sys/geom/linux_lvm/g_linux_lvm.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2008 Andrew Thompson <thompsa@FreeBSD.org>
   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  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/ctype.h>
  33 #include <sys/param.h>
  34 #include <sys/bio.h>
  35 #include <sys/kernel.h>
  36 #include <sys/limits.h>
  37 #include <sys/malloc.h>
  38 #include <sys/queue.h>
  39 #include <sys/sysctl.h>
  40 #include <sys/systm.h>
  41
  42 #include <geom/geom.h>
  43 #include <geom/geom_dbg.h>
  44 #include <sys/endian.h>
  45
  46 #include <geom/linux_lvm/g_linux_lvm.h>
  47
  48 FEATURE(geom_linux_lvm, "GEOM Linux LVM partitioning support");
  49
  50 /* Declare malloc(9) label */
  51 static MALLOC_DEFINE(M_GLLVM, "gllvm", "GEOM_LINUX_LVM Data");
  52
  53 /* GEOM class methods */
  54 static g_access_t g_llvm_access;
  55 static g_init_t g_llvm_init;
  56 static g_orphan_t g_llvm_orphan;
  57 static g_orphan_t g_llvm_taste_orphan;
  58 static g_start_t g_llvm_start;
  59 static g_taste_t g_llvm_taste;
  60 static g_ctl_destroy_geom_t g_llvm_destroy_geom;
  61
  62 static void     g_llvm_done(struct bio *);
  63 static void     g_llvm_remove_disk(struct g_llvm_vg *, struct g_consumer *);
  64 static int      g_llvm_activate_lv(struct g_llvm_vg *, struct g_llvm_lv *);
  65 static int      g_llvm_add_disk(struct g_llvm_vg *, struct g_provider *, char *);
  66 static void     g_llvm_free_vg(struct g_llvm_vg *);
  67 static int      g_llvm_destroy(struct g_llvm_vg *, int);
  68 static int      g_llvm_read_label(struct g_consumer *, struct g_llvm_label *);
  69 static int      g_llvm_read_md(struct g_consumer *, struct g_llvm_metadata *,
  70                     struct g_llvm_label *);
  71
  72 static int      llvm_label_decode(const u_char *, struct g_llvm_label *, int);
  73 static int      llvm_md_decode(const u_char *, struct g_llvm_metadata *,
  74                     struct g_llvm_label *);
  75 static int      llvm_textconf_decode(u_char *, int,
  76                     struct g_llvm_metadata *);
  77 static int      llvm_textconf_decode_pv(char **, char *, struct g_llvm_vg *);
  78 static int      llvm_textconf_decode_lv(char **, char *, struct g_llvm_vg *);
  79 static int      llvm_textconf_decode_sg(char **, char *, struct g_llvm_lv *);
  80
  81 SYSCTL_DECL(_kern_geom);
  82 SYSCTL_NODE(_kern_geom, OID_AUTO, linux_lvm, CTLFLAG_RW, 0,
  83     "GEOM_LINUX_LVM stuff");
  84 static u_int g_llvm_debug = 0;
  85 SYSCTL_UINT(_kern_geom_linux_lvm, OID_AUTO, debug, CTLFLAG_RWTUN, &g_llvm_debug, 0,
  86     "Debug level");
  87
  88 LIST_HEAD(, g_llvm_vg) vg_list;
  89
  90 /*
  91  * Called to notify geom when it's been opened, and for what intent
  92  */
  93 static int
  94 g_llvm_access(struct g_provider *pp, int dr, int dw, int de)
  95 {
  96         struct g_consumer *c;
  97         struct g_llvm_vg *vg;
  98         struct g_geom *gp;
  99         int error;
 100
 101         KASSERT(pp != NULL, ("%s: NULL provider", __func__));
 102         gp = pp->geom;
 103         KASSERT(gp != NULL, ("%s: NULL geom", __func__));
 104         vg = gp->softc;
 105
 106         if (vg == NULL) {
 107                 /* It seems that .access can be called with negative dr,dw,dx
 108                  * in this case but I want to check for myself */
 109                 G_LLVM_DEBUG(0, "access(%d, %d, %d) for %s",
 110                     dr, dw, de, pp->name);
 111
 112                 /* This should only happen when geom is withered so
 113                  * allow only negative requests */
 114                 KASSERT(dr <= 0 && dw <= 0 && de <= 0,
 115                     ("%s: Positive access for %s", __func__, pp->name));
 116                 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
 117                         G_LLVM_DEBUG(0,
 118                             "Device %s definitely destroyed", pp->name);
 119                 return (0);
 120         }
 121
 122         /* Grab an exclusive bit to propagate on our consumers on first open */
 123         if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
 124                 de++;
 125         /* ... drop it on close */
 126         if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
 127                 de--;
 128
 129         error = ENXIO;
 130         LIST_FOREACH(c, &gp->consumer, consumer) {
 131                 KASSERT(c != NULL, ("%s: consumer is NULL", __func__));
 132                 error = g_access(c, dr, dw, de);
 133                 if (error != 0) {
 134                         struct g_consumer *c2;
 135
 136                         /* Backout earlier changes */
 137                         LIST_FOREACH(c2, &gp->consumer, consumer) {
 138                                 if (c2 == c) /* all eariler components fixed */
 139                                         return (error);
 140                                 g_access(c2, -dr, -dw, -de);
 141                         }
 142                 }
 143         }
 144
 145         return (error);
 146 }
 147
 148 /*
 149  * Dismantle bio_queue and destroy its components
 150  */
 151 static void
 152 bioq_dismantle(struct bio_queue_head *bq)
 153 {
 154         struct bio *b;
 155
 156         for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
 157                 bioq_remove(bq, b);
 158                 g_destroy_bio(b);
 159         }
 160 }
 161
 162 /*
 163  * GEOM .done handler
 164  * Can't use standard handler because one requested IO may
 165  * fork into additional data IOs
 166  */
 167 static void
 168 g_llvm_done(struct bio *b)
 169 {
 170         struct bio *parent_b;
 171
 172         parent_b = b->bio_parent;
 173
 174         if (b->bio_error != 0) {
 175                 G_LLVM_DEBUG(0, "Error %d for offset=%ju, length=%ju on %s",
 176                     b->bio_error, b->bio_offset, b->bio_length,
 177                     b->bio_to->name);
 178                 if (parent_b->bio_error == 0)
 179                         parent_b->bio_error = b->bio_error;
 180         }
 181
 182         parent_b->bio_inbed++;
 183         parent_b->bio_completed += b->bio_completed;
 184
 185         if (parent_b->bio_children == parent_b->bio_inbed) {
 186                 parent_b->bio_completed = parent_b->bio_length;
 187                 g_io_deliver(parent_b, parent_b->bio_error);
 188         }
 189         g_destroy_bio(b);
 190 }
 191
 192 static void
 193 g_llvm_start(struct bio *bp)
 194 {
 195         struct g_provider *pp;
 196         struct g_llvm_vg *vg;
 197         struct g_llvm_pv *pv;
 198         struct g_llvm_lv *lv;
 199         struct g_llvm_segment *sg;
 200         struct bio *cb;
 201         struct bio_queue_head bq;
 202         size_t chunk_size;
 203         off_t offset, length;
 204         char *addr;
 205         u_int count;
 206
 207         pp = bp->bio_to;
 208         lv = pp->private;
 209         vg = pp->geom->softc;
 210
 211         switch (bp->bio_cmd) {
 212         case BIO_READ:
 213         case BIO_WRITE:
 214         case BIO_DELETE:
 215         /* XXX BIO_GETATTR allowed? */
 216                 break;
 217         default:
 218                 g_io_deliver(bp, EOPNOTSUPP);
 219                 return;
 220         }
 221
 222         bioq_init(&bq);
 223
 224         chunk_size = vg->vg_extentsize;
 225         addr = bp->bio_data;
 226         offset = bp->bio_offset;        /* virtual offset and length */
 227         length = bp->bio_length;
 228
 229         while (length > 0) {
 230                 size_t chunk_index, in_chunk_offset, in_chunk_length;
 231
 232                 pv = NULL;
 233                 cb = g_clone_bio(bp);
 234                 if (cb == NULL) {
 235                         bioq_dismantle(&bq);
 236                         if (bp->bio_error == 0)
 237                                 bp->bio_error = ENOMEM;
 238                         g_io_deliver(bp, bp->bio_error);
 239                         return;
 240                 }
 241
 242                 /* get the segment and the pv */
 243                 if (lv->lv_sgcount == 1) {
 244                         /* skip much of the calculations for a single sg */
 245                         chunk_index = 0;
 246                         in_chunk_offset = 0;
 247                         in_chunk_length = length;
 248                         sg = lv->lv_firstsg;
 249                         pv = sg->sg_pv;
 250                         cb->bio_offset = offset + sg->sg_pvoffset;
 251                 } else {
 252                         chunk_index = offset / chunk_size; /* round downwards */
 253                         in_chunk_offset = offset % chunk_size;
 254                         in_chunk_length =
 255                             min(length, chunk_size - in_chunk_offset);
 256
 257                         /* XXX could be faster */
 258                         LIST_FOREACH(sg, &lv->lv_segs, sg_next) {
 259                                 if (chunk_index >= sg->sg_start &&
 260                                     chunk_index <= sg->sg_end) {
 261                                         /* adjust chunk index for sg start */
 262                                         chunk_index -= sg->sg_start;
 263                                         pv = sg->sg_pv;
 264                                         break;
 265                                 }
 266                         }
 267                         cb->bio_offset =
 268                             (off_t)chunk_index * (off_t)chunk_size
 269                             + in_chunk_offset + sg->sg_pvoffset;
 270                 }
 271
 272                 KASSERT(pv != NULL, ("Can't find PV for chunk %zu",
 273                     chunk_index));
 274
 275                 cb->bio_to = pv->pv_gprov;
 276                 cb->bio_done = g_llvm_done;
 277                 cb->bio_length = in_chunk_length;
 278                 cb->bio_data = addr;
 279                 cb->bio_caller1 = pv;
 280                 bioq_disksort(&bq, cb);
 281
 282                 G_LLVM_DEBUG(5,
 283                     "Mapped %s(%ju, %ju) on %s to %zu(%zu,%zu) @ %s:%ju",
 284                     bp->bio_cmd == BIO_READ ? "R" : "W",
 285                     offset, length, lv->lv_name,
 286                     chunk_index, in_chunk_offset, in_chunk_length,
 287                     pv->pv_name, cb->bio_offset);
 288
 289                 addr += in_chunk_length;
 290                 length -= in_chunk_length;
 291                 offset += in_chunk_length;
 292         }
 293
 294         /* Fire off bio's here */
 295         count = 0;
 296         for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
 297                 bioq_remove(&bq, cb);
 298                 pv = cb->bio_caller1;
 299                 cb->bio_caller1 = NULL;
 300                 G_LLVM_DEBUG(6, "firing bio to %s, offset=%ju, length=%ju",
 301                     cb->bio_to->name, cb->bio_offset, cb->bio_length);
 302                 g_io_request(cb, pv->pv_gcons);
 303                 count++;
 304         }
 305         if (count == 0) { /* We handled everything locally */
 306                 bp->bio_completed = bp->bio_length;
 307                 g_io_deliver(bp, 0);
 308         }
 309 }
 310
 311 static void
 312 g_llvm_remove_disk(struct g_llvm_vg *vg, struct g_consumer *cp)
 313 {
 314         struct g_llvm_pv *pv;
 315         struct g_llvm_lv *lv;
 316         struct g_llvm_segment *sg;
 317         int found;
 318
 319         KASSERT(cp != NULL, ("Non-valid disk in %s.", __func__));
 320         pv = (struct g_llvm_pv *)cp->private;
 321
 322         G_LLVM_DEBUG(0, "Disk %s removed from %s.", cp->provider->name,
 323             pv->pv_name);
 324
 325         LIST_FOREACH(lv, &vg->vg_lvs, lv_next) {
 326                 /* Find segments that map to this disk */
 327                 found = 0;
 328                 LIST_FOREACH(sg, &lv->lv_segs, sg_next) {
 329                         if (sg->sg_pv == pv) {
 330                                 sg->sg_pv = NULL;
 331                                 lv->lv_sgactive--;
 332                                 found = 1;
 333                                 break;
 334                         }
 335                 }
 336                 if (found) {
 337                         G_LLVM_DEBUG(0, "Device %s removed.",
 338                             lv->lv_gprov->name);
 339                         g_wither_provider(lv->lv_gprov, ENXIO);
 340                         lv->lv_gprov = NULL;
 341                 }
 342         }
 343
 344         if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
 345                 g_access(cp, -cp->acr, -cp->acw, -cp->ace);
 346         g_detach(cp);
 347         g_destroy_consumer(cp);
 348 }
 349
 350 static void
 351 g_llvm_orphan(struct g_consumer *cp)
 352 {
 353         struct g_llvm_vg *vg;
 354         struct g_geom *gp;
 355
 356         g_topology_assert();
 357         gp = cp->geom;
 358         vg = gp->softc;
 359         if (vg == NULL)
 360                 return;
 361
 362         g_llvm_remove_disk(vg, cp);
 363         g_llvm_destroy(vg, 1);
 364 }
 365
 366 static int
 367 g_llvm_activate_lv(struct g_llvm_vg *vg, struct g_llvm_lv *lv)
 368 {
 369         struct g_geom *gp;
 370         struct g_provider *pp;
 371
 372         g_topology_assert();
 373
 374         KASSERT(lv->lv_sgactive == lv->lv_sgcount, ("segment missing"));
 375
 376         gp = vg->vg_geom;
 377         pp = g_new_providerf(gp, "linux_lvm/%s-%s", vg->vg_name, lv->lv_name);
 378         pp->mediasize = vg->vg_extentsize * (off_t)lv->lv_extentcount;
 379         pp->sectorsize = vg->vg_sectorsize;
 380         g_error_provider(pp, 0);
 381         lv->lv_gprov = pp;
 382         pp->private = lv;
 383
 384         G_LLVM_DEBUG(1, "Created %s, %juM", pp->name,
 385             pp->mediasize / (1024*1024));
 386
 387         return (0);
 388 }
 389
 390 static int
 391 g_llvm_add_disk(struct g_llvm_vg *vg, struct g_provider *pp, char *uuid)
 392 {
 393         struct g_geom *gp;
 394         struct g_consumer *cp, *fcp;
 395         struct g_llvm_pv *pv;
 396         struct g_llvm_lv *lv;
 397         struct g_llvm_segment *sg;
 398         int error;
 399
 400         g_topology_assert();
 401
 402         LIST_FOREACH(pv, &vg->vg_pvs, pv_next) {
 403                 if (strcmp(pv->pv_uuid, uuid) == 0)
 404                         break;  /* found it */
 405         }
 406         if (pv == NULL) {
 407                 G_LLVM_DEBUG(3, "uuid %s not found in pv list", uuid);
 408                 return (ENOENT);
 409         }
 410         if (pv->pv_gprov != NULL) {
 411                 G_LLVM_DEBUG(0, "disk %s already initialised in %s",
 412                     pv->pv_name, vg->vg_name);
 413                 return (EEXIST);
 414         }
 415
 416         pv->pv_start *= vg->vg_sectorsize;
 417         gp = vg->vg_geom;
 418         fcp = LIST_FIRST(&gp->consumer);
 419
 420         cp = g_new_consumer(gp);
 421         error = g_attach(cp, pp);
 422         G_LLVM_DEBUG(1, "Attached %s to %s at offset %ju",
 423             pp->name, pv->pv_name, pv->pv_start);
 424
 425         if (error != 0) {
 426                 G_LLVM_DEBUG(0, "cannot attach %s to %s",
 427                     pp->name, vg->vg_name);
 428                 g_destroy_consumer(cp);
 429                 return (error);
 430         }
 431
 432         if (fcp != NULL) {
 433                 if (fcp->provider->sectorsize != pp->sectorsize) {
 434                         G_LLVM_DEBUG(0, "Provider %s of %s has invalid "
 435                             "sector size (%d)", pp->name, vg->vg_name,
 436                             pp->sectorsize);
 437                         return (EINVAL);
 438                 }
 439                 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
 440                         /* Replicate access permissions from first "live"
 441                          * consumer to the new one */
 442                         error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
 443                         if (error != 0) {
 444                                 g_detach(cp);
 445                                 g_destroy_consumer(cp);
 446                                 return (error);
 447                         }
 448                 }
 449         }
 450
 451         cp->private = pv;
 452         pv->pv_gcons = cp;
 453         pv->pv_gprov = pp;
 454
 455         LIST_FOREACH(lv, &vg->vg_lvs, lv_next) {
 456                 /* Find segments that map to this disk */
 457                 LIST_FOREACH(sg, &lv->lv_segs, sg_next) {
 458                         if (strcmp(sg->sg_pvname, pv->pv_name) == 0) {
 459                                 /* avtivate the segment */
 460                                 KASSERT(sg->sg_pv == NULL,
 461                                     ("segment already mapped"));
 462                                 sg->sg_pvoffset =
 463                                     (off_t)sg->sg_pvstart * vg->vg_extentsize
 464                                     + pv->pv_start;
 465                                 sg->sg_pv = pv;
 466                                 lv->lv_sgactive++;
 467
 468                                 G_LLVM_DEBUG(2, "%s: %d to %d @ %s:%d"
 469                                     " offset %ju sector %ju",
 470                                     lv->lv_name, sg->sg_start, sg->sg_end,
 471                                     sg->sg_pvname, sg->sg_pvstart,
 472                                     sg->sg_pvoffset,
 473                                     sg->sg_pvoffset / vg->vg_sectorsize);
 474                         }
 475                 }
 476                 /* Activate any lvs waiting on this disk */
 477                 if (lv->lv_gprov == NULL && lv->lv_sgactive == lv->lv_sgcount) {
 478                         error = g_llvm_activate_lv(vg, lv);
 479                         if (error)
 480                                 break;
 481                 }
 482         }
 483         return (error);
 484 }
 485
 486 static void
 487 g_llvm_init(struct g_class *mp)
 488 {
 489         LIST_INIT(&vg_list);
 490 }
 491
 492 static void
 493 g_llvm_free_vg(struct g_llvm_vg *vg)
 494 {
 495         struct g_llvm_pv *pv;
 496         struct g_llvm_lv *lv;
 497         struct g_llvm_segment *sg;
 498
 499         /* Free all the structures */
 500         while ((pv = LIST_FIRST(&vg->vg_pvs)) != NULL) {
 501                 LIST_REMOVE(pv, pv_next);
 502                 free(pv, M_GLLVM);
 503         }
 504         while ((lv = LIST_FIRST(&vg->vg_lvs)) != NULL) {
 505                 while ((sg = LIST_FIRST(&lv->lv_segs)) != NULL) {
 506                         LIST_REMOVE(sg, sg_next);
 507                         free(sg, M_GLLVM);
 508                 }
 509                 LIST_REMOVE(lv, lv_next);
 510                 free(lv, M_GLLVM);
 511         }
 512         LIST_REMOVE(vg, vg_next);
 513         free(vg, M_GLLVM);
 514 }
 515
 516 static void
 517 g_llvm_taste_orphan(struct g_consumer *cp)
 518 {
 519
 520         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 521             cp->provider->name));
 522 }
 523
 524 static struct g_geom *
 525 g_llvm_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
 526 {
 527         struct g_consumer *cp;
 528         struct g_geom *gp;
 529         struct g_llvm_label ll;
 530         struct g_llvm_metadata md;
 531         struct g_llvm_vg *vg;
 532         int error;
 533
 534         bzero(&md, sizeof(md));
 535
 536         g_topology_assert();
 537         g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
 538         gp = g_new_geomf(mp, "linux_lvm:taste");
 539         /* This orphan function should be never called. */
 540         gp->orphan = g_llvm_taste_orphan;
 541         cp = g_new_consumer(gp);
 542         g_attach(cp, pp);
 543         error = g_llvm_read_label(cp, &ll);
 544         if (!error)
 545                 error = g_llvm_read_md(cp, &md, &ll);
 546         g_detach(cp);
 547         g_destroy_consumer(cp);
 548         g_destroy_geom(gp);
 549         if (error != 0)
 550                 return (NULL);
 551
 552         vg = md.md_vg;
 553         if (vg->vg_geom == NULL) {
 554                 /* new volume group */
 555                 gp = g_new_geomf(mp, "%s", vg->vg_name);
 556                 gp->start = g_llvm_start;
 557                 gp->spoiled = g_llvm_orphan;
 558                 gp->orphan = g_llvm_orphan;
 559                 gp->access = g_llvm_access;
 560                 vg->vg_sectorsize = pp->sectorsize;
 561                 vg->vg_extentsize *= vg->vg_sectorsize;
 562                 vg->vg_geom = gp;
 563                 gp->softc = vg;
 564                 G_LLVM_DEBUG(1, "Created volume %s, extent size %zuK",
 565                     vg->vg_name, vg->vg_extentsize / 1024);
 566         }
 567
 568         /* initialise this disk in the volume group */
 569         g_llvm_add_disk(vg, pp, ll.ll_uuid);
 570         return (vg->vg_geom);
 571 }
 572
 573 static int
 574 g_llvm_destroy(struct g_llvm_vg *vg, int force)
 575 {
 576         struct g_provider *pp;
 577         struct g_geom *gp;
 578
 579         g_topology_assert();
 580         if (vg == NULL)
 581                 return (ENXIO);
 582         gp = vg->vg_geom;
 583
 584         LIST_FOREACH(pp, &gp->provider, provider) {
 585                 if (pp->acr != 0 || pp->acw != 0 || pp->ace != 0) {
 586                         G_LLVM_DEBUG(1, "Device %s is still open (r%dw%de%d)",
 587                             pp->name, pp->acr, pp->acw, pp->ace);
 588                         if (!force)
 589                                 return (EBUSY);
 590                 }
 591         }
 592
 593         g_llvm_free_vg(gp->softc);
 594         gp->softc = NULL;
 595         g_wither_geom(gp, ENXIO);
 596         return (0);
 597 }
 598
 599 static int
 600 g_llvm_destroy_geom(struct gctl_req *req __unused, struct g_class *mp __unused,
 601     struct g_geom *gp)
 602 {
 603         struct g_llvm_vg *vg;
 604
 605         vg = gp->softc;
 606         return (g_llvm_destroy(vg, 0));
 607 }
 608
 609 int
 610 g_llvm_read_label(struct g_consumer *cp, struct g_llvm_label *ll)
 611 {
 612         struct g_provider *pp;
 613         u_char *buf;
 614         int i, error = 0;
 615
 616         g_topology_assert();
 617
 618         /* The LVM label is stored on the first four sectors */
 619         error = g_access(cp, 1, 0, 0);
 620         if (error != 0)
 621                 return (error);
 622         pp = cp->provider;
 623         g_topology_unlock();
 624         buf = g_read_data(cp, 0, pp->sectorsize * 4, &error);
 625         g_topology_lock();
 626         g_access(cp, -1, 0, 0);
 627         if (buf == NULL) {
 628                 G_LLVM_DEBUG(1, "Cannot read metadata from %s (error=%d)",
 629                     pp->name, error);
 630                 return (error);
 631         }
 632
 633         /* Search the four sectors for the LVM label. */
 634         for (i = 0; i < 4; i++) {
 635                 error = llvm_label_decode(&buf[i * pp->sectorsize], ll, i);
 636                 if (error == 0)
 637                         break;  /* found it */
 638         }
 639         g_free(buf);
 640         return (error);
 641 }
 642
 643 int
 644 g_llvm_read_md(struct g_consumer *cp, struct g_llvm_metadata *md,
 645     struct g_llvm_label *ll)
 646 {
 647         struct g_provider *pp;
 648         u_char *buf;
 649         int error;
 650         int size;
 651
 652         g_topology_assert();
 653
 654         error = g_access(cp, 1, 0, 0);
 655         if (error != 0)
 656                 return (error);
 657         pp = cp->provider;
 658         g_topology_unlock();
 659         buf = g_read_data(cp, ll->ll_md_offset, pp->sectorsize, &error);
 660         g_topology_lock();
 661         g_access(cp, -1, 0, 0);
 662         if (buf == NULL) {
 663                 G_LLVM_DEBUG(0, "Cannot read metadata from %s (error=%d)",
 664                     cp->provider->name, error);
 665                 return (error);
 666         }
 667
 668         error = llvm_md_decode(buf, md, ll);
 669         g_free(buf);
 670         if (error != 0) {
 671                 return (error);
 672         }
 673
 674         G_LLVM_DEBUG(1, "reading LVM2 config @ %s:%ju", pp->name,
 675                     ll->ll_md_offset + md->md_reloffset);
 676         error = g_access(cp, 1, 0, 0);
 677         if (error != 0)
 678                 return (error);
 679         pp = cp->provider;
 680         g_topology_unlock();
 681         /* round up to the nearest sector */
 682         size = md->md_relsize +
 683             (pp->sectorsize - md->md_relsize % pp->sectorsize);
 684         buf = g_read_data(cp, ll->ll_md_offset + md->md_reloffset, size, &error);
 685         g_topology_lock();
 686         g_access(cp, -1, 0, 0);
 687         if (buf == NULL) {
 688                 G_LLVM_DEBUG(0, "Cannot read LVM2 config from %s (error=%d)",
 689                     pp->name, error);
 690                 return (error);
 691         }
 692         buf[md->md_relsize] = '\0';
 693         G_LLVM_DEBUG(10, "LVM config:\n%s\n", buf);
 694         error = llvm_textconf_decode(buf, md->md_relsize, md);
 695         g_free(buf);
 696
 697         return (error);
 698 }
 699
 700 static int
 701 llvm_label_decode(const u_char *data, struct g_llvm_label *ll, int sector)
 702 {
 703         uint64_t off;
 704         char *uuid;
 705
 706         /* Magic string */
 707         if (bcmp("LABELONE", data , 8) != 0)
 708                 return (EINVAL);
 709
 710         /* We only support LVM2 text format */
 711         if (bcmp("LVM2 001", data + 24, 8) != 0) {
 712                 G_LLVM_DEBUG(0, "Unsupported LVM format");
 713                 return (EINVAL);
 714         }
 715
 716         ll->ll_sector = le64dec(data + 8);
 717         ll->ll_crc = le32dec(data + 16);
 718         ll->ll_offset = le32dec(data + 20);
 719
 720         if (ll->ll_sector != sector) {
 721                 G_LLVM_DEBUG(0, "Expected sector %ju, found at %d",
 722                     ll->ll_sector, sector);
 723                 return (EINVAL);
 724         }
 725
 726         off = ll->ll_offset;
 727         /*
 728          * convert the binary uuid to string format, the format is
 729          * xxxxxx-xxxx-xxxx-xxxx-xxxx-xxxx-xxxxxx (6-4-4-4-4-4-6)
 730          */
 731         uuid = ll->ll_uuid;
 732         bcopy(data + off, uuid, 6);
 733         off += 6;
 734         uuid += 6;
 735         *uuid++ = '-';
 736         for (int i = 0; i < 5; i++) {
 737                 bcopy(data + off, uuid, 4);
 738                 off += 4;
 739                 uuid += 4;
 740                 *uuid++ = '-';
 741         }
 742         bcopy(data + off, uuid, 6);
 743         off += 6;
 744         uuid += 6;
 745         *uuid++ = '\0';
 746
 747         ll->ll_size = le64dec(data + off);
 748         off += 8;
 749         ll->ll_pestart = le64dec(data + off);
 750         off += 16;
 751
 752         /* Only one data section is supported */
 753         if (le64dec(data + off) != 0) {
 754                 G_LLVM_DEBUG(0, "Only one data section supported");
 755                 return (EINVAL);
 756         }
 757
 758         off += 16;
 759         ll->ll_md_offset = le64dec(data + off);
 760         off += 8;
 761         ll->ll_md_size = le64dec(data + off);
 762         off += 8;
 763
 764         G_LLVM_DEBUG(1, "LVM metadata: offset=%ju, size=%ju", ll->ll_md_offset,
 765             ll->ll_md_size);
 766
 767         /* Only one data section is supported */
 768         if (le64dec(data + off) != 0) {
 769                 G_LLVM_DEBUG(0, "Only one metadata section supported");
 770                 return (EINVAL);
 771         }
 772
 773         G_LLVM_DEBUG(2, "label uuid=%s", ll->ll_uuid);
 774         G_LLVM_DEBUG(2, "sector=%ju, crc=%u, offset=%u, size=%ju, pestart=%ju",
 775             ll->ll_sector, ll->ll_crc, ll->ll_offset, ll->ll_size,
 776             ll->ll_pestart);
 777
 778         return (0);
 779 }
 780
 781 static int
 782 llvm_md_decode(const u_char *data, struct g_llvm_metadata *md,
 783     struct g_llvm_label *ll)
 784 {
 785         uint64_t off;
 786         char magic[16];
 787
 788         off = 0;
 789         md->md_csum = le32dec(data + off);
 790         off += 4;
 791         bcopy(data + off, magic, 16);
 792         off += 16;
 793         md->md_version = le32dec(data + off);
 794         off += 4;
 795         md->md_start = le64dec(data + off);
 796         off += 8;
 797         md->md_size = le64dec(data + off);
 798         off += 8;
 799
 800         if (bcmp(G_LLVM_MAGIC, magic, 16) != 0) {
 801                 G_LLVM_DEBUG(0, "Incorrect md magic number");
 802                 return (EINVAL);
 803         }
 804         if (md->md_version != 1) {
 805                 G_LLVM_DEBUG(0, "Incorrect md version number (%u)",
 806                     md->md_version);
 807                 return (EINVAL);
 808         }
 809         if (md->md_start != ll->ll_md_offset) {
 810                 G_LLVM_DEBUG(0, "Incorrect md offset (%ju)", md->md_start);
 811                 return (EINVAL);
 812         }
 813
 814         /* Aparently only one is ever returned */
 815         md->md_reloffset = le64dec(data + off);
 816         off += 8;
 817         md->md_relsize = le64dec(data + off);
 818         off += 16;      /* XXX skipped checksum */
 819
 820         if (le64dec(data + off) != 0) {
 821                 G_LLVM_DEBUG(0, "Only one reloc supported");
 822                 return (EINVAL);
 823         }
 824
 825         G_LLVM_DEBUG(3, "reloc: offset=%ju, size=%ju",
 826             md->md_reloffset, md->md_relsize);
 827         G_LLVM_DEBUG(3, "md: version=%u, start=%ju, size=%ju",
 828             md->md_version, md->md_start, md->md_size);
 829
 830         return (0);
 831 }
 832
 833 #define GRAB_INT(key, tok1, tok2, v)                                    \
 834         if (tok1 && tok2 && strncmp(tok1, key, sizeof(key)) == 0) {     \
 835                 v = strtol(tok2, &tok1, 10);                            \
 836                 if (tok1 == tok2)                                       \
 837                         /* strtol did not eat any of the buffer */      \
 838                         goto bad;                                       \
 839                 continue;                                               \
 840         }
 841
 842 #define GRAB_STR(key, tok1, tok2, v, len)                               \
 843         if (tok1 && tok2 && strncmp(tok1, key, sizeof(key)) == 0) {     \
 844                 strsep(&tok2, "\"");                                    \
 845                 if (tok2 == NULL)                                       \
 846                         continue;                                       \
 847                 tok1 = strsep(&tok2, "\"");                             \
 848                 if (tok2 == NULL)                                       \
 849                         continue;                                       \
 850                 strncpy(v, tok1, len);                                  \
 851                 continue;                                               \
 852         }
 853
 854 #define SPLIT(key, value, str)                                          \
 855         key = strsep(&value, str);                                      \
 856         /* strip trailing whitespace on the key */                      \
 857         for (char *t = key; *t != '\0'; t++)                            \
 858                 if (isspace(*t)) {                                      \
 859                         *t = '\0';                                      \
 860                         break;                                          \
 861                 }
 862
 863 static size_t
 864 llvm_grab_name(char *name, const char *tok)
 865 {
 866         size_t len;
 867
 868         len = 0;
 869         if (tok == NULL)
 870                 return (0);
 871         if (tok[0] == '-')
 872                 return (0);
 873         if (strcmp(tok, ".") == 0 || strcmp(tok, "..") == 0)
 874                 return (0);
 875         while (tok[len] && (isalpha(tok[len]) || isdigit(tok[len]) ||
 876             tok[len] == '.' || tok[len] == '_' || tok[len] == '-' ||
 877             tok[len] == '+') && len < G_LLVM_NAMELEN - 1)
 878                 len++;
 879         bcopy(tok, name, len);
 880         name[len] = '\0';
 881         return (len);
 882 }
 883
 884 static int
 885 llvm_textconf_decode(u_char *data, int buflen, struct g_llvm_metadata *md)
 886 {
 887         struct g_llvm_vg        *vg;
 888         char *buf = data;
 889         char *tok, *v;
 890         char name[G_LLVM_NAMELEN];
 891         char uuid[G_LLVM_UUIDLEN];
 892         size_t len;
 893
 894         if (buf == NULL || *buf == '\0')
 895                 return (EINVAL);
 896
 897         tok = strsep(&buf, "\n");
 898         if (tok == NULL)
 899                 return (EINVAL);
 900         len = llvm_grab_name(name, tok);
 901         if (len == 0)
 902                 return (EINVAL);
 903
 904         /* check too see if the vg has already been loaded off another disk */
 905         LIST_FOREACH(vg, &vg_list, vg_next) {
 906                 if (strcmp(vg->vg_name, name) == 0) {
 907                         uuid[0] = '\0';
 908                         /* grab the volume group uuid */
 909                         while ((tok = strsep(&buf, "\n")) != NULL) {
 910                                 if (strstr(tok, "{"))
 911                                         break;
 912                                 if (strstr(tok, "=")) {
 913                                         SPLIT(v, tok, "=");
 914                                         GRAB_STR("id", v, tok, uuid,
 915                                             sizeof(uuid));
 916                                 }
 917                         }
 918                         if (strcmp(vg->vg_uuid, uuid) == 0) {
 919                                 /* existing vg */
 920                                 md->md_vg = vg;
 921                                 return (0);
 922                         }
 923                         /* XXX different volume group with name clash! */
 924                         G_LLVM_DEBUG(0,
 925                             "%s already exists, volume group not loaded", name);
 926                         return (EINVAL);
 927                 }
 928         }
 929
 930         vg = malloc(sizeof(*vg), M_GLLVM, M_NOWAIT|M_ZERO);
 931         if (vg == NULL)
 932                 return (ENOMEM);
 933
 934         strncpy(vg->vg_name, name, sizeof(vg->vg_name));
 935         LIST_INIT(&vg->vg_pvs);
 936         LIST_INIT(&vg->vg_lvs);
 937
 938 #define VOL_FOREACH(func, tok, buf, p)                                  \
 939         while ((tok = strsep(buf, "\n")) != NULL) {                     \
 940                 if (strstr(tok, "{")) {                                 \
 941                         func(buf, tok, p);                              \
 942                         continue;                                       \
 943                 }                                                       \
 944                 if (strstr(tok, "}"))                                   \
 945                         break;                                          \
 946         }
 947
 948         while ((tok = strsep(&buf, "\n")) != NULL) {
 949                 if (strcmp(tok, "physical_volumes {") == 0) {
 950                         VOL_FOREACH(llvm_textconf_decode_pv, tok, &buf, vg);
 951                         continue;
 952                 }
 953                 if (strcmp(tok, "logical_volumes {") == 0) {
 954                         VOL_FOREACH(llvm_textconf_decode_lv, tok, &buf, vg);
 955                         continue;
 956                 }
 957                 if (strstr(tok, "{")) {
 958                         G_LLVM_DEBUG(2, "unknown section %s", tok);
 959                         continue;
 960                 }
 961
 962                 /* parse 'key = value' lines */
 963                 if (strstr(tok, "=")) {
 964                         SPLIT(v, tok, "=");
 965                         GRAB_STR("id", v, tok, vg->vg_uuid, sizeof(vg->vg_uuid));
 966                         GRAB_INT("extent_size", v, tok, vg->vg_extentsize);
 967                         continue;
 968                 }
 969         }
 970         /* basic checking */
 971         if (vg->vg_extentsize == 0)
 972                 goto bad;
 973
 974         md->md_vg = vg;
 975         LIST_INSERT_HEAD(&vg_list, vg, vg_next);
 976         G_LLVM_DEBUG(3, "vg: name=%s uuid=%s", vg->vg_name, vg->vg_uuid);
 977         return(0);
 978
 979 bad:
 980         g_llvm_free_vg(vg);
 981         return (-1);
 982 }
 983 #undef  VOL_FOREACH
 984
 985 static int
 986 llvm_textconf_decode_pv(char **buf, char *tok, struct g_llvm_vg *vg)
 987 {
 988         struct g_llvm_pv        *pv;
 989         char *v;
 990         size_t len;
 991
 992         if (*buf == NULL || **buf == '\0')
 993                 return (EINVAL);
 994
 995         pv = malloc(sizeof(*pv), M_GLLVM, M_NOWAIT|M_ZERO);
 996         if (pv == NULL)
 997                 return (ENOMEM);
 998
 999         pv->pv_vg = vg;
1000         len = 0;
1001         if (tok == NULL)
1002                 goto bad;
1003         len = llvm_grab_name(pv->pv_name, tok);
1004         if (len == 0)
1005                 goto bad;
1006
1007         while ((tok = strsep(buf, "\n")) != NULL) {
1008                 if (strstr(tok, "{"))
1009                         goto bad;
1010
1011                 if (strstr(tok, "}"))
1012                         break;
1013
1014                 /* parse 'key = value' lines */
1015                 if (strstr(tok, "=")) {
1016                         SPLIT(v, tok, "=");
1017                         GRAB_STR("id", v, tok, pv->pv_uuid, sizeof(pv->pv_uuid));
1018                         GRAB_INT("pe_start", v, tok, pv->pv_start);
1019                         GRAB_INT("pe_count", v, tok, pv->pv_count);
1020                         continue;
1021                 }
1022         }
1023         if (tok == NULL)
1024                 goto bad;
1025         /* basic checking */
1026         if (pv->pv_count == 0)
1027                 goto bad;
1028
1029         LIST_INSERT_HEAD(&vg->vg_pvs, pv, pv_next);
1030         G_LLVM_DEBUG(3, "pv: name=%s uuid=%s", pv->pv_name, pv->pv_uuid);
1031
1032         return (0);
1033 bad:
1034         free(pv, M_GLLVM);
1035         return (-1);
1036 }
1037
1038 static int
1039 llvm_textconf_decode_lv(char **buf, char *tok, struct g_llvm_vg *vg)
1040 {
1041         struct g_llvm_lv        *lv;
1042         struct g_llvm_segment *sg;
1043         char *v;
1044         size_t len;
1045
1046         if (*buf == NULL || **buf == '\0')
1047                 return (EINVAL);
1048
1049         lv = malloc(sizeof(*lv), M_GLLVM, M_NOWAIT|M_ZERO);
1050         if (lv == NULL)
1051                 return (ENOMEM);
1052
1053         lv->lv_vg = vg;
1054         LIST_INIT(&lv->lv_segs);
1055
1056         if (tok == NULL)
1057                 goto bad;
1058         len = llvm_grab_name(lv->lv_name, tok);
1059         if (len == 0)
1060                 goto bad;
1061
1062         while ((tok = strsep(buf, "\n")) != NULL) {
1063                 if (strstr(tok, "{")) {
1064                         if (strstr(tok, "segment")) {
1065                                 llvm_textconf_decode_sg(buf, tok, lv);
1066                                 continue;
1067                         } else
1068                                 /* unexpected section */
1069                                 goto bad;
1070                 }
1071
1072                 if (strstr(tok, "}"))
1073                         break;
1074
1075                 /* parse 'key = value' lines */
1076                 if (strstr(tok, "=")) {
1077                         SPLIT(v, tok, "=");
1078                         GRAB_STR("id", v, tok, lv->lv_uuid, sizeof(lv->lv_uuid));
1079                         GRAB_INT("segment_count", v, tok, lv->lv_sgcount);
1080                         continue;
1081                 }
1082         }
1083         if (tok == NULL)
1084                 goto bad;
1085         if (lv->lv_sgcount == 0 || lv->lv_sgcount != lv->lv_numsegs)
1086                 /* zero or incomplete segment list */
1087                 goto bad;
1088
1089         /* Optimize for only one segment on the pv */
1090         lv->lv_firstsg = LIST_FIRST(&lv->lv_segs);
1091         LIST_INSERT_HEAD(&vg->vg_lvs, lv, lv_next);
1092         G_LLVM_DEBUG(3, "lv: name=%s uuid=%s", lv->lv_name, lv->lv_uuid);
1093
1094         return (0);
1095 bad:
1096         while ((sg = LIST_FIRST(&lv->lv_segs)) != NULL) {
1097                 LIST_REMOVE(sg, sg_next);
1098                 free(sg, M_GLLVM);
1099         }
1100         free(lv, M_GLLVM);
1101         return (-1);
1102 }
1103
1104 static int
1105 llvm_textconf_decode_sg(char **buf, char *tok, struct g_llvm_lv *lv)
1106 {
1107         struct g_llvm_segment *sg;
1108         char *v;
1109         int count = 0;
1110
1111         if (*buf == NULL || **buf == '\0')
1112                 return (EINVAL);
1113
1114         sg = malloc(sizeof(*sg), M_GLLVM, M_NOWAIT|M_ZERO);
1115         if (sg == NULL)
1116                 return (ENOMEM);
1117
1118         while ((tok = strsep(buf, "\n")) != NULL) {
1119                 /* only a single linear stripe is supported */
1120                 if (strstr(tok, "stripe_count")) {
1121                         SPLIT(v, tok, "=");
1122                         GRAB_INT("stripe_count", v, tok, count);
1123                         if (count != 1)
1124                                 goto bad;
1125                 }
1126
1127                 if (strstr(tok, "{"))
1128                         goto bad;
1129
1130                 if (strstr(tok, "}"))
1131                         break;
1132
1133                 if (strcmp(tok, "stripes = [") == 0) {
1134                         tok = strsep(buf, "\n");
1135                         if (tok == NULL)
1136                                 goto bad;
1137
1138                         strsep(&tok, "\"");
1139                         if (tok == NULL)
1140                                 goto bad;       /* missing open quotes */
1141                         v = strsep(&tok, "\"");
1142                         if (tok == NULL)
1143                                 goto bad;       /* missing close quotes */
1144                         strncpy(sg->sg_pvname, v, sizeof(sg->sg_pvname));
1145                         if (*tok != ',')
1146                                 goto bad;       /* missing comma for stripe */
1147                         tok++;
1148
1149                         sg->sg_pvstart = strtol(tok, &v, 10);
1150                         if (v == tok)
1151                                 /* strtol did not eat any of the buffer */
1152                                 goto bad;
1153
1154                         continue;
1155                 }
1156
1157                 /* parse 'key = value' lines */
1158                 if (strstr(tok, "=")) {
1159                         SPLIT(v, tok, "=");
1160                         GRAB_INT("start_extent", v, tok, sg->sg_start);
1161                         GRAB_INT("extent_count", v, tok, sg->sg_count);
1162                         continue;
1163                 }
1164         }
1165         if (tok == NULL)
1166                 goto bad;
1167         /* basic checking */
1168         if (count != 1 || sg->sg_count == 0)
1169                 goto bad;
1170
1171         sg->sg_end = sg->sg_start + sg->sg_count - 1;
1172         lv->lv_numsegs++;
1173         lv->lv_extentcount += sg->sg_count;
1174         LIST_INSERT_HEAD(&lv->lv_segs, sg, sg_next);
1175
1176         return (0);
1177 bad:
1178         free(sg, M_GLLVM);
1179         return (-1);
1180 }
1181 #undef  GRAB_INT
1182 #undef  GRAB_STR
1183 #undef  SPLIT
1184
1185 static struct g_class g_llvm_class = {
1186         .name = G_LLVM_CLASS_NAME,
1187         .version = G_VERSION,
1188         .init = g_llvm_init,
1189         .taste = g_llvm_taste,
1190         .destroy_geom = g_llvm_destroy_geom
1191 };
1192
1193 DECLARE_GEOM_CLASS(g_llvm_class, g_linux_lvm);
1194 MODULE_VERSION(geom_linux_lvm, 0);