sys/geom/virstor/g_virstor.c

   1 /*-
   2  * Copyright (c) 2006-2007 Ivan Voras <ivoras@freebsd.org>
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 /* Implementation notes:
  28  * - "Components" are wrappers around providers that make up the
  29  *   virtual storage (i.e. a virstor has "physical" components)
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/module.h>
  39 #include <sys/lock.h>
  40 #include <sys/mutex.h>
  41 #include <sys/sx.h>
  42 #include <sys/bio.h>
  43 #include <sys/sbuf.h>
  44 #include <sys/sysctl.h>
  45 #include <sys/malloc.h>
  46 #include <sys/time.h>
  47 #include <sys/proc.h>
  48 #include <sys/kthread.h>
  49 #include <sys/mutex.h>
  50 #include <vm/uma.h>
  51 #include <geom/geom.h>
  52
  53 #include <geom/virstor/g_virstor.h>
  54 #include <geom/virstor/g_virstor_md.h>
  55
  56 FEATURE(g_virstor, "GEOM virtual storage support");
  57
  58 /* Declare malloc(9) label */
  59 static MALLOC_DEFINE(M_GVIRSTOR, "gvirstor", "GEOM_VIRSTOR Data");
  60
  61 /* GEOM class methods */
  62 static g_init_t g_virstor_init;
  63 static g_fini_t g_virstor_fini;
  64 static g_taste_t g_virstor_taste;
  65 static g_ctl_req_t g_virstor_config;
  66 static g_ctl_destroy_geom_t g_virstor_destroy_geom;
  67
  68 /* Declare & initialize class structure ("geom class") */
  69 struct g_class g_virstor_class = {
  70         .name =         G_VIRSTOR_CLASS_NAME,
  71         .version =      G_VERSION,
  72         .init =         g_virstor_init,
  73         .fini =         g_virstor_fini,
  74         .taste =        g_virstor_taste,
  75         .ctlreq =       g_virstor_config,
  76         .destroy_geom = g_virstor_destroy_geom
  77         /* The .dumpconf and the rest are only usable for a geom instance, so
  78          * they will be set when such instance is created. */
  79 };
  80
  81 /* Declare sysctl's and loader tunables */
  82 SYSCTL_DECL(_kern_geom);
  83 static SYSCTL_NODE(_kern_geom, OID_AUTO, virstor, CTLFLAG_RW, 0,
  84     "GEOM_GVIRSTOR information");
  85
  86 static u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */
  87 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RWTUN, &g_virstor_debug,
  88     0, "Debug level (2=production, 5=normal, 15=excessive)");
  89
  90 static u_int g_virstor_chunk_watermark = 100;
  91 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RWTUN,
  92     &g_virstor_chunk_watermark, 0,
  93     "Minimum number of free chunks before issuing administrative warning");
  94
  95 static u_int g_virstor_component_watermark = 1;
  96 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RWTUN,
  97     &g_virstor_component_watermark, 0,
  98     "Minimum number of free components before issuing administrative warning");
  99
 100 static int read_metadata(struct g_consumer *, struct g_virstor_metadata *);
 101 static void write_metadata(struct g_consumer *, struct g_virstor_metadata *);
 102 static int clear_metadata(struct g_virstor_component *);
 103 static int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *,
 104     struct g_virstor_metadata *);
 105 static struct g_geom *create_virstor_geom(struct g_class *,
 106     struct g_virstor_metadata *);
 107 static void virstor_check_and_run(struct g_virstor_softc *);
 108 static u_int virstor_valid_components(struct g_virstor_softc *);
 109 static int virstor_geom_destroy(struct g_virstor_softc *, boolean_t,
 110     boolean_t);
 111 static void remove_component(struct g_virstor_softc *,
 112     struct g_virstor_component *, boolean_t);
 113 static void bioq_dismantle(struct bio_queue_head *);
 114 static int allocate_chunk(struct g_virstor_softc *,
 115     struct g_virstor_component **, u_int *, u_int *);
 116 static void delay_destroy_consumer(void *, int);
 117 static void dump_component(struct g_virstor_component *comp);
 118 #if 0
 119 static void dump_me(struct virstor_map_entry *me, unsigned int nr);
 120 #endif
 121
 122 static void virstor_ctl_stop(struct gctl_req *, struct g_class *);
 123 static void virstor_ctl_add(struct gctl_req *, struct g_class *);
 124 static void virstor_ctl_remove(struct gctl_req *, struct g_class *);
 125 static struct g_virstor_softc * virstor_find_geom(const struct g_class *,
 126     const char *);
 127 static void update_metadata(struct g_virstor_softc *);
 128 static void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *,
 129     u_int, u_int);
 130
 131 static void g_virstor_orphan(struct g_consumer *);
 132 static int g_virstor_access(struct g_provider *, int, int, int);
 133 static void g_virstor_start(struct bio *);
 134 static void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *,
 135     struct g_consumer *, struct g_provider *);
 136 static void g_virstor_done(struct bio *);
 137
 138 static void invalid_call(void);
 139 /*
 140  * Initialise GEOM class (per-class callback)
 141  */
 142 static void
 143 g_virstor_init(struct g_class *mp __unused)
 144 {
 145
 146         /* Catch map struct size mismatch at compile time; Map entries must
 147          * fit into MAXPHYS exactly, with no wasted space. */
 148         CTASSERT(VIRSTOR_MAP_BLOCK_ENTRIES*VIRSTOR_MAP_ENTRY_SIZE == MAXPHYS);
 149
 150         /* Init UMA zones, TAILQ's, other global vars */
 151 }
 152
 153 /*
 154  * Finalise GEOM class (per-class callback)
 155  */
 156 static void
 157 g_virstor_fini(struct g_class *mp __unused)
 158 {
 159
 160         /* Deinit UMA zones & global vars */
 161 }
 162
 163 /*
 164  * Config (per-class callback)
 165  */
 166 static void
 167 g_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb)
 168 {
 169         uint32_t *version;
 170
 171         g_topology_assert();
 172
 173         version = gctl_get_paraml(req, "version", sizeof(*version));
 174         if (version == NULL) {
 175                 gctl_error(req, "Failed to get 'version' argument");
 176                 return;
 177         }
 178         if (*version != G_VIRSTOR_VERSION) {
 179                 gctl_error(req, "Userland and kernel versions out of sync");
 180                 return;
 181         }
 182
 183         g_topology_unlock();
 184         if (strcmp(verb, "add") == 0)
 185                 virstor_ctl_add(req, cp);
 186         else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0)
 187                 virstor_ctl_stop(req, cp);
 188         else if (strcmp(verb, "remove") == 0)
 189                 virstor_ctl_remove(req, cp);
 190         else
 191                 gctl_error(req, "unknown verb: '%s'", verb);
 192         g_topology_lock();
 193 }
 194
 195 /*
 196  * "stop" verb from userland
 197  */
 198 static void
 199 virstor_ctl_stop(struct gctl_req *req, struct g_class *cp)
 200 {
 201         int *force, *nargs;
 202         int i;
 203
 204         nargs = gctl_get_paraml(req, "nargs", sizeof *nargs);
 205         if (nargs == NULL) {
 206                 gctl_error(req, "Error fetching argument '%s'", "nargs");
 207                 return;
 208         }
 209         if (*nargs < 1) {
 210                 gctl_error(req, "Invalid number of arguments");
 211                 return;
 212         }
 213         force = gctl_get_paraml(req, "force", sizeof *force);
 214         if (force == NULL) {
 215                 gctl_error(req, "Error fetching argument '%s'", "force");
 216                 return;
 217         }
 218
 219         g_topology_lock();
 220         for (i = 0; i < *nargs; i++) {
 221                 char param[8];
 222                 const char *name;
 223                 struct g_virstor_softc *sc;
 224                 int error;
 225
 226                 sprintf(param, "arg%d", i);
 227                 name = gctl_get_asciiparam(req, param);
 228                 if (name == NULL) {
 229                         gctl_error(req, "No 'arg%d' argument", i);
 230                         g_topology_unlock();
 231                         return;
 232                 }
 233                 sc = virstor_find_geom(cp, name);
 234                 if (sc == NULL) {
 235                         gctl_error(req, "Don't know anything about '%s'", name);
 236                         g_topology_unlock();
 237                         return;
 238                 }
 239
 240                 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
 241                     sc->geom->name);
 242                 update_metadata(sc);
 243                 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
 244                         LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
 245                             sc->geom->name, error);
 246                 }
 247         }
 248         g_topology_unlock();
 249 }
 250
 251 /*
 252  * "add" verb from userland - add new component(s) to the structure.
 253  * This will be done all at once in here, without going through the
 254  * .taste function for new components.
 255  */
 256 static void
 257 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
 258 {
 259         /* Note: while this is going on, I/O is being done on
 260          * the g_up and g_down threads. The idea is to make changes
 261          * to softc members in a way that can atomically activate
 262          * them all at once. */
 263         struct g_virstor_softc *sc;
 264         int *hardcode, *nargs;
 265         const char *geom_name;  /* geom to add a component to */
 266         struct g_consumer *fcp;
 267         struct g_virstor_bio_q *bq;
 268         u_int added;
 269         int error;
 270         int i;
 271
 272         nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
 273         if (nargs == NULL) {
 274                 gctl_error(req, "Error fetching argument '%s'", "nargs");
 275                 return;
 276         }
 277         if (*nargs < 2) {
 278                 gctl_error(req, "Invalid number of arguments");
 279                 return;
 280         }
 281         hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
 282         if (hardcode == NULL) {
 283                 gctl_error(req, "Error fetching argument '%s'", "hardcode");
 284                 return;
 285         }
 286
 287         /* Find "our" geom */
 288         geom_name = gctl_get_asciiparam(req, "arg0");
 289         if (geom_name == NULL) {
 290                 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
 291                 return;
 292         }
 293         sc = virstor_find_geom(cp, geom_name);
 294         if (sc == NULL) {
 295                 gctl_error(req, "Don't know anything about '%s'", geom_name);
 296                 return;
 297         }
 298
 299         if (virstor_valid_components(sc) != sc->n_components) {
 300                 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
 301                     "virstor %s", sc->geom->name);
 302                 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
 303                 return;
 304         }
 305
 306         fcp = sc->components[0].gcons;
 307         added = 0;
 308         g_topology_lock();
 309         for (i = 1; i < *nargs; i++) {
 310                 struct g_virstor_metadata md;
 311                 char aname[8];
 312                 const char *prov_name;
 313                 struct g_provider *pp;
 314                 struct g_consumer *cp;
 315                 u_int nc;
 316                 u_int j;
 317
 318                 snprintf(aname, sizeof aname, "arg%d", i);
 319                 prov_name = gctl_get_asciiparam(req, aname);
 320                 if (prov_name == NULL) {
 321                         gctl_error(req, "Error fetching argument '%s'", aname);
 322                         g_topology_unlock();
 323                         return;
 324                 }
 325                 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
 326                         prov_name += sizeof(_PATH_DEV) - 1;
 327
 328                 pp = g_provider_by_name(prov_name);
 329                 if (pp == NULL) {
 330                         /* This is the most common error so be verbose about it */
 331                         if (added != 0) {
 332                                 gctl_error(req, "Invalid provider: '%s' (added"
 333                                     " %u components)", prov_name, added);
 334                                 update_metadata(sc);
 335                         } else {
 336                                 gctl_error(req, "Invalid provider: '%s'",
 337                                     prov_name);
 338                         }
 339                         g_topology_unlock();
 340                         return;
 341                 }
 342                 cp = g_new_consumer(sc->geom);
 343                 if (cp == NULL) {
 344                         gctl_error(req, "Cannot create consumer");
 345                         g_topology_unlock();
 346                         return;
 347                 }
 348                 error = g_attach(cp, pp);
 349                 if (error != 0) {
 350                         gctl_error(req, "Cannot attach a consumer to %s",
 351                             pp->name);
 352                         g_destroy_consumer(cp);
 353                         g_topology_unlock();
 354                         return;
 355                 }
 356                 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
 357                         error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
 358                         if (error != 0) {
 359                                 gctl_error(req, "Access request failed for %s",
 360                                     pp->name);
 361                                 g_destroy_consumer(cp);
 362                                 g_topology_unlock();
 363                                 return;
 364                         }
 365                 }
 366                 if (fcp->provider->sectorsize != pp->sectorsize) {
 367                         gctl_error(req, "Sector size doesn't fit for %s",
 368                             pp->name);
 369                         g_destroy_consumer(cp);
 370                         g_topology_unlock();
 371                         return;
 372                 }
 373                 for (j = 0; j < sc->n_components; j++) {
 374                         if (strcmp(sc->components[j].gcons->provider->name,
 375                             pp->name) == 0) {
 376                                 gctl_error(req, "Component %s already in %s",
 377                                     pp->name, sc->geom->name);
 378                                 g_destroy_consumer(cp);
 379                                 g_topology_unlock();
 380                                 return;
 381                         }
 382                 }
 383                 sc->components = realloc(sc->components,
 384                     sizeof(*sc->components) * (sc->n_components + 1),
 385                     M_GVIRSTOR, M_WAITOK);
 386
 387                 nc = sc->n_components;
 388                 sc->components[nc].gcons = cp;
 389                 sc->components[nc].sc = sc;
 390                 sc->components[nc].index = nc;
 391                 sc->components[nc].chunk_count = cp->provider->mediasize /
 392                     sc->chunk_size;
 393                 sc->components[nc].chunk_next = 0;
 394                 sc->components[nc].chunk_reserved = 0;
 395
 396                 if (sc->components[nc].chunk_count < 4) {
 397                         gctl_error(req, "Provider too small: %s",
 398                             cp->provider->name);
 399                         g_destroy_consumer(cp);
 400                         g_topology_unlock();
 401                         return;
 402                 }
 403                 fill_metadata(sc, &md, nc, *hardcode);
 404                 write_metadata(cp, &md);
 405                 /* The new component becomes visible when n_components is
 406                  * incremented */
 407                 sc->n_components++;
 408                 added++;
 409
 410         }
 411         /* This call to update_metadata() is critical. In case there's a
 412          * power failure in the middle of it and some components are updated
 413          * while others are not, there will be trouble on next .taste() iff
 414          * a non-updated component is detected first */
 415         update_metadata(sc);
 416         g_topology_unlock();
 417         LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
 418             sc->geom->name);
 419         /* Fire off BIOs previously queued because there wasn't any
 420          * physical space left. If the BIOs still can't be satisfied
 421          * they will again be added to the end of the queue (during
 422          * which the mutex will be recursed) */
 423         bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
 424         bq->bio = NULL;
 425         mtx_lock(&sc->delayed_bio_q_mtx);
 426         /* First, insert a sentinel to the queue end, so we don't
 427          * end up in an infinite loop if there's still no free
 428          * space available. */
 429         STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
 430         while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
 431                 bq = STAILQ_FIRST(&sc->delayed_bio_q);
 432                 if (bq->bio != NULL) {
 433                         g_virstor_start(bq->bio);
 434                         STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
 435                         free(bq, M_GVIRSTOR);
 436                 } else {
 437                         STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
 438                         free(bq, M_GVIRSTOR);
 439                         break;
 440                 }
 441         }
 442         mtx_unlock(&sc->delayed_bio_q_mtx);
 443
 444 }
 445
 446 /*
 447  * Find a geom handled by the class
 448  */
 449 static struct g_virstor_softc *
 450 virstor_find_geom(const struct g_class *cp, const char *name)
 451 {
 452         struct g_geom *gp;
 453
 454         LIST_FOREACH(gp, &cp->geom, geom) {
 455                 if (strcmp(name, gp->name) == 0)
 456                         return (gp->softc);
 457         }
 458         return (NULL);
 459 }
 460
 461 /*
 462  * Update metadata on all components to reflect the current state
 463  * of these fields:
 464  *    - chunk_next
 465  *    - flags
 466  *    - md_count
 467  * Expects things to be set up so write_metadata() can work, i.e.
 468  * the topology lock must be held.
 469  */
 470 static void
 471 update_metadata(struct g_virstor_softc *sc)
 472 {
 473         struct g_virstor_metadata md;
 474         u_int n;
 475
 476         if (virstor_valid_components(sc) != sc->n_components)
 477                 return; /* Incomplete device */
 478         LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
 479             sc->geom->name);
 480         /* Update metadata on components */
 481         g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
 482             sc->geom->class->name, sc->geom->name);
 483         g_topology_assert();
 484         for (n = 0; n < sc->n_components; n++) {
 485                 read_metadata(sc->components[n].gcons, &md);
 486                 md.chunk_next = sc->components[n].chunk_next;
 487                 md.flags = sc->components[n].flags;
 488                 md.md_count = sc->n_components;
 489                 write_metadata(sc->components[n].gcons, &md);
 490         }
 491 }
 492
 493 /*
 494  * Fills metadata (struct md) from information stored in softc and the nc'th
 495  * component of virstor
 496  */
 497 static void
 498 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
 499     u_int nc, u_int hardcode)
 500 {
 501         struct g_virstor_component *c;
 502
 503         bzero(md, sizeof *md);
 504         c = &sc->components[nc];
 505
 506         strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
 507         md->md_version = G_VIRSTOR_VERSION;
 508         strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
 509         md->md_id = sc->id;
 510         md->md_virsize = sc->virsize;
 511         md->md_chunk_size = sc->chunk_size;
 512         md->md_count = sc->n_components;
 513
 514         if (hardcode) {
 515                 strncpy(md->provider, c->gcons->provider->name,
 516                     sizeof md->provider);
 517         }
 518         md->no = nc;
 519         md->provsize = c->gcons->provider->mediasize;
 520         md->chunk_count = c->chunk_count;
 521         md->chunk_next = c->chunk_next;
 522         md->chunk_reserved = c->chunk_reserved;
 523         md->flags = c->flags;
 524 }
 525
 526 /*
 527  * Remove a component from virstor device.
 528  * Can only be done if the component is unallocated.
 529  */
 530 static void
 531 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
 532 {
 533         /* As this is executed in parallel to I/O, operations on virstor
 534          * structures must be as atomic as possible. */
 535         struct g_virstor_softc *sc;
 536         int *nargs;
 537         const char *geom_name;
 538         u_int removed;
 539         int i;
 540
 541         nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
 542         if (nargs == NULL) {
 543                 gctl_error(req, "Error fetching argument '%s'", "nargs");
 544                 return;
 545         }
 546         if (*nargs < 2) {
 547                 gctl_error(req, "Invalid number of arguments");
 548                 return;
 549         }
 550         /* Find "our" geom */
 551         geom_name = gctl_get_asciiparam(req, "arg0");
 552         if (geom_name == NULL) {
 553                 gctl_error(req, "Error fetching argument '%s'",
 554                     "geom_name (arg0)");
 555                 return;
 556         }
 557         sc = virstor_find_geom(cp, geom_name);
 558         if (sc == NULL) {
 559                 gctl_error(req, "Don't know anything about '%s'", geom_name);
 560                 return;
 561         }
 562
 563         if (virstor_valid_components(sc) != sc->n_components) {
 564                 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
 565                     "virstor %s", sc->geom->name);
 566                 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
 567                 return;
 568         }
 569
 570         removed = 0;
 571         for (i = 1; i < *nargs; i++) {
 572                 char param[8];
 573                 const char *prov_name;
 574                 int j, found;
 575                 struct g_virstor_component *newcomp, *compbak;
 576
 577                 sprintf(param, "arg%d", i);
 578                 prov_name = gctl_get_asciiparam(req, param);
 579                 if (prov_name == NULL) {
 580                         gctl_error(req, "Error fetching argument '%s'", param);
 581                         return;
 582                 }
 583                 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
 584                         prov_name += sizeof(_PATH_DEV) - 1;
 585
 586                 found = -1;
 587                 for (j = 0; j < sc->n_components; j++) {
 588                         if (strcmp(sc->components[j].gcons->provider->name,
 589                             prov_name) == 0) {
 590                                 found = j;
 591                                 break;
 592                         }
 593                 }
 594                 if (found == -1) {
 595                         LOG_MSG(LVL_ERROR, "No %s component in %s",
 596                             prov_name, sc->geom->name);
 597                         continue;
 598                 }
 599
 600                 compbak = sc->components;
 601                 newcomp = malloc(sc->n_components * sizeof(*sc->components),
 602                     M_GVIRSTOR, M_WAITOK | M_ZERO);
 603                 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
 604                 bcopy(&sc->components[found + 1], newcomp + found,
 605                     found * sizeof(*sc->components));
 606                 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
 607                         LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
 608                             "removed from %s",
 609                             prov_name, sc->geom->name);
 610                         free(newcomp, M_GVIRSTOR);
 611                         /* We'll consider this non-fatal error */
 612                         continue;
 613                 }
 614                 /* Renumerate unallocated components */
 615                 for (j = 0; j < sc->n_components-1; j++) {
 616                         if ((sc->components[j].flags &
 617                             VIRSTOR_PROVIDER_ALLOCATED) == 0) {
 618                                 sc->components[j].index = j;
 619                         }
 620                 }
 621                 /* This is the critical section. If a component allocation
 622                  * event happens while both variables are not yet set,
 623                  * there will be trouble. Something will panic on encountering
 624                  * NULL sc->components[x].gcomp member.
 625                  * Luckily, component allocation happens very rarely and
 626                  * removing components is an abnormal action in any case. */
 627                 sc->components = newcomp;
 628                 sc->n_components--;
 629                 /* End critical section */
 630
 631                 g_topology_lock();
 632                 if (clear_metadata(&compbak[found]) != 0) {
 633                         LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
 634                             "metadata on %s", prov_name);
 635                 }
 636                 g_detach(compbak[found].gcons);
 637                 g_destroy_consumer(compbak[found].gcons);
 638                 g_topology_unlock();
 639
 640                 free(compbak, M_GVIRSTOR);
 641
 642                 removed++;
 643         }
 644
 645         /* This call to update_metadata() is critical. In case there's a
 646          * power failure in the middle of it and some components are updated
 647          * while others are not, there will be trouble on next .taste() iff
 648          * a non-updated component is detected first */
 649         g_topology_lock();
 650         update_metadata(sc);
 651         g_topology_unlock();
 652         LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
 653             sc->geom->name);
 654 }
 655
 656 /*
 657  * Clear metadata sector on component
 658  */
 659 static int
 660 clear_metadata(struct g_virstor_component *comp)
 661 {
 662         char *buf;
 663         int error;
 664
 665         LOG_MSG(LVL_INFO, "Clearing metadata on %s",
 666             comp->gcons->provider->name);
 667         g_topology_assert();
 668         error = g_access(comp->gcons, 0, 1, 0);
 669         if (error != 0)
 670                 return (error);
 671         buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
 672             M_WAITOK | M_ZERO);
 673         error = g_write_data(comp->gcons,
 674             comp->gcons->provider->mediasize -
 675             comp->gcons->provider->sectorsize,
 676             buf,
 677             comp->gcons->provider->sectorsize);
 678         free(buf, M_GVIRSTOR);
 679         g_access(comp->gcons, 0, -1, 0);
 680         return (error);
 681 }
 682
 683 /*
 684  * Destroy geom forcibly.
 685  */
 686 static int
 687 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
 688     struct g_geom *gp)
 689 {
 690         struct g_virstor_softc *sc;
 691         int exitval;
 692
 693         sc = gp->softc;
 694         KASSERT(sc != NULL, ("%s: NULL sc", __func__));
 695
 696         exitval = 0;
 697         LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
 698             gp->softc);
 699
 700         if (sc != NULL) {
 701 #ifdef INVARIANTS
 702                 char *buf;
 703                 int error;
 704                 off_t off;
 705                 int isclean, count;
 706                 int n;
 707
 708                 LOG_MSG(LVL_INFO, "INVARIANTS detected");
 709                 LOG_MSG(LVL_INFO, "Verifying allocation "
 710                     "table for %s", sc->geom->name);
 711                 count = 0;
 712                 for (n = 0; n < sc->chunk_count; n++) {
 713                         if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
 714                                 count++;
 715                 }
 716                 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
 717                     sc->geom->name, count);
 718                 n = off = count = 0;
 719                 isclean = 1;
 720                 if (virstor_valid_components(sc) != sc->n_components) {
 721                         /* This is a incomplete virstor device (not all
 722                          * components have been found) */
 723                         LOG_MSG(LVL_ERROR, "Device %s is incomplete",
 724                             sc->geom->name);
 725                         goto bailout;
 726                 }
 727                 error = g_access(sc->components[0].gcons, 1, 0, 0);
 728                 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
 729                     error));
 730                 /* Compare the whole on-disk allocation table with what's
 731                  * currently in memory */
 732                 while (n < sc->chunk_count) {
 733                         buf = g_read_data(sc->components[0].gcons, off,
 734                             sc->sectorsize, &error);
 735                         KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
 736                             "for read at %jd", error, off));
 737                         if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
 738                                 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
 739                                     "entry %d, offset %jd", n, off);
 740                                 isclean = 0;
 741                                 count++;
 742                         }
 743                         n += sc->me_per_sector;
 744                         off += sc->sectorsize;
 745                         g_free(buf);
 746                 }
 747                 error = g_access(sc->components[0].gcons, -1, 0, 0);
 748                 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
 749                     __func__, error));
 750                 if (isclean != 1) {
 751                         LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
 752                             "(%d sectors don't match, max %zu allocations)",
 753                             sc->geom->name, count,
 754                             count * sc->me_per_sector);
 755                 } else {
 756                         LOG_MSG(LVL_INFO, "Allocation table ok for %s",
 757                             sc->geom->name);
 758                 }
 759 bailout:
 760 #endif
 761                 update_metadata(sc);
 762                 virstor_geom_destroy(sc, FALSE, FALSE);
 763                 exitval = EAGAIN;
 764         } else
 765                 exitval = 0;
 766         return (exitval);
 767 }
 768
 769 /*
 770  * Taste event (per-class callback)
 771  * Examines a provider and creates geom instances if needed
 772  */
 773 static struct g_geom *
 774 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
 775 {
 776         struct g_virstor_metadata md;
 777         struct g_geom *gp;
 778         struct g_consumer *cp;
 779         struct g_virstor_softc *sc;
 780         int error;
 781
 782         g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
 783         g_topology_assert();
 784         LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
 785
 786         /* We need a dummy geom to attach a consumer to the given provider */
 787         gp = g_new_geomf(mp, "virstor:taste.helper");
 788         gp->start = (void *)invalid_call;       /* XXX: hacked up so the        */
 789         gp->access = (void *)invalid_call;      /* compiler doesn't complain.   */
 790         gp->orphan = (void *)invalid_call;      /* I really want these to fail. */
 791
 792         cp = g_new_consumer(gp);
 793         g_attach(cp, pp);
 794         error = read_metadata(cp, &md);
 795         g_detach(cp);
 796         g_destroy_consumer(cp);
 797         g_destroy_geom(gp);
 798
 799         if (error != 0)
 800                 return (NULL);
 801
 802         if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
 803                 return (NULL);
 804         if (md.md_version != G_VIRSTOR_VERSION) {
 805                 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
 806                     "to handle %s (%s) : %d should be %d",
 807                     md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
 808                 return (NULL);
 809         }
 810         if (md.provsize != pp->mediasize)
 811                 return (NULL);
 812
 813         /* If the provider name is hardcoded, use the offered provider only
 814          * if it's been offered with its proper name (the one used in
 815          * the label command). */
 816         if (md.provider[0] != '\0' &&
 817             !g_compare_names(md.provider, pp->name))
 818                 return (NULL);
 819
 820         /* Iterate all geoms this class already knows about to see if a new
 821          * geom instance of this class needs to be created (in case the provider
 822          * is first from a (possibly) multi-consumer geom) or it just needs
 823          * to be added to an existing instance. */
 824         sc = NULL;
 825         gp = NULL;
 826         LIST_FOREACH(gp, &mp->geom, geom) {
 827                 sc = gp->softc;
 828                 if (sc == NULL)
 829                         continue;
 830                 if (strcmp(md.md_name, sc->geom->name) != 0)
 831                         continue;
 832                 if (md.md_id != sc->id)
 833                         continue;
 834                 break;
 835         }
 836         if (gp != NULL) { /* We found an existing geom instance; add to it */
 837                 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
 838                 error = add_provider_to_geom(sc, pp, &md);
 839                 if (error != 0) {
 840                         LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
 841                             pp->name, md.md_name, error);
 842                         return (NULL);
 843                 }
 844         } else { /* New geom instance needs to be created */
 845                 gp = create_virstor_geom(mp, &md);
 846                 if (gp == NULL) {
 847                         LOG_MSG(LVL_ERROR, "Error creating new instance of "
 848                             "class %s: %s", mp->name, md.md_name);
 849                         LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
 850                             md.md_name, pp->name);
 851                         return (NULL);
 852                 }
 853                 sc = gp->softc;
 854                 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
 855                     md.md_name);
 856                 error = add_provider_to_geom(sc, pp, &md);
 857                 if (error != 0) {
 858                         LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
 859                             pp->name, md.md_name, error);
 860                         virstor_geom_destroy(sc, TRUE, FALSE);
 861                         return (NULL);
 862                 }
 863         }
 864
 865         return (gp);
 866 }
 867
 868 /*
 869  * Destroyes consumer passed to it in arguments. Used as a callback
 870  * on g_event queue.
 871  */
 872 static void
 873 delay_destroy_consumer(void *arg, int flags __unused)
 874 {
 875         struct g_consumer *c = arg;
 876         KASSERT(c != NULL, ("%s: invalid consumer", __func__));
 877         LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
 878             c->provider->name);
 879         g_detach(c);
 880         g_destroy_consumer(c);
 881 }
 882
 883 /*
 884  * Remove a component (consumer) from geom instance; If it's the first
 885  * component being removed, orphan the provider to announce geom's being
 886  * dismantled
 887  */
 888 static void
 889 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
 890     boolean_t delay)
 891 {
 892         struct g_consumer *c;
 893
 894         KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
 895             sc->geom->name));
 896         c = comp->gcons;
 897
 898         comp->gcons = NULL;
 899         KASSERT(c->provider != NULL, ("%s: no provider", __func__));
 900         LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
 901             sc->geom->name);
 902         if (sc->provider != NULL) {
 903                 LOG_MSG(LVL_INFO, "Removing provider %s", sc->provider->name);
 904                 g_wither_provider(sc->provider, ENXIO);
 905                 sc->provider = NULL;
 906         }
 907
 908         if (c->acr > 0 || c->acw > 0 || c->ace > 0)
 909                 g_access(c, -c->acr, -c->acw, -c->ace);
 910         if (delay) {
 911                 /* Destroy consumer after it's tasted */
 912                 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
 913         } else {
 914                 g_detach(c);
 915                 g_destroy_consumer(c);
 916         }
 917 }
 918
 919 /*
 920  * Destroy geom - called internally
 921  * See g_virstor_destroy_geom for the other one
 922  */
 923 static int
 924 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
 925     boolean_t delay)
 926 {
 927         struct g_provider *pp;
 928         struct g_geom *gp;
 929         u_int n;
 930
 931         g_topology_assert();
 932
 933         if (sc == NULL)
 934                 return (ENXIO);
 935
 936         pp = sc->provider;
 937         if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
 938                 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
 939                     "Device %s is still open.", pp->name);
 940                 if (!force)
 941                         return (EBUSY);
 942         }
 943
 944         for (n = 0; n < sc->n_components; n++) {
 945                 if (sc->components[n].gcons != NULL)
 946                         remove_component(sc, &sc->components[n], delay);
 947         }
 948
 949         gp = sc->geom;
 950         gp->softc = NULL;
 951
 952         KASSERT(sc->provider == NULL, ("Provider still exists for %s",
 953             gp->name));
 954
 955         /* XXX: This might or might not work, since we're called with
 956          * the topology lock held. Also, it might panic the kernel if
 957          * the error'd BIO is in softupdates code. */
 958         mtx_lock(&sc->delayed_bio_q_mtx);
 959         while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
 960                 struct g_virstor_bio_q *bq;
 961                 bq = STAILQ_FIRST(&sc->delayed_bio_q);
 962                 bq->bio->bio_error = ENOSPC;
 963                 g_io_deliver(bq->bio, EIO);
 964                 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
 965                 free(bq, M_GVIRSTOR);
 966         }
 967         mtx_unlock(&sc->delayed_bio_q_mtx);
 968         mtx_destroy(&sc->delayed_bio_q_mtx);
 969
 970         free(sc->map, M_GVIRSTOR);
 971         free(sc->components, M_GVIRSTOR);
 972         bzero(sc, sizeof *sc);
 973         free(sc, M_GVIRSTOR);
 974
 975         pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
 976         if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
 977                 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
 978
 979         g_wither_geom(gp, ENXIO);
 980
 981         return (0);
 982 }
 983
 984 /*
 985  * Utility function: read metadata & decode. Wants topology lock to be
 986  * held.
 987  */
 988 static int
 989 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
 990 {
 991         struct g_provider *pp;
 992         char *buf;
 993         int error;
 994
 995         g_topology_assert();
 996         error = g_access(cp, 1, 0, 0);
 997         if (error != 0)
 998                 return (error);
 999         pp = cp->provider;
1000         g_topology_unlock();
1001         buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
1002             &error);
1003         g_topology_lock();
1004         g_access(cp, -1, 0, 0);
1005         if (buf == NULL)
1006                 return (error);
1007
1008         virstor_metadata_decode(buf, md);
1009         g_free(buf);
1010
1011         return (0);
1012 }
1013
1014 /**
1015  * Utility function: encode & write metadata. Assumes topology lock is
1016  * held.
1017  *
1018  * There is no useful way of recovering from errors in this function,
1019  * not involving panicking the kernel. If the metadata cannot be written
1020  * the most we can do is notify the operator and hope he spots it and
1021  * replaces the broken drive.
1022  */
1023 static void
1024 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1025 {
1026         struct g_provider *pp;
1027         char *buf;
1028         int error;
1029
1030         KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1031             ("Something's fishy in %s", __func__));
1032         LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1033         g_topology_assert();
1034         error = g_access(cp, 0, 1, 0);
1035         if (error != 0) {
1036                 LOG_MSG(LVL_ERROR, "g_access(0,1,0) failed for %s: %d",
1037                     cp->provider->name, error);
1038                 return;
1039         }
1040         pp = cp->provider;
1041
1042         buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1043         virstor_metadata_encode(md, buf);
1044         g_topology_unlock();
1045         error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1046             pp->sectorsize);
1047         g_topology_lock();
1048         g_access(cp, 0, -1, 0);
1049         free(buf, M_GVIRSTOR);
1050
1051         if (error != 0)
1052                 LOG_MSG(LVL_ERROR, "Error %d writing metadata to %s",
1053                     error, cp->provider->name);
1054 }
1055
1056 /*
1057  * Creates a new instance of this GEOM class, initialise softc
1058  */
1059 static struct g_geom *
1060 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1061 {
1062         struct g_geom *gp;
1063         struct g_virstor_softc *sc;
1064
1065         LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1066             md->md_name, md->md_id);
1067
1068         if (md->md_count < 1 || md->md_chunk_size < 1 ||
1069             md->md_virsize < md->md_chunk_size) {
1070                 /* This is bogus configuration, and probably means data is
1071                  * somehow corrupted. Panic, maybe? */
1072                 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1073                     md->md_name);
1074                 return (NULL);
1075         }
1076
1077         /* Check if it's already created */
1078         LIST_FOREACH(gp, &mp->geom, geom) {
1079                 sc = gp->softc;
1080                 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1081                         LOG_MSG(LVL_WARNING, "Geom %s already exists",
1082                             md->md_name);
1083                         if (sc->id != md->md_id) {
1084                                 LOG_MSG(LVL_ERROR,
1085                                     "Some stale or invalid components "
1086                                     "exist for virstor device named %s. "
1087                                     "You will need to <CLEAR> all stale "
1088                                     "components and maybe reconfigure "
1089                                     "the virstor device. Tune "
1090                                     "kern.geom.virstor.debug sysctl up "
1091                                     "for more information.",
1092                                     sc->geom->name);
1093                         }
1094                         return (NULL);
1095                 }
1096         }
1097         gp = g_new_geomf(mp, "%s", md->md_name);
1098         gp->softc = NULL; /* to circumevent races that test softc */
1099
1100         gp->start = g_virstor_start;
1101         gp->spoiled = g_virstor_orphan;
1102         gp->orphan = g_virstor_orphan;
1103         gp->access = g_virstor_access;
1104         gp->dumpconf = g_virstor_dumpconf;
1105
1106         sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1107         sc->id = md->md_id;
1108         sc->n_components = md->md_count;
1109         sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1110             M_GVIRSTOR, M_WAITOK | M_ZERO);
1111         sc->chunk_size = md->md_chunk_size;
1112         sc->virsize = md->md_virsize;
1113         STAILQ_INIT(&sc->delayed_bio_q);
1114         mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1115             "gvirstor", MTX_DEF | MTX_RECURSE);
1116
1117         sc->geom = gp;
1118         sc->provider = NULL; /* virstor_check_and_run will create it */
1119         gp->softc = sc;
1120
1121         LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1122
1123         return (gp);
1124 }
1125
1126 /*
1127  * Add provider to a GEOM class instance
1128  */
1129 static int
1130 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1131     struct g_virstor_metadata *md)
1132 {
1133         struct g_virstor_component *component;
1134         struct g_consumer *cp, *fcp;
1135         struct g_geom *gp;
1136         int error;
1137
1138         if (md->no >= sc->n_components)
1139                 return (EINVAL);
1140
1141         /* "Current" compontent */
1142         component = &(sc->components[md->no]);
1143         if (component->gcons != NULL)
1144                 return (EEXIST);
1145
1146         gp = sc->geom;
1147         fcp = LIST_FIRST(&gp->consumer);
1148
1149         cp = g_new_consumer(gp);
1150         error = g_attach(cp, pp);
1151
1152         if (error != 0) {
1153                 g_destroy_consumer(cp);
1154                 return (error);
1155         }
1156
1157         if (fcp != NULL) {
1158                 if (fcp->provider->sectorsize != pp->sectorsize) {
1159                         /* TODO: this can be made to work */
1160                         LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1161                             "sector size (%d)", pp->name, sc->geom->name,
1162                             pp->sectorsize);
1163                         return (EINVAL);
1164                 }
1165                 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1166                         /* Replicate access permissions from first "live" consumer
1167                          * to the new one */
1168                         error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1169                         if (error != 0) {
1170                                 g_detach(cp);
1171                                 g_destroy_consumer(cp);
1172                                 return (error);
1173                         }
1174                 }
1175         }
1176
1177         /* Bring up a new component */
1178         cp->private = component;
1179         component->gcons = cp;
1180         component->sc = sc;
1181         component->index = md->no;
1182         component->chunk_count = md->chunk_count;
1183         component->chunk_next = md->chunk_next;
1184         component->chunk_reserved = md->chunk_reserved;
1185         component->flags = md->flags;
1186
1187         LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1188
1189         virstor_check_and_run(sc);
1190         return (0);
1191 }
1192
1193 /*
1194  * Check if everything's ready to create the geom provider & device entry,
1195  * create and start provider.
1196  * Called ultimately by .taste, from g_event thread
1197  */
1198 static void
1199 virstor_check_and_run(struct g_virstor_softc *sc)
1200 {
1201         off_t off;
1202         size_t n, count;
1203         int index;
1204         int error;
1205
1206         if (virstor_valid_components(sc) != sc->n_components)
1207                 return;
1208
1209         if (virstor_valid_components(sc) == 0) {
1210                 /* This is actually a candidate for panic() */
1211                 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1212                     sc->provider->name);
1213                 return;
1214         }
1215
1216         sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1217
1218         /* Initialise allocation map from the first consumer */
1219         sc->chunk_count = sc->virsize / sc->chunk_size;
1220         if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1221                 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1222                     sc->provider->name,
1223                     sc->chunk_count * (off_t)sc->chunk_size);
1224         }
1225         sc->map_size = sc->chunk_count * sizeof *(sc->map);
1226         /* The following allocation is in order of 4MB - 8MB */
1227         sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1228         KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1229             __func__, sc->map_size, sc->provider->name));
1230         sc->map_sectors = sc->map_size / sc->sectorsize;
1231
1232         count = 0;
1233         for (n = 0; n < sc->n_components; n++)
1234                 count += sc->components[n].chunk_count;
1235         LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1236             "(%zu KB chunks)",
1237             sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1238
1239         error = g_access(sc->components[0].gcons, 1, 0, 0);
1240         if (error != 0) {
1241                 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1242                     "read allocation map for %s",
1243                     sc->components[0].gcons->provider->name,
1244                     sc->geom->name);
1245                 return;
1246         }
1247         /* Read in the allocation map */
1248         LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1249             sc->components[0].gcons->provider->name);
1250         off = count = n = 0;
1251         while (count < sc->map_size) {
1252                 struct g_virstor_map_entry *mapbuf;
1253                 size_t bs;
1254
1255                 bs = MIN(MAXPHYS, sc->map_size - count);
1256                 if (bs % sc->sectorsize != 0) {
1257                         /* Check for alignment errors */
1258                         bs = rounddown(bs, sc->sectorsize);
1259                         if (bs == 0)
1260                                 break;
1261                         LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1262                             "for %s on %s", sc->geom->name,
1263                             sc->components[0].gcons->provider->name);
1264                 }
1265                 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1266                 if (mapbuf == NULL) {
1267                         free(sc->map, M_GVIRSTOR);
1268                         LOG_MSG(LVL_ERROR, "Error reading allocation map "
1269                             "for %s from %s (offset %ju) (error %d)",
1270                             sc->geom->name,
1271                             sc->components[0].gcons->provider->name,
1272                             off, error);
1273                         return;
1274                 }
1275
1276                 bcopy(mapbuf, &sc->map[n], bs);
1277                 off += bs;
1278                 count += bs;
1279                 n += bs / sizeof *(sc->map);
1280                 g_free(mapbuf);
1281         }
1282         g_access(sc->components[0].gcons, -1, 0, 0);
1283         LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1284
1285         /* find first component with allocatable chunks */
1286         index = -1;
1287         for (n = 0; n < sc->n_components; n++) {
1288                 if (sc->components[n].chunk_next <
1289                     sc->components[n].chunk_count) {
1290                         index = n;
1291                         break;
1292                 }
1293         }
1294         if (index == -1)
1295                 /* not found? set it to the last component and handle it
1296                  * later */
1297                 index = sc->n_components - 1;
1298
1299         if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1300                 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1301                     "(%d/%u: %s)", sc->geom->name,
1302                     index+1,
1303                     sc->n_components,
1304                     sc->components[index].gcons->provider->name);
1305         }
1306         sc->curr_component = index;
1307
1308         if (sc->components[index].chunk_next >=
1309             sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1310                 LOG_MSG(LVL_WARNING,
1311                     "Component %s of %s is running out of free space "
1312                     "(%u chunks left)",
1313                     sc->components[index].gcons->provider->name,
1314                     sc->geom->name, sc->components[index].chunk_count -
1315                     sc->components[index].chunk_next);
1316         }
1317
1318         sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1319         if (sc->sectorsize % sizeof *(sc->map) != 0) {
1320                 LOG_MSG(LVL_ERROR,
1321                     "%s: Map entries don't fit exactly in a sector (%s)",
1322                     __func__, sc->geom->name);
1323                 return;
1324         }
1325
1326         /* Recalculate allocated chunks in components & at the same time
1327          * verify map data is sane. We could trust metadata on this, but
1328          * we want to make sure. */
1329         for (n = 0; n < sc->n_components; n++)
1330                 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1331
1332         for (n = 0; n < sc->chunk_count; n++) {
1333                 if (sc->map[n].provider_no >= sc->n_components ||
1334                         sc->map[n].provider_chunk >=
1335                         sc->components[sc->map[n].provider_no].chunk_count) {
1336                         LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1337                             __func__, (u_int)n, sc->geom->name);
1338                         LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1339                             " provider_chunk: %u, chunk_count: %u", __func__,
1340                             sc->map[n].provider_no, sc->n_components,
1341                             sc->map[n].provider_chunk,
1342                             sc->components[sc->map[n].provider_no].chunk_count);
1343                         return;
1344                 }
1345                 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1346                         sc->components[sc->map[n].provider_no].chunk_next++;
1347         }
1348
1349         sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1350             sc->geom->name);
1351
1352         sc->provider->sectorsize = sc->sectorsize;
1353         sc->provider->mediasize = sc->virsize;
1354         g_error_provider(sc->provider, 0);
1355
1356         LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1357         LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1358             "chunk %u", sc->provider->name, sc->curr_component,
1359             sc->components[sc->curr_component].chunk_next);
1360 }
1361
1362 /*
1363  * Returns count of active providers in this geom instance
1364  */
1365 static u_int
1366 virstor_valid_components(struct g_virstor_softc *sc)
1367 {
1368         unsigned int nc, i;
1369
1370         nc = 0;
1371         KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1372         KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1373         for (i = 0; i < sc->n_components; i++)
1374                 if (sc->components[i].gcons != NULL)
1375                         nc++;
1376         return (nc);
1377 }
1378
1379 /*
1380  * Called when the consumer gets orphaned (?)
1381  */
1382 static void
1383 g_virstor_orphan(struct g_consumer *cp)
1384 {
1385         struct g_virstor_softc *sc;
1386         struct g_virstor_component *comp;
1387         struct g_geom *gp;
1388
1389         g_topology_assert();
1390         gp = cp->geom;
1391         sc = gp->softc;
1392         if (sc == NULL)
1393                 return;
1394
1395         comp = cp->private;
1396         KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1397             __func__));
1398         remove_component(sc, comp, FALSE);
1399         if (virstor_valid_components(sc) == 0)
1400                 virstor_geom_destroy(sc, TRUE, FALSE);
1401 }
1402
1403 /*
1404  * Called to notify geom when it's been opened, and for what intent
1405  */
1406 static int
1407 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1408 {
1409         struct g_consumer *c;
1410         struct g_virstor_softc *sc;
1411         struct g_geom *gp;
1412         int error;
1413
1414         KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1415         gp = pp->geom;
1416         KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1417         sc = gp->softc;
1418
1419         if (sc == NULL) {
1420                 /* It seems that .access can be called with negative dr,dw,dx
1421                  * in this case but I want to check for myself */
1422                 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s",
1423                     dr, dw, de, pp->name);
1424                 /* This should only happen when geom is withered so
1425                  * allow only negative requests */
1426                 KASSERT(dr <= 0 && dw <= 0 && de <= 0,
1427                     ("%s: Positive access for %s", __func__, pp->name));
1428                 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
1429                         LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed",
1430                             pp->name);
1431                 return (0);
1432         }
1433
1434         /* Grab an exclusive bit to propagate on our consumers on first open */
1435         if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1436                 de++;
1437         /* ... drop it on close */
1438         if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1439                 de--;
1440                 update_metadata(sc);    /* Writes statistical information */
1441         }
1442
1443         error = ENXIO;
1444         LIST_FOREACH(c, &gp->consumer, consumer) {
1445                 KASSERT(c != NULL, ("%s: consumer is NULL", __func__));
1446                 error = g_access(c, dr, dw, de);
1447                 if (error != 0) {
1448                         struct g_consumer *c2;
1449
1450                         /* Backout earlier changes */
1451                         LIST_FOREACH(c2, &gp->consumer, consumer) {
1452                                 if (c2 == c) /* all eariler components fixed */
1453                                         return (error);
1454                                 g_access(c2, -dr, -dw, -de);
1455                         }
1456                 }
1457         }
1458
1459         return (error);
1460 }
1461
1462 /*
1463  * Generate XML dump of current state
1464  */
1465 static void
1466 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1467     struct g_consumer *cp, struct g_provider *pp)
1468 {
1469         struct g_virstor_softc *sc;
1470
1471         g_topology_assert();
1472         sc = gp->softc;
1473
1474         if (sc == NULL || pp != NULL)
1475                 return;
1476
1477         if (cp != NULL) {
1478                 /* For each component */
1479                 struct g_virstor_component *comp;
1480
1481                 comp = cp->private;
1482                 if (comp == NULL)
1483                         return;
1484                 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1485                     indent, comp->index);
1486                 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1487                     indent, comp->chunk_count);
1488                 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1489                     indent, comp->chunk_next);
1490                 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1491                     indent, comp->chunk_reserved);
1492                 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1493                     indent,
1494                     comp->chunk_next > 0 ? 100 -
1495                     ((comp->chunk_next + comp->chunk_reserved) * 100) /
1496                     comp->chunk_count : 100);
1497         } else {
1498                 /* For the whole thing */
1499                 u_int count, used, i;
1500                 off_t size;
1501
1502                 count = used = size = 0;
1503                 for (i = 0; i < sc->n_components; i++) {
1504                         if (sc->components[i].gcons != NULL) {
1505                                 count += sc->components[i].chunk_count;
1506                                 used += sc->components[i].chunk_next +
1507                                     sc->components[i].chunk_reserved;
1508                                 size += sc->components[i].gcons->
1509                                     provider->mediasize;
1510                         }
1511                 }
1512
1513                 sbuf_printf(sb, "%s<Status>"
1514                     "Components=%u, Online=%u</Status>\n", indent,
1515                     sc->n_components, virstor_valid_components(sc));
1516                 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1517                     indent, 100-(used * 100) / count);
1518                 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1519                     sc->chunk_size);
1520                 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1521                     indent, used > 0 ? 100 - (used * 100) / count : 100);
1522                 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1523                     indent, count);
1524                 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1525                     indent, sc->chunk_count);
1526                 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1527                     indent,
1528                     (count * 100) / sc->chunk_count);
1529                 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1530                     indent, size);
1531                 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1532                     sc->virsize);
1533         }
1534 }
1535
1536 /*
1537  * GEOM .done handler
1538  * Can't use standard handler because one requested IO may
1539  * fork into additional data IOs
1540  */
1541 static void
1542 g_virstor_done(struct bio *b)
1543 {
1544         struct g_virstor_softc *sc;
1545         struct bio *parent_b;
1546
1547         parent_b = b->bio_parent;
1548         sc = parent_b->bio_to->geom->softc;
1549
1550         if (b->bio_error != 0) {
1551                 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1552                     b->bio_error, b->bio_offset, b->bio_length,
1553                     b->bio_to->name);
1554                 if (parent_b->bio_error == 0)
1555                         parent_b->bio_error = b->bio_error;
1556         }
1557
1558         parent_b->bio_inbed++;
1559         parent_b->bio_completed += b->bio_completed;
1560
1561         if (parent_b->bio_children == parent_b->bio_inbed) {
1562                 parent_b->bio_completed = parent_b->bio_length;
1563                 g_io_deliver(parent_b, parent_b->bio_error);
1564         }
1565         g_destroy_bio(b);
1566 }
1567
1568 /*
1569  * I/O starts here
1570  * Called in g_down thread
1571  */
1572 static void
1573 g_virstor_start(struct bio *b)
1574 {
1575         struct g_virstor_softc *sc;
1576         struct g_virstor_component *comp;
1577         struct bio *cb;
1578         struct g_provider *pp;
1579         char *addr;
1580         off_t offset, length;
1581         struct bio_queue_head bq;
1582         size_t chunk_size;      /* cached for convenience */
1583         u_int count;
1584
1585         pp = b->bio_to;
1586         sc = pp->geom->softc;
1587         KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1588             b->bio_to->error, b->bio_to->name));
1589
1590         LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1591
1592         switch (b->bio_cmd) {
1593         case BIO_READ:
1594         case BIO_WRITE:
1595         case BIO_DELETE:
1596                 break;
1597         default:
1598                 g_io_deliver(b, EOPNOTSUPP);
1599                 return;
1600         }
1601
1602         LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1603         bioq_init(&bq);
1604
1605         chunk_size = sc->chunk_size;
1606         addr = b->bio_data;
1607         offset = b->bio_offset; /* virtual offset and length */
1608         length = b->bio_length;
1609
1610         while (length > 0) {
1611                 size_t chunk_index, in_chunk_offset, in_chunk_length;
1612                 struct virstor_map_entry *me;
1613
1614                 chunk_index = offset / chunk_size; /* round downwards */
1615                 in_chunk_offset = offset % chunk_size;
1616                 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1617                 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1618                     b->bio_cmd == BIO_READ ? "R" : "W",
1619                     offset, length,
1620                     chunk_index, in_chunk_offset, in_chunk_length);
1621                 me = &sc->map[chunk_index];
1622
1623                 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1624                         if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1625                                 /* Reads from unallocated chunks return zeroed
1626                                  * buffers */
1627                                 if (b->bio_cmd == BIO_READ)
1628                                         bzero(addr, in_chunk_length);
1629                         } else {
1630                                 comp = &sc->components[me->provider_no];
1631
1632                                 cb = g_clone_bio(b);
1633                                 if (cb == NULL) {
1634                                         bioq_dismantle(&bq);
1635                                         if (b->bio_error == 0)
1636                                                 b->bio_error = ENOMEM;
1637                                         g_io_deliver(b, b->bio_error);
1638                                         return;
1639                                 }
1640                                 cb->bio_to = comp->gcons->provider;
1641                                 cb->bio_done = g_virstor_done;
1642                                 cb->bio_offset =
1643                                     (off_t)me->provider_chunk * (off_t)chunk_size
1644                                     + in_chunk_offset;
1645                                 cb->bio_length = in_chunk_length;
1646                                 cb->bio_data = addr;
1647                                 cb->bio_caller1 = comp;
1648                                 bioq_disksort(&bq, cb);
1649                         }
1650                 } else { /* handle BIO_WRITE */
1651                         KASSERT(b->bio_cmd == BIO_WRITE,
1652                             ("%s: Unknown command %d", __func__,
1653                             b->bio_cmd));
1654
1655                         if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1656                                 /* We have a virtual chunk, represented by
1657                                  * the "me" entry, but it's not yet allocated
1658                                  * (tied to) a physical chunk. So do it now. */
1659                                 struct virstor_map_entry *data_me;
1660                                 u_int phys_chunk, comp_no;
1661                                 off_t s_offset;
1662                                 int error;
1663
1664                                 error = allocate_chunk(sc, &comp, &comp_no,
1665                                     &phys_chunk);
1666                                 if (error != 0) {
1667                                         /* We cannot allocate a physical chunk
1668                                          * to satisfy this request, so we'll
1669                                          * delay it to when we can...
1670                                          * XXX: this will prevent the fs from
1671                                          * being umounted! */
1672                                         struct g_virstor_bio_q *biq;
1673                                         biq = malloc(sizeof *biq, M_GVIRSTOR,
1674                                             M_NOWAIT);
1675                                         if (biq == NULL) {
1676                                                 bioq_dismantle(&bq);
1677                                                 if (b->bio_error == 0)
1678                                                         b->bio_error = ENOMEM;
1679                                                 g_io_deliver(b, b->bio_error);
1680                                                 return;
1681                                         }
1682                                         biq->bio = b;
1683                                         mtx_lock(&sc->delayed_bio_q_mtx);
1684                                         STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1685                                             biq, linkage);
1686                                         mtx_unlock(&sc->delayed_bio_q_mtx);
1687                                         LOG_MSG(LVL_WARNING, "Delaying BIO "
1688                                             "(size=%ju) until free physical "
1689                                             "space can be found on %s",
1690                                             b->bio_length,
1691                                             sc->provider->name);
1692                                         return;
1693                                 }
1694                                 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1695                                     "for %s",
1696                                     phys_chunk,
1697                                     comp->gcons->provider->name,
1698                                     sc->provider->name);
1699
1700                                 me->provider_no = comp_no;
1701                                 me->provider_chunk = phys_chunk;
1702                                 me->flags |= VIRSTOR_MAP_ALLOCATED;
1703
1704                                 cb = g_clone_bio(b);
1705                                 if (cb == NULL) {
1706                                         me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1707                                         me->provider_no = 0;
1708                                         me->provider_chunk = 0;
1709                                         bioq_dismantle(&bq);
1710                                         if (b->bio_error == 0)
1711                                                 b->bio_error = ENOMEM;
1712                                         g_io_deliver(b, b->bio_error);
1713                                         return;
1714                                 }
1715
1716                                 /* The allocation table is stored continuously
1717                                  * at the start of the drive. We need to
1718                                  * calculate the offset of the sector that holds
1719                                  * this map entry both on the drive and in the
1720                                  * map array.
1721                                  * sc_offset will end up pointing to the drive
1722                                  * sector. */
1723                                 s_offset = chunk_index * sizeof *me;
1724                                 s_offset = rounddown(s_offset, sc->sectorsize);
1725
1726                                 /* data_me points to map entry sector
1727                                  * in memory (analogous to offset) */
1728                                 data_me = &sc->map[rounddown(chunk_index,
1729                                     sc->me_per_sector)];
1730
1731                                 /* Commit sector with map entry to storage */
1732                                 cb->bio_to = sc->components[0].gcons->provider;
1733                                 cb->bio_done = g_virstor_done;
1734                                 cb->bio_offset = s_offset;
1735                                 cb->bio_data = (char *)data_me;
1736                                 cb->bio_length = sc->sectorsize;
1737                                 cb->bio_caller1 = &sc->components[0];
1738                                 bioq_disksort(&bq, cb);
1739                         }
1740
1741                         comp = &sc->components[me->provider_no];
1742                         cb = g_clone_bio(b);
1743                         if (cb == NULL) {
1744                                 bioq_dismantle(&bq);
1745                                 if (b->bio_error == 0)
1746                                         b->bio_error = ENOMEM;
1747                                 g_io_deliver(b, b->bio_error);
1748                                 return;
1749                         }
1750                         /* Finally, handle the data */
1751                         cb->bio_to = comp->gcons->provider;
1752                         cb->bio_done = g_virstor_done;
1753                         cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1754                             in_chunk_offset;
1755                         cb->bio_length = in_chunk_length;
1756                         cb->bio_data = addr;
1757                         cb->bio_caller1 = comp;
1758                         bioq_disksort(&bq, cb);
1759                 }
1760                 addr += in_chunk_length;
1761                 length -= in_chunk_length;
1762                 offset += in_chunk_length;
1763         }
1764
1765         /* Fire off bio's here */
1766         count = 0;
1767         for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1768                 bioq_remove(&bq, cb);
1769                 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1770                 comp = cb->bio_caller1;
1771                 cb->bio_caller1 = NULL;
1772                 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1773                     cb->bio_offset, cb->bio_length);
1774                 g_io_request(cb, comp->gcons);
1775                 count++;
1776         }
1777         if (count == 0) { /* We handled everything locally */
1778                 b->bio_completed = b->bio_length;
1779                 g_io_deliver(b, 0);
1780         }
1781
1782 }
1783
1784 /*
1785  * Allocate a chunk from a physical provider. Returns physical component,
1786  * chunk index relative to the component and the component's index.
1787  */
1788 static int
1789 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1790     u_int *comp_no_p, u_int *chunk)
1791 {
1792         u_int comp_no;
1793
1794         KASSERT(sc->curr_component < sc->n_components,
1795             ("%s: Invalid curr_component: %u",  __func__, sc->curr_component));
1796
1797         comp_no = sc->curr_component;
1798         *comp = &sc->components[comp_no];
1799         dump_component(*comp);
1800         if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1801                 /* This component is full. Allocate next component */
1802                 if (comp_no >= sc->n_components-1) {
1803                         LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1804                             sc->geom->name);
1805                         return (-1);
1806                 }
1807                 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1808                 sc->curr_component = ++comp_no;
1809
1810                 *comp = &sc->components[comp_no];
1811                 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1812                         LOG_MSG(LVL_WARNING, "Device %s running out of components "
1813                             "(switching to %u/%u: %s)", sc->geom->name,
1814                             comp_no+1, sc->n_components,
1815                             (*comp)->gcons->provider->name);
1816                 /* Take care not to overwrite reserved chunks */
1817                 if ( (*comp)->chunk_reserved > 0 &&
1818                     (*comp)->chunk_next < (*comp)->chunk_reserved)
1819                         (*comp)->chunk_next = (*comp)->chunk_reserved;
1820
1821                 (*comp)->flags |=
1822                     VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1823                 dump_component(*comp);
1824                 *comp_no_p = comp_no;
1825                 *chunk = (*comp)->chunk_next++;
1826         } else {
1827                 *comp_no_p = comp_no;
1828                 *chunk = (*comp)->chunk_next++;
1829         }
1830         return (0);
1831 }
1832
1833 /* Dump a component */
1834 static void
1835 dump_component(struct g_virstor_component *comp)
1836 {
1837
1838         if (g_virstor_debug < LVL_DEBUG2)
1839                 return;
1840         printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1841         printf("  chunk_count: %u\n", comp->chunk_count);
1842         printf("   chunk_next: %u\n", comp->chunk_next);
1843         printf("        flags: %u\n", comp->flags);
1844 }
1845
1846 #if 0
1847 /* Dump a map entry */
1848 static void
1849 dump_me(struct virstor_map_entry *me, unsigned int nr)
1850 {
1851         if (g_virstor_debug < LVL_DEBUG)
1852                 return;
1853         printf("VIRT. CHUNK #%d: ", nr);
1854         if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1855                 printf("(unallocated)\n");
1856         else
1857                 printf("allocated at provider %u, provider_chunk %u\n",
1858                     me->provider_no, me->provider_chunk);
1859 }
1860 #endif
1861
1862 /*
1863  * Dismantle bio_queue and destroy its components
1864  */
1865 static void
1866 bioq_dismantle(struct bio_queue_head *bq)
1867 {
1868         struct bio *b;
1869
1870         for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1871                 bioq_remove(bq, b);
1872                 g_destroy_bio(b);
1873         }
1874 }
1875
1876 /*
1877  * The function that shouldn't be called.
1878  * When this is called, the stack is already garbled because of
1879  * argument mismatch. There's nothing to do now but panic, which is
1880  * accidentally the whole purpose of this function.
1881  * Motivation: to guard from accidentally calling geom methods when
1882  * they shouldn't be called. (see g_..._taste)
1883  */
1884 static void
1885 invalid_call(void)
1886 {
1887         panic("invalid_call() has just been called. Something's fishy here.");
1888 }
1889
1890 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */