1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2004-2005 Pawel Jakub Dawidek <pjd@FreeBSD.org> 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 #include <sys/cdefs.h> 30 __FBSDID("$FreeBSD$"); 31 32 #include <sys/param.h> 33 #include <sys/systm.h> 34 #include <sys/kernel.h> 35 #include <sys/module.h> 36 #include <sys/lock.h> 37 #include <sys/mutex.h> 38 #include <sys/bio.h> 39 #include <sys/sbuf.h> 40 #include <sys/sysctl.h> 41 #include <sys/malloc.h> 42 #include <vm/uma.h> 43 #include <geom/geom.h> 44 #include <geom/geom_dbg.h> 45 #include <geom/stripe/g_stripe.h> 46 47 FEATURE(geom_stripe, "GEOM striping support"); 48 49 static MALLOC_DEFINE(M_STRIPE, "stripe_data", "GEOM_STRIPE Data"); 50 51 static uma_zone_t g_stripe_zone; 52 53 static int g_stripe_destroy(struct g_stripe_softc *sc, boolean_t force); 54 static int g_stripe_destroy_geom(struct gctl_req *req, struct g_class *mp, 55 struct g_geom *gp); 56 57 static g_taste_t g_stripe_taste; 58 static g_ctl_req_t g_stripe_config; 59 static g_dumpconf_t g_stripe_dumpconf; 60 static g_init_t g_stripe_init; 61 static g_fini_t g_stripe_fini; 62 63 struct g_class g_stripe_class = { 64 .name = G_STRIPE_CLASS_NAME, 65 .version = G_VERSION, 66 .ctlreq = g_stripe_config, 67 .taste = g_stripe_taste, 68 .destroy_geom = g_stripe_destroy_geom, 69 .init = g_stripe_init, 70 .fini = g_stripe_fini 71 }; 72 73 SYSCTL_DECL(_kern_geom); 74 static SYSCTL_NODE(_kern_geom, OID_AUTO, stripe, CTLFLAG_RW, 0, 75 "GEOM_STRIPE stuff"); 76 static u_int g_stripe_debug = 0; 77 SYSCTL_UINT(_kern_geom_stripe, OID_AUTO, debug, CTLFLAG_RWTUN, &g_stripe_debug, 0, 78 "Debug level"); 79 static int g_stripe_fast = 0; 80 static int 81 g_sysctl_stripe_fast(SYSCTL_HANDLER_ARGS) 82 { 83 int error, fast; 84 85 fast = g_stripe_fast; 86 error = sysctl_handle_int(oidp, &fast, 0, req); 87 if (error == 0 && req->newptr != NULL) 88 g_stripe_fast = fast; 89 return (error); 90 } 91 SYSCTL_PROC(_kern_geom_stripe, OID_AUTO, fast, CTLTYPE_INT | CTLFLAG_RWTUN, 92 NULL, 0, g_sysctl_stripe_fast, "I", "Fast, but memory-consuming, mode"); 93 static u_int g_stripe_maxmem = MAXPHYS * 100; 94 SYSCTL_UINT(_kern_geom_stripe, OID_AUTO, maxmem, CTLFLAG_RDTUN, &g_stripe_maxmem, 95 0, "Maximum memory that can be allocated in \"fast\" mode (in bytes)"); 96 static u_int g_stripe_fast_failed = 0; 97 SYSCTL_UINT(_kern_geom_stripe, OID_AUTO, fast_failed, CTLFLAG_RD, 98 &g_stripe_fast_failed, 0, "How many times \"fast\" mode failed"); 99 100 /* 101 * Greatest Common Divisor. 102 */ 103 static u_int 104 gcd(u_int a, u_int b) 105 { 106 u_int c; 107 108 while (b != 0) { 109 c = a; 110 a = b; 111 b = (c % b); 112 } 113 return (a); 114 } 115 116 /* 117 * Least Common Multiple. 118 */ 119 static u_int 120 lcm(u_int a, u_int b) 121 { 122 123 return ((a * b) / gcd(a, b)); 124 } 125 126 static void 127 g_stripe_init(struct g_class *mp __unused) 128 { 129 130 g_stripe_zone = uma_zcreate("g_stripe_zone", MAXPHYS, NULL, NULL, 131 NULL, NULL, 0, 0); 132 g_stripe_maxmem -= g_stripe_maxmem % MAXPHYS; 133 uma_zone_set_max(g_stripe_zone, g_stripe_maxmem / MAXPHYS); 134 } 135 136 static void 137 g_stripe_fini(struct g_class *mp __unused) 138 { 139 140 uma_zdestroy(g_stripe_zone); 141 } 142 143 /* 144 * Return the number of valid disks. 145 */ 146 static u_int 147 g_stripe_nvalid(struct g_stripe_softc *sc) 148 { 149 u_int i, no; 150 151 no = 0; 152 for (i = 0; i < sc->sc_ndisks; i++) { 153 if (sc->sc_disks[i] != NULL) 154 no++; 155 } 156 157 return (no); 158 } 159 160 static void 161 g_stripe_remove_disk(struct g_consumer *cp) 162 { 163 struct g_stripe_softc *sc; 164 165 g_topology_assert(); 166 KASSERT(cp != NULL, ("Non-valid disk in %s.", __func__)); 167 sc = (struct g_stripe_softc *)cp->geom->softc; 168 KASSERT(sc != NULL, ("NULL sc in %s.", __func__)); 169 170 if (cp->private == NULL) { 171 G_STRIPE_DEBUG(0, "Disk %s removed from %s.", 172 cp->provider->name, sc->sc_name); 173 cp->private = (void *)(uintptr_t)-1; 174 } 175 176 if (sc->sc_provider != NULL) { 177 G_STRIPE_DEBUG(0, "Device %s deactivated.", 178 sc->sc_provider->name); 179 g_wither_provider(sc->sc_provider, ENXIO); 180 sc->sc_provider = NULL; 181 } 182 183 if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0) 184 return; 185 sc->sc_disks[cp->index] = NULL; 186 cp->index = 0; 187 g_detach(cp); 188 g_destroy_consumer(cp); 189 /* If there are no valid disks anymore, remove device. */ 190 if (LIST_EMPTY(&sc->sc_geom->consumer)) 191 g_stripe_destroy(sc, 1); 192 } 193 194 static void 195 g_stripe_orphan(struct g_consumer *cp) 196 { 197 struct g_stripe_softc *sc; 198 struct g_geom *gp; 199 200 g_topology_assert(); 201 gp = cp->geom; 202 sc = gp->softc; 203 if (sc == NULL) 204 return; 205 206 g_stripe_remove_disk(cp); 207 } 208 209 static int 210 g_stripe_access(struct g_provider *pp, int dr, int dw, int de) 211 { 212 struct g_consumer *cp1, *cp2, *tmp; 213 struct g_stripe_softc *sc; 214 struct g_geom *gp; 215 int error; 216 217 g_topology_assert(); 218 gp = pp->geom; 219 sc = gp->softc; 220 KASSERT(sc != NULL, ("NULL sc in %s.", __func__)); 221 222 /* On first open, grab an extra "exclusive" bit */ 223 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0) 224 de++; 225 /* ... and let go of it on last close */ 226 if ((pp->acr + dr) == 0 && (pp->acw + dw) == 0 && (pp->ace + de) == 0) 227 de--; 228 229 LIST_FOREACH_SAFE(cp1, &gp->consumer, consumer, tmp) { 230 error = g_access(cp1, dr, dw, de); 231 if (error != 0) 232 goto fail; 233 if (cp1->acr == 0 && cp1->acw == 0 && cp1->ace == 0 && 234 cp1->private != NULL) { 235 g_stripe_remove_disk(cp1); /* May destroy geom. */ 236 } 237 } 238 return (0); 239 240 fail: 241 LIST_FOREACH(cp2, &gp->consumer, consumer) { 242 if (cp1 == cp2) 243 break; 244 g_access(cp2, -dr, -dw, -de); 245 } 246 return (error); 247 } 248 249 static void 250 g_stripe_copy(struct g_stripe_softc *sc, char *src, char *dst, off_t offset, 251 off_t length, int mode) 252 { 253 off_t stripesize; 254 size_t len; 255 256 stripesize = sc->sc_stripesize; 257 len = (size_t)(stripesize - (offset & (stripesize - 1))); 258 do { 259 bcopy(src, dst, len); 260 if (mode) { 261 dst += len + stripesize * (sc->sc_ndisks - 1); 262 src += len; 263 } else { 264 dst += len; 265 src += len + stripesize * (sc->sc_ndisks - 1); 266 } 267 length -= len; 268 KASSERT(length >= 0, 269 ("Length < 0 (stripesize=%ju, offset=%ju, length=%jd).", 270 (uintmax_t)stripesize, (uintmax_t)offset, (intmax_t)length)); 271 if (length > stripesize) 272 len = stripesize; 273 else 274 len = length; 275 } while (length > 0); 276 } 277 278 static void 279 g_stripe_done(struct bio *bp) 280 { 281 struct g_stripe_softc *sc; 282 struct bio *pbp; 283 284 pbp = bp->bio_parent; 285 sc = pbp->bio_to->geom->softc; 286 if (bp->bio_cmd == BIO_READ && bp->bio_caller1 != NULL) { 287 g_stripe_copy(sc, bp->bio_data, bp->bio_caller1, bp->bio_offset, 288 bp->bio_length, 1); 289 bp->bio_data = bp->bio_caller1; 290 bp->bio_caller1 = NULL; 291 } 292 mtx_lock(&sc->sc_lock); 293 if (pbp->bio_error == 0) 294 pbp->bio_error = bp->bio_error; 295 pbp->bio_completed += bp->bio_completed; 296 pbp->bio_inbed++; 297 if (pbp->bio_children == pbp->bio_inbed) { 298 mtx_unlock(&sc->sc_lock); 299 if (pbp->bio_driver1 != NULL) 300 uma_zfree(g_stripe_zone, pbp->bio_driver1); 301 g_io_deliver(pbp, pbp->bio_error); 302 } else 303 mtx_unlock(&sc->sc_lock); 304 g_destroy_bio(bp); 305 } 306 307 static int 308 g_stripe_start_fast(struct bio *bp, u_int no, off_t offset, off_t length) 309 { 310 TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue); 311 struct g_stripe_softc *sc; 312 char *addr, *data = NULL; 313 struct bio *cbp; 314 off_t stripesize; 315 u_int nparts = 0; 316 int error; 317 318 sc = bp->bio_to->geom->softc; 319 320 addr = bp->bio_data; 321 stripesize = sc->sc_stripesize; 322 323 cbp = g_clone_bio(bp); 324 if (cbp == NULL) { 325 error = ENOMEM; 326 goto failure; 327 } 328 TAILQ_INSERT_TAIL(&queue, cbp, bio_queue); 329 nparts++; 330 /* 331 * Fill in the component buf structure. 332 */ 333 cbp->bio_done = g_stripe_done; 334 cbp->bio_offset = offset; 335 cbp->bio_data = addr; 336 cbp->bio_caller1 = NULL; 337 cbp->bio_length = length; 338 cbp->bio_caller2 = sc->sc_disks[no]; 339 340 /* offset -= offset % stripesize; */ 341 offset -= offset & (stripesize - 1); 342 addr += length; 343 length = bp->bio_length - length; 344 for (no++; length > 0; no++, length -= stripesize, addr += stripesize) { 345 if (no > sc->sc_ndisks - 1) { 346 no = 0; 347 offset += stripesize; 348 } 349 if (nparts >= sc->sc_ndisks) { 350 cbp = TAILQ_NEXT(cbp, bio_queue); 351 if (cbp == NULL) 352 cbp = TAILQ_FIRST(&queue); 353 nparts++; 354 /* 355 * Update bio structure. 356 */ 357 /* 358 * MIN() is in case when 359 * (bp->bio_length % sc->sc_stripesize) != 0. 360 */ 361 cbp->bio_length += MIN(stripesize, length); 362 if (cbp->bio_caller1 == NULL) { 363 cbp->bio_caller1 = cbp->bio_data; 364 cbp->bio_data = NULL; 365 if (data == NULL) { 366 data = uma_zalloc(g_stripe_zone, 367 M_NOWAIT); 368 if (data == NULL) { 369 error = ENOMEM; 370 goto failure; 371 } 372 } 373 } 374 } else { 375 cbp = g_clone_bio(bp); 376 if (cbp == NULL) { 377 error = ENOMEM; 378 goto failure; 379 } 380 TAILQ_INSERT_TAIL(&queue, cbp, bio_queue); 381 nparts++; 382 /* 383 * Fill in the component buf structure. 384 */ 385 cbp->bio_done = g_stripe_done; 386 cbp->bio_offset = offset; 387 cbp->bio_data = addr; 388 cbp->bio_caller1 = NULL; 389 /* 390 * MIN() is in case when 391 * (bp->bio_length % sc->sc_stripesize) != 0. 392 */ 393 cbp->bio_length = MIN(stripesize, length); 394 cbp->bio_caller2 = sc->sc_disks[no]; 395 } 396 } 397 if (data != NULL) 398 bp->bio_driver1 = data; 399 /* 400 * Fire off all allocated requests! 401 */ 402 while ((cbp = TAILQ_FIRST(&queue)) != NULL) { 403 struct g_consumer *cp; 404 405 TAILQ_REMOVE(&queue, cbp, bio_queue); 406 cp = cbp->bio_caller2; 407 cbp->bio_caller2 = NULL; 408 cbp->bio_to = cp->provider; 409 if (cbp->bio_caller1 != NULL) { 410 cbp->bio_data = data; 411 if (bp->bio_cmd == BIO_WRITE) { 412 g_stripe_copy(sc, cbp->bio_caller1, data, 413 cbp->bio_offset, cbp->bio_length, 0); 414 } 415 data += cbp->bio_length; 416 } 417 G_STRIPE_LOGREQ(cbp, "Sending request."); 418 g_io_request(cbp, cp); 419 } 420 return (0); 421 failure: 422 if (data != NULL) 423 uma_zfree(g_stripe_zone, data); 424 while ((cbp = TAILQ_FIRST(&queue)) != NULL) { 425 TAILQ_REMOVE(&queue, cbp, bio_queue); 426 if (cbp->bio_caller1 != NULL) { 427 cbp->bio_data = cbp->bio_caller1; 428 cbp->bio_caller1 = NULL; 429 } 430 bp->bio_children--; 431 g_destroy_bio(cbp); 432 } 433 return (error); 434 } 435 436 static int 437 g_stripe_start_economic(struct bio *bp, u_int no, off_t offset, off_t length) 438 { 439 TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue); 440 struct g_stripe_softc *sc; 441 off_t stripesize; 442 struct bio *cbp; 443 char *addr; 444 int error; 445 446 sc = bp->bio_to->geom->softc; 447 448 stripesize = sc->sc_stripesize; 449 450 cbp = g_clone_bio(bp); 451 if (cbp == NULL) { 452 error = ENOMEM; 453 goto failure; 454 } 455 TAILQ_INSERT_TAIL(&queue, cbp, bio_queue); 456 /* 457 * Fill in the component buf structure. 458 */ 459 if (bp->bio_length == length) 460 cbp->bio_done = g_std_done; /* Optimized lockless case. */ 461 else 462 cbp->bio_done = g_stripe_done; 463 cbp->bio_offset = offset; 464 cbp->bio_length = length; 465 if ((bp->bio_flags & BIO_UNMAPPED) != 0) { 466 bp->bio_ma_n = round_page(bp->bio_ma_offset + 467 bp->bio_length) / PAGE_SIZE; 468 addr = NULL; 469 } else 470 addr = bp->bio_data; 471 cbp->bio_caller2 = sc->sc_disks[no]; 472 473 /* offset -= offset % stripesize; */ 474 offset -= offset & (stripesize - 1); 475 if (bp->bio_cmd != BIO_DELETE) 476 addr += length; 477 length = bp->bio_length - length; 478 for (no++; length > 0; no++, length -= stripesize) { 479 if (no > sc->sc_ndisks - 1) { 480 no = 0; 481 offset += stripesize; 482 } 483 cbp = g_clone_bio(bp); 484 if (cbp == NULL) { 485 error = ENOMEM; 486 goto failure; 487 } 488 TAILQ_INSERT_TAIL(&queue, cbp, bio_queue); 489 490 /* 491 * Fill in the component buf structure. 492 */ 493 cbp->bio_done = g_stripe_done; 494 cbp->bio_offset = offset; 495 /* 496 * MIN() is in case when 497 * (bp->bio_length % sc->sc_stripesize) != 0. 498 */ 499 cbp->bio_length = MIN(stripesize, length); 500 if ((bp->bio_flags & BIO_UNMAPPED) != 0) { 501 cbp->bio_ma_offset += (uintptr_t)addr; 502 cbp->bio_ma += cbp->bio_ma_offset / PAGE_SIZE; 503 cbp->bio_ma_offset %= PAGE_SIZE; 504 cbp->bio_ma_n = round_page(cbp->bio_ma_offset + 505 cbp->bio_length) / PAGE_SIZE; 506 } else 507 cbp->bio_data = addr; 508 509 cbp->bio_caller2 = sc->sc_disks[no]; 510 511 if (bp->bio_cmd != BIO_DELETE) 512 addr += stripesize; 513 } 514 /* 515 * Fire off all allocated requests! 516 */ 517 while ((cbp = TAILQ_FIRST(&queue)) != NULL) { 518 struct g_consumer *cp; 519 520 TAILQ_REMOVE(&queue, cbp, bio_queue); 521 cp = cbp->bio_caller2; 522 cbp->bio_caller2 = NULL; 523 cbp->bio_to = cp->provider; 524 G_STRIPE_LOGREQ(cbp, "Sending request."); 525 g_io_request(cbp, cp); 526 } 527 return (0); 528 failure: 529 while ((cbp = TAILQ_FIRST(&queue)) != NULL) { 530 TAILQ_REMOVE(&queue, cbp, bio_queue); 531 bp->bio_children--; 532 g_destroy_bio(cbp); 533 } 534 return (error); 535 } 536 537 static void 538 g_stripe_flush(struct g_stripe_softc *sc, struct bio *bp) 539 { 540 struct bio_queue_head queue; 541 struct g_consumer *cp; 542 struct bio *cbp; 543 u_int no; 544 545 bioq_init(&queue); 546 for (no = 0; no < sc->sc_ndisks; no++) { 547 cbp = g_clone_bio(bp); 548 if (cbp == NULL) { 549 for (cbp = bioq_first(&queue); cbp != NULL; 550 cbp = bioq_first(&queue)) { 551 bioq_remove(&queue, cbp); 552 g_destroy_bio(cbp); 553 } 554 if (bp->bio_error == 0) 555 bp->bio_error = ENOMEM; 556 g_io_deliver(bp, bp->bio_error); 557 return; 558 } 559 bioq_insert_tail(&queue, cbp); 560 cbp->bio_done = g_stripe_done; 561 cbp->bio_caller2 = sc->sc_disks[no]; 562 cbp->bio_to = sc->sc_disks[no]->provider; 563 } 564 for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) { 565 bioq_remove(&queue, cbp); 566 G_STRIPE_LOGREQ(cbp, "Sending request."); 567 cp = cbp->bio_caller2; 568 cbp->bio_caller2 = NULL; 569 g_io_request(cbp, cp); 570 } 571 } 572 573 static void 574 g_stripe_start(struct bio *bp) 575 { 576 off_t offset, start, length, nstripe, stripesize; 577 struct g_stripe_softc *sc; 578 u_int no; 579 int error, fast = 0; 580 581 sc = bp->bio_to->geom->softc; 582 /* 583 * If sc == NULL, provider's error should be set and g_stripe_start() 584 * should not be called at all. 585 */ 586 KASSERT(sc != NULL, 587 ("Provider's error should be set (error=%d)(device=%s).", 588 bp->bio_to->error, bp->bio_to->name)); 589 590 G_STRIPE_LOGREQ(bp, "Request received."); 591 592 switch (bp->bio_cmd) { 593 case BIO_READ: 594 case BIO_WRITE: 595 case BIO_DELETE: 596 break; 597 case BIO_FLUSH: 598 g_stripe_flush(sc, bp); 599 return; 600 case BIO_GETATTR: 601 /* To which provider it should be delivered? */ 602 default: 603 g_io_deliver(bp, EOPNOTSUPP); 604 return; 605 } 606 607 stripesize = sc->sc_stripesize; 608 609 /* 610 * Calculations are quite messy, but fast I hope. 611 */ 612 613 /* Stripe number. */ 614 /* nstripe = bp->bio_offset / stripesize; */ 615 nstripe = bp->bio_offset >> (off_t)sc->sc_stripebits; 616 /* Disk number. */ 617 no = nstripe % sc->sc_ndisks; 618 /* Start position in stripe. */ 619 /* start = bp->bio_offset % stripesize; */ 620 start = bp->bio_offset & (stripesize - 1); 621 /* Start position in disk. */ 622 /* offset = (nstripe / sc->sc_ndisks) * stripesize + start; */ 623 offset = ((nstripe / sc->sc_ndisks) << sc->sc_stripebits) + start; 624 /* Length of data to operate. */ 625 length = MIN(bp->bio_length, stripesize - start); 626 627 /* 628 * Do use "fast" mode when: 629 * 1. "Fast" mode is ON. 630 * and 631 * 2. Request size is less than or equal to MAXPHYS, 632 * which should always be true. 633 * and 634 * 3. Request size is bigger than stripesize * ndisks. If it isn't, 635 * there will be no need to send more than one I/O request to 636 * a provider, so there is nothing to optmize. 637 * and 638 * 4. Request is not unmapped. 639 * and 640 * 5. It is not a BIO_DELETE. 641 */ 642 if (g_stripe_fast && bp->bio_length <= MAXPHYS && 643 bp->bio_length >= stripesize * sc->sc_ndisks && 644 (bp->bio_flags & BIO_UNMAPPED) == 0 && 645 bp->bio_cmd != BIO_DELETE) { 646 fast = 1; 647 } 648 error = 0; 649 if (fast) { 650 error = g_stripe_start_fast(bp, no, offset, length); 651 if (error != 0) 652 g_stripe_fast_failed++; 653 } 654 /* 655 * Do use "economic" when: 656 * 1. "Economic" mode is ON. 657 * or 658 * 2. "Fast" mode failed. It can only fail if there is no memory. 659 */ 660 if (!fast || error != 0) 661 error = g_stripe_start_economic(bp, no, offset, length); 662 if (error != 0) { 663 if (bp->bio_error == 0) 664 bp->bio_error = error; 665 g_io_deliver(bp, bp->bio_error); 666 } 667 } 668 669 static void 670 g_stripe_check_and_run(struct g_stripe_softc *sc) 671 { 672 struct g_provider *dp; 673 off_t mediasize, ms; 674 u_int no, sectorsize = 0; 675 676 g_topology_assert(); 677 if (g_stripe_nvalid(sc) != sc->sc_ndisks) 678 return; 679 680 sc->sc_provider = g_new_providerf(sc->sc_geom, "stripe/%s", 681 sc->sc_name); 682 sc->sc_provider->flags |= G_PF_DIRECT_SEND | G_PF_DIRECT_RECEIVE; 683 if (g_stripe_fast == 0) 684 sc->sc_provider->flags |= G_PF_ACCEPT_UNMAPPED; 685 /* 686 * Find the smallest disk. 687 */ 688 mediasize = sc->sc_disks[0]->provider->mediasize; 689 if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC) 690 mediasize -= sc->sc_disks[0]->provider->sectorsize; 691 mediasize -= mediasize % sc->sc_stripesize; 692 sectorsize = sc->sc_disks[0]->provider->sectorsize; 693 for (no = 1; no < sc->sc_ndisks; no++) { 694 dp = sc->sc_disks[no]->provider; 695 ms = dp->mediasize; 696 if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC) 697 ms -= dp->sectorsize; 698 ms -= ms % sc->sc_stripesize; 699 if (ms < mediasize) 700 mediasize = ms; 701 sectorsize = lcm(sectorsize, dp->sectorsize); 702 703 /* A provider underneath us doesn't support unmapped */ 704 if ((dp->flags & G_PF_ACCEPT_UNMAPPED) == 0) { 705 G_STRIPE_DEBUG(1, "Cancelling unmapped " 706 "because of %s.", dp->name); 707 sc->sc_provider->flags &= ~G_PF_ACCEPT_UNMAPPED; 708 } 709 } 710 sc->sc_provider->sectorsize = sectorsize; 711 sc->sc_provider->mediasize = mediasize * sc->sc_ndisks; 712 sc->sc_provider->stripesize = sc->sc_stripesize; 713 sc->sc_provider->stripeoffset = 0; 714 g_error_provider(sc->sc_provider, 0); 715 716 G_STRIPE_DEBUG(0, "Device %s activated.", sc->sc_provider->name); 717 } 718 719 static int 720 g_stripe_read_metadata(struct g_consumer *cp, struct g_stripe_metadata *md) 721 { 722 struct g_provider *pp; 723 u_char *buf; 724 int error; 725 726 g_topology_assert(); 727 728 error = g_access(cp, 1, 0, 0); 729 if (error != 0) 730 return (error); 731 pp = cp->provider; 732 g_topology_unlock(); 733 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize, 734 &error); 735 g_topology_lock(); 736 g_access(cp, -1, 0, 0); 737 if (buf == NULL) 738 return (error); 739 740 /* Decode metadata. */ 741 stripe_metadata_decode(buf, md); 742 g_free(buf); 743 744 return (0); 745 } 746 747 /* 748 * Add disk to given device. 749 */ 750 static int 751 g_stripe_add_disk(struct g_stripe_softc *sc, struct g_provider *pp, u_int no) 752 { 753 struct g_consumer *cp, *fcp; 754 struct g_geom *gp; 755 int error; 756 757 g_topology_assert(); 758 /* Metadata corrupted? */ 759 if (no >= sc->sc_ndisks) 760 return (EINVAL); 761 762 /* Check if disk is not already attached. */ 763 if (sc->sc_disks[no] != NULL) 764 return (EEXIST); 765 766 gp = sc->sc_geom; 767 fcp = LIST_FIRST(&gp->consumer); 768 769 cp = g_new_consumer(gp); 770 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 771 cp->private = NULL; 772 cp->index = no; 773 error = g_attach(cp, pp); 774 if (error != 0) { 775 g_destroy_consumer(cp); 776 return (error); 777 } 778 779 if (fcp != NULL && (fcp->acr > 0 || fcp->acw > 0 || fcp->ace > 0)) { 780 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace); 781 if (error != 0) { 782 g_detach(cp); 783 g_destroy_consumer(cp); 784 return (error); 785 } 786 } 787 if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC) { 788 struct g_stripe_metadata md; 789 790 /* Reread metadata. */ 791 error = g_stripe_read_metadata(cp, &md); 792 if (error != 0) 793 goto fail; 794 795 if (strcmp(md.md_magic, G_STRIPE_MAGIC) != 0 || 796 strcmp(md.md_name, sc->sc_name) != 0 || 797 md.md_id != sc->sc_id) { 798 G_STRIPE_DEBUG(0, "Metadata on %s changed.", pp->name); 799 goto fail; 800 } 801 } 802 803 sc->sc_disks[no] = cp; 804 G_STRIPE_DEBUG(0, "Disk %s attached to %s.", pp->name, sc->sc_name); 805 g_stripe_check_and_run(sc); 806 807 return (0); 808 fail: 809 if (fcp != NULL && (fcp->acr > 0 || fcp->acw > 0 || fcp->ace > 0)) 810 g_access(cp, -fcp->acr, -fcp->acw, -fcp->ace); 811 g_detach(cp); 812 g_destroy_consumer(cp); 813 return (error); 814 } 815 816 static struct g_geom * 817 g_stripe_create(struct g_class *mp, const struct g_stripe_metadata *md, 818 u_int type) 819 { 820 struct g_stripe_softc *sc; 821 struct g_geom *gp; 822 u_int no; 823 824 g_topology_assert(); 825 G_STRIPE_DEBUG(1, "Creating device %s (id=%u).", md->md_name, 826 md->md_id); 827 828 /* Two disks is minimum. */ 829 if (md->md_all < 2) { 830 G_STRIPE_DEBUG(0, "Too few disks defined for %s.", md->md_name); 831 return (NULL); 832 } 833 #if 0 834 /* Stripe size have to be grater than or equal to sector size. */ 835 if (md->md_stripesize < sectorsize) { 836 G_STRIPE_DEBUG(0, "Invalid stripe size for %s.", md->md_name); 837 return (NULL); 838 } 839 #endif 840 /* Stripe size have to be power of 2. */ 841 if (!powerof2(md->md_stripesize)) { 842 G_STRIPE_DEBUG(0, "Invalid stripe size for %s.", md->md_name); 843 return (NULL); 844 } 845 846 /* Check for duplicate unit */ 847 LIST_FOREACH(gp, &mp->geom, geom) { 848 sc = gp->softc; 849 if (sc != NULL && strcmp(sc->sc_name, md->md_name) == 0) { 850 G_STRIPE_DEBUG(0, "Device %s already configured.", 851 sc->sc_name); 852 return (NULL); 853 } 854 } 855 gp = g_new_geomf(mp, "%s", md->md_name); 856 sc = malloc(sizeof(*sc), M_STRIPE, M_WAITOK | M_ZERO); 857 gp->start = g_stripe_start; 858 gp->spoiled = g_stripe_orphan; 859 gp->orphan = g_stripe_orphan; 860 gp->access = g_stripe_access; 861 gp->dumpconf = g_stripe_dumpconf; 862 863 sc->sc_id = md->md_id; 864 sc->sc_stripesize = md->md_stripesize; 865 sc->sc_stripebits = bitcount32(sc->sc_stripesize - 1); 866 sc->sc_ndisks = md->md_all; 867 sc->sc_disks = malloc(sizeof(struct g_consumer *) * sc->sc_ndisks, 868 M_STRIPE, M_WAITOK | M_ZERO); 869 for (no = 0; no < sc->sc_ndisks; no++) 870 sc->sc_disks[no] = NULL; 871 sc->sc_type = type; 872 mtx_init(&sc->sc_lock, "gstripe lock", NULL, MTX_DEF); 873 874 gp->softc = sc; 875 sc->sc_geom = gp; 876 sc->sc_provider = NULL; 877 878 G_STRIPE_DEBUG(0, "Device %s created (id=%u).", sc->sc_name, sc->sc_id); 879 880 return (gp); 881 } 882 883 static int 884 g_stripe_destroy(struct g_stripe_softc *sc, boolean_t force) 885 { 886 struct g_provider *pp; 887 struct g_consumer *cp, *cp1; 888 struct g_geom *gp; 889 890 g_topology_assert(); 891 892 if (sc == NULL) 893 return (ENXIO); 894 895 pp = sc->sc_provider; 896 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) { 897 if (force) { 898 G_STRIPE_DEBUG(0, "Device %s is still open, so it " 899 "can't be definitely removed.", pp->name); 900 } else { 901 G_STRIPE_DEBUG(1, 902 "Device %s is still open (r%dw%de%d).", pp->name, 903 pp->acr, pp->acw, pp->ace); 904 return (EBUSY); 905 } 906 } 907 908 gp = sc->sc_geom; 909 LIST_FOREACH_SAFE(cp, &gp->consumer, consumer, cp1) { 910 g_stripe_remove_disk(cp); 911 if (cp1 == NULL) 912 return (0); /* Recursion happened. */ 913 } 914 if (!LIST_EMPTY(&gp->consumer)) 915 return (EINPROGRESS); 916 917 gp->softc = NULL; 918 KASSERT(sc->sc_provider == NULL, ("Provider still exists? (device=%s)", 919 gp->name)); 920 free(sc->sc_disks, M_STRIPE); 921 mtx_destroy(&sc->sc_lock); 922 free(sc, M_STRIPE); 923 G_STRIPE_DEBUG(0, "Device %s destroyed.", gp->name); 924 g_wither_geom(gp, ENXIO); 925 return (0); 926 } 927 928 static int 929 g_stripe_destroy_geom(struct gctl_req *req __unused, 930 struct g_class *mp __unused, struct g_geom *gp) 931 { 932 struct g_stripe_softc *sc; 933 934 sc = gp->softc; 935 return (g_stripe_destroy(sc, 0)); 936 } 937 938 static struct g_geom * 939 g_stripe_taste(struct g_class *mp, struct g_provider *pp, int flags __unused) 940 { 941 struct g_stripe_metadata md; 942 struct g_stripe_softc *sc; 943 struct g_consumer *cp; 944 struct g_geom *gp; 945 int error; 946 947 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name); 948 g_topology_assert(); 949 950 /* Skip providers that are already open for writing. */ 951 if (pp->acw > 0) 952 return (NULL); 953 954 G_STRIPE_DEBUG(3, "Tasting %s.", pp->name); 955 956 gp = g_new_geomf(mp, "stripe:taste"); 957 gp->start = g_stripe_start; 958 gp->access = g_stripe_access; 959 gp->orphan = g_stripe_orphan; 960 cp = g_new_consumer(gp); 961 g_attach(cp, pp); 962 error = g_stripe_read_metadata(cp, &md); 963 g_detach(cp); 964 g_destroy_consumer(cp); 965 g_destroy_geom(gp); 966 if (error != 0) 967 return (NULL); 968 gp = NULL; 969 970 if (strcmp(md.md_magic, G_STRIPE_MAGIC) != 0) 971 return (NULL); 972 if (md.md_version > G_STRIPE_VERSION) { 973 printf("geom_stripe.ko module is too old to handle %s.\n", 974 pp->name); 975 return (NULL); 976 } 977 /* 978 * Backward compatibility: 979 */ 980 /* There was no md_provider field in earlier versions of metadata. */ 981 if (md.md_version < 2) 982 bzero(md.md_provider, sizeof(md.md_provider)); 983 /* There was no md_provsize field in earlier versions of metadata. */ 984 if (md.md_version < 3) 985 md.md_provsize = pp->mediasize; 986 987 if (md.md_provider[0] != '\0' && 988 !g_compare_names(md.md_provider, pp->name)) 989 return (NULL); 990 if (md.md_provsize != pp->mediasize) 991 return (NULL); 992 993 /* 994 * Let's check if device already exists. 995 */ 996 sc = NULL; 997 LIST_FOREACH(gp, &mp->geom, geom) { 998 sc = gp->softc; 999 if (sc == NULL) 1000 continue; 1001 if (sc->sc_type != G_STRIPE_TYPE_AUTOMATIC) 1002 continue; 1003 if (strcmp(md.md_name, sc->sc_name) != 0) 1004 continue; 1005 if (md.md_id != sc->sc_id) 1006 continue; 1007 break; 1008 } 1009 if (gp != NULL) { 1010 G_STRIPE_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name); 1011 error = g_stripe_add_disk(sc, pp, md.md_no); 1012 if (error != 0) { 1013 G_STRIPE_DEBUG(0, 1014 "Cannot add disk %s to %s (error=%d).", pp->name, 1015 gp->name, error); 1016 return (NULL); 1017 } 1018 } else { 1019 gp = g_stripe_create(mp, &md, G_STRIPE_TYPE_AUTOMATIC); 1020 if (gp == NULL) { 1021 G_STRIPE_DEBUG(0, "Cannot create device %s.", 1022 md.md_name); 1023 return (NULL); 1024 } 1025 sc = gp->softc; 1026 G_STRIPE_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name); 1027 error = g_stripe_add_disk(sc, pp, md.md_no); 1028 if (error != 0) { 1029 G_STRIPE_DEBUG(0, 1030 "Cannot add disk %s to %s (error=%d).", pp->name, 1031 gp->name, error); 1032 g_stripe_destroy(sc, 1); 1033 return (NULL); 1034 } 1035 } 1036 1037 return (gp); 1038 } 1039 1040 static void 1041 g_stripe_ctl_create(struct gctl_req *req, struct g_class *mp) 1042 { 1043 u_int attached, no; 1044 struct g_stripe_metadata md; 1045 struct g_provider *pp; 1046 struct g_stripe_softc *sc; 1047 struct g_geom *gp; 1048 struct sbuf *sb; 1049 off_t *stripesize; 1050 const char *name; 1051 char param[16]; 1052 int *nargs; 1053 1054 g_topology_assert(); 1055 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs)); 1056 if (nargs == NULL) { 1057 gctl_error(req, "No '%s' argument.", "nargs"); 1058 return; 1059 } 1060 if (*nargs <= 2) { 1061 gctl_error(req, "Too few arguments."); 1062 return; 1063 } 1064 1065 strlcpy(md.md_magic, G_STRIPE_MAGIC, sizeof(md.md_magic)); 1066 md.md_version = G_STRIPE_VERSION; 1067 name = gctl_get_asciiparam(req, "arg0"); 1068 if (name == NULL) { 1069 gctl_error(req, "No 'arg%u' argument.", 0); 1070 return; 1071 } 1072 strlcpy(md.md_name, name, sizeof(md.md_name)); 1073 md.md_id = arc4random(); 1074 md.md_no = 0; 1075 md.md_all = *nargs - 1; 1076 stripesize = gctl_get_paraml(req, "stripesize", sizeof(*stripesize)); 1077 if (stripesize == NULL) { 1078 gctl_error(req, "No '%s' argument.", "stripesize"); 1079 return; 1080 } 1081 md.md_stripesize = (uint32_t)*stripesize; 1082 bzero(md.md_provider, sizeof(md.md_provider)); 1083 /* This field is not important here. */ 1084 md.md_provsize = 0; 1085 1086 /* Check all providers are valid */ 1087 for (no = 1; no < *nargs; no++) { 1088 snprintf(param, sizeof(param), "arg%u", no); 1089 name = gctl_get_asciiparam(req, param); 1090 if (name == NULL) { 1091 gctl_error(req, "No 'arg%u' argument.", no); 1092 return; 1093 } 1094 if (strncmp(name, "/dev/", strlen("/dev/")) == 0) 1095 name += strlen("/dev/"); 1096 pp = g_provider_by_name(name); 1097 if (pp == NULL) { 1098 G_STRIPE_DEBUG(1, "Disk %s is invalid.", name); 1099 gctl_error(req, "Disk %s is invalid.", name); 1100 return; 1101 } 1102 } 1103 1104 gp = g_stripe_create(mp, &md, G_STRIPE_TYPE_MANUAL); 1105 if (gp == NULL) { 1106 gctl_error(req, "Can't configure %s.", md.md_name); 1107 return; 1108 } 1109 1110 sc = gp->softc; 1111 sb = sbuf_new_auto(); 1112 sbuf_printf(sb, "Can't attach disk(s) to %s:", gp->name); 1113 for (attached = 0, no = 1; no < *nargs; no++) { 1114 snprintf(param, sizeof(param), "arg%u", no); 1115 name = gctl_get_asciiparam(req, param); 1116 if (name == NULL) { 1117 gctl_error(req, "No 'arg%u' argument.", no); 1118 continue; 1119 } 1120 if (strncmp(name, "/dev/", strlen("/dev/")) == 0) 1121 name += strlen("/dev/"); 1122 pp = g_provider_by_name(name); 1123 KASSERT(pp != NULL, ("Provider %s disappear?!", name)); 1124 if (g_stripe_add_disk(sc, pp, no - 1) != 0) { 1125 G_STRIPE_DEBUG(1, "Disk %u (%s) not attached to %s.", 1126 no, pp->name, gp->name); 1127 sbuf_printf(sb, " %s", pp->name); 1128 continue; 1129 } 1130 attached++; 1131 } 1132 sbuf_finish(sb); 1133 if (md.md_all != attached) { 1134 g_stripe_destroy(gp->softc, 1); 1135 gctl_error(req, "%s", sbuf_data(sb)); 1136 } 1137 sbuf_delete(sb); 1138 } 1139 1140 static struct g_stripe_softc * 1141 g_stripe_find_device(struct g_class *mp, const char *name) 1142 { 1143 struct g_stripe_softc *sc; 1144 struct g_geom *gp; 1145 1146 LIST_FOREACH(gp, &mp->geom, geom) { 1147 sc = gp->softc; 1148 if (sc == NULL) 1149 continue; 1150 if (strcmp(sc->sc_name, name) == 0) 1151 return (sc); 1152 } 1153 return (NULL); 1154 } 1155 1156 static void 1157 g_stripe_ctl_destroy(struct gctl_req *req, struct g_class *mp) 1158 { 1159 struct g_stripe_softc *sc; 1160 int *force, *nargs, error; 1161 const char *name; 1162 char param[16]; 1163 u_int i; 1164 1165 g_topology_assert(); 1166 1167 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs)); 1168 if (nargs == NULL) { 1169 gctl_error(req, "No '%s' argument.", "nargs"); 1170 return; 1171 } 1172 if (*nargs <= 0) { 1173 gctl_error(req, "Missing device(s)."); 1174 return; 1175 } 1176 force = gctl_get_paraml(req, "force", sizeof(*force)); 1177 if (force == NULL) { 1178 gctl_error(req, "No '%s' argument.", "force"); 1179 return; 1180 } 1181 1182 for (i = 0; i < (u_int)*nargs; i++) { 1183 snprintf(param, sizeof(param), "arg%u", i); 1184 name = gctl_get_asciiparam(req, param); 1185 if (name == NULL) { 1186 gctl_error(req, "No 'arg%u' argument.", i); 1187 return; 1188 } 1189 sc = g_stripe_find_device(mp, name); 1190 if (sc == NULL) { 1191 gctl_error(req, "No such device: %s.", name); 1192 return; 1193 } 1194 error = g_stripe_destroy(sc, *force); 1195 if (error != 0) { 1196 gctl_error(req, "Cannot destroy device %s (error=%d).", 1197 sc->sc_name, error); 1198 return; 1199 } 1200 } 1201 } 1202 1203 static void 1204 g_stripe_config(struct gctl_req *req, struct g_class *mp, const char *verb) 1205 { 1206 uint32_t *version; 1207 1208 g_topology_assert(); 1209 1210 version = gctl_get_paraml(req, "version", sizeof(*version)); 1211 if (version == NULL) { 1212 gctl_error(req, "No '%s' argument.", "version"); 1213 return; 1214 } 1215 if (*version != G_STRIPE_VERSION) { 1216 gctl_error(req, "Userland and kernel parts are out of sync."); 1217 return; 1218 } 1219 1220 if (strcmp(verb, "create") == 0) { 1221 g_stripe_ctl_create(req, mp); 1222 return; 1223 } else if (strcmp(verb, "destroy") == 0 || 1224 strcmp(verb, "stop") == 0) { 1225 g_stripe_ctl_destroy(req, mp); 1226 return; 1227 } 1228 1229 gctl_error(req, "Unknown verb."); 1230 } 1231 1232 static void 1233 g_stripe_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp, 1234 struct g_consumer *cp, struct g_provider *pp) 1235 { 1236 struct g_stripe_softc *sc; 1237 1238 sc = gp->softc; 1239 if (sc == NULL) 1240 return; 1241 if (pp != NULL) { 1242 /* Nothing here. */ 1243 } else if (cp != NULL) { 1244 sbuf_printf(sb, "%s<Number>%u</Number>\n", indent, 1245 (u_int)cp->index); 1246 } else { 1247 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id); 1248 sbuf_printf(sb, "%s<Stripesize>%ju</Stripesize>\n", indent, 1249 (uintmax_t)sc->sc_stripesize); 1250 sbuf_printf(sb, "%s<Type>", indent); 1251 switch (sc->sc_type) { 1252 case G_STRIPE_TYPE_AUTOMATIC: 1253 sbuf_cat(sb, "AUTOMATIC"); 1254 break; 1255 case G_STRIPE_TYPE_MANUAL: 1256 sbuf_cat(sb, "MANUAL"); 1257 break; 1258 default: 1259 sbuf_cat(sb, "UNKNOWN"); 1260 break; 1261 } 1262 sbuf_cat(sb, "</Type>\n"); 1263 sbuf_printf(sb, "%s<Status>Total=%u, Online=%u</Status>\n", 1264 indent, sc->sc_ndisks, g_stripe_nvalid(sc)); 1265 sbuf_printf(sb, "%s<State>", indent); 1266 if (sc->sc_provider != NULL && sc->sc_provider->error == 0) 1267 sbuf_cat(sb, "UP"); 1268 else 1269 sbuf_cat(sb, "DOWN"); 1270 sbuf_cat(sb, "</State>\n"); 1271 } 1272 } 1273 1274 DECLARE_GEOM_CLASS(g_stripe_class, g_stripe); 1275 MODULE_VERSION(geom_stripe, 0); 1276