Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/dev/vinum/vinumrequest.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
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 * $FreeBSD$ 42 */ 43 44 #include <dev/vinum/vinumhdr.h> 45 #include <dev/vinum/request.h> 46 #include <sys/resourcevar.h> 47 48 #define abs(x) (((x) < 0) ? -(x) : (x)) 49 50 enum requeststatus bre(struct request *rq, 51 int plexno, 52 daddr_t * diskstart, 53 daddr_t diskend); 54 enum requeststatus bre5(struct request *rq, 55 int plexno, 56 daddr_t * diskstart, 57 daddr_t diskend); 58 enum requeststatus build_read_request(struct request *rq, int volplexno); 59 enum requeststatus build_write_request(struct request *rq); 60 enum requeststatus build_rq_buffer(struct rqelement *rqe, struct plex *plex); 61 int find_alternate_sd(struct request *rq); 62 int check_range_covered(struct request *); 63 void complete_rqe(struct buf *bp); 64 void complete_raid5_write(struct rqelement *); 65 int abortrequest(struct request *rq, int error); 66 void sdio_done(struct buf *bp); 67 int vinum_bounds_check(struct buf *bp, struct volume *vol); 68 caddr_t allocdatabuf(struct rqelement *rqe); 69 void freedatabuf(struct rqelement *rqe); 70 71 #ifdef VINUMDEBUG 72 struct rqinfo rqinfo[RQINFO_SIZE]; 73 struct rqinfo *rqip = rqinfo; 74 75 void 76 logrq(enum rqinfo_type type, union rqinfou info, struct buf *ubp) 77 { 78 int s = splhigh(); 79 80 microtime(&rqip->timestamp); /* when did this happen? */ 81 rqip->type = type; 82 rqip->bp = ubp; /* user buffer */ 83 switch (type) { 84 case loginfo_user_bp: 85 case loginfo_user_bpl: 86 case loginfo_sdio: /* subdisk I/O */ 87 case loginfo_sdiol: /* subdisk I/O launch */ 88 case loginfo_sdiodone: /* subdisk I/O complete */ 89 bcopy(info.bp, &rqip->info.b, sizeof(struct buf)); 90 rqip->devmajor = major(info.bp->b_dev); 91 rqip->devminor = minor(info.bp->b_dev); 92 break; 93 94 case loginfo_iodone: 95 case loginfo_rqe: 96 case loginfo_raid5_data: 97 case loginfo_raid5_parity: 98 bcopy(info.rqe, &rqip->info.rqe, sizeof(struct rqelement)); 99 rqip->devmajor = major(info.rqe->b.b_dev); 100 rqip->devminor = minor(info.rqe->b.b_dev); 101 break; 102 103 case loginfo_lockwait: 104 case loginfo_lock: 105 case loginfo_unlock: 106 bcopy(info.lockinfo, &rqip->info.lockinfo, sizeof(struct rangelock)); 107 108 break; 109 110 case loginfo_unused: 111 break; 112 } 113 rqip++; 114 if (rqip >= &rqinfo[RQINFO_SIZE]) /* wrap around */ 115 rqip = rqinfo; 116 splx(s); 117 } 118 119 #endif 120 121 void 122 vinumstrategy(struct buf *bp) 123 { 124 int volno; 125 struct volume *vol = NULL; 126 127 switch (DEVTYPE(bp->b_dev)) { 128 case VINUM_SD_TYPE: 129 case VINUM_RAWSD_TYPE: 130 sdio(bp); 131 return; 132 133 /* 134 * In fact, vinum doesn't handle drives: they're 135 * handled directly by the disk drivers 136 */ 137 case VINUM_DRIVE_TYPE: 138 default: 139 bp->b_error = EIO; /* I/O error */ 140 bp->b_flags |= B_ERROR; 141 biodone(bp); 142 return; 143 144 case VINUM_VOLUME_TYPE: /* volume I/O */ 145 volno = Volno(bp->b_dev); 146 vol = &VOL[volno]; 147 if (vol->state != volume_up) { /* can't access this volume */ 148 bp->b_error = EIO; /* I/O error */ 149 bp->b_flags |= B_ERROR; 150 biodone(bp); 151 return; 152 } 153 if (vinum_bounds_check(bp, vol) <= 0) { /* don't like them bounds */ 154 biodone(bp); 155 return; 156 } 157 /* FALLTHROUGH */ 158 /* 159 * Plex I/O is pretty much the same as volume I/O 160 * for a single plex. Indicate this by passing a NULL 161 * pointer (set above) for the volume 162 */ 163 case VINUM_PLEX_TYPE: 164 case VINUM_RAWPLEX_TYPE: 165 bp->b_resid = bp->b_bcount; /* transfer everything */ 166 vinumstart(bp, 0); 167 return; 168 } 169 } 170 171 /* 172 * Start a transfer. Return -1 on error, 173 * 0 if OK, 1 if we need to retry. 174 * Parameter reviveok is set when doing 175 * transfers for revives: it allows transfers to 176 * be started immediately when a revive is in 177 * progress. During revive, normal transfers 178 * are queued if they share address space with 179 * a currently active revive operation. 180 */ 181 int 182 vinumstart(struct buf *bp, int reviveok) 183 { 184 int plexno; 185 int maxplex; /* maximum number of plexes to handle */ 186 struct volume *vol; 187 struct request *rq; /* build up our request here */ 188 enum requeststatus status; 189 190 #if VINUMDEBUG 191 if (debug & DEBUG_LASTREQS) 192 logrq(loginfo_user_bp, (union rqinfou) bp, bp); 193 #endif 194 195 if ((bp->b_bcount % DEV_BSIZE) != 0) { /* bad length */ 196 bp->b_error = EINVAL; /* invalid size */ 197 bp->b_flags |= B_ERROR; 198 biodone(bp); 199 return -1; 200 } 201 rq = (struct request *) Malloc(sizeof(struct request)); /* allocate a request struct */ 202 if (rq == NULL) { /* can't do it */ 203 bp->b_error = ENOMEM; /* can't get memory */ 204 bp->b_flags |= B_ERROR; 205 biodone(bp); 206 return -1; 207 } 208 bzero(rq, sizeof(struct request)); 209 210 /* 211 * Note the volume ID. This can be NULL, which 212 * the request building functions use as an 213 * indication for single plex I/O 214 */ 215 rq->bp = bp; /* and the user buffer struct */ 216 217 if (DEVTYPE(bp->b_dev) == VINUM_VOLUME_TYPE) { /* it's a volume, */ 218 rq->volplex.volno = Volno(bp->b_dev); /* get the volume number */ 219 vol = &VOL[rq->volplex.volno]; /* and point to it */ 220 vol->active++; /* one more active request */ 221 maxplex = vol->plexes; /* consider all its plexes */ 222 } else { 223 vol = NULL; /* no volume */ 224 rq->volplex.plexno = Plexno(bp->b_dev); /* point to the plex */ 225 rq->isplex = 1; /* note that it's a plex */ 226 maxplex = 1; /* just the one plex */ 227 } 228 229 if (bp->b_flags & B_READ) { 230 /* 231 * This is a read request. Decide 232 * which plex to read from. 233 * 234 * There's a potential race condition here, 235 * since we're not locked, and we could end 236 * up multiply incrementing the round-robin 237 * counter. This doesn't have any serious 238 * effects, however. 239 */ 240 if (vol != NULL) { 241 plexno = vol->preferred_plex; /* get the plex to use */ 242 if (plexno < 0) { /* round robin */ 243 for (plexno = 0; plexno < vol->plexes; plexno++) 244 if (abs(bp->b_blkno - PLEX[vol->plex[plexno]].last_addr) <= ROUNDROBIN_SWITCH) 245 break; 246 if (plexno >= vol->plexes) { 247 vol->last_plex_read++; 248 if (vol->last_plex_read >= vol->plexes) 249 vol->last_plex_read = 0; 250 plexno = vol->last_plex_read; 251 } else { 252 vol->last_plex_read = plexno; 253 }; 254 PLEX[vol->plex[plexno]].last_addr = bp->b_blkno; 255 } 256 status = build_read_request(rq, plexno); /* build a request */ 257 } else { 258 daddr_t diskaddr = bp->b_blkno; /* start offset of transfer */ 259 status = bre(rq, /* build a request list */ 260 rq->volplex.plexno, 261 &diskaddr, 262 diskaddr + (bp->b_bcount / DEV_BSIZE)); 263 } 264 265 if (status > REQUEST_RECOVERED) { /* can't satisfy it */ 266 if (status == REQUEST_DOWN) { /* not enough subdisks */ 267 bp->b_error = EIO; /* I/O error */ 268 bp->b_flags |= B_ERROR; 269 } 270 biodone(bp); 271 freerq(rq); 272 return -1; 273 } 274 return launch_requests(rq, reviveok); /* now start the requests if we can */ 275 } else 276 /* 277 * This is a write operation. We write to all plexes. If this is 278 * a RAID-4 or RAID-5 plex, we must also update the parity stripe. 279 */ 280 { 281 if (vol != NULL) 282 status = build_write_request(rq); /* Not all the subdisks are up */ 283 else { /* plex I/O */ 284 daddr_t diskstart; 285 286 diskstart = bp->b_blkno; /* start offset of transfer */ 287 status = bre(rq, 288 Plexno(bp->b_dev), 289 &diskstart, 290 bp->b_blkno + (bp->b_bcount / DEV_BSIZE)); /* build requests for the plex */ 291 } 292 if (status > REQUEST_RECOVERED) { /* can't satisfy it */ 293 if (status == REQUEST_DOWN) { /* not enough subdisks */ 294 bp->b_error = EIO; /* I/O error */ 295 bp->b_flags |= B_ERROR; 296 } 297 biodone(bp); 298 freerq(rq); 299 return -1; 300 } 301 return launch_requests(rq, reviveok); /* now start the requests if we can */ 302 } 303 } 304 305 /* 306 * Call the low-level strategy routines to 307 * perform the requests in a struct request 308 */ 309 int 310 launch_requests(struct request *rq, int reviveok) 311 { 312 struct rqgroup *rqg; 313 int rqno; /* loop index */ 314 struct rqelement *rqe; /* current element */ 315 struct drive *drive; 316 int rcount; /* request count */ 317 int s; 318 319 /* 320 * First find out whether we're reviving, and the 321 * request contains a conflict. If so, we hang 322 * the request off plex->waitlist of the first 323 * plex we find which is reviving 324 */ 325 326 if ((rq->flags & XFR_REVIVECONFLICT) /* possible revive conflict */ 327 &&(!reviveok)) { /* and we don't want to do it now, */ 328 struct sd *sd; 329 struct request *waitlist; /* point to the waitlist */ 330 331 sd = &SD[rq->sdno]; 332 if (sd->waitlist != NULL) { /* something there already, */ 333 waitlist = sd->waitlist; 334 while (waitlist->next != NULL) /* find the end */ 335 waitlist = waitlist->next; 336 waitlist->next = rq; /* hook our request there */ 337 } else 338 sd->waitlist = rq; /* hook our request at the front */ 339 340 #if VINUMDEBUG 341 if (debug & DEBUG_REVIVECONFLICT) 342 log(LOG_DEBUG, 343 "Revive conflict sd %d: %p\n%s dev %d.%d, offset 0x%x, length %ld\n", 344 rq->sdno, 345 rq, 346 rq->bp->b_flags & B_READ ? "Read" : "Write", 347 major(rq->bp->b_dev), 348 minor(rq->bp->b_dev), 349 rq->bp->b_blkno, 350 rq->bp->b_bcount); 351 #endif 352 return 0; /* and get out of here */ 353 } 354 rq->active = 0; /* nothing yet */ 355 #if VINUMDEBUG 356 if (debug & DEBUG_ADDRESSES) 357 log(LOG_DEBUG, 358 "Request: %p\n%s dev %d.%d, offset 0x%x, length %ld\n", 359 rq, 360 rq->bp->b_flags & B_READ ? "Read" : "Write", 361 major(rq->bp->b_dev), 362 minor(rq->bp->b_dev), 363 rq->bp->b_blkno, 364 rq->bp->b_bcount); 365 vinum_conf.lastrq = rq; 366 vinum_conf.lastbuf = rq->bp; 367 if (debug & DEBUG_LASTREQS) 368 logrq(loginfo_user_bpl, (union rqinfou) rq->bp, rq->bp); 369 #endif 370 371 /* 372 * We used to have an splbio() here anyway, out 373 * of superstition. With the division of labour 374 * below (first count the requests, then issue 375 * them), it looks as if we don't need this 376 * splbio() protection. In fact, as dillon 377 * points out, there's a race condition 378 * incrementing and decrementing rq->active and 379 * rqg->active. This splbio() didn't help 380 * there, because the device strategy routine 381 * can sleep. Solve this by putting shorter 382 * duration locks on the code. 383 */ 384 /* 385 * This loop happens without any participation 386 * of the bottom half, so it requires no 387 * protection. 388 */ 389 for (rqg = rq->rqg; rqg != NULL; rqg = rqg->next) { /* through the whole request chain */ 390 rqg->active = rqg->count; /* they're all active */ 391 for (rqno = 0; rqno < rqg->count; rqno++) { 392 rqe = &rqg->rqe[rqno]; 393 if (rqe->flags & XFR_BAD_SUBDISK) /* this subdisk is bad, */ 394 rqg->active--; /* one less active request */ 395 } 396 if (rqg->active) /* we have at least one active request, */ 397 rq->active++; /* one more active request group */ 398 } 399 400 /* 401 * Now fire off the requests. In this loop the 402 * bottom half could be completing requests 403 * before we finish, so we need splbio() protection. 404 */ 405 s = splbio (); 406 for (rqg = rq->rqg; rqg != NULL;) { /* through the whole request chain */ 407 if (rqg->lockbase >= 0) /* this rqg needs a lock first */ 408 rqg->lock = lockrange(rqg->lockbase, rqg->rq->bp, &PLEX[rqg->plexno]); 409 rcount = rqg->count; 410 for (rqno = 0; rqno < rcount;) { 411 rqe = &rqg->rqe[rqno]; 412 413 /* 414 * Point to next rqg before the bottom end 415 * changes the structures. 416 */ 417 if (++rqno >= rcount) 418 rqg = rqg->next; 419 if ((rqe->flags & XFR_BAD_SUBDISK) == 0) { /* this subdisk is good, */ 420 drive = &DRIVE[rqe->driveno]; /* look at drive */ 421 drive->active++; 422 if (drive->active >= drive->maxactive) 423 drive->maxactive = drive->active; 424 vinum_conf.active++; 425 if (vinum_conf.active >= vinum_conf.maxactive) 426 vinum_conf.maxactive = vinum_conf.active; 427 428 #ifdef VINUMDEBUG 429 if (debug & DEBUG_ADDRESSES) 430 log(LOG_DEBUG, 431 " %s dev %d.%d, sd %d, offset 0x%x, devoffset 0x%x, length %ld\n", 432 rqe->b.b_flags & B_READ ? "Read" : "Write", 433 major(rqe->b.b_dev), 434 minor(rqe->b.b_dev), 435 rqe->sdno, 436 (u_int) (rqe->b.b_blkno - SD[rqe->sdno].driveoffset), 437 rqe->b.b_blkno, 438 rqe->b.b_bcount); 439 if (debug & DEBUG_LASTREQS) 440 logrq(loginfo_rqe, (union rqinfou) rqe, rq->bp); 441 #endif 442 /* fire off the request */ 443 BUF_STRATEGY(&rqe->b, 0); 444 } 445 } 446 } 447 splx (s); 448 return 0; 449 } 450 451 /* 452 * define the low-level requests needed to perform a 453 * high-level I/O operation for a specific plex 'plexno'. 454 * 455 * Return REQUEST_OK if all subdisks involved in the request are up, 456 * REQUEST_DOWN if some subdisks are not up, and REQUEST_EOF if the 457 * request is at least partially outside the bounds of the subdisks. 458 * 459 * Modify the pointer *diskstart to point to the end address. On 460 * read, return on the first bad subdisk, so that the caller 461 * (build_read_request) can try alternatives. 462 * 463 * On entry to this routine, the rqg structures are not assigned. The 464 * assignment is performed by expandrq(). Strictly speaking, the 465 * elements rqe->sdno of all entries should be set to -1, since 0 466 * (from bzero) is a valid subdisk number. We avoid this problem by 467 * initializing the ones we use, and not looking at the others (index 468 * >= rqg->requests). 469 */ 470 enum requeststatus 471 bre(struct request *rq, 472 int plexno, 473 daddr_t * diskaddr, 474 daddr_t diskend) 475 { 476 int sdno; 477 struct sd *sd; 478 struct rqgroup *rqg; 479 struct buf *bp; /* user's bp */ 480 struct plex *plex; 481 enum requeststatus status; /* return value */ 482 daddr_t plexoffset; /* offset of transfer in plex */ 483 daddr_t stripebase; /* base address of stripe (1st subdisk) */ 484 daddr_t stripeoffset; /* offset in stripe */ 485 daddr_t blockoffset; /* offset in stripe on subdisk */ 486 struct rqelement *rqe; /* point to this request information */ 487 daddr_t diskstart = *diskaddr; /* remember where this transfer starts */ 488 enum requeststatus s; /* temp return value */ 489 490 bp = rq->bp; /* buffer pointer */ 491 status = REQUEST_OK; /* return value: OK until proven otherwise */ 492 plex = &PLEX[plexno]; /* point to the plex */ 493 494 switch (plex->organization) { 495 case plex_concat: 496 sd = NULL; /* (keep compiler quiet) */ 497 for (sdno = 0; sdno < plex->subdisks; sdno++) { 498 sd = &SD[plex->sdnos[sdno]]; 499 if (*diskaddr < sd->plexoffset) /* we must have a hole, */ 500 status = REQUEST_DEGRADED; /* note the fact */ 501 if (*diskaddr < (sd->plexoffset + sd->sectors)) { /* the request starts in this subdisk */ 502 rqg = allocrqg(rq, 1); /* space for the request */ 503 if (rqg == NULL) { /* malloc failed */ 504 bp->b_error = ENOMEM; 505 bp->b_flags |= B_ERROR; 506 return REQUEST_ENOMEM; 507 } 508 rqg->plexno = plexno; 509 510 rqe = &rqg->rqe[0]; /* point to the element */ 511 rqe->rqg = rqg; /* group */ 512 rqe->sdno = sd->sdno; /* put in the subdisk number */ 513 plexoffset = *diskaddr; /* start offset in plex */ 514 rqe->sdoffset = plexoffset - sd->plexoffset; /* start offset in subdisk */ 515 rqe->useroffset = plexoffset - diskstart; /* start offset in user buffer */ 516 rqe->dataoffset = 0; 517 rqe->datalen = min(diskend - *diskaddr, /* number of sectors to transfer in this sd */ 518 sd->sectors - rqe->sdoffset); 519 rqe->groupoffset = 0; /* no groups for concatenated plexes */ 520 rqe->grouplen = 0; 521 rqe->buflen = rqe->datalen; /* buffer length is data buffer length */ 522 rqe->flags = 0; 523 rqe->driveno = sd->driveno; 524 if (sd->state != sd_up) { /* *now* we find the sd is down */ 525 s = checksdstate(sd, rq, *diskaddr, diskend); /* do we need to change state? */ 526 if (s == REQUEST_DOWN) { /* down? */ 527 rqe->flags = XFR_BAD_SUBDISK; /* yup */ 528 if (rq->bp->b_flags & B_READ) /* read request, */ 529 return REQUEST_DEGRADED; /* give up here */ 530 /* 531 * If we're writing, don't give up 532 * because of a bad subdisk. Go 533 * through to the bitter end, but note 534 * which ones we can't access. 535 */ 536 status = REQUEST_DEGRADED; /* can't do it all */ 537 } 538 } 539 *diskaddr += rqe->datalen; /* bump the address */ 540 if (build_rq_buffer(rqe, plex)) { /* build the buffer */ 541 deallocrqg(rqg); 542 bp->b_error = ENOMEM; 543 bp->b_flags |= B_ERROR; 544 return REQUEST_ENOMEM; /* can't do it */ 545 } 546 } 547 if (*diskaddr == diskend) /* we're finished, */ 548 break; /* get out of here */ 549 } 550 /* 551 * We've got to the end of the plex. Have we got to the end of 552 * the transfer? It would seem that having an offset beyond the 553 * end of the subdisk is an error, but in fact it can happen if 554 * the volume has another plex of different size. There's a valid 555 * question as to why you would want to do this, but currently 556 * it's allowed. 557 * 558 * In a previous version, I returned REQUEST_DOWN here. I think 559 * REQUEST_EOF is more appropriate now. 560 */ 561 if (diskend > sd->sectors + sd->plexoffset) /* pointing beyond EOF? */ 562 status = REQUEST_EOF; 563 break; 564 565 case plex_striped: 566 { 567 while (*diskaddr < diskend) { /* until we get it all sorted out */ 568 if (*diskaddr >= plex->length) /* beyond the end of the plex */ 569 return REQUEST_EOF; /* can't continue */ 570 571 /* The offset of the start address from the start of the stripe. */ 572 stripeoffset = *diskaddr % (plex->stripesize * plex->subdisks); 573 574 /* The plex-relative address of the start of the stripe. */ 575 stripebase = *diskaddr - stripeoffset; 576 577 /* The number of the subdisk in which the start is located. */ 578 sdno = stripeoffset / plex->stripesize; 579 580 /* The offset from the beginning of the stripe on this subdisk. */ 581 blockoffset = stripeoffset % plex->stripesize; 582 583 sd = &SD[plex->sdnos[sdno]]; /* the subdisk in question */ 584 rqg = allocrqg(rq, 1); /* space for the request */ 585 if (rqg == NULL) { /* malloc failed */ 586 bp->b_error = ENOMEM; 587 bp->b_flags |= B_ERROR; 588 return REQUEST_ENOMEM; 589 } 590 rqg->plexno = plexno; 591 592 rqe = &rqg->rqe[0]; /* point to the element */ 593 rqe->rqg = rqg; 594 rqe->sdoffset = stripebase / plex->subdisks + blockoffset; /* start offset in this subdisk */ 595 rqe->useroffset = *diskaddr - diskstart; /* The offset of the start in the user buffer */ 596 rqe->dataoffset = 0; 597 rqe->datalen = min(diskend - *diskaddr, /* the amount remaining to transfer */ 598 plex->stripesize - blockoffset); /* and the amount left in this stripe */ 599 rqe->groupoffset = 0; /* no groups for striped plexes */ 600 rqe->grouplen = 0; 601 rqe->buflen = rqe->datalen; /* buffer length is data buffer length */ 602 rqe->flags = 0; 603 rqe->sdno = sd->sdno; /* put in the subdisk number */ 604 rqe->driveno = sd->driveno; 605 606 if (sd->state != sd_up) { /* *now* we find the sd is down */ 607 s = checksdstate(sd, rq, *diskaddr, diskend); /* do we need to change state? */ 608 if (s == REQUEST_DOWN) { /* down? */ 609 rqe->flags = XFR_BAD_SUBDISK; /* yup */ 610 if (rq->bp->b_flags & B_READ) /* read request, */ 611 return REQUEST_DEGRADED; /* give up here */ 612 /* 613 * If we're writing, don't give up 614 * because of a bad subdisk. Go through 615 * to the bitter end, but note which 616 * ones we can't access. 617 */ 618 status = REQUEST_DEGRADED; /* can't do it all */ 619 } 620 } 621 /* 622 * It would seem that having an offset 623 * beyond the end of the subdisk is an 624 * error, but in fact it can happen if the 625 * volume has another plex of different 626 * size. There's a valid question as to why 627 * you would want to do this, but currently 628 * it's allowed. 629 */ 630 if (rqe->sdoffset + rqe->datalen > sd->sectors) { /* ends beyond the end of the subdisk? */ 631 rqe->datalen = sd->sectors - rqe->sdoffset; /* truncate */ 632 #if VINUMDEBUG 633 if (debug & DEBUG_EOFINFO) { /* tell on the request */ 634 log(LOG_DEBUG, 635 "vinum: EOF on plex %s, sd %s offset %x (user offset %x)\n", 636 plex->name, 637 sd->name, 638 (u_int) sd->sectors, 639 bp->b_blkno); 640 log(LOG_DEBUG, 641 "vinum: stripebase %x, stripeoffset %x, blockoffset %x\n", 642 stripebase, 643 stripeoffset, 644 blockoffset); 645 } 646 #endif 647 } 648 if (build_rq_buffer(rqe, plex)) { /* build the buffer */ 649 deallocrqg(rqg); 650 bp->b_error = ENOMEM; 651 bp->b_flags |= B_ERROR; 652 return REQUEST_ENOMEM; /* can't do it */ 653 } 654 *diskaddr += rqe->datalen; /* look at the remainder */ 655 if ((*diskaddr < diskend) /* didn't finish the request on this stripe */ 656 &&(*diskaddr < plex->length)) { /* and there's more to come */ 657 plex->multiblock++; /* count another one */ 658 if (sdno == plex->subdisks - 1) /* last subdisk, */ 659 plex->multistripe++; /* another stripe as well */ 660 } 661 } 662 } 663 break; 664 665 /* 666 * RAID-4 and RAID-5 are complicated enough to have their own 667 * function. 668 */ 669 case plex_raid4: 670 case plex_raid5: 671 status = bre5(rq, plexno, diskaddr, diskend); 672 break; 673 674 default: 675 log(LOG_ERR, "vinum: invalid plex type %d in bre\n", plex->organization); 676 status = REQUEST_DOWN; /* can't access it */ 677 } 678 679 return status; 680 } 681 682 /* 683 * Build up a request structure for reading volumes. 684 * This function is not needed for plex reads, since there's 685 * no recovery if a plex read can't be satisified. 686 */ 687 enum requeststatus 688 build_read_request(struct request *rq, /* request */ 689 int plexindex) 690 { /* index in the volume's plex table */ 691 struct buf *bp; 692 daddr_t startaddr; /* offset of previous part of transfer */ 693 daddr_t diskaddr; /* offset of current part of transfer */ 694 daddr_t diskend; /* and end offset of transfer */ 695 int plexno; /* plex index in vinum_conf */ 696 struct rqgroup *rqg; /* point to the request we're working on */ 697 struct volume *vol; /* volume in question */ 698 int recovered = 0; /* set if we recover a read */ 699 enum requeststatus status = REQUEST_OK; 700 int plexmask; /* bit mask of plexes, for recovery */ 701 702 bp = rq->bp; /* buffer pointer */ 703 diskaddr = bp->b_blkno; /* start offset of transfer */ 704 diskend = diskaddr + (bp->b_bcount / DEV_BSIZE); /* and end offset of transfer */ 705 rqg = &rq->rqg[plexindex]; /* plex request */ 706 vol = &VOL[rq->volplex.volno]; /* point to volume */ 707 708 while (diskaddr < diskend) { /* build up request components */ 709 startaddr = diskaddr; 710 status = bre(rq, vol->plex[plexindex], &diskaddr, diskend); /* build up a request */ 711 switch (status) { 712 case REQUEST_OK: 713 continue; 714 715 case REQUEST_RECOVERED: 716 /* 717 * XXX FIXME if we have more than one plex, and we can 718 * satisfy the request from another, don't use the 719 * recovered request, since it's more expensive. 720 */ 721 recovered = 1; 722 break; 723 724 case REQUEST_ENOMEM: 725 return status; 726 /* 727 * If we get here, our request is not complete. Try 728 * to fill in the missing parts from another plex. 729 * This can happen multiple times in this function, 730 * and we reinitialize the plex mask each time, since 731 * we could have a hole in our plexes. 732 */ 733 case REQUEST_EOF: 734 case REQUEST_DOWN: /* can't access the plex */ 735 case REQUEST_DEGRADED: /* can't access the plex */ 736 plexmask = ((1 << vol->plexes) - 1) /* all plexes in the volume */ 737 &~(1 << plexindex); /* except for the one we were looking at */ 738 for (plexno = 0; plexno < vol->plexes; plexno++) { 739 if (plexmask == 0) /* no plexes left to try */ 740 return REQUEST_DOWN; /* failed */ 741 diskaddr = startaddr; /* start at the beginning again */ 742 if (plexmask & (1 << plexno)) { /* we haven't tried this plex yet */ 743 bre(rq, vol->plex[plexno], &diskaddr, diskend); /* try a request */ 744 if (diskaddr > startaddr) { /* we satisfied another part */ 745 recovered = 1; /* we recovered from the problem */ 746 status = REQUEST_OK; /* don't complain about it */ 747 break; 748 } 749 } 750 } 751 if (diskaddr == startaddr) /* didn't get any further, */ 752 return status; 753 } 754 if (recovered) 755 vol->recovered_reads += recovered; /* adjust our recovery count */ 756 } 757 return status; 758 } 759 760 /* 761 * Build up a request structure for writes. 762 * Return 0 if all subdisks involved in the request are up, 1 if some 763 * subdisks are not up, and -1 if the request is at least partially 764 * outside the bounds of the subdisks. 765 */ 766 enum requeststatus 767 build_write_request(struct request *rq) 768 { /* request */ 769 struct buf *bp; 770 daddr_t diskstart; /* offset of current part of transfer */ 771 daddr_t diskend; /* and end offset of transfer */ 772 int plexno; /* plex index in vinum_conf */ 773 struct volume *vol; /* volume in question */ 774 enum requeststatus status; 775 776 bp = rq->bp; /* buffer pointer */ 777 vol = &VOL[rq->volplex.volno]; /* point to volume */ 778 diskend = bp->b_blkno + (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 = bp->b_blkno; /* 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 */ 803 804 vol = &VOL[rqe->rqg->rq->volplex.volno]; 805 sd = &SD[rqe->sdno]; /* point to subdisk */ 806 bp = &rqe->b; 807 ubp = rqe->rqg->rq->bp; /* pointer to user buffer header */ 808 809 /* Initialize the buf struct */ 810 /* copy these flags from user bp */ 811 bp->b_flags = ubp->b_flags & (B_ORDERED | B_NOCACHE | B_READ | B_ASYNC); 812 bp->b_flags |= B_CALL; /* inform us when it's done */ 813 #ifdef VINUMDEBUG 814 if (rqe->flags & XFR_BUFLOCKED) /* paranoia */ 815 panic("build_rq_buffer: rqe already locked"); /* XXX remove this when we're sure */ 816 #endif 817 BUF_LOCKINIT(bp); /* get a lock for the buffer */ 818 BUF_LOCK(bp, LK_EXCLUSIVE); /* and lock it */ 819 BUF_KERNPROC(bp); 820 rqe->flags |= XFR_BUFLOCKED; 821 bp->b_iodone = complete_rqe; 822 /* 823 * You'd think that we wouldn't need to even 824 * build the request buffer for a dead subdisk, 825 * but in some cases we need information like 826 * the user buffer address. Err on the side of 827 * generosity and supply what we can. That 828 * obviously doesn't include drive information 829 * when the drive is dead. 830 */ 831 if ((rqe->flags & XFR_BAD_SUBDISK) == 0) /* subdisk is accessible, */ 832 bp->b_dev = DRIVE[rqe->driveno].dev; /* drive device */ 833 bp->b_blkno = rqe->sdoffset + sd->driveoffset; /* start address */ 834 bp->b_bcount = rqe->buflen << DEV_BSHIFT; /* number of bytes to transfer */ 835 bp->b_resid = bp->b_bcount; /* and it's still all waiting */ 836 bp->b_bufsize = bp->b_bcount; /* and buffer size */ 837 bp->b_rcred = FSCRED; /* we have the file system credentials */ 838 bp->b_wcred = FSCRED; /* we have the file system credentials */ 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->bp; /* 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 buf *bp) 898 { 899 int s; /* spl */ 900 struct sd *sd; 901 struct sdbuf *sbp; 902 daddr_t endoffset; 903 struct drive *drive; 904 905 #if VINUMDEBUG 906 if (debug & DEBUG_LASTREQS) 907 logrq(loginfo_sdio, (union rqinfou) bp, bp); 908 #endif 909 sd = &SD[Sdno(bp->b_dev)]; /* point to the subdisk */ 910 drive = &DRIVE[sd->driveno]; 911 912 if (drive->state != drive_up) { 913 if (sd->state >= sd_crashed) { 914 if ((bp->b_flags & B_READ) == 0) /* writing, */ 915 set_sd_state(sd->sdno, sd_stale, setstate_force); 916 else 917 set_sd_state(sd->sdno, sd_crashed, setstate_force); 918 } 919 bp->b_error = EIO; 920 bp->b_flags |= B_ERROR; 921 biodone(bp); 922 return; 923 } 924 /* 925 * We allow access to any kind of subdisk as long as we can expect 926 * to get the I/O performed. 927 */ 928 if (sd->state < sd_empty) { /* nothing to talk to, */ 929 bp->b_error = EIO; 930 bp->b_flags |= B_ERROR; 931 biodone(bp); 932 return; 933 } 934 /* Get a buffer */ 935 sbp = (struct sdbuf *) Malloc(sizeof(struct sdbuf)); 936 if (sbp == NULL) { 937 bp->b_error = ENOMEM; 938 bp->b_flags |= B_ERROR; 939 biodone(bp); 940 return; 941 } 942 bzero(sbp, sizeof(struct sdbuf)); /* start with nothing */ 943 sbp->b.b_flags = bp->b_flags | B_CALL; /* inform us when it's done */ 944 sbp->b.b_bufsize = bp->b_bufsize; /* buffer size */ 945 sbp->b.b_bcount = bp->b_bcount; /* number of bytes to transfer */ 946 sbp->b.b_resid = bp->b_resid; /* and amount waiting */ 947 sbp->b.b_dev = DRIVE[sd->driveno].dev; /* device */ 948 sbp->b.b_data = bp->b_data; /* data buffer */ 949 sbp->b.b_blkno = bp->b_blkno + sd->driveoffset; 950 sbp->b.b_iodone = sdio_done; /* come here on completion */ 951 BUF_LOCKINIT(&sbp->b); /* get a lock for the buffer */ 952 BUF_LOCK(&sbp->b, LK_EXCLUSIVE); /* and lock it */ 953 BUF_KERNPROC(&sbp->b); 954 sbp->bp = bp; /* note the address of the original header */ 955 sbp->sdno = sd->sdno; /* note for statistics */ 956 sbp->driveno = sd->driveno; 957 endoffset = bp->b_blkno + sbp->b.b_bcount / DEV_BSIZE; /* final sector offset */ 958 if (endoffset > sd->sectors) { /* beyond the end */ 959 sbp->b.b_bcount -= (endoffset - sd->sectors) * DEV_BSIZE; /* trim */ 960 if (sbp->b.b_bcount <= 0) { /* nothing to transfer */ 961 bp->b_resid = bp->b_bcount; /* nothing transferred */ 962 biodone(bp); 963 BUF_UNLOCK(&sbp->b); 964 BUF_LOCKFREE(&sbp->b); 965 Free(sbp); 966 return; 967 } 968 } 969 #if VINUMDEBUG 970 if (debug & DEBUG_ADDRESSES) 971 log(LOG_DEBUG, 972 " %s dev %d.%d, sd %d, offset 0x%x, devoffset 0x%x, length %ld\n", 973 sbp->b.b_flags & B_READ ? "Read" : "Write", 974 major(sbp->b.b_dev), 975 minor(sbp->b.b_dev), 976 sbp->sdno, 977 (u_int) (sbp->b.b_blkno - SD[sbp->sdno].driveoffset), 978 (int) sbp->b.b_blkno, 979 sbp->b.b_bcount); 980 #endif 981 s = splbio(); 982 #if VINUMDEBUG 983 if (debug & DEBUG_LASTREQS) 984 logrq(loginfo_sdiol, (union rqinfou) &sbp->b, &sbp->b); 985 #endif 986 BUF_STRATEGY(&sbp->b, 0); 987 splx(s); 988 } 989 990 /* 991 * Simplified version of bounds_check_with_label 992 * Determine the size of the transfer, and make sure it is 993 * within the boundaries of the partition. Adjust transfer 994 * if needed, and signal errors or early completion. 995 * 996 * Volumes are simpler than disk slices: they only contain 997 * one component (though we call them a, b and c to make 998 * system utilities happy), and they always take up the 999 * complete space of the "partition". 1000 * 1001 * I'm still not happy with this: why should the label be 1002 * protected? If it weren't so damned difficult to write 1003 * one in the first pleace (because it's protected), it wouldn't 1004 * be a problem. 1005 */ 1006 int 1007 vinum_bounds_check(struct buf *bp, struct volume *vol) 1008 { 1009 int maxsize = vol->size; /* size of the partition (sectors) */ 1010 int size = (bp->b_bcount + DEV_BSIZE - 1) >> DEV_BSHIFT; /* size of this request (sectors) */ 1011 1012 /* Would this transfer overwrite the disk label? */ 1013 if (bp->b_blkno <= LABELSECTOR /* starts before or at the label */ 1014 #if LABELSECTOR != 0 1015 && bp->b_blkno + size > LABELSECTOR /* and finishes after */ 1016 #endif 1017 && (!(vol->flags & VF_RAW)) /* and it's not raw */ 1018 &&((bp->b_flags & B_READ) == 0) /* and it's a write */ 1019 &&(!vol->flags & (VF_WLABEL | VF_LABELLING))) { /* and we're not allowed to write the label */ 1020 bp->b_error = EROFS; /* read-only */ 1021 bp->b_flags |= B_ERROR; 1022 return -1; 1023 } 1024 if (size == 0) /* no transfer specified, */ 1025 return 0; /* treat as EOF */ 1026 /* beyond partition? */ 1027 if (bp->b_blkno < 0 /* negative start */ 1028 || bp->b_blkno + size > maxsize) { /* or goes beyond the end of the partition */ 1029 /* if exactly at end of disk, return an EOF */ 1030 if (bp->b_blkno == maxsize) { 1031 bp->b_resid = bp->b_bcount; 1032 return 0; 1033 } 1034 /* or truncate if part of it fits */ 1035 size = maxsize - bp->b_blkno; 1036 if (size <= 0) { /* nothing to transfer */ 1037 bp->b_error = EINVAL; 1038 bp->b_flags |= B_ERROR; 1039 return -1; 1040 } 1041 bp->b_bcount = size << DEV_BSHIFT; 1042 } 1043 bp->b_pblkno = bp->b_blkno; 1044 return 1; 1045 } 1046 1047 /* 1048 * Allocate a request group and hook 1049 * it in in the list for rq 1050 */ 1051 struct rqgroup * 1052 allocrqg(struct request *rq, int elements) 1053 { 1054 struct rqgroup *rqg; /* the one we're going to allocate */ 1055 int size = sizeof(struct rqgroup) + elements * sizeof(struct rqelement); 1056 1057 rqg = (struct rqgroup *) Malloc(size); 1058 if (rqg != NULL) { /* malloc OK, */ 1059 if (rq->rqg) /* we already have requests */ 1060 rq->lrqg->next = rqg; /* hang it off the end */ 1061 else /* first request */ 1062 rq->rqg = rqg; /* at the start */ 1063 rq->lrqg = rqg; /* this one is the last in the list */ 1064 1065 bzero(rqg, size); /* no old junk */ 1066 rqg->rq = rq; /* point back to the parent request */ 1067 rqg->count = elements; /* number of requests in the group */ 1068 rqg->lockbase = -1; /* no lock required yet */ 1069 } 1070 return rqg; 1071 } 1072 1073 /* 1074 * Deallocate a request group out of a chain. We do 1075 * this by linear search: the chain is short, this 1076 * almost never happens, and currently it can only 1077 * happen to the first member of the chain. 1078 */ 1079 void 1080 deallocrqg(struct rqgroup *rqg) 1081 { 1082 struct rqgroup *rqgc = rqg->rq->rqg; /* point to the request chain */ 1083 1084 if (rqg->lock) /* got a lock? */ 1085 unlockrange(rqg->plexno, rqg->lock); /* yes, free it */ 1086 if (rqgc == rqg) /* we're first in line */ 1087 rqg->rq->rqg = rqg->next; /* unhook ourselves */ 1088 else { 1089 while ((rqgc->next != NULL) /* find the group */ 1090 &&(rqgc->next != rqg)) 1091 rqgc = rqgc->next; 1092 if (rqgc->next == NULL) 1093 log(LOG_ERR, 1094 "vinum deallocrqg: rqg %p not found in request %p\n", 1095 rqg->rq, 1096 rqg); 1097 else 1098 rqgc->next = rqg->next; /* make the chain jump over us */ 1099 } 1100 Free(rqg); 1101 } 1102 1103 /* Local Variables: */ 1104 /* fill-column: 50 */ 1105 /* End: */
Cache object: 9cdafce95a3a184d8033eae2809e7df1
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]