sys/dev/raid/vinum/vinumrequest.c

   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 1.19 2007/06/07 22:58:38 corecode 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     cdev_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 int
 126 vinumstrategy(struct dev_strategy_args *ap)
 127 {
 128     cdev_t dev = ap->a_head.a_dev;
 129     struct bio *bio = ap->a_bio;
 130     struct buf *bp = bio->bio_buf;
 131     struct bio *nbio = bio;
 132     struct volume *vol = NULL;
 133     int volno;
 134
 135     switch (DEVTYPE(dev)) {
 136     case VINUM_SD_TYPE:
 137     case VINUM_RAWSD_TYPE:
 138         bio->bio_driver_info = dev;
 139         sdio(bio);
 140         break;
 141     case VINUM_DRIVE_TYPE:
 142     default:
 143         /*
 144          * In fact, vinum doesn't handle drives: they're
 145          * handled directly by the disk drivers
 146          */
 147         bp->b_error = EIO;                                  /* I/O error */
 148         bp->b_flags |= B_ERROR;
 149         biodone(bio);
 150         break;
 151
 152     case VINUM_VOLUME_TYPE:                                 /* volume I/O */
 153         volno = Volno(dev);
 154         vol = &VOL[volno];
 155         if (vol->state != volume_up) {                      /* can't access this volume */
 156             bp->b_error = EIO;                              /* I/O error */
 157             bp->b_flags |= B_ERROR;
 158             biodone(bio);
 159             break;
 160         }
 161         nbio = vinum_bounds_check(bio, vol);
 162         if (nbio == NULL) {
 163             biodone(bio);
 164             break;
 165         }
 166         /* FALLTHROUGH */
 167     case VINUM_PLEX_TYPE:
 168     case VINUM_RAWPLEX_TYPE:
 169         /*
 170          * Plex I/O is pretty much the same as volume I/O
 171          * for a single plex.  Indicate this by passing a NULL
 172          * pointer (set above) for the volume
 173          */
 174         bp->b_resid = bp->b_bcount;                         /* transfer everything */
 175         vinumstart(dev, nbio, 0);
 176         break;
 177     }
 178     return(0);
 179 }
 180
 181 /*
 182  * Start a transfer.  Return -1 on error,
 183  * 0 if OK, 1 if we need to retry.
 184  * Parameter reviveok is set when doing
 185  * transfers for revives: it allows transfers to
 186  * be started immediately when a revive is in
 187  * progress.  During revive, normal transfers
 188  * are queued if they share address space with
 189  * a currently active revive operation.
 190  */
 191 int
 192 vinumstart(cdev_t dev, struct bio *bio, int reviveok)
 193 {
 194     struct buf *bp = bio->bio_buf;
 195     int plexno;
 196     int maxplex;                                            /* maximum number of plexes to handle */
 197     struct volume *vol;
 198     struct request *rq;                                     /* build up our request here */
 199     enum requeststatus status;
 200
 201     bio->bio_driver_info = dev;
 202
 203 #if VINUMDEBUG
 204     if (debug & DEBUG_LASTREQS)
 205         logrq(loginfo_user_bp, (union rqinfou) bio, bio);
 206 #endif
 207
 208     if ((bp->b_bcount % DEV_BSIZE) != 0) {                  /* bad length */
 209         bp->b_error = EINVAL;                               /* invalid size */
 210         bp->b_flags |= B_ERROR;
 211         biodone(bio);
 212         return -1;
 213     }
 214     rq = (struct request *) Malloc(sizeof(struct request)); /* allocate a request struct */
 215     if (rq == NULL) {                                       /* can't do it */
 216         bp->b_error = ENOMEM;                               /* can't get memory */
 217         bp->b_flags |= B_ERROR;
 218         biodone(bio);
 219         return -1;
 220     }
 221     bzero(rq, sizeof(struct request));
 222
 223     /*
 224      * Note the volume ID.  This can be NULL, which
 225      * the request building functions use as an
 226      * indication for single plex I/O
 227      */
 228     rq->bio = bio;                                          /* and the user buffer struct */
 229
 230     if (DEVTYPE(dev) == VINUM_VOLUME_TYPE) {        /* it's a volume, */
 231         rq->volplex.volno = Volno(dev);             /* get the volume number */
 232         vol = &VOL[rq->volplex.volno];                      /* and point to it */
 233         vol->active++;                                      /* one more active request */
 234         maxplex = vol->plexes;                              /* consider all its plexes */
 235     } else {
 236         vol = NULL;                                         /* no volume */
 237         rq->volplex.plexno = Plexno(dev);                   /* point to the plex */
 238         rq->isplex = 1;                                     /* note that it's a plex */
 239         maxplex = 1;                                        /* just the one plex */
 240     }
 241
 242     if (bp->b_cmd == BUF_CMD_READ) {
 243         /*
 244          * This is a read request.  Decide
 245          * which plex to read from.
 246          *
 247          * There's a potential race condition here,
 248          * since we're not locked, and we could end
 249          * up multiply incrementing the round-robin
 250          * counter.  This doesn't have any serious
 251          * effects, however.
 252          */
 253         if (vol != NULL) {
 254             plexno = vol->preferred_plex;                   /* get the plex to use */
 255             if (plexno < 0) {                               /* round robin */
 256                 plexno = vol->last_plex_read;
 257                 vol->last_plex_read++;
 258                 if (vol->last_plex_read >= vol->plexes)     /* got the the end? */
 259                     vol->last_plex_read = 0;                /* wrap around */
 260             }
 261             status = build_read_request(rq, plexno);        /* build a request */
 262         } else {
 263             daddr_t diskaddr = (daddr_t)(bio->bio_offset >> DEV_BSHIFT);
 264                                                             /* start offset of transfer */
 265             status = bre(rq,                                /* build a request list */
 266                 rq->volplex.plexno,
 267                 &diskaddr,
 268                 diskaddr + (bp->b_bcount / DEV_BSIZE));
 269         }
 270
 271         if (status > REQUEST_RECOVERED) {                   /* can't satisfy it */
 272             if (status == REQUEST_DOWN) {                   /* not enough subdisks */
 273                 bp->b_error = EIO;                          /* I/O error */
 274                 bp->b_flags |= B_ERROR;
 275             }
 276             biodone(bio);
 277             freerq(rq);
 278             return -1;
 279         }
 280         return launch_requests(rq, reviveok);               /* now start the requests if we can */
 281     } else
 282         /*
 283          * This is a write operation.  We write to all plexes.  If this is
 284          * a RAID-4 or RAID-5 plex, we must also update the parity stripe.
 285          */
 286     {
 287         if (vol != NULL)
 288             status = build_write_request(rq);               /* Not all the subdisks are up */
 289         else {                                              /* plex I/O */
 290             daddr_t diskstart;
 291             daddr_t diskend;
 292
 293             diskstart = (daddr_t)(bio->bio_offset >> DEV_BSHIFT); /* start offset of transfer */
 294             diskend = diskstart + bp->b_bcount / DEV_BSIZE;
 295             status = bre(rq, Plexno(dev),
 296                 &diskstart, diskend);  /* build requests for the plex */
 297         }
 298         if (status > REQUEST_RECOVERED) {                   /* can't satisfy it */
 299             if (status == REQUEST_DOWN) {                   /* not enough subdisks */
 300                 bp->b_error = EIO;                          /* I/O error */
 301                 bp->b_flags |= B_ERROR;
 302             }
 303             biodone(bio);
 304             freerq(rq);
 305             return -1;
 306         }
 307         return launch_requests(rq, reviveok);               /* now start the requests if we can */
 308     }
 309 }
 310
 311 /*
 312  * Call the low-level strategy routines to
 313  * perform the requests in a struct request
 314  */
 315 int
 316 launch_requests(struct request *rq, int reviveok)
 317 {
 318     struct rqgroup *rqg;
 319     int rqno;                                               /* loop index */
 320     struct rqelement *rqe;                                  /* current element */
 321     struct drive *drive;
 322     int rcount;                                             /* request count */
 323
 324     /*
 325      * First find out whether we're reviving, and the
 326      * request contains a conflict.  If so, we hang
 327      * the request off plex->waitlist of the first
 328      * plex we find which is reviving
 329      */
 330
 331     if ((rq->flags & XFR_REVIVECONFLICT)                    /* possible revive conflict */
 332     &&(!reviveok)) {                                        /* and we don't want to do it now, */
 333         struct sd *sd;
 334         struct request *waitlist;                           /* point to the waitlist */
 335
 336         sd = &SD[rq->sdno];
 337         if (sd->waitlist != NULL) {                         /* something there already, */
 338             waitlist = sd->waitlist;
 339             while (waitlist->next != NULL)                  /* find the end */
 340                 waitlist = waitlist->next;
 341             waitlist->next = rq;                            /* hook our request there */
 342         } else
 343             sd->waitlist = rq;                              /* hook our request at the front */
 344
 345 #if VINUMDEBUG
 346         if (debug & DEBUG_REVIVECONFLICT) {
 347             log(LOG_DEBUG,
 348                 "Revive conflict sd %d: %p\n%s dev %d.%d, offset 0x%llx, length %d\n",
 349                 rq->sdno,
 350                 rq,
 351                 (rq->bio->bio_buf->b_cmd & BUF_CMD_READ) ? "Read" : "Write",
 352                 major(((cdev_t)rq->bio->bio_driver_info)),
 353                 minor(((cdev_t)rq->bio->bio_driver_info)),
 354                 rq->bio->bio_offset,
 355                 rq->bio->bio_buf->b_bcount);
 356         }
 357 #endif
 358         return 0;                                           /* and get out of here */
 359     }
 360     rq->active = 0;                                         /* nothing yet */
 361 #if VINUMDEBUG
 362     if (debug & DEBUG_ADDRESSES)
 363         log(LOG_DEBUG,
 364             "Request: %p\n%s dev %d.%d, offset 0x%llx, length %d\n",
 365             rq,
 366             (rq->bio->bio_buf->b_cmd == BUF_CMD_READ) ? "Read" : "Write",
 367             major(((cdev_t)rq->bio->bio_driver_info)),
 368             minor(((cdev_t)rq->bio->bio_driver_info)),
 369             rq->bio->bio_offset,
 370             rq->bio->bio_buf->b_bcount);
 371     vinum_conf.lastrq = rq;
 372     vinum_conf.lastbio = rq->bio;
 373     if (debug & DEBUG_LASTREQS)
 374         logrq(loginfo_user_bpl, (union rqinfou) rq->bio, rq->bio);
 375 #endif
 376
 377     /*
 378      * This loop happens without any participation
 379      * of the bottom half, so it requires no
 380      * protection.
 381      */
 382     for (rqg = rq->rqg; rqg != NULL; rqg = rqg->next) {     /* through the whole request chain */
 383         rqg->active = rqg->count;                           /* they're all active */
 384         for (rqno = 0; rqno < rqg->count; rqno++) {
 385             rqe = &rqg->rqe[rqno];
 386             if (rqe->flags & XFR_BAD_SUBDISK)               /* this subdisk is bad, */
 387                 rqg->active--;                              /* one less active request */
 388         }
 389         if (rqg->active)                                    /* we have at least one active request, */
 390             rq->active++;                                   /* one more active request group */
 391     }
 392
 393     /*
 394      * Now fire off the requests.  In this loop the
 395      * bottom half could be completing requests
 396      * before we finish, so we need critical section protection.
 397      */
 398     crit_enter();
 399     for (rqg = rq->rqg; rqg != NULL;) {                     /* through the whole request chain */
 400         if (rqg->lockbase >= 0)                             /* this rqg needs a lock first */
 401             rqg->lock = lockrange(rqg->lockbase, rqg->rq->bio->bio_buf, &PLEX[rqg->plexno]);
 402         rcount = rqg->count;
 403         for (rqno = 0; rqno < rcount;) {
 404             cdev_t dev;
 405
 406             rqe = &rqg->rqe[rqno];
 407
 408             /*
 409              * Point to next rqg before the bottom end
 410              * changes the structures.
 411              */
 412             if (++rqno >= rcount)
 413                 rqg = rqg->next;
 414             if ((rqe->flags & XFR_BAD_SUBDISK) == 0) {      /* this subdisk is good, */
 415                 drive = &DRIVE[rqe->driveno];               /* look at drive */
 416                 drive->active++;
 417                 if (drive->active >= drive->maxactive)
 418                     drive->maxactive = drive->active;
 419                 vinum_conf.active++;
 420                 if (vinum_conf.active >= vinum_conf.maxactive)
 421                     vinum_conf.maxactive = vinum_conf.active;
 422
 423                 dev = rqe->b.b_bio1.bio_driver_info;
 424 #ifdef VINUMDEBUG
 425                 if (debug & DEBUG_ADDRESSES)
 426                     log(LOG_DEBUG,
 427                         "  %s dev %d.%d, sd %d, offset 0x%llx, devoffset 0x%llx, length %d\n",
 428                         (rqe->b.b_cmd == BUF_CMD_READ) ? "Read" : "Write",
 429                         major(dev),
 430                         minor(dev),
 431                         rqe->sdno,
 432                         rqe->b.b_bio1.bio_offset - ((off_t)SD[rqe->sdno].driveoffset << DEV_BSHIFT),
 433                         rqe->b.b_bio1.bio_offset,
 434                         rqe->b.b_bcount);
 435                 if (debug & DEBUG_LASTREQS)
 436                     logrq(loginfo_rqe, (union rqinfou) rqe, rq->bio);
 437 #endif
 438                 /* fire off the request */
 439                 dev_dstrategy(dev, &rqe->b.b_bio1);
 440             }
 441         }
 442     }
 443     crit_exit();
 444     return 0;
 445 }
 446
 447 /*
 448  * define the low-level requests needed to perform a
 449  * high-level I/O operation for a specific plex 'plexno'.
 450  *
 451  * Return REQUEST_OK if all subdisks involved in the request are up,
 452  * REQUEST_DOWN if some subdisks are not up, and REQUEST_EOF if the
 453  * request is at least partially outside the bounds of the subdisks.
 454  *
 455  * Modify the pointer *diskstart to point to the end address.  On
 456  * read, return on the first bad subdisk, so that the caller
 457  * (build_read_request) can try alternatives.
 458  *
 459  * On entry to this routine, the rqg structures are not assigned.  The
 460  * assignment is performed by expandrq().  Strictly speaking, the
 461  * elements rqe->sdno of all entries should be set to -1, since 0
 462  * (from bzero) is a valid subdisk number.  We avoid this problem by
 463  * initializing the ones we use, and not looking at the others (index
 464  * >= rqg->requests).
 465  */
 466 enum requeststatus
 467 bre(struct request *rq,
 468     int plexno,
 469     daddr_t * diskaddr,
 470     daddr_t diskend)
 471 {
 472     int sdno;
 473     struct sd *sd;
 474     struct rqgroup *rqg;
 475     struct bio *bio;
 476     struct buf *bp;                                         /* user's bp */
 477     struct plex *plex;
 478     enum requeststatus status;                              /* return value */
 479     daddr_t plexoffset;                                     /* offset of transfer in plex */
 480     daddr_t stripebase;                                     /* base address of stripe (1st subdisk) */
 481     daddr_t stripeoffset;                                   /* offset in stripe */
 482     daddr_t blockoffset;                                    /* offset in stripe on subdisk */
 483     struct rqelement *rqe;                                  /* point to this request information */
 484     daddr_t diskstart = *diskaddr;                          /* remember where this transfer starts */
 485     enum requeststatus s;                                   /* temp return value */
 486
 487     bio = rq->bio;                                          /* buffer pointer */
 488     bp = bio->bio_buf;
 489     status = REQUEST_OK;                                    /* return value: OK until proven otherwise */
 490     plex = &PLEX[plexno];                                   /* point to the plex */
 491
 492     switch (plex->organization) {
 493     case plex_concat:
 494         sd = NULL;                                          /* (keep compiler quiet) */
 495         for (sdno = 0; sdno < plex->subdisks; sdno++) {
 496             sd = &SD[plex->sdnos[sdno]];
 497             if (*diskaddr < sd->plexoffset)                 /* we must have a hole, */
 498                 status = REQUEST_DEGRADED;                  /* note the fact */
 499             if (*diskaddr < (sd->plexoffset + sd->sectors)) { /* the request starts in this subdisk */
 500                 rqg = allocrqg(rq, 1);                      /* space for the request */
 501                 if (rqg == NULL) {                          /* malloc failed */
 502                     bp->b_error = ENOMEM;
 503                     bp->b_flags |= B_ERROR;
 504                     return REQUEST_ENOMEM;
 505                 }
 506                 rqg->plexno = plexno;
 507
 508                 rqe = &rqg->rqe[0];                         /* point to the element */
 509                 rqe->rqg = rqg;                             /* group */
 510                 rqe->sdno = sd->sdno;                       /* put in the subdisk number */
 511                 plexoffset = *diskaddr;                     /* start offset in plex */
 512                 rqe->sdoffset = plexoffset - sd->plexoffset; /* start offset in subdisk */
 513                 rqe->useroffset = plexoffset - diskstart;   /* start offset in user buffer */
 514                 rqe->dataoffset = 0;
 515                 rqe->datalen = min(diskend - *diskaddr,     /* number of sectors to transfer in this sd */
 516                     sd->sectors - rqe->sdoffset);
 517                 rqe->groupoffset = 0;                       /* no groups for concatenated plexes */
 518                 rqe->grouplen = 0;
 519                 rqe->buflen = rqe->datalen;                 /* buffer length is data buffer length */
 520                 rqe->flags = 0;
 521                 rqe->driveno = sd->driveno;
 522                 if (sd->state != sd_up) {                   /* *now* we find the sd is down */
 523                     s = checksdstate(sd, rq, *diskaddr, diskend); /* do we need to change state? */
 524                     if (s == REQUEST_DOWN) {                /* down? */
 525                         rqe->flags = XFR_BAD_SUBDISK;       /* yup */
 526                         if (rq->bio->bio_buf->b_cmd == BUF_CMD_READ)    /* read request, */
 527                             return REQUEST_DEGRADED;        /* give up here */
 528                         /*
 529                          * If we're writing, don't give up
 530                          * because of a bad subdisk.  Go
 531                          * through to the bitter end, but note
 532                          * which ones we can't access.
 533                          */
 534                         status = REQUEST_DEGRADED;          /* can't do it all */
 535                     }
 536                 }
 537                 *diskaddr += rqe->datalen;                  /* bump the address */
 538                 if (build_rq_buffer(rqe, plex)) {           /* build the buffer */
 539                     deallocrqg(rqg);
 540                     bp->b_error = ENOMEM;
 541                     bp->b_flags |= B_ERROR;
 542                     return REQUEST_ENOMEM;                  /* can't do it */
 543                 }
 544             }
 545             if (*diskaddr == diskend)                       /* we're finished, */
 546                 break;                                      /* get out of here */
 547         }
 548         /*
 549          * We've got to the end of the plex.  Have we got to the end of
 550          * the transfer?  It would seem that having an offset beyond the
 551          * end of the subdisk is an error, but in fact it can happen if
 552          * the volume has another plex of different size.  There's a valid
 553          * question as to why you would want to do this, but currently
 554          * it's allowed.
 555          *
 556          * In a previous version, I returned REQUEST_DOWN here.  I think
 557          * REQUEST_EOF is more appropriate now.
 558          */
 559         if (diskend > sd->sectors + sd->plexoffset)         /* pointing beyond EOF? */
 560             status = REQUEST_EOF;
 561         break;
 562
 563     case plex_striped:
 564         {
 565             while (*diskaddr < diskend) {                   /* until we get it all sorted out */
 566                 if (*diskaddr >= plex->length)              /* beyond the end of the plex */
 567                     return REQUEST_EOF;                     /* can't continue */
 568
 569                 /* The offset of the start address from the start of the stripe. */
 570                 stripeoffset = *diskaddr % (plex->stripesize * plex->subdisks);
 571
 572                 /* The plex-relative address of the start of the stripe. */
 573                 stripebase = *diskaddr - stripeoffset;
 574
 575                 /* The number of the subdisk in which the start is located. */
 576                 sdno = stripeoffset / plex->stripesize;
 577
 578                 /* The offset from the beginning of the stripe on this subdisk. */
 579                 blockoffset = stripeoffset % plex->stripesize;
 580
 581                 sd = &SD[plex->sdnos[sdno]];                /* the subdisk in question */
 582                 rqg = allocrqg(rq, 1);                      /* space for the request */
 583                 if (rqg == NULL) {                          /* malloc failed */
 584                     bp->b_error = ENOMEM;
 585                     bp->b_flags |= B_ERROR;
 586                     return REQUEST_ENOMEM;
 587                 }
 588                 rqg->plexno = plexno;
 589
 590                 rqe = &rqg->rqe[0];                         /* point to the element */
 591                 rqe->rqg = rqg;
 592                 rqe->sdoffset = stripebase / plex->subdisks + blockoffset; /* start offset in this subdisk */
 593                 rqe->useroffset = *diskaddr - diskstart;    /* The offset of the start in the user buffer */
 594                 rqe->dataoffset = 0;
 595                 rqe->datalen = min(diskend - *diskaddr,     /* the amount remaining to transfer */
 596                     plex->stripesize - blockoffset);        /* and the amount left in this stripe */
 597                 rqe->groupoffset = 0;                       /* no groups for striped plexes */
 598                 rqe->grouplen = 0;
 599                 rqe->buflen = rqe->datalen;                 /* buffer length is data buffer length */
 600                 rqe->flags = 0;
 601                 rqe->sdno = sd->sdno;                       /* put in the subdisk number */
 602                 rqe->driveno = sd->driveno;
 603
 604                 if (sd->state != sd_up) {                   /* *now* we find the sd is down */
 605                     s = checksdstate(sd, rq, *diskaddr, diskend); /* do we need to change state? */
 606                     if (s == REQUEST_DOWN) {                /* down? */
 607                         rqe->flags = XFR_BAD_SUBDISK;       /* yup */
 608                         if (rq->bio->bio_buf->b_cmd == BUF_CMD_READ)        /* read request, */
 609                             return REQUEST_DEGRADED;        /* give up here */
 610                         /*
 611                          * If we're writing, don't give up
 612                          * because of a bad subdisk.  Go through
 613                          * to the bitter end, but note which
 614                          * ones we can't access.
 615                          */
 616                         status = REQUEST_DEGRADED;          /* can't do it all */
 617                     }
 618                 }
 619                 /*
 620                  * It would seem that having an offset
 621                  * beyond the end of the subdisk is an
 622                  * error, but in fact it can happen if the
 623                  * volume has another plex of different
 624                  * size.  There's a valid question as to why
 625                  * you would want to do this, but currently
 626                  * it's allowed.
 627                  */
 628                 if (rqe->sdoffset + rqe->datalen > sd->sectors) { /* ends beyond the end of the subdisk? */
 629                     rqe->datalen = sd->sectors - rqe->sdoffset; /* truncate */
 630 #if VINUMDEBUG
 631                     if (debug & DEBUG_EOFINFO) {            /* tell on the request */
 632                         log(LOG_DEBUG,
 633                             "vinum: EOF on plex %s, sd %s offset %llx (user offset %x)\n",
 634                             plex->name,
 635                             sd->name,
 636                             (u_int) sd->sectors,
 637                             bp->b_bio1.bio_offset);
 638                         log(LOG_DEBUG,
 639                             "vinum: stripebase %x, stripeoffset %x, blockoffset %x\n",
 640                             stripebase,
 641                             stripeoffset,
 642                             blockoffset);
 643                     }
 644 #endif
 645                 }
 646                 if (build_rq_buffer(rqe, plex)) {           /* build the buffer */
 647                     deallocrqg(rqg);
 648                     bp->b_error = ENOMEM;
 649                     bp->b_flags |= B_ERROR;
 650                     return REQUEST_ENOMEM;                  /* can't do it */
 651                 }
 652                 *diskaddr += rqe->datalen;                  /* look at the remainder */
 653                 if ((*diskaddr < diskend)                   /* didn't finish the request on this stripe */
 654                 &&(*diskaddr < plex->length)) {             /* and there's more to come */
 655                     plex->multiblock++;                     /* count another one */
 656                     if (sdno == plex->subdisks - 1)         /* last subdisk, */
 657                         plex->multistripe++;                /* another stripe as well */
 658                 }
 659             }
 660         }
 661         break;
 662
 663         /*
 664          * RAID-4 and RAID-5 are complicated enough to have their own
 665          * function.
 666          */
 667     case plex_raid4:
 668     case plex_raid5:
 669         status = bre5(rq, plexno, diskaddr, diskend);
 670         break;
 671
 672     default:
 673         log(LOG_ERR, "vinum: invalid plex type %d in bre\n", plex->organization);
 674         status = REQUEST_DOWN;                              /* can't access it */
 675     }
 676
 677     return status;
 678 }
 679
 680 /*
 681  * Build up a request structure for reading volumes.
 682  * This function is not needed for plex reads, since there's
 683  * no recovery if a plex read can't be satisified.
 684  */
 685 enum requeststatus
 686 build_read_request(struct request *rq,                      /* request */
 687     int plexindex)
 688 {                                                           /* index in the volume's plex table */
 689     struct bio *bio;
 690     struct buf *bp;
 691     daddr_t startaddr;                                      /* offset of previous part of transfer */
 692     daddr_t diskaddr;                                       /* offset of current part of transfer */
 693     daddr_t diskend;                                        /* and end offset of transfer */
 694     int plexno;                                             /* plex index in vinum_conf */
 695     struct rqgroup *rqg;                                    /* point to the request we're working on */
 696     struct volume *vol;                                     /* volume in question */
 697     int recovered = 0;                                      /* set if we recover a read */
 698     enum requeststatus status = REQUEST_OK;
 699     int plexmask;                                           /* bit mask of plexes, for recovery */
 700
 701     bio = rq->bio;                                          /* buffer pointer */
 702     bp = bio->bio_buf;
 703     diskaddr = bio->bio_offset >> DEV_BSHIFT;               /* 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 bio *bio;
 770     struct buf *bp;
 771     daddr_t diskstart;                                      /* offset of current part of transfer */
 772     daddr_t diskend;                                        /* and end offset of transfer */
 773     int plexno;                                             /* plex index in vinum_conf */
 774     struct volume *vol;                                     /* volume in question */
 775     enum requeststatus status;
 776
 777     bio = rq->bio;                                          /* buffer pointer */
 778     bp = bio->bio_buf;
 779     vol = &VOL[rq->volplex.volno];                          /* point to volume */
 780     diskend = (daddr_t)(bio->bio_offset >> DEV_BSHIFT) + (bp->b_bcount / DEV_BSIZE);        /* end offset of transfer */
 781     status = REQUEST_DOWN;                                  /* assume the worst */
 782     for (plexno = 0; plexno < vol->plexes; plexno++) {
 783         diskstart = (daddr_t)(bio->bio_offset >> DEV_BSHIFT);                       /* start offset of transfer */
 784         /*
 785          * Build requests for the plex.
 786          * We take the best possible result here (min,
 787          * not max): we're happy if we can write at all
 788          */
 789         status = min(status, bre(rq,
 790                 vol->plex[plexno],
 791                 &diskstart,
 792                 diskend));
 793     }
 794     return status;
 795 }
 796
 797 /* Fill in the struct buf part of a request element. */
 798 enum requeststatus
 799 build_rq_buffer(struct rqelement *rqe, struct plex *plex)
 800 {
 801     struct sd *sd;                                          /* point to subdisk */
 802     struct volume *vol;
 803     struct buf *bp;
 804     struct buf *ubp;                                        /* user (high level) buffer header */
 805     struct bio *ubio;
 806
 807     vol = &VOL[rqe->rqg->rq->volplex.volno];
 808     sd = &SD[rqe->sdno];                                    /* point to subdisk */
 809     bp = &rqe->b;
 810     ubio = rqe->rqg->rq->bio;                               /* pointer to user buffer header */
 811     ubp = ubio->bio_buf;
 812
 813     /* Initialize the buf struct */
 814     /* copy these flags from user bp */
 815     bp->b_flags = ubp->b_flags & (B_ORDERED | B_NOCACHE | B_ASYNC);
 816     bp->b_cmd = ubp->b_cmd;
 817 #ifdef VINUMDEBUG
 818     if (rqe->flags & XFR_BUFLOCKED)                         /* paranoia */
 819         panic("build_rq_buffer: rqe already locked");       /* XXX remove this when we're sure */
 820 #endif
 821     BUF_LOCKINIT(bp);                                       /* get a lock for the buffer */
 822     BUF_LOCK(bp, LK_EXCLUSIVE);                             /* and lock it */
 823     BUF_KERNPROC(bp);
 824     initbufbio(bp);
 825     rqe->flags |= XFR_BUFLOCKED;
 826     bp->b_bio1.bio_done = complete_rqe;
 827     /*
 828      * You'd think that we wouldn't need to even
 829      * build the request buffer for a dead subdisk,
 830      * but in some cases we need information like
 831      * the user buffer address.  Err on the side of
 832      * generosity and supply what we can.  That
 833      * obviously doesn't include drive information
 834      * when the drive is dead.
 835      */
 836     if ((rqe->flags & XFR_BAD_SUBDISK) == 0)                /* subdisk is accessible, */
 837         bp->b_bio1.bio_driver_info = DRIVE[rqe->driveno].dev; /* drive device */
 838     bp->b_bio1.bio_offset = (off_t)(rqe->sdoffset + sd->driveoffset) << DEV_BSHIFT;     /* start address */
 839     bp->b_bcount = rqe->buflen << DEV_BSHIFT;               /* number of bytes to transfer */
 840     bp->b_resid = bp->b_bcount;                             /* and it's still all waiting */
 841
 842     if (rqe->flags & XFR_MALLOCED) {                        /* this operation requires a malloced buffer */
 843         bp->b_data = Malloc(bp->b_bcount);                  /* get a buffer to put it in */
 844         if (bp->b_data == NULL) {                           /* failed */
 845             abortrequest(rqe->rqg->rq, ENOMEM);
 846             return REQUEST_ENOMEM;                          /* no memory */
 847         }
 848     } else
 849         /*
 850          * Point directly to user buffer data.  This means
 851          * that we don't need to do anything when we have
 852          * finished the transfer
 853          */
 854         bp->b_data = ubp->b_data + rqe->useroffset * DEV_BSIZE;
 855     /*
 856      * On a recovery read, we perform an XOR of
 857      * all blocks to the user buffer.  To make
 858      * this work, we first clean out the buffer
 859      */
 860     if ((rqe->flags & (XFR_RECOVERY_READ | XFR_BAD_SUBDISK))
 861         == (XFR_RECOVERY_READ | XFR_BAD_SUBDISK)) {         /* bad subdisk of a recovery read */
 862         int length = rqe->grouplen << DEV_BSHIFT;           /* and count involved */
 863         char *data = (char *) &rqe->b.b_data[rqe->groupoffset << DEV_BSHIFT]; /* destination */
 864
 865         bzero(data, length);                                /* clean it out */
 866     }
 867     return 0;
 868 }
 869
 870 /*
 871  * Abort a request: free resources and complete the
 872  * user request with the specified error
 873  */
 874 int
 875 abortrequest(struct request *rq, int error)
 876 {
 877     struct buf *bp = rq->bio->bio_buf;                      /* user buffer */
 878
 879     bp->b_error = error;
 880     freerq(rq);                                             /* free everything we're doing */
 881     bp->b_flags |= B_ERROR;
 882     return error;                                           /* and give up */
 883 }
 884
 885 /*
 886  * Check that our transfer will cover the
 887  * complete address space of the user request.
 888  *
 889  * Return 1 if it can, otherwise 0
 890  */
 891 int
 892 check_range_covered(struct request *rq)
 893 {
 894     return 1;
 895 }
 896
 897 /* Perform I/O on a subdisk */
 898 void
 899 sdio(struct bio *bio)
 900 {
 901     cdev_t dev;
 902     cdev_t sddev;
 903     struct sd *sd;
 904     struct sdbuf *sbp;
 905     daddr_t endoffset;
 906     struct drive *drive;
 907     struct buf *bp = bio->bio_buf;
 908
 909     dev = bio->bio_driver_info;
 910
 911 #if VINUMDEBUG
 912     if (debug & DEBUG_LASTREQS)
 913         logrq(loginfo_sdio, (union rqinfou) bio, bio);
 914 #endif
 915     sd = &SD[Sdno(dev)];                                    /* point to the subdisk */
 916     drive = &DRIVE[sd->driveno];
 917
 918     if (drive->state != drive_up) {
 919         if (sd->state >= sd_crashed) {
 920             if (bp->b_cmd != BUF_CMD_READ)                  /* writing, */
 921                 set_sd_state(sd->sdno, sd_stale, setstate_force);
 922             else
 923                 set_sd_state(sd->sdno, sd_crashed, setstate_force);
 924         }
 925         bp->b_error = EIO;
 926         bp->b_flags |= B_ERROR;
 927         biodone(bio);
 928         return;
 929     }
 930     /*
 931      * We allow access to any kind of subdisk as long as we can expect
 932      * to get the I/O performed.
 933      */
 934     if (sd->state < sd_empty) {                             /* nothing to talk to, */
 935         bp->b_error = EIO;
 936         bp->b_flags |= B_ERROR;
 937         biodone(bio);
 938         return;
 939     }
 940     /* Get a buffer */
 941     sbp = (struct sdbuf *) Malloc(sizeof(struct sdbuf));
 942     if (sbp == NULL) {
 943         bp->b_error = ENOMEM;
 944         bp->b_flags |= B_ERROR;
 945         biodone(bio);
 946         return;
 947     }
 948     sddev = DRIVE[sd->driveno].dev;                 /* device */
 949     bzero(sbp, sizeof(struct sdbuf));                       /* start with nothing */
 950     sbp->b.b_cmd = bp->b_cmd;
 951     sbp->b.b_bcount = bp->b_bcount;                         /* number of bytes to transfer */
 952     sbp->b.b_resid = bp->b_resid;                           /* and amount waiting */
 953     sbp->b.b_data = bp->b_data;                             /* data buffer */
 954     BUF_LOCKINIT(&sbp->b);                                  /* get a lock for the buffer */
 955     BUF_LOCK(&sbp->b, LK_EXCLUSIVE);                        /* and lock it */
 956     BUF_KERNPROC(&sbp->b);
 957     initbufbio(&sbp->b);
 958     sbp->b.b_bio1.bio_offset = bio->bio_offset + ((off_t)sd->driveoffset << DEV_BSHIFT);
 959     sbp->b.b_bio1.bio_done = sdio_done;                     /* come here on completion */
 960     sbp->bio = bio;                                         /* note the address of the original header */
 961     sbp->sdno = sd->sdno;                                   /* note for statistics */
 962     sbp->driveno = sd->driveno;
 963     endoffset = (daddr_t)(bio->bio_offset >> DEV_BSHIFT) + sbp->b.b_bcount / DEV_BSIZE;  /* final sector offset */
 964     if (endoffset > sd->sectors) {                          /* beyond the end */
 965         sbp->b.b_bcount -= (endoffset - sd->sectors) * DEV_BSIZE; /* trim */
 966         if (sbp->b.b_bcount <= 0) {                         /* nothing to transfer */
 967             bp->b_resid = bp->b_bcount;                     /* nothing transferred */
 968             biodone(bio);
 969             BUF_UNLOCK(&sbp->b);
 970             BUF_LOCKFREE(&sbp->b);
 971             Free(sbp);
 972             return;
 973         }
 974     }
 975 #if VINUMDEBUG
 976     if (debug & DEBUG_ADDRESSES)
 977         log(LOG_DEBUG,
 978             "  %s dev %d.%d, sd %d, offset 0x%llx, devoffset 0x%llx, length %d\n",
 979             (sbp->b.b_cmd == BUF_CMD_READ) ? "Read" : "Write",
 980             major(sddev),
 981             minor(sddev),
 982             sbp->sdno,
 983             sbp->b.b_bio1.bio_offset - ((off_t)SD[sbp->sdno].driveoffset << DEV_BSHIFT),
 984             sbp->b.b_bio1.bio_offset,
 985             sbp->b.b_bcount);
 986 #endif
 987     crit_enter();
 988 #if VINUMDEBUG
 989     if (debug & DEBUG_LASTREQS)
 990         logrq(loginfo_sdiol, (union rqinfou) &sbp->b.b_bio1, &sbp->b.b_bio1);
 991 #endif
 992     dev_dstrategy(sddev, &sbp->b.b_bio1);
 993     crit_exit();
 994 }
 995
 996 /*
 997  * Determine the size of the transfer, and make sure it is
 998  * within the boundaries of the partition. Adjust transfer
 999  * if needed, and signal errors or early completion.
1000  *
1001  * Volumes are simpler than disk slices: they only contain
1002  * one component (though we call them a, b and c to make
1003  * system utilities happy), and they always take up the
1004  * complete space of the "partition".
1005  *
1006  * I'm still not happy with this: why should the label be
1007  * protected?  If it weren't so damned difficult to write
1008  * one in the first pleace (because it's protected), it wouldn't
1009  * be a problem.
1010  */
1011 struct bio *
1012 vinum_bounds_check(struct bio *bio, struct volume *vol)
1013 {
1014     struct buf *bp = bio->bio_buf;
1015     struct bio *nbio;
1016     int maxsize = vol->size;                                /* size of the partition (sectors) */
1017     int size = (bp->b_bcount + DEV_BSIZE - 1) >> DEV_BSHIFT; /* size of this request (sectors) */
1018     daddr_t blkno = (daddr_t)(bio->bio_offset >> DEV_BSHIFT);
1019
1020     if (size == 0)                                          /* no transfer specified, */
1021         return 0;                                           /* treat as EOF */
1022     /* beyond partition? */
1023     if (bio->bio_offset < 0                                 /* negative start */
1024         || blkno + size > maxsize) {                /* or goes beyond the end of the partition */
1025         /* if exactly at end of disk, return an EOF */
1026         if (blkno == maxsize) {
1027             bp->b_resid = bp->b_bcount;
1028             return (NULL);
1029         }
1030         /* or truncate if part of it fits */
1031         size = maxsize - blkno;
1032         if (size <= 0) {                                    /* nothing to transfer */
1033             bp->b_error = EINVAL;
1034             bp->b_flags |= B_ERROR;
1035             return (NULL);
1036         }
1037         bp->b_bcount = size << DEV_BSHIFT;
1038     }
1039     nbio = push_bio(bio);
1040     nbio->bio_offset = bio->bio_offset;
1041     return (nbio);
1042 }
1043
1044 /*
1045  * Allocate a request group and hook
1046  * it in in the list for rq
1047  */
1048 struct rqgroup *
1049 allocrqg(struct request *rq, int elements)
1050 {
1051     struct rqgroup *rqg;                                    /* the one we're going to allocate */
1052     int size = sizeof(struct rqgroup) + elements * sizeof(struct rqelement);
1053
1054     rqg = (struct rqgroup *) Malloc(size);
1055     if (rqg != NULL) {                                      /* malloc OK, */
1056         if (rq->rqg)                                        /* we already have requests */
1057             rq->lrqg->next = rqg;                           /* hang it off the end */
1058         else                                                /* first request */
1059             rq->rqg = rqg;                                  /* at the start */
1060         rq->lrqg = rqg;                                     /* this one is the last in the list */
1061
1062         bzero(rqg, size);                                   /* no old junk */
1063         rqg->rq = rq;                                       /* point back to the parent request */
1064         rqg->count = elements;                              /* number of requests in the group */
1065         rqg->lockbase = -1;                                 /* no lock required yet */
1066     }
1067     return rqg;
1068 }
1069
1070 /*
1071  * Deallocate a request group out of a chain.  We do
1072  * this by linear search: the chain is short, this
1073  * almost never happens, and currently it can only
1074  * happen to the first member of the chain.
1075  */
1076 void
1077 deallocrqg(struct rqgroup *rqg)
1078 {
1079     struct rqgroup *rqgc = rqg->rq->rqg;                    /* point to the request chain */
1080
1081     if (rqg->lock)                                          /* got a lock? */
1082         unlockrange(rqg->plexno, rqg->lock);                /* yes, free it */
1083     if (rqgc == rqg)                                        /* we're first in line */
1084         rqg->rq->rqg = rqg->next;                           /* unhook ourselves */
1085     else {
1086         while ((rqgc->next != NULL)                         /* find the group */
1087         &&(rqgc->next != rqg))
1088             rqgc = rqgc->next;
1089         if (rqgc->next == NULL)
1090             log(LOG_ERR,
1091                 "vinum deallocrqg: rqg %p not found in request %p\n",
1092                 rqg->rq,
1093                 rqg);
1094         else
1095             rqgc->next = rqg->next;                         /* make the chain jump over us */
1096     }
1097     Free(rqg);
1098 }
1099
1100 /* Local Variables: */
1101 /* fill-column: 50 */
1102 /* End: */