1 /*- 2 * Copyright (c) 1997, 1998, 1999 3 * Nan Yang Computer Services Limited. All rights reserved. 4 * 5 * Parts copyright (c) 1997, 1998 Cybernet Corporation, NetMAX project. 6 * 7 * Written by Greg Lehey 8 * 9 * This software is distributed under the so-called ``Berkeley 10 * License'': 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by Nan Yang Computer 23 * Services Limited. 24 * 4. Neither the name of the Company nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * This software is provided ``as is'', and any express or implied 29 * warranties, including, but not limited to, the implied warranties of 30 * merchantability and fitness for a particular purpose are disclaimed. 31 * In no event shall the company or contributors be liable for any 32 * direct, indirect, incidental, special, exemplary, or consequential 33 * damages (including, but not limited to, procurement of substitute 34 * goods or services; loss of use, data, or profits; or business 35 * interruption) however caused and on any theory of liability, whether 36 * in contract, strict liability, or tort (including negligence or 37 * otherwise) arising in any way out of the use of this software, even if 38 * advised of the possibility of such damage. 39 * 40 * $Id: vinumrequest.c,v 1.30 2001/01/09 04:20:55 grog Exp grog $ 41 * $FreeBSD: src/sys/dev/vinum/vinumrequest.c,v 1.44.2.5 2002/08/28 04:30:56 grog Exp $ 42 * $DragonFly: src/sys/dev/raid/vinum/vinumrequest.c,v 1.12 2006/05/03 20:44:49 dillon Exp $ 43 */ 44 45 #include "vinumhdr.h" 46 #include "request.h" 47 #include <sys/resourcevar.h> 48 49 enum requeststatus bre(struct request *rq, 50 int plexno, 51 daddr_t * diskstart, 52 daddr_t diskend); 53 enum requeststatus bre5(struct request *rq, 54 int plexno, 55 daddr_t * diskstart, 56 daddr_t diskend); 57 enum requeststatus build_read_request(struct request *rq, int volplexno); 58 enum requeststatus build_write_request(struct request *rq); 59 enum requeststatus build_rq_buffer(struct rqelement *rqe, struct plex *plex); 60 int find_alternate_sd(struct request *rq); 61 int check_range_covered(struct request *); 62 void complete_rqe(struct bio *bio); 63 void complete_raid5_write(struct rqelement *); 64 int abortrequest(struct request *rq, int error); 65 void sdio_done(struct bio *bio); 66 struct bio *vinum_bounds_check(struct bio *bio, struct volume *vol); 67 caddr_t allocdatabuf(struct rqelement *rqe); 68 void freedatabuf(struct rqelement *rqe); 69 70 #ifdef VINUMDEBUG 71 struct rqinfo rqinfo[RQINFO_SIZE]; 72 struct rqinfo *rqip = rqinfo; 73 74 void 75 logrq(enum rqinfo_type type, union rqinfou info, struct bio *ubio) 76 { 77 dev_t dev; 78 79 crit_enter(); 80 81 microtime(&rqip->timestamp); /* when did this happen? */ 82 rqip->type = type; 83 rqip->bio = ubio; /* user buffer */ 84 85 switch (type) { 86 case loginfo_user_bp: 87 case loginfo_user_bpl: 88 case loginfo_sdio: /* subdisk I/O */ 89 case loginfo_sdiol: /* subdisk I/O launch */ 90 case loginfo_sdiodone: /* subdisk I/O complete */ 91 bcopy(info.bio, &rqip->info.bio, sizeof(struct bio)); 92 dev = info.bio->bio_driver_info; 93 rqip->devmajor = major(dev); 94 rqip->devminor = minor(dev); 95 break; 96 97 case loginfo_iodone: 98 case loginfo_rqe: 99 case loginfo_raid5_data: 100 case loginfo_raid5_parity: 101 bcopy(info.rqe, &rqip->info.rqe, sizeof(struct rqelement)); 102 dev = info.rqe->b.b_bio1.bio_driver_info; 103 rqip->devmajor = major(dev); 104 rqip->devminor = minor(dev); 105 break; 106 107 case loginfo_lockwait: 108 case loginfo_lock: 109 case loginfo_unlock: 110 bcopy(info.lockinfo, &rqip->info.lockinfo, sizeof(struct rangelock)); 111 112 break; 113 114 case loginfo_unused: 115 break; 116 } 117 rqip++; 118 if (rqip >= &rqinfo[RQINFO_SIZE]) /* wrap around */ 119 rqip = rqinfo; 120 crit_exit(); 121 } 122 123 #endif 124 125 void 126 vinumstrategy(dev_t dev, struct bio *bio) 127 { 128 struct buf *bp = bio->bio_buf; 129 struct bio *nbio = bio; 130 struct volume *vol = NULL; 131 int volno; 132 133 switch (DEVTYPE(dev)) { 134 case VINUM_SD_TYPE: 135 case VINUM_RAWSD_TYPE: 136 bio->bio_driver_info = dev; 137 sdio(bio); 138 return; 139 140 /* 141 * In fact, vinum doesn't handle drives: they're 142 * handled directly by the disk drivers 143 */ 144 case VINUM_DRIVE_TYPE: 145 default: 146 bp->b_error = EIO; /* I/O error */ 147 bp->b_flags |= B_ERROR; 148 biodone(bio); 149 return; 150 151 case VINUM_VOLUME_TYPE: /* volume I/O */ 152 volno = Volno(dev); 153 vol = &VOL[volno]; 154 if (vol->state != volume_up) { /* can't access this volume */ 155 bp->b_error = EIO; /* I/O error */ 156 bp->b_flags |= B_ERROR; 157 biodone(bio); 158 return; 159 } 160 nbio = vinum_bounds_check(bio, vol); 161 if (nbio == NULL) { 162 biodone(bio); 163 return; 164 } 165 /* FALLTHROUGH */ 166 /* 167 * Plex I/O is pretty much the same as volume I/O 168 * for a single plex. Indicate this by passing a NULL 169 * pointer (set above) for the volume 170 */ 171 case VINUM_PLEX_TYPE: 172 case VINUM_RAWPLEX_TYPE: 173 bp->b_resid = bp->b_bcount; /* transfer everything */ 174 vinumstart(dev, nbio, 0); 175 return; 176 } 177 } 178 179 /* 180 * Start a transfer. Return -1 on error, 181 * 0 if OK, 1 if we need to retry. 182 * Parameter reviveok is set when doing 183 * transfers for revives: it allows transfers to 184 * be started immediately when a revive is in 185 * progress. During revive, normal transfers 186 * are queued if they share address space with 187 * a currently active revive operation. 188 */ 189 int 190 vinumstart(dev_t dev, struct bio *bio, int reviveok) 191 { 192 struct buf *bp = bio->bio_buf; 193 int plexno; 194 int maxplex; /* maximum number of plexes to handle */ 195 struct volume *vol; 196 struct request *rq; /* build up our request here */ 197 enum requeststatus status; 198 199 bio->bio_driver_info = dev; 200 201 #if VINUMDEBUG 202 if (debug & DEBUG_LASTREQS) 203 logrq(loginfo_user_bp, (union rqinfou) bio, bio); 204 #endif 205 206 if ((bp->b_bcount % DEV_BSIZE) != 0) { /* bad length */ 207 bp->b_error = EINVAL; /* invalid size */ 208 bp->b_flags |= B_ERROR; 209 biodone(bio); 210 return -1; 211 } 212 rq = (struct request *) Malloc(sizeof(struct request)); /* allocate a request struct */ 213 if (rq == NULL) { /* can't do it */ 214 bp->b_error = ENOMEM; /* can't get memory */ 215 bp->b_flags |= B_ERROR; 216 biodone(bio); 217 return -1; 218 } 219 bzero(rq, sizeof(struct request)); 220 221 /* 222 * Note the volume ID. This can be NULL, which 223 * the request building functions use as an 224 * indication for single plex I/O 225 */ 226 rq->bio = bio; /* and the user buffer struct */ 227 228 if (DEVTYPE(dev) == VINUM_VOLUME_TYPE) { /* it's a volume, */ 229 rq->volplex.volno = Volno(dev); /* get the volume number */ 230 vol = &VOL[rq->volplex.volno]; /* and point to it */ 231 vol->active++; /* one more active request */ 232 maxplex = vol->plexes; /* consider all its plexes */ 233 } else { 234 vol = NULL; /* no volume */ 235 rq->volplex.plexno = Plexno(dev); /* point to the plex */ 236 rq->isplex = 1; /* note that it's a plex */ 237 maxplex = 1; /* just the one plex */ 238 } 239 240 if (bp->b_cmd == BUF_CMD_READ) { 241 /* 242 * This is a read request. Decide 243 * which plex to read from. 244 * 245 * There's a potential race condition here, 246 * since we're not locked, and we could end 247 * up multiply incrementing the round-robin 248 * counter. This doesn't have any serious 249 * effects, however. 250 */ 251 if (vol != NULL) { 252 plexno = vol->preferred_plex; /* get the plex to use */ 253 if (plexno < 0) { /* round robin */ 254 plexno = vol->last_plex_read; 255 vol->last_plex_read++; 256 if (vol->last_plex_read >= vol->plexes) /* got the the end? */ 257 vol->last_plex_read = 0; /* wrap around */ 258 } 259 status = build_read_request(rq, plexno); /* build a request */ 260 } else { 261 daddr_t diskaddr = (daddr_t)(bio->bio_offset >> DEV_BSHIFT); 262 /* start offset of transfer */ 263 status = bre(rq, /* build a request list */ 264 rq->volplex.plexno, 265 &diskaddr, 266 diskaddr + (bp->b_bcount / DEV_BSIZE)); 267 } 268 269 if (status > REQUEST_RECOVERED) { /* can't satisfy it */ 270 if (status == REQUEST_DOWN) { /* not enough subdisks */ 271 bp->b_error = EIO; /* I/O error */ 272 bp->b_flags |= B_ERROR; 273 } 274 biodone(bio); 275 freerq(rq); 276 return -1; 277 } 278 return launch_requests(rq, reviveok); /* now start the requests if we can */ 279 } else 280 /* 281 * This is a write operation. We write to all plexes. If this is 282 * a RAID-4 or RAID-5 plex, we must also update the parity stripe. 283 */ 284 { 285 if (vol != NULL) 286 status = build_write_request(rq); /* Not all the subdisks are up */ 287 else { /* plex I/O */ 288 daddr_t diskstart; 289 daddr_t diskend; 290 291 diskstart = (daddr_t)(bio->bio_offset >> DEV_BSHIFT); /* start offset of transfer */ 292 diskend = diskstart + bp->b_bcount / DEV_BSIZE; 293 status = bre(rq, Plexno(dev), 294 &diskstart, diskend); /* build requests for the plex */ 295 } 296 if (status > REQUEST_RECOVERED) { /* can't satisfy it */ 297 if (status == REQUEST_DOWN) { /* not enough subdisks */ 298 bp->b_error = EIO; /* I/O error */ 299 bp->b_flags |= B_ERROR; 300 } 301 biodone(bio); 302 freerq(rq); 303 return -1; 304 } 305 return launch_requests(rq, reviveok); /* now start the requests if we can */ 306 } 307 } 308 309 /* 310 * Call the low-level strategy routines to 311 * perform the requests in a struct request 312 */ 313 int 314 launch_requests(struct request *rq, int reviveok) 315 { 316 struct rqgroup *rqg; 317 int rqno; /* loop index */ 318 struct rqelement *rqe; /* current element */ 319 struct drive *drive; 320 int rcount; /* request count */ 321 322 /* 323 * First find out whether we're reviving, and the 324 * request contains a conflict. If so, we hang 325 * the request off plex->waitlist of the first 326 * plex we find which is reviving 327 */ 328 329 if ((rq->flags & XFR_REVIVECONFLICT) /* possible revive conflict */ 330 &&(!reviveok)) { /* and we don't want to do it now, */ 331 struct sd *sd; 332 struct request *waitlist; /* point to the waitlist */ 333 334 sd = &SD[rq->sdno]; 335 if (sd->waitlist != NULL) { /* something there already, */ 336 waitlist = sd->waitlist; 337 while (waitlist->next != NULL) /* find the end */ 338 waitlist = waitlist->next; 339 waitlist->next = rq; /* hook our request there */ 340 } else 341 sd->waitlist = rq; /* hook our request at the front */ 342 343 #if VINUMDEBUG 344 if (debug & DEBUG_REVIVECONFLICT) { 345 log(LOG_DEBUG, 346 "Revive conflict sd %d: %p\n%s dev %d.%d, offset 0x%llx, length %d\n", 347 rq->sdno, 348 rq, 349 (rq->bio->bio_buf->b_cmd & BUF_CMD_READ) ? "Read" : "Write", 350 major(((dev_t)rq->bio->bio_driver_info)), 351 minor(((dev_t)rq->bio->bio_driver_info)), 352 rq->bio->bio_offset, 353 rq->bio->bio_buf->b_bcount); 354 } 355 #endif 356 return 0; /* and get out of here */ 357 } 358 rq->active = 0; /* nothing yet */ 359 #if VINUMDEBUG 360 if (debug & DEBUG_ADDRESSES) 361 log(LOG_DEBUG, 362 "Request: %p\n%s dev %d.%d, offset 0x%llx, length %d\n", 363 rq, 364 (rq->bio->bio_buf->b_cmd == BUF_CMD_READ) ? "Read" : "Write", 365 major(((dev_t)rq->bio->bio_driver_info)), 366 minor(((dev_t)rq->bio->bio_driver_info)), 367 rq->bio->bio_offset, 368 rq->bio->bio_buf->b_bcount); 369 vinum_conf.lastrq = rq; 370 vinum_conf.lastbio = rq->bio; 371 if (debug & DEBUG_LASTREQS) 372 logrq(loginfo_user_bpl, (union rqinfou) rq->bio, rq->bio); 373 #endif 374 375 /* 376 * This loop happens without any participation 377 * of the bottom half, so it requires no 378 * protection. 379 */ 380 for (rqg = rq->rqg; rqg != NULL; rqg = rqg->next) { /* through the whole request chain */ 381 rqg->active = rqg->count; /* they're all active */ 382 for (rqno = 0; rqno < rqg->count; rqno++) { 383 rqe = &rqg->rqe[rqno]; 384 if (rqe->flags & XFR_BAD_SUBDISK) /* this subdisk is bad, */ 385 rqg->active--; /* one less active request */ 386 } 387 if (rqg->active) /* we have at least one active request, */ 388 rq->active++; /* one more active request group */ 389 } 390 391 /* 392 * Now fire off the requests. In this loop the 393 * bottom half could be completing requests 394 * before we finish, so we need critical section protection. 395 */ 396 crit_enter(); 397 for (rqg = rq->rqg; rqg != NULL;) { /* through the whole request chain */ 398 if (rqg->lockbase >= 0) /* this rqg needs a lock first */ 399 rqg->lock = lockrange(rqg->lockbase, rqg->rq->bio->bio_buf, &PLEX[rqg->plexno]); 400 rcount = rqg->count; 401 for (rqno = 0; rqno < rcount;) { 402 dev_t dev; 403 404 rqe = &rqg->rqe[rqno]; 405 406 /* 407 * Point to next rqg before the bottom end 408 * changes the structures. 409 */ 410 if (++rqno >= rcount) 411 rqg = rqg->next; 412 if ((rqe->flags & XFR_BAD_SUBDISK) == 0) { /* this subdisk is good, */ 413 drive = &DRIVE[rqe->driveno]; /* look at drive */ 414 drive->active++; 415 if (drive->active >= drive->maxactive) 416 drive->maxactive = drive->active; 417 vinum_conf.active++; 418 if (vinum_conf.active >= vinum_conf.maxactive) 419 vinum_conf.maxactive = vinum_conf.active; 420 421 dev = rqe->b.b_bio1.bio_driver_info; 422 #ifdef VINUMDEBUG 423 if (debug & DEBUG_ADDRESSES) 424 log(LOG_DEBUG, 425 " %s dev %d.%d, sd %d, offset 0x%llx, devoffset 0x%llx, length %d\n", 426 (rqe->b.b_cmd == BUF_CMD_READ) ? "Read" : "Write", 427 major(dev), 428 minor(dev), 429 rqe->sdno, 430 rqe->b.b_bio1.bio_offset - ((off_t)SD[rqe->sdno].driveoffset << DEV_BSHIFT), 431 rqe->b.b_bio1.bio_offset, 432 rqe->b.b_bcount); 433 if (debug & DEBUG_LASTREQS) 434 logrq(loginfo_rqe, (union rqinfou) rqe, rq->bio); 435 #endif 436 /* fire off the request */ 437 dev_dstrategy(dev, &rqe->b.b_bio1); 438 } 439 } 440 } 441 crit_exit(); 442 return 0; 443 } 444 445 /* 446 * define the low-level requests needed to perform a 447 * high-level I/O operation for a specific plex 'plexno'. 448 * 449 * Return REQUEST_OK if all subdisks involved in the request are up, 450 * REQUEST_DOWN if some subdisks are not up, and REQUEST_EOF if the 451 * request is at least partially outside the bounds of the subdisks. 452 * 453 * Modify the pointer *diskstart to point to the end address. On 454 * read, return on the first bad subdisk, so that the caller 455 * (build_read_request) can try alternatives. 456 * 457 * On entry to this routine, the rqg structures are not assigned. The 458 * assignment is performed by expandrq(). Strictly speaking, the 459 * elements rqe->sdno of all entries should be set to -1, since 0 460 * (from bzero) is a valid subdisk number. We avoid this problem by 461 * initializing the ones we use, and not looking at the others (index 462 * >= rqg->requests). 463 */ 464 enum requeststatus 465 bre(struct request *rq, 466 int plexno, 467 daddr_t * diskaddr, 468 daddr_t diskend) 469 { 470 int sdno; 471 struct sd *sd; 472 struct rqgroup *rqg; 473 struct bio *bio; 474 struct buf *bp; /* user's bp */ 475 struct plex *plex; 476 enum requeststatus status; /* return value */ 477 daddr_t plexoffset; /* offset of transfer in plex */ 478 daddr_t stripebase; /* base address of stripe (1st subdisk) */ 479 daddr_t stripeoffset; /* offset in stripe */ 480 daddr_t blockoffset; /* offset in stripe on subdisk */ 481 struct rqelement *rqe; /* point to this request information */ 482 daddr_t diskstart = *diskaddr; /* remember where this transfer starts */ 483 enum requeststatus s; /* temp return value */ 484 485 bio = rq->bio; /* buffer pointer */ 486 bp = bio->bio_buf; 487 status = REQUEST_OK; /* return value: OK until proven otherwise */ 488 plex = &PLEX[plexno]; /* point to the plex */ 489 490 switch (plex->organization) { 491 case plex_concat: 492 sd = NULL; /* (keep compiler quiet) */ 493 for (sdno = 0; sdno < plex->subdisks; sdno++) { 494 sd = &SD[plex->sdnos[sdno]]; 495 if (*diskaddr < sd->plexoffset) /* we must have a hole, */ 496 status = REQUEST_DEGRADED; /* note the fact */ 497 if (*diskaddr < (sd->plexoffset + sd->sectors)) { /* the request starts in this subdisk */ 498 rqg = allocrqg(rq, 1); /* space for the request */ 499 if (rqg == NULL) { /* malloc failed */ 500 bp->b_error = ENOMEM; 501 bp->b_flags |= B_ERROR; 502 return REQUEST_ENOMEM; 503 } 504 rqg->plexno = plexno; 505 506 rqe = &rqg->rqe[0]; /* point to the element */ 507 rqe->rqg = rqg; /* group */ 508 rqe->sdno = sd->sdno; /* put in the subdisk number */ 509 plexoffset = *diskaddr; /* start offset in plex */ 510 rqe->sdoffset = plexoffset - sd->plexoffset; /* start offset in subdisk */ 511 rqe->useroffset = plexoffset - diskstart; /* start offset in user buffer */ 512 rqe->dataoffset = 0; 513 rqe->datalen = min(diskend - *diskaddr, /* number of sectors to transfer in this sd */ 514 sd->sectors - rqe->sdoffset); 515 rqe->groupoffset = 0; /* no groups for concatenated plexes */ 516 rqe->grouplen = 0; 517 rqe->buflen = rqe->datalen; /* buffer length is data buffer length */ 518 rqe->flags = 0; 519 rqe->driveno = sd->driveno; 520 if (sd->state != sd_up) { /* *now* we find the sd is down */ 521 s = checksdstate(sd, rq, *diskaddr, diskend); /* do we need to change state? */ 522 if (s == REQUEST_DOWN) { /* down? */ 523 rqe->flags = XFR_BAD_SUBDISK; /* yup */ 524 if (rq->bio->bio_buf->b_cmd == BUF_CMD_READ) /* read request, */ 525 return REQUEST_DEGRADED; /* give up here */ 526 /* 527 * If we're writing, don't give up 528 * because of a bad subdisk. Go 529 * through to the bitter end, but note 530 * which ones we can't access. 531 */ 532 status = REQUEST_DEGRADED; /* can't do it all */ 533 } 534 } 535 *diskaddr += rqe->datalen; /* bump the address */ 536 if (build_rq_buffer(rqe, plex)) { /* build the buffer */ 537 deallocrqg(rqg); 538 bp->b_error = ENOMEM; 539 bp->b_flags |= B_ERROR; 540 return REQUEST_ENOMEM; /* can't do it */ 541 } 542 } 543 if (*diskaddr == diskend) /* we're finished, */ 544 break; /* get out of here */ 545 } 546 /* 547 * We've got to the end of the plex. Have we got to the end of 548 * the transfer? It would seem that having an offset beyond the 549 * end of the subdisk is an error, but in fact it can happen if 550 * the volume has another plex of different size. There's a valid 551 * question as to why you would want to do this, but currently 552 * it's allowed. 553 * 554 * In a previous version, I returned REQUEST_DOWN here. I think 555 * REQUEST_EOF is more appropriate now. 556 */ 557 if (diskend > sd->sectors + sd->plexoffset) /* pointing beyond EOF? */ 558 status = REQUEST_EOF; 559 break; 560 561 case plex_striped: 562 { 563 while (*diskaddr < diskend) { /* until we get it all sorted out */ 564 if (*diskaddr >= plex->length) /* beyond the end of the plex */ 565 return REQUEST_EOF; /* can't continue */ 566 567 /* The offset of the start address from the start of the stripe. */ 568 stripeoffset = *diskaddr % (plex->stripesize * plex->subdisks); 569 570 /* The plex-relative address of the start of the stripe. */ 571 stripebase = *diskaddr - stripeoffset; 572 573 /* The number of the subdisk in which the start is located. */ 574 sdno = stripeoffset / plex->stripesize; 575 576 /* The offset from the beginning of the stripe on this subdisk. */ 577 blockoffset = stripeoffset % plex->stripesize; 578 579 sd = &SD[plex->sdnos[sdno]]; /* the subdisk in question */ 580 rqg = allocrqg(rq, 1); /* space for the request */ 581 if (rqg == NULL) { /* malloc failed */ 582 bp->b_error = ENOMEM; 583 bp->b_flags |= B_ERROR; 584 return REQUEST_ENOMEM; 585 } 586 rqg->plexno = plexno; 587 588 rqe = &rqg->rqe[0]; /* point to the element */ 589 rqe->rqg = rqg; 590 rqe->sdoffset = stripebase / plex->subdisks + blockoffset; /* start offset in this subdisk */ 591 rqe->useroffset = *diskaddr - diskstart; /* The offset of the start in the user buffer */ 592 rqe->dataoffset = 0; 593 rqe->datalen = min(diskend - *diskaddr, /* the amount remaining to transfer */ 594 plex->stripesize - blockoffset); /* and the amount left in this stripe */ 595 rqe->groupoffset = 0; /* no groups for striped plexes */ 596 rqe->grouplen = 0; 597 rqe->buflen = rqe->datalen; /* buffer length is data buffer length */ 598 rqe->flags = 0; 599 rqe->sdno = sd->sdno; /* put in the subdisk number */ 600 rqe->driveno = sd->driveno; 601 602 if (sd->state != sd_up) { /* *now* we find the sd is down */ 603 s = checksdstate(sd, rq, *diskaddr, diskend); /* do we need to change state? */ 604 if (s == REQUEST_DOWN) { /* down? */ 605 rqe->flags = XFR_BAD_SUBDISK; /* yup */ 606 if (rq->bio->bio_buf->b_cmd == BUF_CMD_READ) /* read request, */ 607 return REQUEST_DEGRADED; /* give up here */ 608 /* 609 * If we're writing, don't give up 610 * because of a bad subdisk. Go through 611 * to the bitter end, but note which 612 * ones we can't access. 613 */ 614 status = REQUEST_DEGRADED; /* can't do it all */ 615 } 616 } 617 /* 618 * It would seem that having an offset 619 * beyond the end of the subdisk is an 620 * error, but in fact it can happen if the 621 * volume has another plex of different 622 * size. There's a valid question as to why 623 * you would want to do this, but currently 624 * it's allowed. 625 */ 626 if (rqe->sdoffset + rqe->datalen > sd->sectors) { /* ends beyond the end of the subdisk? */ 627 rqe->datalen = sd->sectors - rqe->sdoffset; /* truncate */ 628 #if VINUMDEBUG 629 if (debug & DEBUG_EOFINFO) { /* tell on the request */ 630 log(LOG_DEBUG, 631 "vinum: EOF on plex %s, sd %s offset %llx (user offset %x)\n", 632 plex->name, 633 sd->name, 634 (u_int) sd->sectors, 635 bp->b_bio1.bio_offset); 636 log(LOG_DEBUG, 637 "vinum: stripebase %x, stripeoffset %x, blockoffset %x\n", 638 stripebase, 639 stripeoffset, 640 blockoffset); 641 } 642 #endif 643 } 644 if (build_rq_buffer(rqe, plex)) { /* build the buffer */ 645 deallocrqg(rqg); 646 bp->b_error = ENOMEM; 647 bp->b_flags |= B_ERROR; 648 return REQUEST_ENOMEM; /* can't do it */ 649 } 650 *diskaddr += rqe->datalen; /* look at the remainder */ 651 if ((*diskaddr < diskend) /* didn't finish the request on this stripe */ 652 &&(*diskaddr < plex->length)) { /* and there's more to come */ 653 plex->multiblock++; /* count another one */ 654 if (sdno == plex->subdisks - 1) /* last subdisk, */ 655 plex->multistripe++; /* another stripe as well */ 656 } 657 } 658 } 659 break; 660 661 /* 662 * RAID-4 and RAID-5 are complicated enough to have their own 663 * function. 664 */ 665 case plex_raid4: 666 case plex_raid5: 667 status = bre5(rq, plexno, diskaddr, diskend); 668 break; 669 670 default: 671 log(LOG_ERR, "vinum: invalid plex type %d in bre\n", plex->organization); 672 status = REQUEST_DOWN; /* can't access it */ 673 } 674 675 return status; 676 } 677 678 /* 679 * Build up a request structure for reading volumes. 680 * This function is not needed for plex reads, since there's 681 * no recovery if a plex read can't be satisified. 682 */ 683 enum requeststatus 684 build_read_request(struct request *rq, /* request */ 685 int plexindex) 686 { /* index in the volume's plex table */ 687 struct bio *bio; 688 struct buf *bp; 689 daddr_t startaddr; /* offset of previous part of transfer */ 690 daddr_t diskaddr; /* offset of current part of transfer */ 691 daddr_t diskend; /* and end offset of transfer */ 692 int plexno; /* plex index in vinum_conf */ 693 struct rqgroup *rqg; /* point to the request we're working on */ 694 struct volume *vol; /* volume in question */ 695 int recovered = 0; /* set if we recover a read */ 696 enum requeststatus status = REQUEST_OK; 697 int plexmask; /* bit mask of plexes, for recovery */ 698 699 bio = rq->bio; /* buffer pointer */ 700 bp = bio->bio_buf; 701 diskaddr = bio->bio_offset >> DEV_BSHIFT; /* start offset of transfer */ 702 diskend = diskaddr + (bp->b_bcount / DEV_BSIZE); /* and end offset of transfer */ 703 rqg = &rq->rqg[plexindex]; /* plex request */ 704 vol = &VOL[rq->volplex.volno]; /* point to volume */ 705 706 while (diskaddr < diskend) { /* build up request components */ 707 startaddr = diskaddr; 708 status = bre(rq, vol->plex[plexindex], &diskaddr, diskend); /* build up a request */ 709 switch (status) { 710 case REQUEST_OK: 711 continue; 712 713 case REQUEST_RECOVERED: 714 /* 715 * XXX FIXME if we have more than one plex, and we can 716 * satisfy the request from another, don't use the 717 * recovered request, since it's more expensive. 718 */ 719 recovered = 1; 720 break; 721 722 case REQUEST_ENOMEM: 723 return status; 724 /* 725 * If we get here, our request is not complete. Try 726 * to fill in the missing parts from another plex. 727 * This can happen multiple times in this function, 728 * and we reinitialize the plex mask each time, since 729 * we could have a hole in our plexes. 730 */ 731 case REQUEST_EOF: 732 case REQUEST_DOWN: /* can't access the plex */ 733 case REQUEST_DEGRADED: /* can't access the plex */ 734 plexmask = ((1 << vol->plexes) - 1) /* all plexes in the volume */ 735 &~(1 << plexindex); /* except for the one we were looking at */ 736 for (plexno = 0; plexno < vol->plexes; plexno++) { 737 if (plexmask == 0) /* no plexes left to try */ 738 return REQUEST_DOWN; /* failed */ 739 diskaddr = startaddr; /* start at the beginning again */ 740 if (plexmask & (1 << plexno)) { /* we haven't tried this plex yet */ 741 bre(rq, vol->plex[plexno], &diskaddr, diskend); /* try a request */ 742 if (diskaddr > startaddr) { /* we satisfied another part */ 743 recovered = 1; /* we recovered from the problem */ 744 status = REQUEST_OK; /* don't complain about it */ 745 break; 746 } 747 } 748 } 749 if (diskaddr == startaddr) /* didn't get any further, */ 750 return status; 751 } 752 if (recovered) 753 vol->recovered_reads += recovered; /* adjust our recovery count */ 754 } 755 return status; 756 } 757 758 /* 759 * Build up a request structure for writes. 760 * Return 0 if all subdisks involved in the request are up, 1 if some 761 * subdisks are not up, and -1 if the request is at least partially 762 * outside the bounds of the subdisks. 763 */ 764 enum requeststatus 765 build_write_request(struct request *rq) 766 { /* request */ 767 struct bio *bio; 768 struct buf *bp; 769 daddr_t diskstart; /* offset of current part of transfer */ 770 daddr_t diskend; /* and end offset of transfer */ 771 int plexno; /* plex index in vinum_conf */ 772 struct volume *vol; /* volume in question */ 773 enum requeststatus status; 774 775 bio = rq->bio; /* buffer pointer */ 776 bp = bio->bio_buf; 777 vol = &VOL[rq->volplex.volno]; /* point to volume */ 778 diskend = (daddr_t)(bio->bio_offset >> DEV_BSHIFT) + (bp->b_bcount / DEV_BSIZE); /* end offset of transfer */ 779 status = REQUEST_DOWN; /* assume the worst */ 780 for (plexno = 0; plexno < vol->plexes; plexno++) { 781 diskstart = (daddr_t)(bio->bio_offset >> DEV_BSHIFT); /* start offset of transfer */ 782 /* 783 * Build requests for the plex. 784 * We take the best possible result here (min, 785 * not max): we're happy if we can write at all 786 */ 787 status = min(status, bre(rq, 788 vol->plex[plexno], 789 &diskstart, 790 diskend)); 791 } 792 return status; 793 } 794 795 /* Fill in the struct buf part of a request element. */ 796 enum requeststatus 797 build_rq_buffer(struct rqelement *rqe, struct plex *plex) 798 { 799 struct sd *sd; /* point to subdisk */ 800 struct volume *vol; 801 struct buf *bp; 802 struct buf *ubp; /* user (high level) buffer header */ 803 struct bio *ubio; 804 805 vol = &VOL[rqe->rqg->rq->volplex.volno]; 806 sd = &SD[rqe->sdno]; /* point to subdisk */ 807 bp = &rqe->b; 808 ubio = rqe->rqg->rq->bio; /* pointer to user buffer header */ 809 ubp = ubio->bio_buf; 810 811 /* Initialize the buf struct */ 812 /* copy these flags from user bp */ 813 bp->b_flags = ubp->b_flags & (B_ORDERED | B_NOCACHE | B_ASYNC); 814 bp->b_cmd = ubp->b_cmd; 815 bp->b_flags |= B_PAGING; 816 #ifdef VINUMDEBUG 817 if (rqe->flags & XFR_BUFLOCKED) /* paranoia */ 818 panic("build_rq_buffer: rqe already locked"); /* XXX remove this when we're sure */ 819 #endif 820 BUF_LOCKINIT(bp); /* get a lock for the buffer */ 821 BUF_LOCK(bp, LK_EXCLUSIVE); /* and lock it */ 822 BUF_KERNPROC(bp); 823 rqe->flags |= XFR_BUFLOCKED; 824 bp->b_bio1.bio_done = complete_rqe; 825 /* 826 * You'd think that we wouldn't need to even 827 * build the request buffer for a dead subdisk, 828 * but in some cases we need information like 829 * the user buffer address. Err on the side of 830 * generosity and supply what we can. That 831 * obviously doesn't include drive information 832 * when the drive is dead. 833 */ 834 if ((rqe->flags & XFR_BAD_SUBDISK) == 0) /* subdisk is accessible, */ 835 bp->b_bio1.bio_driver_info = DRIVE[rqe->driveno].dev; /* drive device */ 836 bp->b_bio1.bio_offset = (off_t)(rqe->sdoffset + sd->driveoffset) << DEV_BSHIFT; /* start address */ 837 bp->b_bcount = rqe->buflen << DEV_BSHIFT; /* number of bytes to transfer */ 838 bp->b_resid = bp->b_bcount; /* and it's still all waiting */ 839 840 if (rqe->flags & XFR_MALLOCED) { /* this operation requires a malloced buffer */ 841 bp->b_data = Malloc(bp->b_bcount); /* get a buffer to put it in */ 842 if (bp->b_data == NULL) { /* failed */ 843 abortrequest(rqe->rqg->rq, ENOMEM); 844 return REQUEST_ENOMEM; /* no memory */ 845 } 846 } else 847 /* 848 * Point directly to user buffer data. This means 849 * that we don't need to do anything when we have 850 * finished the transfer 851 */ 852 bp->b_data = ubp->b_data + rqe->useroffset * DEV_BSIZE; 853 /* 854 * On a recovery read, we perform an XOR of 855 * all blocks to the user buffer. To make 856 * this work, we first clean out the buffer 857 */ 858 if ((rqe->flags & (XFR_RECOVERY_READ | XFR_BAD_SUBDISK)) 859 == (XFR_RECOVERY_READ | XFR_BAD_SUBDISK)) { /* bad subdisk of a recovery read */ 860 int length = rqe->grouplen << DEV_BSHIFT; /* and count involved */ 861 char *data = (char *) &rqe->b.b_data[rqe->groupoffset << DEV_BSHIFT]; /* destination */ 862 863 bzero(data, length); /* clean it out */ 864 } 865 return 0; 866 } 867 868 /* 869 * Abort a request: free resources and complete the 870 * user request with the specified error 871 */ 872 int 873 abortrequest(struct request *rq, int error) 874 { 875 struct buf *bp = rq->bio->bio_buf; /* user buffer */ 876 877 bp->b_error = error; 878 freerq(rq); /* free everything we're doing */ 879 bp->b_flags |= B_ERROR; 880 return error; /* and give up */ 881 } 882 883 /* 884 * Check that our transfer will cover the 885 * complete address space of the user request. 886 * 887 * Return 1 if it can, otherwise 0 888 */ 889 int 890 check_range_covered(struct request *rq) 891 { 892 return 1; 893 } 894 895 /* Perform I/O on a subdisk */ 896 void 897 sdio(struct bio *bio) 898 { 899 dev_t dev; 900 dev_t sddev; 901 struct sd *sd; 902 struct sdbuf *sbp; 903 daddr_t endoffset; 904 struct drive *drive; 905 struct buf *bp = bio->bio_buf; 906 907 dev = bio->bio_driver_info; 908 909 #if VINUMDEBUG 910 if (debug & DEBUG_LASTREQS) 911 logrq(loginfo_sdio, (union rqinfou) bio, bio); 912 #endif 913 sd = &SD[Sdno(dev)]; /* point to the subdisk */ 914 drive = &DRIVE[sd->driveno]; 915 916 if (drive->state != drive_up) { 917 if (sd->state >= sd_crashed) { 918 if (bp->b_cmd != BUF_CMD_READ) /* writing, */ 919 set_sd_state(sd->sdno, sd_stale, setstate_force); 920 else 921 set_sd_state(sd->sdno, sd_crashed, setstate_force); 922 } 923 bp->b_error = EIO; 924 bp->b_flags |= B_ERROR; 925 biodone(bio); 926 return; 927 } 928 /* 929 * We allow access to any kind of subdisk as long as we can expect 930 * to get the I/O performed. 931 */ 932 if (sd->state < sd_empty) { /* nothing to talk to, */ 933 bp->b_error = EIO; 934 bp->b_flags |= B_ERROR; 935 biodone(bio); 936 return; 937 } 938 /* Get a buffer */ 939 sbp = (struct sdbuf *) Malloc(sizeof(struct sdbuf)); 940 if (sbp == NULL) { 941 bp->b_error = ENOMEM; 942 bp->b_flags |= B_ERROR; 943 biodone(bio); 944 return; 945 } 946 sddev = DRIVE[sd->driveno].dev; /* device */ 947 bzero(sbp, sizeof(struct sdbuf)); /* start with nothing */ 948 sbp->b.b_flags = bp->b_flags | B_PAGING; 949 sbp->b.b_bcount = bp->b_bcount; /* number of bytes to transfer */ 950 sbp->b.b_resid = bp->b_resid; /* and amount waiting */ 951 sbp->b.b_data = bp->b_data; /* data buffer */ 952 BUF_LOCKINIT(&sbp->b); /* get a lock for the buffer */ 953 BUF_LOCK(&sbp->b, LK_EXCLUSIVE); /* and lock it */ 954 BUF_KERNPROC(&sbp->b); 955 initbufbio(&sbp->b); 956 sbp->b.b_bio1.bio_offset = bio->bio_offset + ((off_t)sd->driveoffset << DEV_BSHIFT); 957 sbp->b.b_bio1.bio_done = sdio_done; /* come here on completion */ 958 sbp->bio = bio; /* note the address of the original header */ 959 sbp->sdno = sd->sdno; /* note for statistics */ 960 sbp->driveno = sd->driveno; 961 endoffset = (daddr_t)(bio->bio_offset >> DEV_BSHIFT) + sbp->b.b_bcount / DEV_BSIZE; /* final sector offset */ 962 if (endoffset > sd->sectors) { /* beyond the end */ 963 sbp->b.b_bcount -= (endoffset - sd->sectors) * DEV_BSIZE; /* trim */ 964 if (sbp->b.b_bcount <= 0) { /* nothing to transfer */ 965 bp->b_resid = bp->b_bcount; /* nothing transferred */ 966 biodone(bio); 967 BUF_UNLOCK(&sbp->b); 968 BUF_LOCKFREE(&sbp->b); 969 Free(sbp); 970 return; 971 } 972 } 973 #if VINUMDEBUG 974 if (debug & DEBUG_ADDRESSES) 975 log(LOG_DEBUG, 976 " %s dev %d.%d, sd %d, offset 0x%llx, devoffset 0x%llx, length %d\n", 977 (sbp->b.b_cmd == BUF_CMD_READ) ? "Read" : "Write", 978 major(sddev), 979 minor(sddev), 980 sbp->sdno, 981 sbp->b.b_bio1.bio_offset - ((off_t)SD[sbp->sdno].driveoffset << DEV_BSHIFT), 982 sbp->b.b_bio1.bio_offset, 983 sbp->b.b_bcount); 984 #endif 985 crit_enter(); 986 #if VINUMDEBUG 987 if (debug & DEBUG_LASTREQS) 988 logrq(loginfo_sdiol, (union rqinfou) &sbp->b.b_bio1, &sbp->b.b_bio1); 989 #endif 990 dev_dstrategy(sddev, &sbp->b.b_bio1); 991 crit_exit(); 992 } 993 994 /* 995 * Simplified version of bounds_check_with_label 996 * Determine the size of the transfer, and make sure it is 997 * within the boundaries of the partition. Adjust transfer 998 * if needed, and signal errors or early completion. 999 * 1000 * Volumes are simpler than disk slices: they only contain 1001 * one component (though we call them a, b and c to make 1002 * system utilities happy), and they always take up the 1003 * complete space of the "partition". 1004 * 1005 * I'm still not happy with this: why should the label be 1006 * protected? If it weren't so damned difficult to write 1007 * one in the first pleace (because it's protected), it wouldn't 1008 * be a problem. 1009 */ 1010 struct bio * 1011 vinum_bounds_check(struct bio *bio, struct volume *vol) 1012 { 1013 struct buf *bp = bio->bio_buf; 1014 struct bio *nbio; 1015 int maxsize = vol->size; /* size of the partition (sectors) */ 1016 int size = (bp->b_bcount + DEV_BSIZE - 1) >> DEV_BSHIFT; /* size of this request (sectors) */ 1017 daddr_t blkno = (daddr_t)(bio->bio_offset >> DEV_BSHIFT); 1018 1019 /* Would this transfer overwrite the disk label? */ 1020 if (blkno <= LABELSECTOR /* starts before or at the label */ 1021 #if LABELSECTOR != 0 1022 && blkno + size > LABELSECTOR /* and finishes after */ 1023 #endif 1024 && (!(vol->flags & VF_RAW)) /* and it's not raw */ 1025 && (bp->b_cmd != BUF_CMD_READ) /* and it's a write */ 1026 &&(!vol->flags & (VF_WLABEL | VF_LABELLING))) { /* and we're not allowed to write the label */ 1027 bp->b_error = EROFS; /* read-only */ 1028 bp->b_flags |= B_ERROR; 1029 return (NULL); 1030 } 1031 if (size == 0) /* no transfer specified, */ 1032 return 0; /* treat as EOF */ 1033 /* beyond partition? */ 1034 if (bio->bio_offset < 0 /* negative start */ 1035 || blkno + size > maxsize) { /* or goes beyond the end of the partition */ 1036 /* if exactly at end of disk, return an EOF */ 1037 if (blkno == maxsize) { 1038 bp->b_resid = bp->b_bcount; 1039 return (NULL); 1040 } 1041 /* or truncate if part of it fits */ 1042 size = maxsize - blkno; 1043 if (size <= 0) { /* nothing to transfer */ 1044 bp->b_error = EINVAL; 1045 bp->b_flags |= B_ERROR; 1046 return (NULL); 1047 } 1048 bp->b_bcount = size << DEV_BSHIFT; 1049 } 1050 nbio = push_bio(bio); 1051 nbio->bio_offset = bio->bio_offset; 1052 return (nbio); 1053 } 1054 1055 /* 1056 * Allocate a request group and hook 1057 * it in in the list for rq 1058 */ 1059 struct rqgroup * 1060 allocrqg(struct request *rq, int elements) 1061 { 1062 struct rqgroup *rqg; /* the one we're going to allocate */ 1063 int size = sizeof(struct rqgroup) + elements * sizeof(struct rqelement); 1064 1065 rqg = (struct rqgroup *) Malloc(size); 1066 if (rqg != NULL) { /* malloc OK, */ 1067 if (rq->rqg) /* we already have requests */ 1068 rq->lrqg->next = rqg; /* hang it off the end */ 1069 else /* first request */ 1070 rq->rqg = rqg; /* at the start */ 1071 rq->lrqg = rqg; /* this one is the last in the list */ 1072 1073 bzero(rqg, size); /* no old junk */ 1074 rqg->rq = rq; /* point back to the parent request */ 1075 rqg->count = elements; /* number of requests in the group */ 1076 rqg->lockbase = -1; /* no lock required yet */ 1077 } 1078 return rqg; 1079 } 1080 1081 /* 1082 * Deallocate a request group out of a chain. We do 1083 * this by linear search: the chain is short, this 1084 * almost never happens, and currently it can only 1085 * happen to the first member of the chain. 1086 */ 1087 void 1088 deallocrqg(struct rqgroup *rqg) 1089 { 1090 struct rqgroup *rqgc = rqg->rq->rqg; /* point to the request chain */ 1091 1092 if (rqg->lock) /* got a lock? */ 1093 unlockrange(rqg->plexno, rqg->lock); /* yes, free it */ 1094 if (rqgc == rqg) /* we're first in line */ 1095 rqg->rq->rqg = rqg->next; /* unhook ourselves */ 1096 else { 1097 while ((rqgc->next != NULL) /* find the group */ 1098 &&(rqgc->next != rqg)) 1099 rqgc = rqgc->next; 1100 if (rqgc->next == NULL) 1101 log(LOG_ERR, 1102 "vinum deallocrqg: rqg %p not found in request %p\n", 1103 rqg->rq, 1104 rqg); 1105 else 1106 rqgc->next = rqg->next; /* make the chain jump over us */ 1107 } 1108 Free(rqg); 1109 } 1110 1111 /* Local Variables: */ 1112 /* fill-column: 50 */ 1113 /* End: */ 1114