_PROVIDER | G_PART_PARM_SCHEME;
1644                         oparms |= G_PART_PARM_ENTRIES;
1645                 }
1646                 break;
1647         case 'd':
1648                 if (!strcmp(verb, "delete")) {
1649                         ctlreq = G_PART_CTL_DELETE;
1650                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_INDEX;
1651                 } else if (!strcmp(verb, "destroy")) {
1652                         ctlreq = G_PART_CTL_DESTROY;
1653                         mparms |= G_PART_PARM_GEOM;
1654                         oparms |= G_PART_PARM_FORCE;
1655                 }
1656                 break;
1657         case 'm':
1658                 if (!strcmp(verb, "modify")) {
1659                         ctlreq = G_PART_CTL_MODIFY;
1660                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_INDEX;
1661                         oparms |= G_PART_PARM_LABEL | G_PART_PARM_TYPE;
1662                 } else if (!strcmp(verb, "move")) {
1663                         ctlreq = G_PART_CTL_MOVE;
1664                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_INDEX;
1665                 }
1666                 break;
1667         case 'r':
1668                 if (!strcmp(verb, "recover")) {
1669                         ctlreq = G_PART_CTL_RECOVER;
1670                         mparms |= G_PART_PARM_GEOM;
1671                 } else if (!strcmp(verb, "resize")) {
1672                         ctlreq = G_PART_CTL_RESIZE;
1673                         mparms |= G_PART_PARM_GEOM | G_PART_PARM_INDEX |
1674                             G_PART_PARM_SIZE;
1675                 }
1676                 break;
1677         case 's':
1678                 if (!strcmp(verb, "set")) {
1679                         ctlreq = G_PART_CTL_SET;
1680                         mparms |= G_PART_PARM_ATTRIB | G_PART_PARM_GEOM;
1681                         oparms |= G_PART_PARM_INDEX;
1682                 }
1683                 break;
1684         case 'u':
1685                 if (!strcmp(verb, "undo")) {
1686                         ctlreq = G_PART_CTL_UNDO;
1687                         mparms |= G_PART_PARM_GEOM;
1688                         modifies = 0;
1689                 } else if (!strcmp(verb, "unset")) {
1690                         ctlreq = G_PART_CTL_UNSET;
1691                         mparms |= G_PART_PARM_ATTRIB | G_PART_PARM_GEOM;
1692                         oparms |= G_PART_PARM_INDEX;
1693                 }
1694                 break;
1695         }
1696         if (ctlreq == G_PART_CTL_NONE) {
1697                 gctl_error(req, "%d verb '%s'", EINVAL, verb);
1698                 return;
1699         }
1700
1701         bzero(&gpp, sizeof(gpp));
1702         for (i = 0; i < req->narg; i++) {
1703                 ap = &req->arg[i];
1704                 parm = 0;
1705                 switch (ap->name[0]) {
1706                 case 'a':
1707                         if (!strcmp(ap->name, "arg0")) {
1708                                 parm = mparms &
1709                                     (G_PART_PARM_GEOM | G_PART_PARM_PROVIDER);
1710                         }
1711                         if (!strcmp(ap->name, "attrib"))
1712                                 parm = G_PART_PARM_ATTRIB;
1713                         break;
1714                 case 'b':
1715                         if (!strcmp(ap->name, "bootcode"))
1716                                 parm = G_PART_PARM_BOOTCODE;
1717                         break;
1718                 case 'c':
1719                         if (!strcmp(ap->name, "class"))
1720                                 continue;
1721                         break;
1722                 case 'e':
1723                         if (!strcmp(ap->name, "entries"))
1724                                 parm = G_PART_PARM_ENTRIES;
1725                         break;
1726                 case 'f':
1727                         if (!strcmp(ap->name, "flags"))
1728                                 parm = G_PART_PARM_FLAGS;
1729                         else if (!strcmp(ap->name, "force"))
1730                                 parm = G_PART_PARM_FORCE;
1731                         break;
1732                 case 'i':
1733                         if (!strcmp(ap->name, "index"))
1734                                 parm = G_PART_PARM_INDEX;
1735                         break;
1736                 case 'l':
1737                         if (!strcmp(ap->name, "label"))
1738                                 parm = G_PART_PARM_LABEL;
1739                         break;
1740                 case 'o':
1741                         if (!strcmp(ap->name, "output"))
1742                                 parm = G_PART_PARM_OUTPUT;
1743                         break;
1744                 case 's':
1745                         if (!strcmp(ap->name, "scheme"))
1746                                 parm = G_PART_PARM_SCHEME;
1747                         else if (!strcmp(ap->name, "size"))
1748                                 parm = G_PART_PARM_SIZE;
1749                         else if (!strcmp(ap->name, "start"))
1750                                 parm = G_PART_PARM_START;
1751                         else if (!strcmp(ap->name, "skip_dsn"))
1752                                 parm = G_PART_PARM_SKIP_DSN;
1753                         break;
1754                 case 't':
1755                         if (!strcmp(ap->name, "type"))
1756                                 parm = G_PART_PARM_TYPE;
1757                         break;
1758                 case 'v':
1759                         if (!strcmp(ap->name, "verb"))
1760                                 continue;
1761                         else if (!strcmp(ap->name, "version"))
1762                                 parm = G_PART_PARM_VERSION;
1763                         break;
1764                 }
1765                 if ((parm & (mparms | oparms)) == 0) {
1766                         gctl_error(req, "%d param '%s'", EINVAL, ap->name);
1767                         return;
1768                 }
1769                 switch (parm) {
1770                 case G_PART_PARM_ATTRIB:
1771                         error = g_part_parm_str(req, ap->name,
1772                             &gpp.gpp_attrib);
1773                         break;
1774                 case G_PART_PARM_BOOTCODE:
1775                         error = g_part_parm_bootcode(req, ap->name,
1776                             &gpp.gpp_codeptr, &gpp.gpp_codesize);
1777                         break;
1778                 case G_PART_PARM_ENTRIES:
1779                         error = g_part_parm_intmax(req, ap->name,
1780                             &gpp.gpp_entries);
1781                         break;
1782                 case G_PART_PARM_FLAGS:
1783                         error = g_part_parm_str(req, ap->name, &gpp.gpp_flags);
1784                         break;
1785                 case G_PART_PARM_FORCE:
1786                         error = g_part_parm_uint32(req, ap->name,
1787                             &gpp.gpp_force);
1788                         break;
1789                 case G_PART_PARM_GEOM:
1790                         error = g_part_parm_geom(req, ap->name, &gpp.gpp_geom);
1791                         break;
1792                 case G_PART_PARM_INDEX:
1793                         error = g_part_parm_intmax(req, ap->name,
1794                             &gpp.gpp_index);
1795                         break;
1796                 case G_PART_PARM_LABEL:
1797                         error = g_part_parm_str(req, ap->name, &gpp.gpp_label);
1798                         break;
1799                 case G_PART_PARM_OUTPUT:
1800                         error = 0;      /* Write-only parameter */
1801                         break;
1802                 case G_PART_PARM_PROVIDER:
1803                         error = g_part_parm_provider(req, ap->name,
1804                             &gpp.gpp_provider);
1805                         break;
1806                 case G_PART_PARM_SCHEME:
1807                         error = g_part_parm_scheme(req, ap->name,
1808                             &gpp.gpp_scheme);
1809                         break;
1810                 case G_PART_PARM_SIZE:
1811                         error = g_part_parm_quad(req, ap->name, &gpp.gpp_size);
1812                         break;
1813                 case G_PART_PARM_SKIP_DSN:
1814                         error = g_part_parm_uint32(req, ap->name,
1815                             &gpp.gpp_skip_dsn);
1816                         break;
1817                 case G_PART_PARM_START:
1818                         error = g_part_parm_quad(req, ap->name,
1819                             &gpp.gpp_start);
1820                         break;
1821                 case G_PART_PARM_TYPE:
1822                         error = g_part_parm_str(req, ap->name, &gpp.gpp_type);
1823                         break;
1824                 case G_PART_PARM_VERSION:
1825                         error = g_part_parm_uint32(req, ap->name,
1826                             &gpp.gpp_version);
1827                         break;
1828                 default:
1829                         error = EDOOFUS;
1830                         gctl_error(req, "%d %s", error, ap->name);
1831                         break;
1832                 }
1833                 if (error != 0) {
1834                         if (error == ENOATTR) {
1835                                 gctl_error(req, "%d param '%s'", error,
1836                                     ap->name);
1837                         }
1838                         return;
1839                 }
1840                 gpp.gpp_parms |= parm;
1841         }
1842         if ((gpp.gpp_parms & mparms) != mparms) {
1843                 parm = mparms - (gpp.gpp_parms & mparms);
1844                 gctl_error(req, "%d param '%x'", ENOATTR, parm);
1845                 return;
1846         }
1847
1848         /* Obtain permissions if possible/necessary. */
1849         close_on_error = 0;
1850         table = NULL;
1851         if (modifies && (gpp.gpp_parms & G_PART_PARM_GEOM)) {
1852                 table = gpp.gpp_geom->softc;
1853                 if (table != NULL && table->gpt_corrupt &&
1854                     ctlreq != G_PART_CTL_DESTROY &&
1855                     ctlreq != G_PART_CTL_RECOVER &&
1856                     geom_part_check_integrity) {
1857                         gctl_error(req, "%d table '%s' is corrupt",
1858                             EPERM, gpp.gpp_geom->name);
1859                         return;
1860                 }
1861                 if (table != NULL && !table->gpt_opened) {
1862                         error = g_access(LIST_FIRST(&gpp.gpp_geom->consumer),
1863                             1, 1, 1);
1864                         if (error) {
1865                                 gctl_error(req, "%d geom '%s'", error,
1866                                     gpp.gpp_geom->name);
1867                                 return;
1868                         }
1869                         table->gpt_opened = 1;
1870                         close_on_error = 1;
1871                 }
1872         }
1873
1874         /* Allow the scheme to check or modify the parameters. */
1875         if (table != NULL) {
1876                 error = G_PART_PRECHECK(table, ctlreq, &gpp);
1877                 if (error) {
1878                         gctl_error(req, "%d pre-check failed", error);
1879                         goto out;
1880                 }
1881         } else
1882                 error = EDOOFUS;        /* Prevent bogus uninit. warning. */
1883
1884         switch (ctlreq) {
1885         case G_PART_CTL_NONE:
1886                 panic("%s", __func__);
1887         case G_PART_CTL_ADD:
1888                 error = g_part_ctl_add(req, &gpp);
1889                 break;
1890         case G_PART_CTL_BOOTCODE:
1891                 error = g_part_ctl_bootcode(req, &gpp);
1892                 break;
1893         case G_PART_CTL_COMMIT:
1894                 error = g_part_ctl_commit(req, &gpp);
1895                 break;
1896         case G_PART_CTL_CREATE:
1897                 error = g_part_ctl_create(req, &gpp);
1898                 break;
1899         case G_PART_CTL_DELETE:
1900                 error = g_part_ctl_delete(req, &gpp);
1901                 break;
1902         case G_PART_CTL_DESTROY:
1903                 error = g_part_ctl_destroy(req, &gpp);
1904                 break;
1905         case G_PART_CTL_MODIFY:
1906                 error = g_part_ctl_modify(req, &gpp);
1907                 break;
1908         case G_PART_CTL_MOVE:
1909                 error = g_part_ctl_move(req, &gpp);
1910                 break;
1911         case G_PART_CTL_RECOVER:
1912                 error = g_part_ctl_recover(req, &gpp);
1913                 break;
1914         case G_PART_CTL_RESIZE:
1915                 error = g_part_ctl_resize(req, &gpp);
1916                 break;
1917         case G_PART_CTL_SET:
1918                 error = g_part_ctl_setunset(req, &gpp, 1);
1919                 break;
1920         case G_PART_CTL_UNDO:
1921                 error = g_part_ctl_undo(req, &gpp);
1922                 break;
1923         case G_PART_CTL_UNSET:
1924                 error = g_part_ctl_setunset(req, &gpp, 0);
1925                 break;
1926         }
1927
1928         /* Implement automatic commit. */
1929         if (!error) {
1930                 auto_commit = (modifies &&
1931                     (gpp.gpp_parms & G_PART_PARM_FLAGS) &&
1932                     strchr(gpp.gpp_flags, 'C') != NULL) ? 1 : 0;
1933                 if (auto_commit) {
1934                         KASSERT(gpp.gpp_parms & G_PART_PARM_GEOM, ("%s",
1935                             __func__));
1936                         error = g_part_ctl_commit(req, &gpp);
1937                 }
1938         }
1939
1940  out:
1941         if (error && close_on_error) {
1942                 g_access(LIST_FIRST(&gpp.gpp_geom->consumer), -1, -1, -1);
1943                 table->gpt_opened = 0;
1944         }
1945 }
1946
1947 static int
1948 g_part_destroy_geom(struct gctl_req *req, struct g_class *mp,
1949     struct g_geom *gp)
1950 {
1951
1952         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s,%s)", __func__, mp->name, gp->name));
1953         g_topology_assert();
1954
1955         g_part_wither(gp, EINVAL);
1956         return (0);
1957 }
1958
1959 static struct g_geom *
1960 g_part_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
1961 {
1962         struct g_consumer *cp;
1963         struct g_geom *gp;
1964         struct g_part_entry *entry;
1965         struct g_part_table *table;
1966         struct root_hold_token *rht;
1967         int attr, depth;
1968         int error;
1969
1970         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s,%s)", __func__, mp->name, pp->name));
1971         g_topology_assert();
1972
1973         /* Skip providers that are already open for writing. */
1974         if (pp->acw > 0)
1975                 return (NULL);
1976
1977         /*
1978          * Create a GEOM with consumer and hook it up to the provider.
1979          * With that we become part of the topology. Obtain read access
1980          * to the provider.
1981          */
1982         gp = g_new_geomf(mp, "%s", pp->name);
1983         cp = g_new_consumer(gp);
1984         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
1985         error = g_attach(cp, pp);
1986         if (error == 0)
1987                 error = g_access(cp, 1, 0, 0);
1988         if (error != 0) {
1989                 if (cp->provider)
1990                         g_detach(cp);
1991                 g_destroy_consumer(cp);
1992                 g_destroy_geom(gp);
1993                 return (NULL);
1994         }
1995
1996         rht = root_mount_hold(mp->name);
1997         g_topology_unlock();
1998
1999         /*
2000          * Short-circuit the whole probing galore when there's no
2001          * media present.
2002          */
2003         if (pp->mediasize == 0 || pp->sectorsize == 0) {
2004                 error = ENODEV;
2005                 goto fail;
2006         }
2007
2008         /* Make sure we can nest and if so, determine our depth. */
2009         error = g_getattr("PART::isleaf", cp, &attr);
2010         if (!error && attr) {
2011                 error = ENODEV;
2012                 goto fail;
2013         }
2014         error = g_getattr("PART::depth", cp, &attr);
2015         depth = (!error) ? attr + 1 : 0;
2016
2017         error = g_part_probe(gp, cp, depth);
2018         if (error)
2019                 goto fail;
2020
2021         table = gp->softc;
2022
2023         /*
2024          * Synthesize a disk geometry. Some partitioning schemes
2025          * depend on it and since some file systems need it even
2026          * when the partitition scheme doesn't, we do it here in
2027          * scheme-independent code.
2028          */
2029         g_part_geometry(table, cp, pp->mediasize / pp->sectorsize);
2030
2031         error = G_PART_READ(table, cp);
2032         if (error)
2033                 goto fail;
2034         error = g_part_check_integrity(table, cp);
2035         if (error)
2036                 goto fail;
2037
2038         g_topology_lock();
2039         LIST_FOREACH(entry, &table->gpt_entry, gpe_entry) {
2040                 if (!entry->gpe_internal)
2041                         g_part_new_provider(gp, table, entry);
2042         }
2043
2044         root_mount_rel(rht);
2045         g_access(cp, -1, 0, 0);
2046         return (gp);
2047
2048  fail:
2049         g_topology_lock();
2050         root_mount_rel(rht);
2051         g_access(cp, -1, 0, 0);
2052         g_detach(cp);
2053         g_destroy_consumer(cp);
2054         g_destroy_geom(gp);
2055         return (NULL);
2056 }
2057
2058 /*
2059  * Geom methods.
2060  */
2061
2062 static int
2063 g_part_access(struct g_provider *pp, int dr, int dw, int de)
2064 {
2065         struct g_consumer *cp;
2066
2067         G_PART_TRACE((G_T_ACCESS, "%s(%s,%d,%d,%d)", __func__, pp->name, dr,
2068             dw, de));
2069
2070         cp = LIST_FIRST(&pp->geom->consumer);
2071
2072         /* We always gain write-exclusive access. */
2073         return (g_access(cp, dr, dw, dw + de));
2074 }
2075
2076 static void
2077 g_part_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
2078     struct g_consumer *cp, struct g_provider *pp)
2079 {
2080         char buf[64];
2081         struct g_part_entry *entry;
2082         struct g_part_table *table;
2083
2084         KASSERT(sb != NULL && gp != NULL, ("%s", __func__));
2085         table = gp->softc;
2086
2087         if (indent == NULL) {
2088                 KASSERT(cp == NULL && pp != NULL, ("%s", __func__));
2089                 entry = pp->private;
2090                 if (entry == NULL)
2091                         return;
2092                 sbuf_printf(sb, " i %u o %ju ty %s", entry->gpe_index,
2093                     (uintmax_t)entry->gpe_offset,
2094                     G_PART_TYPE(table, entry, buf, sizeof(buf)));
2095                 /*
2096                  * libdisk compatibility quirk - the scheme dumps the
2097                  * slicer name and partition type in a way that is
2098                  * compatible with libdisk. When libdisk is not used
2099                  * anymore, this should go away.
2100                  */
2101                 G_PART_DUMPCONF(table, entry, sb, indent);
2102         } else if (cp != NULL) {        /* Consumer configuration. */
2103                 KASSERT(pp == NULL, ("%s", __func__));
2104                 /* none */
2105         } else if (pp != NULL) {        /* Provider configuration. */
2106                 entry = pp->private;
2107                 if (entry == NULL)
2108                         return;
2109                 sbuf_printf(sb, "%s<start>%ju</start>\n", indent,
2110                     (uintmax_t)entry->gpe_start);
2111                 sbuf_printf(sb, "%s<end>%ju</end>\n", indent,
2112                     (uintmax_t)entry->gpe_end);
2113                 sbuf_printf(sb, "%s<index>%u</index>\n", indent,
2114                     entry->gpe_index);
2115                 sbuf_printf(sb, "%s<type>%s</type>\n", indent,
2116                     G_PART_TYPE(table, entry, buf, sizeof(buf)));
2117                 sbuf_printf(sb, "%s<offset>%ju</offset>\n", indent,
2118                     (uintmax_t)entry->gpe_offset);
2119                 sbuf_printf(sb, "%s<length>%ju</length>\n", indent,
2120                     (uintmax_t)pp->mediasize);
2121                 G_PART_DUMPCONF(table, entry, sb, indent);
2122         } else {                        /* Geom configuration. */
2123                 sbuf_printf(sb, "%s<scheme>%s</scheme>\n", indent,
2124                     table->gpt_scheme->name);
2125                 sbuf_printf(sb, "%s<entries>%u</entries>\n", indent,
2126                     table->gpt_entries);
2127                 sbuf_printf(sb, "%s<first>%ju</first>\n", indent,
2128                     (uintmax_t)table->gpt_first);
2129                 sbuf_printf(sb, "%s<last>%ju</last>\n", indent,
2130                     (uintmax_t)table->gpt_last);
2131                 sbuf_printf(sb, "%s<fwsectors>%u</fwsectors>\n", indent,
2132                     table->gpt_sectors);
2133                 sbuf_printf(sb, "%s<fwheads>%u</fwheads>\n", indent,
2134                     table->gpt_heads);
2135                 sbuf_printf(sb, "%s<state>%s</state>\n", indent,
2136                     table->gpt_corrupt ? "CORRUPT": "OK");
2137                 sbuf_printf(sb, "%s<modified>%s</modified>\n", indent,
2138                     table->gpt_opened ? "true": "false");
2139                 G_PART_DUMPCONF(table, NULL, sb, indent);
2140         }
2141 }
2142
2143 /*-
2144  * This start routine is only called for non-trivial requests, all the
2145  * trivial ones are handled autonomously by the slice code.
2146  * For requests we handle here, we must call the g_io_deliver() on the
2147  * bio, and return non-zero to indicate to the slice code that we did so.
2148  * This code executes in the "DOWN" I/O path, this means:
2149  *    * No sleeping.
2150  *    * Don't grab the topology lock.
2151  *    * Don't call biowait, g_getattr(), g_setattr() or g_read_data()
2152  */
2153 static int
2154 g_part_ioctl(struct g_provider *pp, u_long cmd, void *data, int fflag, struct thread *td)
2155 {
2156         struct g_part_table *table;
2157
2158         table = pp->geom->softc;
2159         return G_PART_IOCTL(table, pp, cmd, data, fflag, td);
2160 }
2161
2162 static void
2163 g_part_resize(struct g_consumer *cp)
2164 {
2165         struct g_part_table *table;
2166
2167         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, cp->provider->name));
2168         g_topology_assert();
2169
2170         if (auto_resize == 0)
2171                 return;
2172
2173         table = cp->geom->softc;
2174         if (table->gpt_opened == 0) {
2175                 if (g_access(cp, 1, 1, 1) != 0)
2176                         return;
2177                 table->gpt_opened = 1;
2178         }
2179         if (G_PART_RESIZE(table, NULL, NULL) == 0)
2180                 printf("GEOM_PART: %s was automatically resized.\n"
2181                     "  Use `gpart commit %s` to save changes or "
2182                     "`gpart undo %s` to revert them.\n", cp->geom->name,
2183                     cp->geom->name, cp->geom->name);
2184         if (g_part_check_integrity(table, cp) != 0) {
2185                 g_access(cp, -1, -1, -1);
2186                 table->gpt_opened = 0;
2187                 g_part_wither(table->gpt_gp, ENXIO);
2188         }
2189 }
2190
2191 static void
2192 g_part_orphan(struct g_consumer *cp)
2193 {
2194         struct g_provider *pp;
2195         struct g_part_table *table;
2196
2197         pp = cp->provider;
2198         KASSERT(pp != NULL, ("%s", __func__));
2199         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, pp->name));
2200         g_topology_assert();
2201
2202         KASSERT(pp->error != 0, ("%s", __func__));
2203         table = cp->geom->softc;
2204         if (table != NULL && table->gpt_opened)
2205                 g_access(cp, -1, -1, -1);
2206         g_part_wither(cp->geom, pp->error);
2207 }
2208
2209 static void
2210 g_part_spoiled(struct g_consumer *cp)
2211 {
2212
2213         G_PART_TRACE((G_T_TOPOLOGY, "%s(%s)", __func__, cp->provider->name));
2214         g_topology_assert();
2215
2216         cp->flags |= G_CF_ORPHAN;
2217         g_part_wither(cp->geom, ENXIO);
2218 }
2219
2220 static void
2221 g_part_start(struct bio *bp)
2222 {
2223         struct bio *bp2;
2224         struct g_consumer *cp;
2225         struct g_geom *gp;
2226         struct g_part_entry *entry;
2227         struct g_part_table *table;
2228         struct g_kerneldump *gkd;
2229         struct g_provider *pp;
2230         void (*done_func)(struct bio *) = g_std_done;
2231         char buf[64];
2232
2233         biotrack(bp, __func__);
2234
2235         pp = bp->bio_to;
2236         gp = pp->geom;
2237         table = gp->softc;
2238         cp = LIST_FIRST(&gp->consumer);
2239
2240         G_PART_TRACE((G_T_BIO, "%s: cmd=%d, provider=%s", __func__, bp->bio_cmd,
2241             pp->name));
2242
2243         entry = pp->private;
2244         if (entry == NULL) {
2245                 g_io_deliver(bp, ENXIO);
2246                 return;
2247         }
2248
2249         switch(bp->bio_cmd) {
2250         case BIO_DELETE:
2251         case BIO_READ:
2252         case BIO_WRITE:
2253                 if (bp->bio_offset >= pp->mediasize) {
2254                         g_io_deliver(bp, EIO);
2255                         return;
2256                 }
2257                 bp2 = g_clone_bio(bp);
2258                 if (bp2 == NULL) {
2259                         g_io_deliver(bp, ENOMEM);
2260                         return;
2261                 }
2262                 if (bp2->bio_offset + bp2->bio_length > pp->mediasize)
2263                         bp2->bio_length = pp->mediasize - bp2->bio_offset;
2264                 bp2->bio_done = g_std_done;
2265                 bp2->bio_offset += entry->gpe_offset;
2266                 g_io_request(bp2, cp);
2267                 return;
2268         case BIO_SPEEDUP:
2269         case BIO_FLUSH:
2270                 break;
2271         case BIO_GETATTR:
2272                 if (g_handleattr_int(bp, "GEOM::fwheads", table->gpt_heads))
2273                         return;
2274                 if (g_handleattr_int(bp, "GEOM::fwsectors", table->gpt_sectors))
2275                         return;
2276                 /*
2277                  * allow_nesting overrides "isleaf" to false _unless_ the
2278                  * provider offset is zero, since otherwise we would recurse.
2279                  */
2280                 if (g_handleattr_int(bp, "PART::isleaf",
2281                         table->gpt_isleaf &&
2282                         (allow_nesting == 0 || entry->gpe_offset == 0)))
2283                         return;
2284                 if (g_handleattr_int(bp, "PART::depth", table->gpt_depth))
2285                         return;
2286                 if (g_handleattr_str(bp, "PART::scheme",
2287                     table->gpt_scheme->name))
2288                         return;
2289                 if (g_handleattr_str(bp, "PART::type",
2290                     G_PART_TYPE(table, entry, buf, sizeof(buf))))
2291                         return;
2292                 if (!strcmp("GEOM::physpath", bp->bio_attribute)) {
2293                         done_func = g_part_get_physpath_done;
2294                         break;
2295                 }
2296                 if (!strcmp("GEOM::kerneldump", bp->bio_attribute)) {
2297                         /*
2298                          * Check that the partition is suitable for kernel
2299                          * dumps. Typically only swap partitions should be
2300                          * used. If the request comes from the nested scheme
2301                          * we allow dumping there as well.
2302                          */
2303                         if ((bp->bio_from == NULL ||
2304                             bp->bio_from->geom->class != &g_part_class) &&
2305                             G_PART_DUMPTO(table, entry) == 0) {
2306                                 g_io_deliver(bp, ENODEV);
2307                                 printf("GEOM_PART: Partition '%s' not suitable"
2308                                     " for kernel dumps (wrong type?)\n",
2309                                     pp->name);
2310                                 return;
2311                         }
2312                         gkd = (struct g_kerneldump *)bp->bio_data;
2313                         if (gkd->offset >= pp->mediasize) {
2314                                 g_io_deliver(bp, EIO);
2315                                 return;
2316                         }
2317                         if (gkd->offset + gkd->length > pp->mediasize)
2318                                 gkd->length = pp->mediasize - gkd->offset;
2319                         gkd->offset += entry->gpe_offset;
2320                 }
2321                 break;
2322         default:
2323                 g_io_deliver(bp, EOPNOTSUPP);
2324                 return;
2325         }
2326
2327         bp2 = g_clone_bio(bp);
2328         if (bp2 == NULL) {
2329                 g_io_deliver(bp, ENOMEM);
2330                 return;
2331         }
2332         bp2->bio_done = done_func;
2333         g_io_request(bp2, cp);
2334 }
2335
2336 static void
2337 g_part_init(struct g_class *mp)
2338 {
2339
2340         TAILQ_INSERT_HEAD(&g_part_schemes, &g_part_null_scheme, scheme_list);
2341 }
2342
2343 static void
2344 g_part_fini(struct g_class *mp)
2345 {
2346
2347         TAILQ_REMOVE(&g_part_schemes, &g_part_null_scheme, scheme_list);
2348 }
2349
2350 static void
2351 g_part_unload_event(void *arg, int flag)
2352 {
2353         struct g_consumer *cp;
2354         struct g_geom *gp;
2355         struct g_provider *pp;
2356         struct g_part_scheme *scheme;
2357         struct g_part_table *table;
2358         uintptr_t *xchg;
2359         int acc, error;
2360
2361         if (flag == EV_CANCEL)
2362                 return;
2363
2364         xchg = arg;
2365         error = 0;
2366         scheme = (void *)(*xchg);
2367
2368         g_topology_assert();
2369
2370         LIST_FOREACH(gp, &g_part_class.geom, geom) {
2371                 table = gp->softc;
2372                 if (table->gpt_scheme != scheme)
2373                         continue;
2374
2375                 acc = 0;
2376                 LIST_FOREACH(pp, &gp->provider, provider)
2377                         acc += pp->acr + pp->acw + pp->ace;
2378                 LIST_FOREACH(cp, &gp->consumer, consumer)
2379                         acc += cp->acr + cp->acw + cp->ace;
2380
2381                 if (!acc)
2382                         g_part_wither(gp, ENOSYS);
2383                 else
2384                         error = EBUSY;
2385         }
2386
2387         if (!error)
2388                 TAILQ_REMOVE(&g_part_schemes, scheme, scheme_list);
2389
2390         *xchg = error;
2391 }
2392
2393 int
2394 g_part_modevent(module_t mod, int type, struct g_part_scheme *scheme)
2395 {
2396         struct g_part_scheme *iter;
2397         uintptr_t arg;
2398         int error;
2399
2400         error = 0;
2401         switch (type) {
2402         case MOD_LOAD:
2403                 TAILQ_FOREACH(iter, &g_part_schemes, scheme_list) {
2404                         if (scheme == iter) {
2405                                 printf("GEOM_PART: scheme %s is already "
2406                                     "registered!\n", scheme->name);
2407                                 break;
2408                         }
2409                 }
2410                 if (iter == NULL) {
2411                         TAILQ_INSERT_TAIL(&g_part_schemes, scheme,
2412                             scheme_list);
2413                         g_retaste(&g_part_class);
2414                 }
2415                 break;
2416         case MOD_UNLOAD:
2417                 arg = (uintptr_t)scheme;
2418                 error = g_waitfor_event(g_part_unload_event, &arg, M_WAITOK,
2419                     NULL);
2420                 if (error == 0)
2421                         error = arg;
2422                 break;
2423         default:
2424                 error = EOPNOTSUPP;
2425                 break;
2426         }
2427
2428         return (error);
2429 }