geom_io.c revision 287405
1/*- 2 * Copyright (c) 2002 Poul-Henning Kamp 3 * Copyright (c) 2002 Networks Associates Technology, Inc. 4 * Copyright (c) 2013 The FreeBSD Foundation 5 * All rights reserved. 6 * 7 * This software was developed for the FreeBSD Project by Poul-Henning Kamp 8 * and NAI Labs, the Security Research Division of Network Associates, Inc. 9 * under DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 10 * DARPA CHATS research program. 11 * 12 * Portions of this software were developed by Konstantin Belousov 13 * under sponsorship from the FreeBSD Foundation. 14 * 15 * Redistribution and use in source and binary forms, with or without 16 * modification, are permitted provided that the following conditions 17 * are met: 18 * 1. Redistributions of source code must retain the above copyright 19 * notice, this list of conditions and the following disclaimer. 20 * 2. Redistributions in binary form must reproduce the above copyright 21 * notice, this list of conditions and the following disclaimer in the 22 * documentation and/or other materials provided with the distribution. 23 * 3. The names of the authors may not be used to endorse or promote 24 * products derived from this software without specific prior written 25 * permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 30 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 37 * SUCH DAMAGE. 38 */ 39 40#include <sys/cdefs.h> 41__FBSDID("$FreeBSD: head/sys/geom/geom_io.c 287405 2015-09-02 17:29:30Z imp $"); 42 43#include <sys/param.h> 44#include <sys/systm.h> 45#include <sys/kernel.h> 46#include <sys/malloc.h> 47#include <sys/bio.h> 48#include <sys/ktr.h> 49#include <sys/proc.h> 50#include <sys/stack.h> 51#include <sys/sysctl.h> 52#include <sys/vmem.h> 53 54#include <sys/errno.h> 55#include <geom/geom.h> 56#include <geom/geom_int.h> 57#include <sys/devicestat.h> 58 59#include <vm/uma.h> 60#include <vm/vm.h> 61#include <vm/vm_param.h> 62#include <vm/vm_kern.h> 63#include <vm/vm_page.h> 64#include <vm/vm_object.h> 65#include <vm/vm_extern.h> 66#include <vm/vm_map.h> 67 68static int g_io_transient_map_bio(struct bio *bp); 69 70static struct g_bioq g_bio_run_down; 71static struct g_bioq g_bio_run_up; 72static struct g_bioq g_bio_run_task; 73 74/* 75 * Pace is a hint that we've had some trouble recently allocating 76 * bios, so we should back off trying to send I/O down the stack 77 * a bit to let the problem resolve. When pacing, we also turn 78 * off direct dispatch to also reduce memory pressure from I/Os 79 * there, at the expxense of some added latency while the memory 80 * pressures exist. See g_io_schedule_down() for more details 81 * and limitations. 82 */ 83static volatile u_int pace; 84 85static uma_zone_t biozone; 86 87/* 88 * The head of the list of classifiers used in g_io_request. 89 * Use g_register_classifier() and g_unregister_classifier() 90 * to add/remove entries to the list. 91 * Classifiers are invoked in registration order. 92 */ 93static TAILQ_HEAD(g_classifier_tailq, g_classifier_hook) 94 g_classifier_tailq = TAILQ_HEAD_INITIALIZER(g_classifier_tailq); 95 96#include <machine/atomic.h> 97 98static void 99g_bioq_lock(struct g_bioq *bq) 100{ 101 102 mtx_lock(&bq->bio_queue_lock); 103} 104 105static void 106g_bioq_unlock(struct g_bioq *bq) 107{ 108 109 mtx_unlock(&bq->bio_queue_lock); 110} 111 112#if 0 113static void 114g_bioq_destroy(struct g_bioq *bq) 115{ 116 117 mtx_destroy(&bq->bio_queue_lock); 118} 119#endif 120 121static void 122g_bioq_init(struct g_bioq *bq) 123{ 124 125 TAILQ_INIT(&bq->bio_queue); 126 mtx_init(&bq->bio_queue_lock, "bio queue", NULL, MTX_DEF); 127} 128 129static struct bio * 130g_bioq_first(struct g_bioq *bq) 131{ 132 struct bio *bp; 133 134 bp = TAILQ_FIRST(&bq->bio_queue); 135 if (bp != NULL) { 136 KASSERT((bp->bio_flags & BIO_ONQUEUE), 137 ("Bio not on queue bp=%p target %p", bp, bq)); 138 bp->bio_flags &= ~BIO_ONQUEUE; 139 TAILQ_REMOVE(&bq->bio_queue, bp, bio_queue); 140 bq->bio_queue_length--; 141 } 142 return (bp); 143} 144 145struct bio * 146g_new_bio(void) 147{ 148 struct bio *bp; 149 150 bp = uma_zalloc(biozone, M_NOWAIT | M_ZERO); 151#ifdef KTR 152 if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) { 153 struct stack st; 154 155 CTR1(KTR_GEOM, "g_new_bio(): %p", bp); 156 stack_save(&st); 157 CTRSTACK(KTR_GEOM, &st, 3, 0); 158 } 159#endif 160 return (bp); 161} 162 163struct bio * 164g_alloc_bio(void) 165{ 166 struct bio *bp; 167 168 bp = uma_zalloc(biozone, M_WAITOK | M_ZERO); 169#ifdef KTR 170 if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) { 171 struct stack st; 172 173 CTR1(KTR_GEOM, "g_alloc_bio(): %p", bp); 174 stack_save(&st); 175 CTRSTACK(KTR_GEOM, &st, 3, 0); 176 } 177#endif 178 return (bp); 179} 180 181void 182g_destroy_bio(struct bio *bp) 183{ 184#ifdef KTR 185 if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) { 186 struct stack st; 187 188 CTR1(KTR_GEOM, "g_destroy_bio(): %p", bp); 189 stack_save(&st); 190 CTRSTACK(KTR_GEOM, &st, 3, 0); 191 } 192#endif 193 uma_zfree(biozone, bp); 194} 195 196struct bio * 197g_clone_bio(struct bio *bp) 198{ 199 struct bio *bp2; 200 201 bp2 = uma_zalloc(biozone, M_NOWAIT | M_ZERO); 202 if (bp2 != NULL) { 203 bp2->bio_parent = bp; 204 bp2->bio_cmd = bp->bio_cmd; 205 /* 206 * BIO_ORDERED flag may be used by disk drivers to enforce 207 * ordering restrictions, so this flag needs to be cloned. 208 * BIO_UNMAPPED should be inherited, to properly indicate 209 * which way the buffer is passed. 210 * Other bio flags are not suitable for cloning. 211 */ 212 bp2->bio_flags = bp->bio_flags & (BIO_ORDERED | BIO_UNMAPPED); 213 bp2->bio_length = bp->bio_length; 214 bp2->bio_offset = bp->bio_offset; 215 bp2->bio_data = bp->bio_data; 216 bp2->bio_ma = bp->bio_ma; 217 bp2->bio_ma_n = bp->bio_ma_n; 218 bp2->bio_ma_offset = bp->bio_ma_offset; 219 bp2->bio_attribute = bp->bio_attribute; 220 /* Inherit classification info from the parent */ 221 bp2->bio_classifier1 = bp->bio_classifier1; 222 bp2->bio_classifier2 = bp->bio_classifier2; 223 bp->bio_children++; 224 } 225#ifdef KTR 226 if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) { 227 struct stack st; 228 229 CTR2(KTR_GEOM, "g_clone_bio(%p): %p", bp, bp2); 230 stack_save(&st); 231 CTRSTACK(KTR_GEOM, &st, 3, 0); 232 } 233#endif 234 return(bp2); 235} 236 237struct bio * 238g_duplicate_bio(struct bio *bp) 239{ 240 struct bio *bp2; 241 242 bp2 = uma_zalloc(biozone, M_WAITOK | M_ZERO); 243 bp2->bio_flags = bp->bio_flags & BIO_UNMAPPED; 244 bp2->bio_parent = bp; 245 bp2->bio_cmd = bp->bio_cmd; 246 bp2->bio_length = bp->bio_length; 247 bp2->bio_offset = bp->bio_offset; 248 bp2->bio_data = bp->bio_data; 249 bp2->bio_ma = bp->bio_ma; 250 bp2->bio_ma_n = bp->bio_ma_n; 251 bp2->bio_ma_offset = bp->bio_ma_offset; 252 bp2->bio_attribute = bp->bio_attribute; 253 bp->bio_children++; 254#ifdef KTR 255 if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) { 256 struct stack st; 257 258 CTR2(KTR_GEOM, "g_duplicate_bio(%p): %p", bp, bp2); 259 stack_save(&st); 260 CTRSTACK(KTR_GEOM, &st, 3, 0); 261 } 262#endif 263 return(bp2); 264} 265 266void 267g_io_init() 268{ 269 270 g_bioq_init(&g_bio_run_down); 271 g_bioq_init(&g_bio_run_up); 272 g_bioq_init(&g_bio_run_task); 273 biozone = uma_zcreate("g_bio", sizeof (struct bio), 274 NULL, NULL, 275 NULL, NULL, 276 0, 0); 277} 278 279int 280g_io_getattr(const char *attr, struct g_consumer *cp, int *len, void *ptr) 281{ 282 struct bio *bp; 283 int error; 284 285 g_trace(G_T_BIO, "bio_getattr(%s)", attr); 286 bp = g_alloc_bio(); 287 bp->bio_cmd = BIO_GETATTR; 288 bp->bio_done = NULL; 289 bp->bio_attribute = attr; 290 bp->bio_length = *len; 291 bp->bio_data = ptr; 292 g_io_request(bp, cp); 293 error = biowait(bp, "ggetattr"); 294 *len = bp->bio_completed; 295 g_destroy_bio(bp); 296 return (error); 297} 298 299int 300g_io_flush(struct g_consumer *cp) 301{ 302 struct bio *bp; 303 int error; 304 305 g_trace(G_T_BIO, "bio_flush(%s)", cp->provider->name); 306 bp = g_alloc_bio(); 307 bp->bio_cmd = BIO_FLUSH; 308 bp->bio_flags |= BIO_ORDERED; 309 bp->bio_done = NULL; 310 bp->bio_attribute = NULL; 311 bp->bio_offset = cp->provider->mediasize; 312 bp->bio_length = 0; 313 bp->bio_data = NULL; 314 g_io_request(bp, cp); 315 error = biowait(bp, "gflush"); 316 g_destroy_bio(bp); 317 return (error); 318} 319 320static int 321g_io_check(struct bio *bp) 322{ 323 struct g_consumer *cp; 324 struct g_provider *pp; 325 off_t excess; 326 int error; 327 328 cp = bp->bio_from; 329 pp = bp->bio_to; 330 331 /* Fail if access counters dont allow the operation */ 332 switch(bp->bio_cmd) { 333 case BIO_READ: 334 case BIO_GETATTR: 335 if (cp->acr == 0) 336 return (EPERM); 337 break; 338 case BIO_WRITE: 339 case BIO_DELETE: 340 case BIO_FLUSH: 341 if (cp->acw == 0) 342 return (EPERM); 343 break; 344 default: 345 return (EPERM); 346 } 347 /* if provider is marked for error, don't disturb. */ 348 if (pp->error) 349 return (pp->error); 350 if (cp->flags & G_CF_ORPHAN) 351 return (ENXIO); 352 353 switch(bp->bio_cmd) { 354 case BIO_READ: 355 case BIO_WRITE: 356 case BIO_DELETE: 357 /* Zero sectorsize or mediasize is probably a lack of media. */ 358 if (pp->sectorsize == 0 || pp->mediasize == 0) 359 return (ENXIO); 360 /* Reject I/O not on sector boundary */ 361 if (bp->bio_offset % pp->sectorsize) 362 return (EINVAL); 363 /* Reject I/O not integral sector long */ 364 if (bp->bio_length % pp->sectorsize) 365 return (EINVAL); 366 /* Reject requests before or past the end of media. */ 367 if (bp->bio_offset < 0) 368 return (EIO); 369 if (bp->bio_offset > pp->mediasize) 370 return (EIO); 371 372 /* Truncate requests to the end of providers media. */ 373 excess = bp->bio_offset + bp->bio_length; 374 if (excess > bp->bio_to->mediasize) { 375 KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 || 376 round_page(bp->bio_ma_offset + 377 bp->bio_length) / PAGE_SIZE == bp->bio_ma_n, 378 ("excess bio %p too short", bp)); 379 excess -= bp->bio_to->mediasize; 380 bp->bio_length -= excess; 381 if ((bp->bio_flags & BIO_UNMAPPED) != 0) { 382 bp->bio_ma_n = round_page(bp->bio_ma_offset + 383 bp->bio_length) / PAGE_SIZE; 384 } 385 if (excess > 0) 386 CTR3(KTR_GEOM, "g_down truncated bio " 387 "%p provider %s by %d", bp, 388 bp->bio_to->name, excess); 389 } 390 391 /* Deliver zero length transfers right here. */ 392 if (bp->bio_length == 0) { 393 CTR2(KTR_GEOM, "g_down terminated 0-length " 394 "bp %p provider %s", bp, bp->bio_to->name); 395 return (0); 396 } 397 398 if ((bp->bio_flags & BIO_UNMAPPED) != 0 && 399 (bp->bio_to->flags & G_PF_ACCEPT_UNMAPPED) == 0 && 400 (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE)) { 401 if ((error = g_io_transient_map_bio(bp)) >= 0) 402 return (error); 403 } 404 break; 405 default: 406 break; 407 } 408 return (EJUSTRETURN); 409} 410 411/* 412 * bio classification support. 413 * 414 * g_register_classifier() and g_unregister_classifier() 415 * are used to add/remove a classifier from the list. 416 * The list is protected using the g_bio_run_down lock, 417 * because the classifiers are called in this path. 418 * 419 * g_io_request() passes bio's that are not already classified 420 * (i.e. those with bio_classifier1 == NULL) to g_run_classifiers(). 421 * Classifiers can store their result in the two fields 422 * bio_classifier1 and bio_classifier2. 423 * A classifier that updates one of the fields should 424 * return a non-zero value. 425 * If no classifier updates the field, g_run_classifiers() sets 426 * bio_classifier1 = BIO_NOTCLASSIFIED to avoid further calls. 427 */ 428 429int 430g_register_classifier(struct g_classifier_hook *hook) 431{ 432 433 g_bioq_lock(&g_bio_run_down); 434 TAILQ_INSERT_TAIL(&g_classifier_tailq, hook, link); 435 g_bioq_unlock(&g_bio_run_down); 436 437 return (0); 438} 439 440void 441g_unregister_classifier(struct g_classifier_hook *hook) 442{ 443 struct g_classifier_hook *entry; 444 445 g_bioq_lock(&g_bio_run_down); 446 TAILQ_FOREACH(entry, &g_classifier_tailq, link) { 447 if (entry == hook) { 448 TAILQ_REMOVE(&g_classifier_tailq, hook, link); 449 break; 450 } 451 } 452 g_bioq_unlock(&g_bio_run_down); 453} 454 455static void 456g_run_classifiers(struct bio *bp) 457{ 458 struct g_classifier_hook *hook; 459 int classified = 0; 460 461 TAILQ_FOREACH(hook, &g_classifier_tailq, link) 462 classified |= hook->func(hook->arg, bp); 463 464 if (!classified) 465 bp->bio_classifier1 = BIO_NOTCLASSIFIED; 466} 467 468void 469g_io_request(struct bio *bp, struct g_consumer *cp) 470{ 471 struct g_provider *pp; 472 struct mtx *mtxp; 473 int direct, error, first; 474 475 KASSERT(cp != NULL, ("NULL cp in g_io_request")); 476 KASSERT(bp != NULL, ("NULL bp in g_io_request")); 477 pp = cp->provider; 478 KASSERT(pp != NULL, ("consumer not attached in g_io_request")); 479#ifdef DIAGNOSTIC 480 KASSERT(bp->bio_driver1 == NULL, 481 ("bio_driver1 used by the consumer (geom %s)", cp->geom->name)); 482 KASSERT(bp->bio_driver2 == NULL, 483 ("bio_driver2 used by the consumer (geom %s)", cp->geom->name)); 484 KASSERT(bp->bio_pflags == 0, 485 ("bio_pflags used by the consumer (geom %s)", cp->geom->name)); 486 /* 487 * Remember consumer's private fields, so we can detect if they were 488 * modified by the provider. 489 */ 490 bp->_bio_caller1 = bp->bio_caller1; 491 bp->_bio_caller2 = bp->bio_caller2; 492 bp->_bio_cflags = bp->bio_cflags; 493#endif 494 495 if (bp->bio_cmd & (BIO_READ|BIO_WRITE|BIO_GETATTR)) { 496 KASSERT(bp->bio_data != NULL, 497 ("NULL bp->data in g_io_request(cmd=%hhu)", bp->bio_cmd)); 498 } 499 if (bp->bio_cmd & (BIO_DELETE|BIO_FLUSH)) { 500 KASSERT(bp->bio_data == NULL, 501 ("non-NULL bp->data in g_io_request(cmd=%hhu)", 502 bp->bio_cmd)); 503 } 504 if (bp->bio_cmd & (BIO_READ|BIO_WRITE|BIO_DELETE)) { 505 KASSERT(bp->bio_offset % cp->provider->sectorsize == 0, 506 ("wrong offset %jd for sectorsize %u", 507 bp->bio_offset, cp->provider->sectorsize)); 508 KASSERT(bp->bio_length % cp->provider->sectorsize == 0, 509 ("wrong length %jd for sectorsize %u", 510 bp->bio_length, cp->provider->sectorsize)); 511 } 512 513 g_trace(G_T_BIO, "bio_request(%p) from %p(%s) to %p(%s) cmd %d", 514 bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd); 515 516 bp->bio_from = cp; 517 bp->bio_to = pp; 518 bp->bio_error = 0; 519 bp->bio_completed = 0; 520 521 KASSERT(!(bp->bio_flags & BIO_ONQUEUE), 522 ("Bio already on queue bp=%p", bp)); 523 if ((g_collectstats & G_STATS_CONSUMERS) != 0 || 524 ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL)) 525 binuptime(&bp->bio_t0); 526 else 527 getbinuptime(&bp->bio_t0); 528 529#ifdef GET_STACK_USAGE 530 direct = (cp->flags & G_CF_DIRECT_SEND) != 0 && 531 (pp->flags & G_PF_DIRECT_RECEIVE) != 0 && 532 !g_is_geom_thread(curthread) && 533 ((pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 || 534 (bp->bio_flags & BIO_UNMAPPED) == 0 || THREAD_CAN_SLEEP()) && 535 pace == 0; 536 if (direct) { 537 /* Block direct execution if less then half of stack left. */ 538 size_t st, su; 539 GET_STACK_USAGE(st, su); 540 if (su * 2 > st) 541 direct = 0; 542 } 543#else 544 direct = 0; 545#endif 546 547 if (!TAILQ_EMPTY(&g_classifier_tailq) && !bp->bio_classifier1) { 548 g_bioq_lock(&g_bio_run_down); 549 g_run_classifiers(bp); 550 g_bioq_unlock(&g_bio_run_down); 551 } 552 553 /* 554 * The statistics collection is lockless, as such, but we 555 * can not update one instance of the statistics from more 556 * than one thread at a time, so grab the lock first. 557 */ 558 mtxp = mtx_pool_find(mtxpool_sleep, pp); 559 mtx_lock(mtxp); 560 if (g_collectstats & G_STATS_PROVIDERS) 561 devstat_start_transaction(pp->stat, &bp->bio_t0); 562 if (g_collectstats & G_STATS_CONSUMERS) 563 devstat_start_transaction(cp->stat, &bp->bio_t0); 564 pp->nstart++; 565 cp->nstart++; 566 mtx_unlock(mtxp); 567 568 if (direct) { 569 error = g_io_check(bp); 570 if (error >= 0) { 571 CTR3(KTR_GEOM, "g_io_request g_io_check on bp %p " 572 "provider %s returned %d", bp, bp->bio_to->name, 573 error); 574 g_io_deliver(bp, error); 575 return; 576 } 577 bp->bio_to->geom->start(bp); 578 } else { 579 g_bioq_lock(&g_bio_run_down); 580 first = TAILQ_EMPTY(&g_bio_run_down.bio_queue); 581 TAILQ_INSERT_TAIL(&g_bio_run_down.bio_queue, bp, bio_queue); 582 bp->bio_flags |= BIO_ONQUEUE; 583 g_bio_run_down.bio_queue_length++; 584 g_bioq_unlock(&g_bio_run_down); 585 /* Pass it on down. */ 586 if (first) 587 wakeup(&g_wait_down); 588 } 589} 590 591void 592g_io_deliver(struct bio *bp, int error) 593{ 594 struct bintime now; 595 struct g_consumer *cp; 596 struct g_provider *pp; 597 struct mtx *mtxp; 598 int direct, first; 599 600 KASSERT(bp != NULL, ("NULL bp in g_io_deliver")); 601 pp = bp->bio_to; 602 KASSERT(pp != NULL, ("NULL bio_to in g_io_deliver")); 603 cp = bp->bio_from; 604 if (cp == NULL) { 605 bp->bio_error = error; 606 bp->bio_done(bp); 607 return; 608 } 609 KASSERT(cp != NULL, ("NULL bio_from in g_io_deliver")); 610 KASSERT(cp->geom != NULL, ("NULL bio_from->geom in g_io_deliver")); 611#ifdef DIAGNOSTIC 612 /* 613 * Some classes - GJournal in particular - can modify bio's 614 * private fields while the bio is in transit; G_GEOM_VOLATILE_BIO 615 * flag means it's an expected behaviour for that particular geom. 616 */ 617 if ((cp->geom->flags & G_GEOM_VOLATILE_BIO) == 0) { 618 KASSERT(bp->bio_caller1 == bp->_bio_caller1, 619 ("bio_caller1 used by the provider %s", pp->name)); 620 KASSERT(bp->bio_caller2 == bp->_bio_caller2, 621 ("bio_caller2 used by the provider %s", pp->name)); 622 KASSERT(bp->bio_cflags == bp->_bio_cflags, 623 ("bio_cflags used by the provider %s", pp->name)); 624 } 625#endif 626 KASSERT(bp->bio_completed >= 0, ("bio_completed can't be less than 0")); 627 KASSERT(bp->bio_completed <= bp->bio_length, 628 ("bio_completed can't be greater than bio_length")); 629 630 g_trace(G_T_BIO, 631"g_io_deliver(%p) from %p(%s) to %p(%s) cmd %d error %d off %jd len %jd", 632 bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd, error, 633 (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length); 634 635 KASSERT(!(bp->bio_flags & BIO_ONQUEUE), 636 ("Bio already on queue bp=%p", bp)); 637 638 /* 639 * XXX: next two doesn't belong here 640 */ 641 bp->bio_bcount = bp->bio_length; 642 bp->bio_resid = bp->bio_bcount - bp->bio_completed; 643 644#ifdef GET_STACK_USAGE 645 direct = (pp->flags & G_PF_DIRECT_SEND) && 646 (cp->flags & G_CF_DIRECT_RECEIVE) && 647 !g_is_geom_thread(curthread); 648 if (direct) { 649 /* Block direct execution if less then half of stack left. */ 650 size_t st, su; 651 GET_STACK_USAGE(st, su); 652 if (su * 2 > st) 653 direct = 0; 654 } 655#else 656 direct = 0; 657#endif 658 659 /* 660 * The statistics collection is lockless, as such, but we 661 * can not update one instance of the statistics from more 662 * than one thread at a time, so grab the lock first. 663 */ 664 if ((g_collectstats & G_STATS_CONSUMERS) != 0 || 665 ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL)) 666 binuptime(&now); 667 mtxp = mtx_pool_find(mtxpool_sleep, cp); 668 mtx_lock(mtxp); 669 if (g_collectstats & G_STATS_PROVIDERS) 670 devstat_end_transaction_bio_bt(pp->stat, bp, &now); 671 if (g_collectstats & G_STATS_CONSUMERS) 672 devstat_end_transaction_bio_bt(cp->stat, bp, &now); 673 cp->nend++; 674 pp->nend++; 675 mtx_unlock(mtxp); 676 677 if (error != ENOMEM) { 678 bp->bio_error = error; 679 if (direct) { 680 biodone(bp); 681 } else { 682 g_bioq_lock(&g_bio_run_up); 683 first = TAILQ_EMPTY(&g_bio_run_up.bio_queue); 684 TAILQ_INSERT_TAIL(&g_bio_run_up.bio_queue, bp, bio_queue); 685 bp->bio_flags |= BIO_ONQUEUE; 686 g_bio_run_up.bio_queue_length++; 687 g_bioq_unlock(&g_bio_run_up); 688 if (first) 689 wakeup(&g_wait_up); 690 } 691 return; 692 } 693 694 if (bootverbose) 695 printf("ENOMEM %p on %p(%s)\n", bp, pp, pp->name); 696 bp->bio_children = 0; 697 bp->bio_inbed = 0; 698 bp->bio_driver1 = NULL; 699 bp->bio_driver2 = NULL; 700 bp->bio_pflags = 0; 701 g_io_request(bp, cp); 702 pace = 1; 703 return; 704} 705 706SYSCTL_DECL(_kern_geom); 707 708static long transient_maps; 709SYSCTL_LONG(_kern_geom, OID_AUTO, transient_maps, CTLFLAG_RD, 710 &transient_maps, 0, 711 "Total count of the transient mapping requests"); 712u_int transient_map_retries = 10; 713SYSCTL_UINT(_kern_geom, OID_AUTO, transient_map_retries, CTLFLAG_RW, 714 &transient_map_retries, 0, 715 "Max count of retries used before giving up on creating transient map"); 716int transient_map_hard_failures; 717SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_hard_failures, CTLFLAG_RD, 718 &transient_map_hard_failures, 0, 719 "Failures to establish the transient mapping due to retry attempts " 720 "exhausted"); 721int transient_map_soft_failures; 722SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_soft_failures, CTLFLAG_RD, 723 &transient_map_soft_failures, 0, 724 "Count of retried failures to establish the transient mapping"); 725int inflight_transient_maps; 726SYSCTL_INT(_kern_geom, OID_AUTO, inflight_transient_maps, CTLFLAG_RD, 727 &inflight_transient_maps, 0, 728 "Current count of the active transient maps"); 729 730static int 731g_io_transient_map_bio(struct bio *bp) 732{ 733 vm_offset_t addr; 734 long size; 735 u_int retried; 736 737 KASSERT(unmapped_buf_allowed, ("unmapped disabled")); 738 739 size = round_page(bp->bio_ma_offset + bp->bio_length); 740 KASSERT(size / PAGE_SIZE == bp->bio_ma_n, ("Bio too short %p", bp)); 741 addr = 0; 742 retried = 0; 743 atomic_add_long(&transient_maps, 1); 744retry: 745 if (vmem_alloc(transient_arena, size, M_BESTFIT | M_NOWAIT, &addr)) { 746 if (transient_map_retries != 0 && 747 retried >= transient_map_retries) { 748 CTR2(KTR_GEOM, "g_down cannot map bp %p provider %s", 749 bp, bp->bio_to->name); 750 atomic_add_int(&transient_map_hard_failures, 1); 751 return (EDEADLK/* XXXKIB */); 752 } else { 753 /* 754 * Naive attempt to quisce the I/O to get more 755 * in-flight requests completed and defragment 756 * the transient_arena. 757 */ 758 CTR3(KTR_GEOM, "g_down retrymap bp %p provider %s r %d", 759 bp, bp->bio_to->name, retried); 760 pause("g_d_tra", hz / 10); 761 retried++; 762 atomic_add_int(&transient_map_soft_failures, 1); 763 goto retry; 764 } 765 } 766 atomic_add_int(&inflight_transient_maps, 1); 767 pmap_qenter((vm_offset_t)addr, bp->bio_ma, OFF_TO_IDX(size)); 768 bp->bio_data = (caddr_t)addr + bp->bio_ma_offset; 769 bp->bio_flags |= BIO_TRANSIENT_MAPPING; 770 bp->bio_flags &= ~BIO_UNMAPPED; 771 return (EJUSTRETURN); 772} 773 774void 775g_io_schedule_down(struct thread *tp __unused) 776{ 777 struct bio *bp; 778 int error; 779 780 for(;;) { 781 g_bioq_lock(&g_bio_run_down); 782 bp = g_bioq_first(&g_bio_run_down); 783 if (bp == NULL) { 784 CTR0(KTR_GEOM, "g_down going to sleep"); 785 msleep(&g_wait_down, &g_bio_run_down.bio_queue_lock, 786 PRIBIO | PDROP, "-", 0); 787 continue; 788 } 789 CTR0(KTR_GEOM, "g_down has work to do"); 790 g_bioq_unlock(&g_bio_run_down); 791 if (pace != 0) { 792 /* 793 * There has been at least one memory allocation 794 * failure since the last I/O completed. Pause 1ms to 795 * give the system a chance to free up memory. We only 796 * do this once because a large number of allocations 797 * can fail in the direct dispatch case and there's no 798 * relationship between the number of these failures and 799 * the length of the outage. If there's still an outage, 800 * we'll pause again and again until it's 801 * resolved. Older versions paused longer and once per 802 * allocation failure. This was OK for a single threaded 803 * g_down, but with direct dispatch would lead to max of 804 * 10 IOPs for minutes at a time when transient memory 805 * issues prevented allocation for a batch of requests 806 * from the upper layers. 807 * 808 * XXX This pacing is really lame. It needs to be solved 809 * by other methods. This is OK only because the worst 810 * case scenario is so rare. In the worst case scenario 811 * all memory is tied up waiting for I/O to complete 812 * which can never happen since we can't allocate bios 813 * for that I/O. 814 */ 815 CTR0(KTR_GEOM, "g_down pacing self"); 816 pause("g_down", min(hz/1000, 1)); 817 pace = 0; 818 } 819 CTR2(KTR_GEOM, "g_down processing bp %p provider %s", bp, 820 bp->bio_to->name); 821 error = g_io_check(bp); 822 if (error >= 0) { 823 CTR3(KTR_GEOM, "g_down g_io_check on bp %p provider " 824 "%s returned %d", bp, bp->bio_to->name, error); 825 g_io_deliver(bp, error); 826 continue; 827 } 828 THREAD_NO_SLEEPING(); 829 CTR4(KTR_GEOM, "g_down starting bp %p provider %s off %ld " 830 "len %ld", bp, bp->bio_to->name, bp->bio_offset, 831 bp->bio_length); 832 bp->bio_to->geom->start(bp); 833 THREAD_SLEEPING_OK(); 834 } 835} 836 837void 838bio_taskqueue(struct bio *bp, bio_task_t *func, void *arg) 839{ 840 bp->bio_task = func; 841 bp->bio_task_arg = arg; 842 /* 843 * The taskqueue is actually just a second queue off the "up" 844 * queue, so we use the same lock. 845 */ 846 g_bioq_lock(&g_bio_run_up); 847 KASSERT(!(bp->bio_flags & BIO_ONQUEUE), 848 ("Bio already on queue bp=%p target taskq", bp)); 849 bp->bio_flags |= BIO_ONQUEUE; 850 TAILQ_INSERT_TAIL(&g_bio_run_task.bio_queue, bp, bio_queue); 851 g_bio_run_task.bio_queue_length++; 852 wakeup(&g_wait_up); 853 g_bioq_unlock(&g_bio_run_up); 854} 855 856 857void 858g_io_schedule_up(struct thread *tp __unused) 859{ 860 struct bio *bp; 861 for(;;) { 862 g_bioq_lock(&g_bio_run_up); 863 bp = g_bioq_first(&g_bio_run_task); 864 if (bp != NULL) { 865 g_bioq_unlock(&g_bio_run_up); 866 THREAD_NO_SLEEPING(); 867 CTR1(KTR_GEOM, "g_up processing task bp %p", bp); 868 bp->bio_task(bp->bio_task_arg); 869 THREAD_SLEEPING_OK(); 870 continue; 871 } 872 bp = g_bioq_first(&g_bio_run_up); 873 if (bp != NULL) { 874 g_bioq_unlock(&g_bio_run_up); 875 THREAD_NO_SLEEPING(); 876 CTR4(KTR_GEOM, "g_up biodone bp %p provider %s off " 877 "%jd len %ld", bp, bp->bio_to->name, 878 bp->bio_offset, bp->bio_length); 879 biodone(bp); 880 THREAD_SLEEPING_OK(); 881 continue; 882 } 883 CTR0(KTR_GEOM, "g_up going to sleep"); 884 msleep(&g_wait_up, &g_bio_run_up.bio_queue_lock, 885 PRIBIO | PDROP, "-", 0); 886 } 887} 888 889void * 890g_read_data(struct g_consumer *cp, off_t offset, off_t length, int *error) 891{ 892 struct bio *bp; 893 void *ptr; 894 int errorc; 895 896 KASSERT(length > 0 && length >= cp->provider->sectorsize && 897 length <= MAXPHYS, ("g_read_data(): invalid length %jd", 898 (intmax_t)length)); 899 900 bp = g_alloc_bio(); 901 bp->bio_cmd = BIO_READ; 902 bp->bio_done = NULL; 903 bp->bio_offset = offset; 904 bp->bio_length = length; 905 ptr = g_malloc(length, M_WAITOK); 906 bp->bio_data = ptr; 907 g_io_request(bp, cp); 908 errorc = biowait(bp, "gread"); 909 if (error != NULL) 910 *error = errorc; 911 g_destroy_bio(bp); 912 if (errorc) { 913 g_free(ptr); 914 ptr = NULL; 915 } 916 return (ptr); 917} 918 919int 920g_write_data(struct g_consumer *cp, off_t offset, void *ptr, off_t length) 921{ 922 struct bio *bp; 923 int error; 924 925 KASSERT(length > 0 && length >= cp->provider->sectorsize && 926 length <= MAXPHYS, ("g_write_data(): invalid length %jd", 927 (intmax_t)length)); 928 929 bp = g_alloc_bio(); 930 bp->bio_cmd = BIO_WRITE; 931 bp->bio_done = NULL; 932 bp->bio_offset = offset; 933 bp->bio_length = length; 934 bp->bio_data = ptr; 935 g_io_request(bp, cp); 936 error = biowait(bp, "gwrite"); 937 g_destroy_bio(bp); 938 return (error); 939} 940 941int 942g_delete_data(struct g_consumer *cp, off_t offset, off_t length) 943{ 944 struct bio *bp; 945 int error; 946 947 KASSERT(length > 0 && length >= cp->provider->sectorsize, 948 ("g_delete_data(): invalid length %jd", (intmax_t)length)); 949 950 bp = g_alloc_bio(); 951 bp->bio_cmd = BIO_DELETE; 952 bp->bio_done = NULL; 953 bp->bio_offset = offset; 954 bp->bio_length = length; 955 bp->bio_data = NULL; 956 g_io_request(bp, cp); 957 error = biowait(bp, "gdelete"); 958 g_destroy_bio(bp); 959 return (error); 960} 961 962void 963g_print_bio(struct bio *bp) 964{ 965 const char *pname, *cmd = NULL; 966 967 if (bp->bio_to != NULL) 968 pname = bp->bio_to->name; 969 else 970 pname = "[unknown]"; 971 972 switch (bp->bio_cmd) { 973 case BIO_GETATTR: 974 cmd = "GETATTR"; 975 printf("%s[%s(attr=%s)]", pname, cmd, bp->bio_attribute); 976 return; 977 case BIO_FLUSH: 978 cmd = "FLUSH"; 979 printf("%s[%s]", pname, cmd); 980 return; 981 case BIO_READ: 982 cmd = "READ"; 983 break; 984 case BIO_WRITE: 985 cmd = "WRITE"; 986 break; 987 case BIO_DELETE: 988 cmd = "DELETE"; 989 break; 990 default: 991 cmd = "UNKNOWN"; 992 printf("%s[%s()]", pname, cmd); 993 return; 994 } 995 printf("%s[%s(offset=%jd, length=%jd)]", pname, cmd, 996 (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length); 997} 998