g_virstor.c revision 330737
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: stable/10/sys/geom/virstor/g_virstor.c 330737 2018-03-10 04:17:01Z asomers $"); 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 56FEATURE(g_virstor, "GEOM virtual storage support"); 57 58/* Declare malloc(9) label */ 59static MALLOC_DEFINE(M_GVIRSTOR, "gvirstor", "GEOM_VIRSTOR Data"); 60 61/* GEOM class methods */ 62static g_init_t g_virstor_init; 63static g_fini_t g_virstor_fini; 64static g_taste_t g_virstor_taste; 65static g_ctl_req_t g_virstor_config; 66static g_ctl_destroy_geom_t g_virstor_destroy_geom; 67 68/* Declare & initialize class structure ("geom class") */ 69struct 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 */ 82SYSCTL_DECL(_kern_geom); 83static SYSCTL_NODE(_kern_geom, OID_AUTO, virstor, CTLFLAG_RW, 0, 84 "GEOM_GVIRSTOR information"); 85 86static u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */ 87TUNABLE_INT("kern.geom.virstor.debug", &g_virstor_debug); 88SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RW, &g_virstor_debug, 89 0, "Debug level (2=production, 5=normal, 15=excessive)"); 90 91static u_int g_virstor_chunk_watermark = 100; 92TUNABLE_INT("kern.geom.virstor.chunk_watermark", &g_virstor_chunk_watermark); 93SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RW, 94 &g_virstor_chunk_watermark, 0, 95 "Minimum number of free chunks before issuing administrative warning"); 96 97static u_int g_virstor_component_watermark = 1; 98TUNABLE_INT("kern.geom.virstor.component_watermark", 99 &g_virstor_component_watermark); 100SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RW, 101 &g_virstor_component_watermark, 0, 102 "Minimum number of free components before issuing administrative warning"); 103 104static int read_metadata(struct g_consumer *, struct g_virstor_metadata *); 105static void write_metadata(struct g_consumer *, struct g_virstor_metadata *); 106static int clear_metadata(struct g_virstor_component *); 107static int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *, 108 struct g_virstor_metadata *); 109static struct g_geom *create_virstor_geom(struct g_class *, 110 struct g_virstor_metadata *); 111static void virstor_check_and_run(struct g_virstor_softc *); 112static u_int virstor_valid_components(struct g_virstor_softc *); 113static int virstor_geom_destroy(struct g_virstor_softc *, boolean_t, 114 boolean_t); 115static void remove_component(struct g_virstor_softc *, 116 struct g_virstor_component *, boolean_t); 117static void bioq_dismantle(struct bio_queue_head *); 118static int allocate_chunk(struct g_virstor_softc *, 119 struct g_virstor_component **, u_int *, u_int *); 120static void delay_destroy_consumer(void *, int); 121static void dump_component(struct g_virstor_component *comp); 122#if 0 123static void dump_me(struct virstor_map_entry *me, unsigned int nr); 124#endif 125 126static void virstor_ctl_stop(struct gctl_req *, struct g_class *); 127static void virstor_ctl_add(struct gctl_req *, struct g_class *); 128static void virstor_ctl_remove(struct gctl_req *, struct g_class *); 129static struct g_virstor_softc * virstor_find_geom(const struct g_class *, 130 const char *); 131static void update_metadata(struct g_virstor_softc *); 132static void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *, 133 u_int, u_int); 134 135static void g_virstor_orphan(struct g_consumer *); 136static int g_virstor_access(struct g_provider *, int, int, int); 137static void g_virstor_start(struct bio *); 138static void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *, 139 struct g_consumer *, struct g_provider *); 140static void g_virstor_done(struct bio *); 141 142static void invalid_call(void); 143/* 144 * Initialise GEOM class (per-class callback) 145 */ 146static void 147g_virstor_init(struct g_class *mp __unused) 148{ 149 150 /* Catch map struct size mismatch at compile time; Map entries must 151 * fit into MAXPHYS exactly, with no wasted space. */ 152 CTASSERT(VIRSTOR_MAP_BLOCK_ENTRIES*VIRSTOR_MAP_ENTRY_SIZE == MAXPHYS); 153 154 /* Init UMA zones, TAILQ's, other global vars */ 155} 156 157/* 158 * Finalise GEOM class (per-class callback) 159 */ 160static void 161g_virstor_fini(struct g_class *mp __unused) 162{ 163 164 /* Deinit UMA zones & global vars */ 165} 166 167/* 168 * Config (per-class callback) 169 */ 170static void 171g_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb) 172{ 173 uint32_t *version; 174 175 g_topology_assert(); 176 177 version = gctl_get_paraml(req, "version", sizeof(*version)); 178 if (version == NULL) { 179 gctl_error(req, "Failed to get 'version' argument"); 180 return; 181 } 182 if (*version != G_VIRSTOR_VERSION) { 183 gctl_error(req, "Userland and kernel versions out of sync"); 184 return; 185 } 186 187 g_topology_unlock(); 188 if (strcmp(verb, "add") == 0) 189 virstor_ctl_add(req, cp); 190 else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0) 191 virstor_ctl_stop(req, cp); 192 else if (strcmp(verb, "remove") == 0) 193 virstor_ctl_remove(req, cp); 194 else 195 gctl_error(req, "unknown verb: '%s'", verb); 196 g_topology_lock(); 197} 198 199/* 200 * "stop" verb from userland 201 */ 202static void 203virstor_ctl_stop(struct gctl_req *req, struct g_class *cp) 204{ 205 int *force, *nargs; 206 int i; 207 208 nargs = gctl_get_paraml(req, "nargs", sizeof *nargs); 209 if (nargs == NULL) { 210 gctl_error(req, "Error fetching argument '%s'", "nargs"); 211 return; 212 } 213 if (*nargs < 1) { 214 gctl_error(req, "Invalid number of arguments"); 215 return; 216 } 217 force = gctl_get_paraml(req, "force", sizeof *force); 218 if (force == NULL) { 219 gctl_error(req, "Error fetching argument '%s'", "force"); 220 return; 221 } 222 223 g_topology_lock(); 224 for (i = 0; i < *nargs; i++) { 225 char param[8]; 226 const char *name; 227 struct g_virstor_softc *sc; 228 int error; 229 230 sprintf(param, "arg%d", i); 231 name = gctl_get_asciiparam(req, param); 232 if (name == NULL) { 233 gctl_error(req, "No 'arg%d' argument", i); 234 g_topology_unlock(); 235 return; 236 } 237 sc = virstor_find_geom(cp, name); 238 if (sc == NULL) { 239 gctl_error(req, "Don't know anything about '%s'", name); 240 g_topology_unlock(); 241 return; 242 } 243 244 LOG_MSG(LVL_INFO, "Stopping %s by the userland command", 245 sc->geom->name); 246 update_metadata(sc); 247 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) { 248 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d", 249 sc->geom->name, error); 250 } 251 } 252 g_topology_unlock(); 253} 254 255/* 256 * "add" verb from userland - add new component(s) to the structure. 257 * This will be done all at once in here, without going through the 258 * .taste function for new components. 259 */ 260static void 261virstor_ctl_add(struct gctl_req *req, struct g_class *cp) 262{ 263 /* Note: while this is going on, I/O is being done on 264 * the g_up and g_down threads. The idea is to make changes 265 * to softc members in a way that can atomically activate 266 * them all at once. */ 267 struct g_virstor_softc *sc; 268 int *hardcode, *nargs; 269 const char *geom_name; /* geom to add a component to */ 270 struct g_consumer *fcp; 271 struct g_virstor_bio_q *bq; 272 u_int added; 273 int error; 274 int i; 275 276 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs)); 277 if (nargs == NULL) { 278 gctl_error(req, "Error fetching argument '%s'", "nargs"); 279 return; 280 } 281 if (*nargs < 2) { 282 gctl_error(req, "Invalid number of arguments"); 283 return; 284 } 285 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode)); 286 if (hardcode == NULL) { 287 gctl_error(req, "Error fetching argument '%s'", "hardcode"); 288 return; 289 } 290 291 /* Find "our" geom */ 292 geom_name = gctl_get_asciiparam(req, "arg0"); 293 if (geom_name == NULL) { 294 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)"); 295 return; 296 } 297 sc = virstor_find_geom(cp, geom_name); 298 if (sc == NULL) { 299 gctl_error(req, "Don't know anything about '%s'", geom_name); 300 return; 301 } 302 303 if (virstor_valid_components(sc) != sc->n_components) { 304 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete " 305 "virstor %s", sc->geom->name); 306 gctl_error(req, "Virstor %s is incomplete", sc->geom->name); 307 return; 308 } 309 310 fcp = sc->components[0].gcons; 311 added = 0; 312 g_topology_lock(); 313 for (i = 1; i < *nargs; i++) { 314 struct g_virstor_metadata md; 315 char aname[8]; 316 const char *prov_name; 317 struct g_provider *pp; 318 struct g_consumer *cp; 319 u_int nc; 320 u_int j; 321 322 snprintf(aname, sizeof aname, "arg%d", i); 323 prov_name = gctl_get_asciiparam(req, aname); 324 if (prov_name == NULL) { 325 gctl_error(req, "Error fetching argument '%s'", aname); 326 g_topology_unlock(); 327 return; 328 } 329 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0) 330 prov_name += sizeof(_PATH_DEV) - 1; 331 332 pp = g_provider_by_name(prov_name); 333 if (pp == NULL) { 334 /* This is the most common error so be verbose about it */ 335 if (added != 0) { 336 gctl_error(req, "Invalid provider: '%s' (added" 337 " %u components)", prov_name, added); 338 update_metadata(sc); 339 } else { 340 gctl_error(req, "Invalid provider: '%s'", 341 prov_name); 342 } 343 g_topology_unlock(); 344 return; 345 } 346 cp = g_new_consumer(sc->geom); 347 if (cp == NULL) { 348 gctl_error(req, "Cannot create consumer"); 349 g_topology_unlock(); 350 return; 351 } 352 error = g_attach(cp, pp); 353 if (error != 0) { 354 gctl_error(req, "Cannot attach a consumer to %s", 355 pp->name); 356 g_destroy_consumer(cp); 357 g_topology_unlock(); 358 return; 359 } 360 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) { 361 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace); 362 if (error != 0) { 363 gctl_error(req, "Access request failed for %s", 364 pp->name); 365 g_destroy_consumer(cp); 366 g_topology_unlock(); 367 return; 368 } 369 } 370 if (fcp->provider->sectorsize != pp->sectorsize) { 371 gctl_error(req, "Sector size doesn't fit for %s", 372 pp->name); 373 g_destroy_consumer(cp); 374 g_topology_unlock(); 375 return; 376 } 377 for (j = 0; j < sc->n_components; j++) { 378 if (strcmp(sc->components[j].gcons->provider->name, 379 pp->name) == 0) { 380 gctl_error(req, "Component %s already in %s", 381 pp->name, sc->geom->name); 382 g_destroy_consumer(cp); 383 g_topology_unlock(); 384 return; 385 } 386 } 387 sc->components = realloc(sc->components, 388 sizeof(*sc->components) * (sc->n_components + 1), 389 M_GVIRSTOR, M_WAITOK); 390 391 nc = sc->n_components; 392 sc->components[nc].gcons = cp; 393 sc->components[nc].sc = sc; 394 sc->components[nc].index = nc; 395 sc->components[nc].chunk_count = cp->provider->mediasize / 396 sc->chunk_size; 397 sc->components[nc].chunk_next = 0; 398 sc->components[nc].chunk_reserved = 0; 399 400 if (sc->components[nc].chunk_count < 4) { 401 gctl_error(req, "Provider too small: %s", 402 cp->provider->name); 403 g_destroy_consumer(cp); 404 g_topology_unlock(); 405 return; 406 } 407 fill_metadata(sc, &md, nc, *hardcode); 408 write_metadata(cp, &md); 409 /* The new component becomes visible when n_components is 410 * incremented */ 411 sc->n_components++; 412 added++; 413 414 } 415 /* This call to update_metadata() is critical. In case there's a 416 * power failure in the middle of it and some components are updated 417 * while others are not, there will be trouble on next .taste() iff 418 * a non-updated component is detected first */ 419 update_metadata(sc); 420 g_topology_unlock(); 421 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added, 422 sc->geom->name); 423 /* Fire off BIOs previously queued because there wasn't any 424 * physical space left. If the BIOs still can't be satisfied 425 * they will again be added to the end of the queue (during 426 * which the mutex will be recursed) */ 427 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK); 428 bq->bio = NULL; 429 mtx_lock(&sc->delayed_bio_q_mtx); 430 /* First, insert a sentinel to the queue end, so we don't 431 * end up in an infinite loop if there's still no free 432 * space available. */ 433 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage); 434 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) { 435 bq = STAILQ_FIRST(&sc->delayed_bio_q); 436 if (bq->bio != NULL) { 437 g_virstor_start(bq->bio); 438 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage); 439 free(bq, M_GVIRSTOR); 440 } else { 441 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage); 442 free(bq, M_GVIRSTOR); 443 break; 444 } 445 } 446 mtx_unlock(&sc->delayed_bio_q_mtx); 447 448} 449 450/* 451 * Find a geom handled by the class 452 */ 453static struct g_virstor_softc * 454virstor_find_geom(const struct g_class *cp, const char *name) 455{ 456 struct g_geom *gp; 457 458 LIST_FOREACH(gp, &cp->geom, geom) { 459 if (strcmp(name, gp->name) == 0) 460 return (gp->softc); 461 } 462 return (NULL); 463} 464 465/* 466 * Update metadata on all components to reflect the current state 467 * of these fields: 468 * - chunk_next 469 * - flags 470 * - md_count 471 * Expects things to be set up so write_metadata() can work, i.e. 472 * the topology lock must be held. 473 */ 474static void 475update_metadata(struct g_virstor_softc *sc) 476{ 477 struct g_virstor_metadata md; 478 u_int n; 479 480 if (virstor_valid_components(sc) != sc->n_components) 481 return; /* Incomplete device */ 482 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s", 483 sc->geom->name); 484 /* Update metadata on components */ 485 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, 486 sc->geom->class->name, sc->geom->name); 487 g_topology_assert(); 488 for (n = 0; n < sc->n_components; n++) { 489 read_metadata(sc->components[n].gcons, &md); 490 md.chunk_next = sc->components[n].chunk_next; 491 md.flags = sc->components[n].flags; 492 md.md_count = sc->n_components; 493 write_metadata(sc->components[n].gcons, &md); 494 } 495} 496 497/* 498 * Fills metadata (struct md) from information stored in softc and the nc'th 499 * component of virstor 500 */ 501static void 502fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md, 503 u_int nc, u_int hardcode) 504{ 505 struct g_virstor_component *c; 506 507 bzero(md, sizeof *md); 508 c = &sc->components[nc]; 509 510 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic); 511 md->md_version = G_VIRSTOR_VERSION; 512 strncpy(md->md_name, sc->geom->name, sizeof md->md_name); 513 md->md_id = sc->id; 514 md->md_virsize = sc->virsize; 515 md->md_chunk_size = sc->chunk_size; 516 md->md_count = sc->n_components; 517 518 if (hardcode) { 519 strncpy(md->provider, c->gcons->provider->name, 520 sizeof md->provider); 521 } 522 md->no = nc; 523 md->provsize = c->gcons->provider->mediasize; 524 md->chunk_count = c->chunk_count; 525 md->chunk_next = c->chunk_next; 526 md->chunk_reserved = c->chunk_reserved; 527 md->flags = c->flags; 528} 529 530/* 531 * Remove a component from virstor device. 532 * Can only be done if the component is unallocated. 533 */ 534static void 535virstor_ctl_remove(struct gctl_req *req, struct g_class *cp) 536{ 537 /* As this is executed in parallel to I/O, operations on virstor 538 * structures must be as atomic as possible. */ 539 struct g_virstor_softc *sc; 540 int *nargs; 541 const char *geom_name; 542 u_int removed; 543 int i; 544 545 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs)); 546 if (nargs == NULL) { 547 gctl_error(req, "Error fetching argument '%s'", "nargs"); 548 return; 549 } 550 if (*nargs < 2) { 551 gctl_error(req, "Invalid number of arguments"); 552 return; 553 } 554 /* Find "our" geom */ 555 geom_name = gctl_get_asciiparam(req, "arg0"); 556 if (geom_name == NULL) { 557 gctl_error(req, "Error fetching argument '%s'", 558 "geom_name (arg0)"); 559 return; 560 } 561 sc = virstor_find_geom(cp, geom_name); 562 if (sc == NULL) { 563 gctl_error(req, "Don't know anything about '%s'", geom_name); 564 return; 565 } 566 567 if (virstor_valid_components(sc) != sc->n_components) { 568 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete " 569 "virstor %s", sc->geom->name); 570 gctl_error(req, "Virstor %s is incomplete", sc->geom->name); 571 return; 572 } 573 574 removed = 0; 575 for (i = 1; i < *nargs; i++) { 576 char param[8]; 577 const char *prov_name; 578 int j, found; 579 struct g_virstor_component *newcomp, *compbak; 580 581 sprintf(param, "arg%d", i); 582 prov_name = gctl_get_asciiparam(req, param); 583 if (prov_name == NULL) { 584 gctl_error(req, "Error fetching argument '%s'", param); 585 return; 586 } 587 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0) 588 prov_name += sizeof(_PATH_DEV) - 1; 589 590 found = -1; 591 for (j = 0; j < sc->n_components; j++) { 592 if (strcmp(sc->components[j].gcons->provider->name, 593 prov_name) == 0) { 594 found = j; 595 break; 596 } 597 } 598 if (found == -1) { 599 LOG_MSG(LVL_ERROR, "No %s component in %s", 600 prov_name, sc->geom->name); 601 continue; 602 } 603 604 compbak = sc->components; 605 newcomp = malloc(sc->n_components * sizeof(*sc->components), 606 M_GVIRSTOR, M_WAITOK | M_ZERO); 607 bcopy(sc->components, newcomp, found * sizeof(*sc->components)); 608 bcopy(&sc->components[found + 1], newcomp + found, 609 found * sizeof(*sc->components)); 610 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) { 611 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be " 612 "removed from %s", 613 prov_name, sc->geom->name); 614 free(newcomp, M_GVIRSTOR); 615 /* We'll consider this non-fatal error */ 616 continue; 617 } 618 /* Renumerate unallocated components */ 619 for (j = 0; j < sc->n_components-1; j++) { 620 if ((sc->components[j].flags & 621 VIRSTOR_PROVIDER_ALLOCATED) == 0) { 622 sc->components[j].index = j; 623 } 624 } 625 /* This is the critical section. If a component allocation 626 * event happens while both variables are not yet set, 627 * there will be trouble. Something will panic on encountering 628 * NULL sc->components[x].gcomp member. 629 * Luckily, component allocation happens very rarely and 630 * removing components is an abnormal action in any case. */ 631 sc->components = newcomp; 632 sc->n_components--; 633 /* End critical section */ 634 635 g_topology_lock(); 636 if (clear_metadata(&compbak[found]) != 0) { 637 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear " 638 "metadata on %s", prov_name); 639 } 640 g_detach(compbak[found].gcons); 641 g_destroy_consumer(compbak[found].gcons); 642 g_topology_unlock(); 643 644 free(compbak, M_GVIRSTOR); 645 646 removed++; 647 } 648 649 /* This call to update_metadata() is critical. In case there's a 650 * power failure in the middle of it and some components are updated 651 * while others are not, there will be trouble on next .taste() iff 652 * a non-updated component is detected first */ 653 g_topology_lock(); 654 update_metadata(sc); 655 g_topology_unlock(); 656 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed, 657 sc->geom->name); 658} 659 660/* 661 * Clear metadata sector on component 662 */ 663static int 664clear_metadata(struct g_virstor_component *comp) 665{ 666 char *buf; 667 int error; 668 669 LOG_MSG(LVL_INFO, "Clearing metadata on %s", 670 comp->gcons->provider->name); 671 g_topology_assert(); 672 error = g_access(comp->gcons, 0, 1, 0); 673 if (error != 0) 674 return (error); 675 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR, 676 M_WAITOK | M_ZERO); 677 error = g_write_data(comp->gcons, 678 comp->gcons->provider->mediasize - 679 comp->gcons->provider->sectorsize, 680 buf, 681 comp->gcons->provider->sectorsize); 682 free(buf, M_GVIRSTOR); 683 g_access(comp->gcons, 0, -1, 0); 684 return (error); 685} 686 687/* 688 * Destroy geom forcibly. 689 */ 690static int 691g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp, 692 struct g_geom *gp) 693{ 694 struct g_virstor_softc *sc; 695 int exitval; 696 697 sc = gp->softc; 698 KASSERT(sc != NULL, ("%s: NULL sc", __func__)); 699 700 exitval = 0; 701 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name, 702 gp->softc); 703 704 if (sc != NULL) { 705#ifdef INVARIANTS 706 char *buf; 707 int error; 708 off_t off; 709 int isclean, count; 710 int n; 711 712 LOG_MSG(LVL_INFO, "INVARIANTS detected"); 713 LOG_MSG(LVL_INFO, "Verifying allocation " 714 "table for %s", sc->geom->name); 715 count = 0; 716 for (n = 0; n < sc->chunk_count; n++) { 717 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0) 718 count++; 719 } 720 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks", 721 sc->geom->name, count); 722 n = off = count = 0; 723 isclean = 1; 724 if (virstor_valid_components(sc) != sc->n_components) { 725 /* This is a incomplete virstor device (not all 726 * components have been found) */ 727 LOG_MSG(LVL_ERROR, "Device %s is incomplete", 728 sc->geom->name); 729 goto bailout; 730 } 731 error = g_access(sc->components[0].gcons, 1, 0, 0); 732 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__, 733 error)); 734 /* Compare the whole on-disk allocation table with what's 735 * currently in memory */ 736 while (n < sc->chunk_count) { 737 buf = g_read_data(sc->components[0].gcons, off, 738 sc->sectorsize, &error); 739 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) " 740 "for read at %jd", error, off)); 741 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) { 742 LOG_MSG(LVL_ERROR, "ERROR in allocation table, " 743 "entry %d, offset %jd", n, off); 744 isclean = 0; 745 count++; 746 } 747 n += sc->me_per_sector; 748 off += sc->sectorsize; 749 g_free(buf); 750 } 751 error = g_access(sc->components[0].gcons, -1, 0, 0); 752 KASSERT(error == 0, ("%s: g_access failed (%d) on exit", 753 __func__, error)); 754 if (isclean != 1) { 755 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s " 756 "(%d sectors don't match, max %zu allocations)", 757 sc->geom->name, count, 758 count * sc->me_per_sector); 759 } else { 760 LOG_MSG(LVL_INFO, "Allocation table ok for %s", 761 sc->geom->name); 762 } 763bailout: 764#endif 765 update_metadata(sc); 766 virstor_geom_destroy(sc, FALSE, FALSE); 767 exitval = EAGAIN; 768 } else 769 exitval = 0; 770 return (exitval); 771} 772 773/* 774 * Taste event (per-class callback) 775 * Examines a provider and creates geom instances if needed 776 */ 777static struct g_geom * 778g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags) 779{ 780 struct g_virstor_metadata md; 781 struct g_geom *gp; 782 struct g_consumer *cp; 783 struct g_virstor_softc *sc; 784 int error; 785 786 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name); 787 g_topology_assert(); 788 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name); 789 790 /* We need a dummy geom to attach a consumer to the given provider */ 791 gp = g_new_geomf(mp, "virstor:taste.helper"); 792 gp->start = (void *)invalid_call; /* XXX: hacked up so the */ 793 gp->access = (void *)invalid_call; /* compiler doesn't complain. */ 794 gp->orphan = (void *)invalid_call; /* I really want these to fail. */ 795 796 cp = g_new_consumer(gp); 797 g_attach(cp, pp); 798 error = read_metadata(cp, &md); 799 g_detach(cp); 800 g_destroy_consumer(cp); 801 g_destroy_geom(gp); 802 803 if (error != 0) 804 return (NULL); 805 806 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0) 807 return (NULL); 808 if (md.md_version != G_VIRSTOR_VERSION) { 809 LOG_MSG(LVL_ERROR, "Kernel module version invalid " 810 "to handle %s (%s) : %d should be %d", 811 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION); 812 return (NULL); 813 } 814 if (md.provsize != pp->mediasize) 815 return (NULL); 816 817 /* If the provider name is hardcoded, use the offered provider only 818 * if it's been offered with its proper name (the one used in 819 * the label command). */ 820 if (md.provider[0] != '\0' && 821 !g_compare_names(md.provider, pp->name)) 822 return (NULL); 823 824 /* Iterate all geoms this class already knows about to see if a new 825 * geom instance of this class needs to be created (in case the provider 826 * is first from a (possibly) multi-consumer geom) or it just needs 827 * to be added to an existing instance. */ 828 sc = NULL; 829 gp = NULL; 830 LIST_FOREACH(gp, &mp->geom, geom) { 831 sc = gp->softc; 832 if (sc == NULL) 833 continue; 834 if (strcmp(md.md_name, sc->geom->name) != 0) 835 continue; 836 if (md.md_id != sc->id) 837 continue; 838 break; 839 } 840 if (gp != NULL) { /* We found an existing geom instance; add to it */ 841 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name); 842 error = add_provider_to_geom(sc, pp, &md); 843 if (error != 0) { 844 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)", 845 pp->name, md.md_name, error); 846 return (NULL); 847 } 848 } else { /* New geom instance needs to be created */ 849 gp = create_virstor_geom(mp, &md); 850 if (gp == NULL) { 851 LOG_MSG(LVL_ERROR, "Error creating new instance of " 852 "class %s: %s", mp->name, md.md_name); 853 LOG_MSG(LVL_DEBUG, "Error creating %s at %s", 854 md.md_name, pp->name); 855 return (NULL); 856 } 857 sc = gp->softc; 858 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name, 859 md.md_name); 860 error = add_provider_to_geom(sc, pp, &md); 861 if (error != 0) { 862 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)", 863 pp->name, md.md_name, error); 864 virstor_geom_destroy(sc, TRUE, FALSE); 865 return (NULL); 866 } 867 } 868 869 return (gp); 870} 871 872/* 873 * Destroyes consumer passed to it in arguments. Used as a callback 874 * on g_event queue. 875 */ 876static void 877delay_destroy_consumer(void *arg, int flags __unused) 878{ 879 struct g_consumer *c = arg; 880 KASSERT(c != NULL, ("%s: invalid consumer", __func__)); 881 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay", 882 c->provider->name); 883 g_detach(c); 884 g_destroy_consumer(c); 885} 886 887/* 888 * Remove a component (consumer) from geom instance; If it's the first 889 * component being removed, orphan the provider to announce geom's being 890 * dismantled 891 */ 892static void 893remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp, 894 boolean_t delay) 895{ 896 struct g_consumer *c; 897 898 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s", 899 sc->geom->name)); 900 c = comp->gcons; 901 902 comp->gcons = NULL; 903 KASSERT(c->provider != NULL, ("%s: no provider", __func__)); 904 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name, 905 sc->geom->name); 906 if (sc->provider != NULL) { 907 LOG_MSG(LVL_INFO, "Removing provider %s", sc->provider->name); 908 g_wither_provider(sc->provider, ENXIO); 909 sc->provider = NULL; 910 } 911 912 if (c->acr > 0 || c->acw > 0 || c->ace > 0) 913 g_access(c, -c->acr, -c->acw, -c->ace); 914 if (delay) { 915 /* Destroy consumer after it's tasted */ 916 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL); 917 } else { 918 g_detach(c); 919 g_destroy_consumer(c); 920 } 921} 922 923/* 924 * Destroy geom - called internally 925 * See g_virstor_destroy_geom for the other one 926 */ 927static int 928virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force, 929 boolean_t delay) 930{ 931 struct g_provider *pp; 932 struct g_geom *gp; 933 u_int n; 934 935 g_topology_assert(); 936 937 if (sc == NULL) 938 return (ENXIO); 939 940 pp = sc->provider; 941 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) { 942 LOG_MSG(force ? LVL_WARNING : LVL_ERROR, 943 "Device %s is still open.", pp->name); 944 if (!force) 945 return (EBUSY); 946 } 947 948 for (n = 0; n < sc->n_components; n++) { 949 if (sc->components[n].gcons != NULL) 950 remove_component(sc, &sc->components[n], delay); 951 } 952 953 gp = sc->geom; 954 gp->softc = NULL; 955 956 KASSERT(sc->provider == NULL, ("Provider still exists for %s", 957 gp->name)); 958 959 /* XXX: This might or might not work, since we're called with 960 * the topology lock held. Also, it might panic the kernel if 961 * the error'd BIO is in softupdates code. */ 962 mtx_lock(&sc->delayed_bio_q_mtx); 963 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) { 964 struct g_virstor_bio_q *bq; 965 bq = STAILQ_FIRST(&sc->delayed_bio_q); 966 bq->bio->bio_error = ENOSPC; 967 g_io_deliver(bq->bio, EIO); 968 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage); 969 free(bq, M_GVIRSTOR); 970 } 971 mtx_unlock(&sc->delayed_bio_q_mtx); 972 mtx_destroy(&sc->delayed_bio_q_mtx); 973 974 free(sc->map, M_GVIRSTOR); 975 free(sc->components, M_GVIRSTOR); 976 bzero(sc, sizeof *sc); 977 free(sc, M_GVIRSTOR); 978 979 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */ 980 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)) 981 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name); 982 983 g_wither_geom(gp, ENXIO); 984 985 return (0); 986} 987 988/* 989 * Utility function: read metadata & decode. Wants topology lock to be 990 * held. 991 */ 992static int 993read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md) 994{ 995 struct g_provider *pp; 996 char *buf; 997 int error; 998 999 g_topology_assert(); 1000 error = g_access(cp, 1, 0, 0); 1001 if (error != 0) 1002 return (error); 1003 pp = cp->provider; 1004 g_topology_unlock(); 1005 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize, 1006 &error); 1007 g_topology_lock(); 1008 g_access(cp, -1, 0, 0); 1009 if (buf == NULL) 1010 return (error); 1011 1012 virstor_metadata_decode(buf, md); 1013 g_free(buf); 1014 1015 return (0); 1016} 1017 1018/** 1019 * Utility function: encode & write metadata. Assumes topology lock is 1020 * held. 1021 * 1022 * There is no useful way of recovering from errors in this function, 1023 * not involving panicking the kernel. If the metadata cannot be written 1024 * the most we can do is notify the operator and hope he spots it and 1025 * replaces the broken drive. 1026 */ 1027static void 1028write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md) 1029{ 1030 struct g_provider *pp; 1031 char *buf; 1032 int error; 1033 1034 KASSERT(cp != NULL && md != NULL && cp->provider != NULL, 1035 ("Something's fishy in %s", __func__)); 1036 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name); 1037 g_topology_assert(); 1038 error = g_access(cp, 0, 1, 0); 1039 if (error != 0) { 1040 LOG_MSG(LVL_ERROR, "g_access(0,1,0) failed for %s: %d", 1041 cp->provider->name, error); 1042 return; 1043 } 1044 pp = cp->provider; 1045 1046 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK); 1047 bzero(buf, pp->sectorsize); 1048 virstor_metadata_encode(md, buf); 1049 g_topology_unlock(); 1050 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf, 1051 pp->sectorsize); 1052 g_topology_lock(); 1053 g_access(cp, 0, -1, 0); 1054 free(buf, M_GVIRSTOR); 1055 1056 if (error != 0) 1057 LOG_MSG(LVL_ERROR, "Error %d writing metadata to %s", 1058 error, cp->provider->name); 1059} 1060 1061/* 1062 * Creates a new instance of this GEOM class, initialise softc 1063 */ 1064static struct g_geom * 1065create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md) 1066{ 1067 struct g_geom *gp; 1068 struct g_virstor_softc *sc; 1069 1070 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)", 1071 md->md_name, md->md_id); 1072 1073 if (md->md_count < 1 || md->md_chunk_size < 1 || 1074 md->md_virsize < md->md_chunk_size) { 1075 /* This is bogus configuration, and probably means data is 1076 * somehow corrupted. Panic, maybe? */ 1077 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s", 1078 md->md_name); 1079 return (NULL); 1080 } 1081 1082 /* Check if it's already created */ 1083 LIST_FOREACH(gp, &mp->geom, geom) { 1084 sc = gp->softc; 1085 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) { 1086 LOG_MSG(LVL_WARNING, "Geom %s already exists", 1087 md->md_name); 1088 if (sc->id != md->md_id) { 1089 LOG_MSG(LVL_ERROR, 1090 "Some stale or invalid components " 1091 "exist for virstor device named %s. " 1092 "You will need to <CLEAR> all stale " 1093 "components and maybe reconfigure " 1094 "the virstor device. Tune " 1095 "kern.geom.virstor.debug sysctl up " 1096 "for more information.", 1097 sc->geom->name); 1098 } 1099 return (NULL); 1100 } 1101 } 1102 gp = g_new_geomf(mp, "%s", md->md_name); 1103 gp->softc = NULL; /* to circumevent races that test softc */ 1104 1105 gp->start = g_virstor_start; 1106 gp->spoiled = g_virstor_orphan; 1107 gp->orphan = g_virstor_orphan; 1108 gp->access = g_virstor_access; 1109 gp->dumpconf = g_virstor_dumpconf; 1110 1111 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO); 1112 sc->id = md->md_id; 1113 sc->n_components = md->md_count; 1114 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count, 1115 M_GVIRSTOR, M_WAITOK | M_ZERO); 1116 sc->chunk_size = md->md_chunk_size; 1117 sc->virsize = md->md_virsize; 1118 STAILQ_INIT(&sc->delayed_bio_q); 1119 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx", 1120 "gvirstor", MTX_DEF | MTX_RECURSE); 1121 1122 sc->geom = gp; 1123 sc->provider = NULL; /* virstor_check_and_run will create it */ 1124 gp->softc = sc; 1125 1126 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name); 1127 1128 return (gp); 1129} 1130 1131/* 1132 * Add provider to a GEOM class instance 1133 */ 1134static int 1135add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp, 1136 struct g_virstor_metadata *md) 1137{ 1138 struct g_virstor_component *component; 1139 struct g_consumer *cp, *fcp; 1140 struct g_geom *gp; 1141 int error; 1142 1143 if (md->no >= sc->n_components) 1144 return (EINVAL); 1145 1146 /* "Current" compontent */ 1147 component = &(sc->components[md->no]); 1148 if (component->gcons != NULL) 1149 return (EEXIST); 1150 1151 gp = sc->geom; 1152 fcp = LIST_FIRST(&gp->consumer); 1153 1154 cp = g_new_consumer(gp); 1155 error = g_attach(cp, pp); 1156 1157 if (error != 0) { 1158 g_destroy_consumer(cp); 1159 return (error); 1160 } 1161 1162 if (fcp != NULL) { 1163 if (fcp->provider->sectorsize != pp->sectorsize) { 1164 /* TODO: this can be made to work */ 1165 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid " 1166 "sector size (%d)", pp->name, sc->geom->name, 1167 pp->sectorsize); 1168 return (EINVAL); 1169 } 1170 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) { 1171 /* Replicate access permissions from first "live" consumer 1172 * to the new one */ 1173 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace); 1174 if (error != 0) { 1175 g_detach(cp); 1176 g_destroy_consumer(cp); 1177 return (error); 1178 } 1179 } 1180 } 1181 1182 /* Bring up a new component */ 1183 cp->private = component; 1184 component->gcons = cp; 1185 component->sc = sc; 1186 component->index = md->no; 1187 component->chunk_count = md->chunk_count; 1188 component->chunk_next = md->chunk_next; 1189 component->chunk_reserved = md->chunk_reserved; 1190 component->flags = md->flags; 1191 1192 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name); 1193 1194 virstor_check_and_run(sc); 1195 return (0); 1196} 1197 1198/* 1199 * Check if everything's ready to create the geom provider & device entry, 1200 * create and start provider. 1201 * Called ultimately by .taste, from g_event thread 1202 */ 1203static void 1204virstor_check_and_run(struct g_virstor_softc *sc) 1205{ 1206 off_t off; 1207 size_t n, count; 1208 int index; 1209 int error; 1210 1211 if (virstor_valid_components(sc) != sc->n_components) 1212 return; 1213 1214 if (virstor_valid_components(sc) == 0) { 1215 /* This is actually a candidate for panic() */ 1216 LOG_MSG(LVL_ERROR, "No valid components for %s?", 1217 sc->provider->name); 1218 return; 1219 } 1220 1221 sc->sectorsize = sc->components[0].gcons->provider->sectorsize; 1222 1223 /* Initialise allocation map from the first consumer */ 1224 sc->chunk_count = sc->virsize / sc->chunk_size; 1225 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) { 1226 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes", 1227 sc->provider->name, 1228 sc->chunk_count * (off_t)sc->chunk_size); 1229 } 1230 sc->map_size = sc->chunk_count * sizeof *(sc->map); 1231 /* The following allocation is in order of 4MB - 8MB */ 1232 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK); 1233 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s", 1234 __func__, sc->map_size, sc->provider->name)); 1235 sc->map_sectors = sc->map_size / sc->sectorsize; 1236 1237 count = 0; 1238 for (n = 0; n < sc->n_components; n++) 1239 count += sc->components[n].chunk_count; 1240 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual " 1241 "(%zu KB chunks)", 1242 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024); 1243 1244 error = g_access(sc->components[0].gcons, 1, 0, 0); 1245 if (error != 0) { 1246 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to " 1247 "read allocation map for %s", 1248 sc->components[0].gcons->provider->name, 1249 sc->geom->name); 1250 return; 1251 } 1252 /* Read in the allocation map */ 1253 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name, 1254 sc->components[0].gcons->provider->name); 1255 off = count = n = 0; 1256 while (count < sc->map_size) { 1257 struct g_virstor_map_entry *mapbuf; 1258 size_t bs; 1259 1260 bs = MIN(MAXPHYS, sc->map_size - count); 1261 if (bs % sc->sectorsize != 0) { 1262 /* Check for alignment errors */ 1263 bs = (bs / sc->sectorsize) * sc->sectorsize; 1264 if (bs == 0) 1265 break; 1266 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned " 1267 "for %s on %s", sc->geom->name, 1268 sc->components[0].gcons->provider->name); 1269 } 1270 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error); 1271 if (mapbuf == NULL) { 1272 free(sc->map, M_GVIRSTOR); 1273 LOG_MSG(LVL_ERROR, "Error reading allocation map " 1274 "for %s from %s (offset %ju) (error %d)", 1275 sc->geom->name, 1276 sc->components[0].gcons->provider->name, 1277 off, error); 1278 return; 1279 } 1280 1281 bcopy(mapbuf, &sc->map[n], bs); 1282 off += bs; 1283 count += bs; 1284 n += bs / sizeof *(sc->map); 1285 g_free(mapbuf); 1286 } 1287 g_access(sc->components[0].gcons, -1, 0, 0); 1288 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name); 1289 1290 /* find first component with allocatable chunks */ 1291 index = -1; 1292 for (n = 0; n < sc->n_components; n++) { 1293 if (sc->components[n].chunk_next < 1294 sc->components[n].chunk_count) { 1295 index = n; 1296 break; 1297 } 1298 } 1299 if (index == -1) 1300 /* not found? set it to the last component and handle it 1301 * later */ 1302 index = sc->n_components - 1; 1303 1304 if (index >= sc->n_components - g_virstor_component_watermark - 1) { 1305 LOG_MSG(LVL_WARNING, "Device %s running out of components " 1306 "(%d/%u: %s)", sc->geom->name, 1307 index+1, 1308 sc->n_components, 1309 sc->components[index].gcons->provider->name); 1310 } 1311 sc->curr_component = index; 1312 1313 if (sc->components[index].chunk_next >= 1314 sc->components[index].chunk_count - g_virstor_chunk_watermark) { 1315 LOG_MSG(LVL_WARNING, 1316 "Component %s of %s is running out of free space " 1317 "(%u chunks left)", 1318 sc->components[index].gcons->provider->name, 1319 sc->geom->name, sc->components[index].chunk_count - 1320 sc->components[index].chunk_next); 1321 } 1322 1323 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map); 1324 if (sc->sectorsize % sizeof *(sc->map) != 0) { 1325 LOG_MSG(LVL_ERROR, 1326 "%s: Map entries don't fit exactly in a sector (%s)", 1327 __func__, sc->geom->name); 1328 return; 1329 } 1330 1331 /* Recalculate allocated chunks in components & at the same time 1332 * verify map data is sane. We could trust metadata on this, but 1333 * we want to make sure. */ 1334 for (n = 0; n < sc->n_components; n++) 1335 sc->components[n].chunk_next = sc->components[n].chunk_reserved; 1336 1337 for (n = 0; n < sc->chunk_count; n++) { 1338 if (sc->map[n].provider_no >= sc->n_components || 1339 sc->map[n].provider_chunk >= 1340 sc->components[sc->map[n].provider_no].chunk_count) { 1341 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s", 1342 __func__, (u_int)n, sc->geom->name); 1343 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u" 1344 " provider_chunk: %u, chunk_count: %u", __func__, 1345 sc->map[n].provider_no, sc->n_components, 1346 sc->map[n].provider_chunk, 1347 sc->components[sc->map[n].provider_no].chunk_count); 1348 return; 1349 } 1350 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED) 1351 sc->components[sc->map[n].provider_no].chunk_next++; 1352 } 1353 1354 sc->provider = g_new_providerf(sc->geom, "virstor/%s", 1355 sc->geom->name); 1356 1357 sc->provider->sectorsize = sc->sectorsize; 1358 sc->provider->mediasize = sc->virsize; 1359 g_error_provider(sc->provider, 0); 1360 1361 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name); 1362 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting " 1363 "chunk %u", sc->provider->name, sc->curr_component, 1364 sc->components[sc->curr_component].chunk_next); 1365} 1366 1367/* 1368 * Returns count of active providers in this geom instance 1369 */ 1370static u_int 1371virstor_valid_components(struct g_virstor_softc *sc) 1372{ 1373 unsigned int nc, i; 1374 1375 nc = 0; 1376 KASSERT(sc != NULL, ("%s: softc is NULL", __func__)); 1377 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__)); 1378 for (i = 0; i < sc->n_components; i++) 1379 if (sc->components[i].gcons != NULL) 1380 nc++; 1381 return (nc); 1382} 1383 1384/* 1385 * Called when the consumer gets orphaned (?) 1386 */ 1387static void 1388g_virstor_orphan(struct g_consumer *cp) 1389{ 1390 struct g_virstor_softc *sc; 1391 struct g_virstor_component *comp; 1392 struct g_geom *gp; 1393 1394 g_topology_assert(); 1395 gp = cp->geom; 1396 sc = gp->softc; 1397 if (sc == NULL) 1398 return; 1399 1400 comp = cp->private; 1401 KASSERT(comp != NULL, ("%s: No component in private part of consumer", 1402 __func__)); 1403 remove_component(sc, comp, FALSE); 1404 if (virstor_valid_components(sc) == 0) 1405 virstor_geom_destroy(sc, TRUE, FALSE); 1406} 1407 1408/* 1409 * Called to notify geom when it's been opened, and for what intent 1410 */ 1411static int 1412g_virstor_access(struct g_provider *pp, int dr, int dw, int de) 1413{ 1414 struct g_consumer *c; 1415 struct g_virstor_softc *sc; 1416 struct g_geom *gp; 1417 int error; 1418 1419 KASSERT(pp != NULL, ("%s: NULL provider", __func__)); 1420 gp = pp->geom; 1421 KASSERT(gp != NULL, ("%s: NULL geom", __func__)); 1422 sc = gp->softc; 1423 1424 if (sc == NULL) { 1425 /* It seems that .access can be called with negative dr,dw,dx 1426 * in this case but I want to check for myself */ 1427 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s", 1428 dr, dw, de, pp->name); 1429 /* This should only happen when geom is withered so 1430 * allow only negative requests */ 1431 KASSERT(dr <= 0 && dw <= 0 && de <= 0, 1432 ("%s: Positive access for %s", __func__, pp->name)); 1433 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) 1434 LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed", 1435 pp->name); 1436 return (0); 1437 } 1438 1439 /* Grab an exclusive bit to propagate on our consumers on first open */ 1440 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0) 1441 de++; 1442 /* ... drop it on close */ 1443 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) { 1444 de--; 1445 update_metadata(sc); /* Writes statistical information */ 1446 } 1447 1448 error = ENXIO; 1449 LIST_FOREACH(c, &gp->consumer, consumer) { 1450 KASSERT(c != NULL, ("%s: consumer is NULL", __func__)); 1451 error = g_access(c, dr, dw, de); 1452 if (error != 0) { 1453 struct g_consumer *c2; 1454 1455 /* Backout earlier changes */ 1456 LIST_FOREACH(c2, &gp->consumer, consumer) { 1457 if (c2 == c) /* all eariler components fixed */ 1458 return (error); 1459 g_access(c2, -dr, -dw, -de); 1460 } 1461 } 1462 } 1463 1464 return (error); 1465} 1466 1467/* 1468 * Generate XML dump of current state 1469 */ 1470static void 1471g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp, 1472 struct g_consumer *cp, struct g_provider *pp) 1473{ 1474 struct g_virstor_softc *sc; 1475 1476 g_topology_assert(); 1477 sc = gp->softc; 1478 1479 if (sc == NULL || pp != NULL) 1480 return; 1481 1482 if (cp != NULL) { 1483 /* For each component */ 1484 struct g_virstor_component *comp; 1485 1486 comp = cp->private; 1487 if (comp == NULL) 1488 return; 1489 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n", 1490 indent, comp->index); 1491 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n", 1492 indent, comp->chunk_count); 1493 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n", 1494 indent, comp->chunk_next); 1495 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n", 1496 indent, comp->chunk_reserved); 1497 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n", 1498 indent, 1499 comp->chunk_next > 0 ? 100 - 1500 ((comp->chunk_next + comp->chunk_reserved) * 100) / 1501 comp->chunk_count : 100); 1502 } else { 1503 /* For the whole thing */ 1504 u_int count, used, i; 1505 off_t size; 1506 1507 count = used = size = 0; 1508 for (i = 0; i < sc->n_components; i++) { 1509 if (sc->components[i].gcons != NULL) { 1510 count += sc->components[i].chunk_count; 1511 used += sc->components[i].chunk_next + 1512 sc->components[i].chunk_reserved; 1513 size += sc->components[i].gcons-> 1514 provider->mediasize; 1515 } 1516 } 1517 1518 sbuf_printf(sb, "%s<Status>" 1519 "Components=%u, Online=%u</Status>\n", indent, 1520 sc->n_components, virstor_valid_components(sc)); 1521 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n", 1522 indent, 100-(used * 100) / count); 1523 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent, 1524 sc->chunk_size); 1525 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n", 1526 indent, used > 0 ? 100 - (used * 100) / count : 100); 1527 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n", 1528 indent, count); 1529 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n", 1530 indent, sc->chunk_count); 1531 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n", 1532 indent, 1533 (count * 100) / sc->chunk_count); 1534 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n", 1535 indent, size); 1536 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent, 1537 sc->virsize); 1538 } 1539} 1540 1541/* 1542 * GEOM .done handler 1543 * Can't use standard handler because one requested IO may 1544 * fork into additional data IOs 1545 */ 1546static void 1547g_virstor_done(struct bio *b) 1548{ 1549 struct g_virstor_softc *sc; 1550 struct bio *parent_b; 1551 1552 parent_b = b->bio_parent; 1553 sc = parent_b->bio_to->geom->softc; 1554 1555 if (b->bio_error != 0) { 1556 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s", 1557 b->bio_error, b->bio_offset, b->bio_length, 1558 b->bio_to->name); 1559 if (parent_b->bio_error == 0) 1560 parent_b->bio_error = b->bio_error; 1561 } 1562 1563 parent_b->bio_inbed++; 1564 parent_b->bio_completed += b->bio_completed; 1565 1566 if (parent_b->bio_children == parent_b->bio_inbed) { 1567 parent_b->bio_completed = parent_b->bio_length; 1568 g_io_deliver(parent_b, parent_b->bio_error); 1569 } 1570 g_destroy_bio(b); 1571} 1572 1573/* 1574 * I/O starts here 1575 * Called in g_down thread 1576 */ 1577static void 1578g_virstor_start(struct bio *b) 1579{ 1580 struct g_virstor_softc *sc; 1581 struct g_virstor_component *comp; 1582 struct bio *cb; 1583 struct g_provider *pp; 1584 char *addr; 1585 off_t offset, length; 1586 struct bio_queue_head bq; 1587 size_t chunk_size; /* cached for convenience */ 1588 u_int count; 1589 1590 pp = b->bio_to; 1591 sc = pp->geom->softc; 1592 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__, 1593 b->bio_to->error, b->bio_to->name)); 1594 1595 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__); 1596 1597 switch (b->bio_cmd) { 1598 case BIO_READ: 1599 case BIO_WRITE: 1600 case BIO_DELETE: 1601 break; 1602 default: 1603 g_io_deliver(b, EOPNOTSUPP); 1604 return; 1605 } 1606 1607 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length); 1608 bioq_init(&bq); 1609 1610 chunk_size = sc->chunk_size; 1611 addr = b->bio_data; 1612 offset = b->bio_offset; /* virtual offset and length */ 1613 length = b->bio_length; 1614 1615 while (length > 0) { 1616 size_t chunk_index, in_chunk_offset, in_chunk_length; 1617 struct virstor_map_entry *me; 1618 1619 chunk_index = offset / chunk_size; /* round downwards */ 1620 in_chunk_offset = offset % chunk_size; 1621 in_chunk_length = min(length, chunk_size - in_chunk_offset); 1622 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)", 1623 b->bio_cmd == BIO_READ ? "R" : "W", 1624 offset, length, 1625 chunk_index, in_chunk_offset, in_chunk_length); 1626 me = &sc->map[chunk_index]; 1627 1628 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) { 1629 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) { 1630 /* Reads from unallocated chunks return zeroed 1631 * buffers */ 1632 if (b->bio_cmd == BIO_READ) 1633 bzero(addr, in_chunk_length); 1634 } else { 1635 comp = &sc->components[me->provider_no]; 1636 1637 cb = g_clone_bio(b); 1638 if (cb == NULL) { 1639 bioq_dismantle(&bq); 1640 if (b->bio_error == 0) 1641 b->bio_error = ENOMEM; 1642 g_io_deliver(b, b->bio_error); 1643 return; 1644 } 1645 cb->bio_to = comp->gcons->provider; 1646 cb->bio_done = g_virstor_done; 1647 cb->bio_offset = 1648 (off_t)me->provider_chunk * (off_t)chunk_size 1649 + in_chunk_offset; 1650 cb->bio_length = in_chunk_length; 1651 cb->bio_data = addr; 1652 cb->bio_caller1 = comp; 1653 bioq_disksort(&bq, cb); 1654 } 1655 } else { /* handle BIO_WRITE */ 1656 KASSERT(b->bio_cmd == BIO_WRITE, 1657 ("%s: Unknown command %d", __func__, 1658 b->bio_cmd)); 1659 1660 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) { 1661 /* We have a virtual chunk, represented by 1662 * the "me" entry, but it's not yet allocated 1663 * (tied to) a physical chunk. So do it now. */ 1664 struct virstor_map_entry *data_me; 1665 u_int phys_chunk, comp_no; 1666 off_t s_offset; 1667 int error; 1668 1669 error = allocate_chunk(sc, &comp, &comp_no, 1670 &phys_chunk); 1671 if (error != 0) { 1672 /* We cannot allocate a physical chunk 1673 * to satisfy this request, so we'll 1674 * delay it to when we can... 1675 * XXX: this will prevent the fs from 1676 * being umounted! */ 1677 struct g_virstor_bio_q *biq; 1678 biq = malloc(sizeof *biq, M_GVIRSTOR, 1679 M_NOWAIT); 1680 if (biq == NULL) { 1681 bioq_dismantle(&bq); 1682 if (b->bio_error == 0) 1683 b->bio_error = ENOMEM; 1684 g_io_deliver(b, b->bio_error); 1685 return; 1686 } 1687 biq->bio = b; 1688 mtx_lock(&sc->delayed_bio_q_mtx); 1689 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, 1690 biq, linkage); 1691 mtx_unlock(&sc->delayed_bio_q_mtx); 1692 LOG_MSG(LVL_WARNING, "Delaying BIO " 1693 "(size=%ju) until free physical " 1694 "space can be found on %s", 1695 b->bio_length, 1696 sc->provider->name); 1697 return; 1698 } 1699 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s " 1700 "for %s", 1701 phys_chunk, 1702 comp->gcons->provider->name, 1703 sc->provider->name); 1704 1705 me->provider_no = comp_no; 1706 me->provider_chunk = phys_chunk; 1707 me->flags |= VIRSTOR_MAP_ALLOCATED; 1708 1709 cb = g_clone_bio(b); 1710 if (cb == NULL) { 1711 me->flags &= ~VIRSTOR_MAP_ALLOCATED; 1712 me->provider_no = 0; 1713 me->provider_chunk = 0; 1714 bioq_dismantle(&bq); 1715 if (b->bio_error == 0) 1716 b->bio_error = ENOMEM; 1717 g_io_deliver(b, b->bio_error); 1718 return; 1719 } 1720 1721 /* The allocation table is stored continuously 1722 * at the start of the drive. We need to 1723 * calculate the offset of the sector that holds 1724 * this map entry both on the drive and in the 1725 * map array. 1726 * sc_offset will end up pointing to the drive 1727 * sector. */ 1728 s_offset = chunk_index * sizeof *me; 1729 s_offset = (s_offset / sc->sectorsize) * 1730 sc->sectorsize; 1731 1732 /* data_me points to map entry sector 1733 * in memory (analoguos to offset) */ 1734 data_me = &sc->map[(chunk_index / 1735 sc->me_per_sector) * sc->me_per_sector]; 1736 1737 /* Commit sector with map entry to storage */ 1738 cb->bio_to = sc->components[0].gcons->provider; 1739 cb->bio_done = g_virstor_done; 1740 cb->bio_offset = s_offset; 1741 cb->bio_data = (char *)data_me; 1742 cb->bio_length = sc->sectorsize; 1743 cb->bio_caller1 = &sc->components[0]; 1744 bioq_disksort(&bq, cb); 1745 } 1746 1747 comp = &sc->components[me->provider_no]; 1748 cb = g_clone_bio(b); 1749 if (cb == NULL) { 1750 bioq_dismantle(&bq); 1751 if (b->bio_error == 0) 1752 b->bio_error = ENOMEM; 1753 g_io_deliver(b, b->bio_error); 1754 return; 1755 } 1756 /* Finally, handle the data */ 1757 cb->bio_to = comp->gcons->provider; 1758 cb->bio_done = g_virstor_done; 1759 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size + 1760 in_chunk_offset; 1761 cb->bio_length = in_chunk_length; 1762 cb->bio_data = addr; 1763 cb->bio_caller1 = comp; 1764 bioq_disksort(&bq, cb); 1765 } 1766 addr += in_chunk_length; 1767 length -= in_chunk_length; 1768 offset += in_chunk_length; 1769 } 1770 1771 /* Fire off bio's here */ 1772 count = 0; 1773 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) { 1774 bioq_remove(&bq, cb); 1775 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request"); 1776 comp = cb->bio_caller1; 1777 cb->bio_caller1 = NULL; 1778 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju", 1779 cb->bio_offset, cb->bio_length); 1780 g_io_request(cb, comp->gcons); 1781 count++; 1782 } 1783 if (count == 0) { /* We handled everything locally */ 1784 b->bio_completed = b->bio_length; 1785 g_io_deliver(b, 0); 1786 } 1787 1788} 1789 1790/* 1791 * Allocate a chunk from a physical provider. Returns physical component, 1792 * chunk index relative to the component and the component's index. 1793 */ 1794static int 1795allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp, 1796 u_int *comp_no_p, u_int *chunk) 1797{ 1798 u_int comp_no; 1799 1800 KASSERT(sc->curr_component < sc->n_components, 1801 ("%s: Invalid curr_component: %u", __func__, sc->curr_component)); 1802 1803 comp_no = sc->curr_component; 1804 *comp = &sc->components[comp_no]; 1805 dump_component(*comp); 1806 if ((*comp)->chunk_next >= (*comp)->chunk_count) { 1807 /* This component is full. Allocate next component */ 1808 if (comp_no >= sc->n_components-1) { 1809 LOG_MSG(LVL_ERROR, "All physical space allocated for %s", 1810 sc->geom->name); 1811 return (-1); 1812 } 1813 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT; 1814 sc->curr_component = ++comp_no; 1815 1816 *comp = &sc->components[comp_no]; 1817 if (comp_no >= sc->n_components - g_virstor_component_watermark-1) 1818 LOG_MSG(LVL_WARNING, "Device %s running out of components " 1819 "(switching to %u/%u: %s)", sc->geom->name, 1820 comp_no+1, sc->n_components, 1821 (*comp)->gcons->provider->name); 1822 /* Take care not to overwrite reserved chunks */ 1823 if ( (*comp)->chunk_reserved > 0 && 1824 (*comp)->chunk_next < (*comp)->chunk_reserved) 1825 (*comp)->chunk_next = (*comp)->chunk_reserved; 1826 1827 (*comp)->flags |= 1828 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT; 1829 dump_component(*comp); 1830 *comp_no_p = comp_no; 1831 *chunk = (*comp)->chunk_next++; 1832 } else { 1833 *comp_no_p = comp_no; 1834 *chunk = (*comp)->chunk_next++; 1835 } 1836 return (0); 1837} 1838 1839/* Dump a component */ 1840static void 1841dump_component(struct g_virstor_component *comp) 1842{ 1843 1844 if (g_virstor_debug < LVL_DEBUG2) 1845 return; 1846 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name); 1847 printf(" chunk_count: %u\n", comp->chunk_count); 1848 printf(" chunk_next: %u\n", comp->chunk_next); 1849 printf(" flags: %u\n", comp->flags); 1850} 1851 1852#if 0 1853/* Dump a map entry */ 1854static void 1855dump_me(struct virstor_map_entry *me, unsigned int nr) 1856{ 1857 if (g_virstor_debug < LVL_DEBUG) 1858 return; 1859 printf("VIRT. CHUNK #%d: ", nr); 1860 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) 1861 printf("(unallocated)\n"); 1862 else 1863 printf("allocated at provider %u, provider_chunk %u\n", 1864 me->provider_no, me->provider_chunk); 1865} 1866#endif 1867 1868/* 1869 * Dismantle bio_queue and destroy its components 1870 */ 1871static void 1872bioq_dismantle(struct bio_queue_head *bq) 1873{ 1874 struct bio *b; 1875 1876 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) { 1877 bioq_remove(bq, b); 1878 g_destroy_bio(b); 1879 } 1880} 1881 1882/* 1883 * The function that shouldn't be called. 1884 * When this is called, the stack is already garbled because of 1885 * argument mismatch. There's nothing to do now but panic, which is 1886 * accidentally the whole purpose of this function. 1887 * Motivation: to guard from accidentally calling geom methods when 1888 * they shouldn't be called. (see g_..._taste) 1889 */ 1890static void 1891invalid_call(void) 1892{ 1893 panic("invalid_call() has just been called. Something's fishy here."); 1894} 1895 1896DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */ 1897