sys/kern/subr_diskslice.c

   1 /*-
   2  * Copyright (c) 1994 Bruce D. Evans.
   3  * All rights reserved.
   4  *
   5  * Copyright (c) 1990 The Regents of the University of California.
   6  * All rights reserved.
   7  *
   8  * This code is derived from software contributed to Berkeley by
   9  * William Jolitz.
  10  *
  11  * Copyright (c) 1982, 1986, 1988 Regents of the University of California.
  12  * All rights reserved.
  13  *
  14  * Redistribution and use in source and binary forms, with or without
  15  * modification, are permitted provided that the following conditions
  16  * are met:
  17  * 1. Redistributions of source code must retain the above copyright
  18  *    notice, this list of conditions and the following disclaimer.
  19  * 2. Redistributions in binary form must reproduce the above copyright
  20  *    notice, this list of conditions and the following disclaimer in the
  21  *    documentation and/or other materials provided with the distribution.
  22  * 3. All advertising materials mentioning features or use of this software
  23  *    must display the following acknowledgement:
  24  *      This product includes software developed by the University of
  25  *      California, Berkeley and its contributors.
  26  * 4. Neither the name of the University nor the names of its contributors
  27  *    may be used to endorse or promote products derived from this software
  28  *    without specific prior written permission.
  29  *
  30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  40  * SUCH DAMAGE.
  41  *
  42  *      from: @(#)wd.c  7.2 (Berkeley) 5/9/91
  43  *      from: wd.c,v 1.55 1994/10/22 01:57:12 phk Exp $
  44  *      from: @(#)ufs_disksubr.c        7.16 (Berkeley) 5/4/91
  45  *      from: ufs_disksubr.c,v 1.8 1994/06/07 01:21:39 phk Exp $
  46  * $FreeBSD: src/sys/kern/subr_diskslice.c,v 1.82.2.6 2001/07/24 09:49:41 dd Exp $
  47  * $DragonFly: src/sys/kern/subr_diskslice.c,v 1.50.2.1 2008/09/25 01:44:52 dillon Exp $
  48  */
  49
  50 #include <sys/param.h>
  51 #include <sys/systm.h>
  52 #include <sys/buf.h>
  53 #include <sys/conf.h>
  54 #include <sys/disklabel.h>
  55 #include <sys/disklabel32.h>
  56 #include <sys/disklabel64.h>
  57 #include <sys/diskslice.h>
  58 #include <sys/disk.h>
  59 #include <sys/diskmbr.h>
  60 #include <sys/fcntl.h>
  61 #include <sys/malloc.h>
  62 #include <sys/stat.h>
  63 #include <sys/syslog.h>
  64 #include <sys/proc.h>
  65 #include <sys/vnode.h>
  66 #include <sys/device.h>
  67 #include <sys/thread2.h>
  68
  69 #include <vfs/ufs/dinode.h>     /* XXX used only for fs.h */
  70 #include <vfs/ufs/fs.h>         /* XXX used only to get BBSIZE/SBSIZE */
  71
  72 static int  dsreadandsetlabel(cdev_t dev, u_int flags,
  73                            struct diskslices *ssp, struct diskslice *sp,
  74                            struct disk_info *info);
  75 static void free_ds_label (struct diskslices *ssp, int slice);
  76 static void set_ds_label (struct diskslices *ssp, int slice, disklabel_t lp,
  77                            disklabel_ops_t ops);
  78 static void set_ds_wlabel (struct diskslices *ssp, int slice, int wlabel);
  79
  80 /*
  81  * Determine the size of the transfer, and make sure it is
  82  * within the boundaries of the partition. Adjust transfer
  83  * if needed, and signal errors or early completion.
  84  *
  85  * XXX TODO:
  86  *      o Split buffers that are too big for the device.
  87  *      o Check for overflow.
  88  *      o Finish cleaning this up.
  89  *
  90  * This function returns 1 on success, 0 if transfer equates
  91  * to EOF (end of disk) or -1 on failure.  The appropriate
  92  * 'errno' value is also set in bp->b_error and bp->b_flags
  93  * is marked with B_ERROR.
  94  */
  95 struct bio *
  96 dscheck(cdev_t dev, struct bio *bio, struct diskslices *ssp)
  97 {
  98         struct buf *bp = bio->bio_buf;
  99         struct bio *nbio;
 100         disklabel_t lp;
 101         disklabel_ops_t ops;
 102         long nsec;
 103         u_int64_t secno;
 104         u_int64_t endsecno;
 105         u_int64_t slicerel_secno;
 106         struct diskslice *sp;
 107         u_int32_t part;
 108         u_int32_t slice;
 109         int shift;
 110         int mask;
 111
 112         slice = dkslice(dev);
 113         part  = dkpart(dev);
 114
 115         if (bio->bio_offset < 0) {
 116                 kprintf("dscheck(%s): negative bio_offset %lld\n",
 117                         devtoname(dev), bio->bio_offset);
 118                 goto bad;
 119         }
 120         if (slice >= ssp->dss_nslices) {
 121                 kprintf("dscheck(%s): slice too large %d/%d\n",
 122                         devtoname(dev), slice, ssp->dss_nslices);
 123                 goto bad;
 124         }
 125         sp = &ssp->dss_slices[slice];
 126
 127         /*
 128          * Calculate secno and nsec
 129          */
 130         if (ssp->dss_secmult == 1) {
 131                 shift = DEV_BSHIFT;
 132                 goto doshift;
 133         } else if (ssp->dss_secshift != -1) {
 134                 shift = DEV_BSHIFT + ssp->dss_secshift;
 135 doshift:
 136                 mask = (1 << shift) - 1;
 137                 if ((int)bp->b_bcount & mask)
 138                         goto bad_bcount;
 139                 if ((int)bio->bio_offset & mask)
 140                         goto bad_blkno;
 141                 secno = bio->bio_offset >> shift;
 142                 nsec = bp->b_bcount >> shift;
 143         } else {
 144                 if (bp->b_bcount % ssp->dss_secsize)
 145                         goto bad_bcount;
 146                 if (bio->bio_offset % ssp->dss_secsize)
 147                         goto bad_blkno;
 148                 secno = bio->bio_offset / ssp->dss_secsize;
 149                 nsec = bp->b_bcount / ssp->dss_secsize;
 150         }
 151
 152         /*
 153          * Calculate slice-relative sector number end slice-relative
 154          * limit.
 155          */
 156         if (slice == WHOLE_DISK_SLICE) {
 157                 /*
 158                  * Labels have not been allowed on whole-disks for a while.
 159                  * This really puts the nail in the coffin.
 160                  *
 161                  * Accesses to the WHOLE_DISK_SLICE do not use a disklabel
 162                  * and partition numbers are special-cased.  Currently numbers
 163                  * less then 128 are not allowed.  Partition numbers >= 128
 164                  * are encoded in the high 8 bits of the 64 bit buffer offset
 165                  * and are fed directly through to the device with no
 166                  * further interpretation.  In particular, no sector
 167                  * translation interpretation should occur because the
 168                  * sector size for the special raw access may not be the
 169                  * same as the nominal sector size for the device.
 170                  */
 171                 lp.opaque = NULL;
 172                 if (part < 128) {
 173                         kprintf("dscheck(%s): illegal partition number (%d) "
 174                                 "for WHOLE_DISK_SLICE access\n",
 175                                 devtoname(dev), part);
 176                         goto bad;
 177                 } else if (part != WHOLE_SLICE_PART) {
 178                         nbio = push_bio(bio);
 179                         nbio->bio_offset = bio->bio_offset |
 180                                            (u_int64_t)part << 56;
 181                         return(nbio);
 182                 }
 183
 184                 /*
 185                  * sp->ds_size is for the whole disk in the WHOLE_DISK_SLICE,
 186                  * there are no reserved areas.
 187                  */
 188                 endsecno = sp->ds_size;
 189                 slicerel_secno = secno;
 190         } else if (part == WHOLE_SLICE_PART) {
 191                 /*
 192                  * NOTE! opens on a whole-slice partition will not attempt
 193                  * to read a disklabel in, so there may not be an in-core
 194                  * disklabel even if there is one on the disk.
 195                  */
 196                 endsecno = sp->ds_size;
 197                 slicerel_secno = secno;
 198         } else if ((lp = sp->ds_label).opaque != NULL) {
 199                 /*
 200                  * A label is present, extract the partition.  Snooping of
 201                  * the disklabel is not supported even if accessible.  Of
 202                  * course, the reserved area is still write protected.
 203                  */
 204                 ops = sp->ds_ops;
 205                 if (ops->op_getpartbounds(ssp, lp, part,
 206                                           &slicerel_secno, &endsecno)) {
 207                         kprintf("dscheck(%s): partition %d out of bounds\n",
 208                                 devtoname(dev), part);
 209                         goto bad;
 210                 }
 211                 slicerel_secno += secno;
 212         } else {
 213                 /*
 214                  * Attempt to access partition when no disklabel present
 215                  */
 216                 kprintf("dscheck(%s): attempt to access non-existent partition\n",
 217                         devtoname(dev));
 218                 goto bad;
 219         }
 220
 221         /*
 222          * Disallow writes to reserved areas unless ds_wlabel allows it.
 223          */
 224         if (slicerel_secno < sp->ds_reserved && nsec &&
 225             bp->b_cmd == BUF_CMD_WRITE && sp->ds_wlabel == 0) {
 226                 bp->b_error = EROFS;
 227                 goto error;
 228         }
 229
 230         /*
 231          * If we get here, bio_offset must be on a block boundary and
 232          * the sector size must be a power of 2.
 233          */
 234         if ((bio->bio_offset & (ssp->dss_secsize - 1)) ||
 235             (ssp->dss_secsize ^ (ssp->dss_secsize - 1)) !=
 236             ((ssp->dss_secsize << 1) - 1)) {
 237                 kprintf("%s: invalid BIO offset, not sector aligned or"
 238                         " invalid sector size (not power of 2) %08llx %d\n",
 239                         devtoname(dev), bio->bio_offset, ssp->dss_secsize);
 240                 goto bad;
 241         }
 242
 243         /*
 244          * EOF handling
 245          */
 246         if (secno + nsec > endsecno) {
 247                 /*
 248                  * Return an error if beyond the end of the disk, or
 249                  * if B_BNOCLIP is set.  Tell the system that we do not
 250                  * need to keep the buffer around.
 251                  */
 252                 if (secno > endsecno || (bp->b_flags & B_BNOCLIP))
 253                         goto bad;
 254
 255                 /*
 256                  * If exactly at end of disk, return an EOF.  Throw away
 257                  * the buffer contents, if any, by setting B_INVAL.
 258                  */
 259                 if (secno == endsecno) {
 260                         bp->b_resid = bp->b_bcount;
 261                         bp->b_flags |= B_INVAL;
 262                         goto done;
 263                 }
 264
 265                 /*
 266                  * Else truncate
 267                  */
 268                 nsec = endsecno - secno;
 269                 bp->b_bcount = nsec * ssp->dss_secsize;
 270         }
 271
 272         nbio = push_bio(bio);
 273         nbio->bio_offset = (off_t)(sp->ds_offset + slicerel_secno) *
 274                            ssp->dss_secsize;
 275         return (nbio);
 276
 277 bad_bcount:
 278         kprintf(
 279         "dscheck(%s): b_bcount %d is not on a sector boundary (ssize %d)\n",
 280             devtoname(dev), bp->b_bcount, ssp->dss_secsize);
 281         goto bad;
 282
 283 bad_blkno:
 284         kprintf(
 285         "dscheck(%s): bio_offset %lld is not on a sector boundary (ssize %d)\n",
 286             devtoname(dev), bio->bio_offset, ssp->dss_secsize);
 287 bad:
 288         bp->b_error = EINVAL;
 289         /* fall through */
 290 error:
 291         /*
 292          * Terminate the I/O with a ranging error.  Since the buffer is
 293          * either illegal or beyond the file EOF, mark it B_INVAL as well.
 294          */
 295         bp->b_resid = bp->b_bcount;
 296         bp->b_flags |= B_ERROR | B_INVAL;
 297 done:
 298         /*
 299          * Caller must biodone() the originally passed bio if NULL is
 300          * returned.
 301          */
 302         return (NULL);
 303 }
 304
 305 void
 306 dsclose(cdev_t dev, int mode, struct diskslices *ssp)
 307 {
 308         u_int32_t part;
 309         u_int32_t slice;
 310         struct diskslice *sp;
 311
 312         slice = dkslice(dev);
 313         part  = dkpart(dev);
 314         if (slice < ssp->dss_nslices) {
 315                 sp = &ssp->dss_slices[slice];
 316                 dsclrmask(sp, part);
 317         }
 318 }
 319
 320 void
 321 dsgone(struct diskslices **sspp)
 322 {
 323         int slice;
 324         struct diskslice *sp;
 325         struct diskslices *ssp;
 326
 327         for (slice = 0, ssp = *sspp; slice < ssp->dss_nslices; slice++) {
 328                 sp = &ssp->dss_slices[slice];
 329                 free_ds_label(ssp, slice);
 330         }
 331         kfree(ssp, M_DEVBUF);
 332         *sspp = NULL;
 333 }
 334
 335 /*
 336  * For the "write" commands (DIOCSDINFO and DIOCWDINFO), this
 337  * is subject to the same restriction as dsopen().
 338  */
 339 int
 340 dsioctl(cdev_t dev, u_long cmd, caddr_t data, int flags,
 341         struct diskslices **sspp, struct disk_info *info)
 342 {
 343         int error;
 344         disklabel_t lp;
 345         disklabel_t lptmp;
 346         disklabel_ops_t ops;
 347         int old_wlabel;
 348         u_int32_t openmask[DKMAXPARTITIONS/(sizeof(u_int32_t)*8)];
 349         int part;
 350         int slice;
 351         struct diskslice *sp;
 352         struct diskslices *ssp;
 353
 354         slice = dkslice(dev);
 355         part = dkpart(dev);
 356         ssp = *sspp;
 357         if (slice >= ssp->dss_nslices)
 358                 return (EINVAL);
 359         sp = &ssp->dss_slices[slice];
 360         lp = sp->ds_label;
 361         ops = sp->ds_ops;       /* may be NULL if no label */
 362
 363         switch (cmd) {
 364         case DIOCGDVIRGIN32:
 365                 ops = &disklabel32_ops;
 366                 /* fall through */
 367         case DIOCGDVIRGIN64:
 368                 if (cmd != DIOCGDVIRGIN32)
 369                         ops = &disklabel64_ops;
 370                 /*
 371                  * You can only retrieve a virgin disklabel on the whole
 372                  * disk slice or whole-slice partition.
 373                  */
 374                 if (slice != WHOLE_DISK_SLICE &&
 375                     part != WHOLE_SLICE_PART) {
 376                         return(EINVAL);
 377                 }
 378
 379                 lp.opaque = data;
 380                 ops->op_makevirginlabel(lp, ssp, sp, info);
 381                 return (0);
 382
 383         case DIOCGDINFO32:
 384         case DIOCGDINFO64:
 385                 /*
 386                  * You can only retrieve a disklabel on the whole
 387                  * slice partition.
 388                  *
 389                  * We do not support labels directly on whole-disks
 390                  * any more (that is, disks without slices), unless the
 391                  * device driver has asked for a compatible label (e.g.
 392                  * for a CD) to allow booting off of storage that is
 393                  * otherwise unlabeled.
 394                  */
 395                 error = 0;
 396                 if (part != WHOLE_SLICE_PART)
 397                         return(EINVAL);
 398                 if (slice == WHOLE_DISK_SLICE &&
 399                     (info->d_dsflags & DSO_COMPATLABEL) == 0) {
 400                         return (ENODEV);
 401                 }
 402                 if (sp->ds_label.opaque == NULL) {
 403                         error = dsreadandsetlabel(dev, info->d_dsflags,
 404                                                   ssp, sp, info);
 405                         ops = sp->ds_ops;       /* may be NULL */
 406                 }
 407
 408                 /*
 409                  * The type of label we found must match the type of
 410                  * label requested.
 411                  */
 412                 if (error == 0 && IOCPARM_LEN(cmd) != ops->labelsize)
 413                         error = ENOATTR;
 414                 if (error == 0)
 415                         bcopy(sp->ds_label.opaque, data, ops->labelsize);
 416                 return (error);
 417
 418         case DIOCGPART:
 419                 {
 420                         struct partinfo *dpart = (void *)data;
 421
 422                         /*
 423                          * The disk management layer may not have read the
 424                          * disklabel yet because simply opening a slice no
 425                          * longer 'probes' the disk that way.  Be sure we
 426                          * have tried.
 427                          *
 428                          * We ignore any error.
 429                          */
 430                         if (sp->ds_label.opaque == NULL &&
 431                             part == WHOLE_SLICE_PART &&
 432                             slice != WHOLE_DISK_SLICE) {
 433                                 dsreadandsetlabel(dev, info->d_dsflags,
 434                                                   ssp, sp, info);
 435                                 ops = sp->ds_ops;       /* may be NULL */
 436                         }
 437
 438                         bzero(dpart, sizeof(*dpart));
 439                         dpart->media_offset   = (u_int64_t)sp->ds_offset *
 440                                                 info->d_media_blksize;
 441                         dpart->media_size     = (u_int64_t)sp->ds_size *
 442                                                 info->d_media_blksize;
 443                         dpart->media_blocks   = sp->ds_size;
 444                         dpart->media_blksize  = info->d_media_blksize;
 445                         dpart->reserved_blocks= sp->ds_reserved;
 446                         dpart->fstype_uuid = sp->ds_type_uuid;
 447                         dpart->storage_uuid = sp->ds_stor_uuid;
 448
 449                         if (slice != WHOLE_DISK_SLICE &&
 450                             part != WHOLE_SLICE_PART) {
 451                                 u_int64_t start;
 452                                 u_int64_t blocks;
 453                                 if (lp.opaque == NULL)
 454                                         return(EINVAL);
 455                                 if (ops->op_getpartbounds(ssp, lp, part,
 456                                                           &start, &blocks)) {
 457                                         return(EINVAL);
 458                                 }
 459                                 ops->op_loadpartinfo(lp, part, dpart);
 460                                 dpart->media_offset += start *
 461                                                        info->d_media_blksize;
 462                                 dpart->media_size = blocks *
 463                                                     info->d_media_blksize;
 464                                 dpart->media_blocks = blocks;
 465
 466                                 /*
 467                                  * partition starting sector (p_offset)
 468                                  * requires slice's reserved areas to be
 469                                  * adjusted.
 470                                  */
 471                                 if (dpart->reserved_blocks > start)
 472                                         dpart->reserved_blocks -= start;
 473                                 else
 474                                         dpart->reserved_blocks = 0;
 475                         }
 476
 477                         /*
 478                          * Load remaining fields from the info structure
 479                          */
 480                         dpart->d_nheads =       info->d_nheads;
 481                         dpart->d_ncylinders =   info->d_ncylinders;
 482                         dpart->d_secpertrack =  info->d_secpertrack;
 483                         dpart->d_secpercyl =    info->d_secpercyl;
 484                 }
 485                 return (0);
 486
 487         case DIOCGSLICEINFO:
 488                 bcopy(ssp, data, (char *)&ssp->dss_slices[ssp->dss_nslices] -
 489                                  (char *)ssp);
 490                 return (0);
 491
 492         case DIOCSDINFO32:
 493                 ops = &disklabel32_ops;
 494                 /* fall through */
 495         case DIOCSDINFO64:
 496                 if (cmd != DIOCSDINFO32)
 497                         ops = &disklabel64_ops;
 498                 /*
 499                  * You can write a disklabel on the whole disk slice or
 500                  * whole-slice partition.
 501                  */
 502                 if (slice != WHOLE_DISK_SLICE &&
 503                     part != WHOLE_SLICE_PART) {
 504                         return(EINVAL);
 505                 }
 506
 507                 /*
 508                  * We no longer support writing disklabels directly to media
 509                  * without there being a slice.  Keep this as a separate
 510                  * conditional.
 511                  */
 512                 if (slice == WHOLE_DISK_SLICE)
 513                         return (ENODEV);
 514                 if (!(flags & FWRITE))
 515                         return (EBADF);
 516
 517                 /*
 518                  * If an existing label is present it must be the same
 519                  * type as the label being passed by the ioctl.
 520                  */
 521                 if (sp->ds_label.opaque && sp->ds_ops != ops)
 522                         return (ENOATTR);
 523
 524                 /*
 525                  * Create a temporary copy of the existing label
 526                  * (if present) so setdisklabel can compare it against
 527                  * the new label.
 528                  */
 529                 lp.opaque = kmalloc(ops->labelsize, M_DEVBUF, M_WAITOK);
 530                 if (sp->ds_label.opaque == NULL)
 531                         bzero(lp.opaque, ops->labelsize);
 532                 else
 533                         bcopy(sp->ds_label.opaque, lp.opaque, ops->labelsize);
 534                 if (sp->ds_label.opaque == NULL) {
 535                         bzero(openmask, sizeof(openmask));
 536                 } else {
 537                         bcopy(sp->ds_openmask, openmask, sizeof(openmask));
 538                 }
 539                 lptmp.opaque = data;
 540                 error = ops->op_setdisklabel(lp, lptmp, ssp, sp, openmask);
 541                 if (error != 0) {
 542                         kfree(lp.opaque, M_DEVBUF);
 543                         return (error);
 544                 }
 545                 free_ds_label(ssp, slice);
 546                 set_ds_label(ssp, slice, lp, ops);
 547                 return (0);
 548
 549         case DIOCSYNCSLICEINFO:
 550                 /*
 551                  * This ioctl can only be done on the whole disk
 552                  */
 553                 if (slice != WHOLE_DISK_SLICE || part != WHOLE_SLICE_PART)
 554                         return (EINVAL);
 555
 556                 if (*(int *)data == 0) {
 557                         for (slice = 0; slice < ssp->dss_nslices; slice++) {
 558                                 struct diskslice *ds = &ssp->dss_slices[slice];
 559
 560                                 switch(dscountmask(ds)) {
 561                                 case 0:
 562                                         break;
 563                                 case 1:
 564                                         if (slice != WHOLE_DISK_SLICE)
 565                                                 return (EBUSY);
 566                                         if (!dschkmask(ds, RAW_PART))
 567                                                 return (EBUSY);
 568                                         break;
 569                                 default:
 570                                         return (EBUSY);
 571                                 }
 572                         }
 573                 }
 574
 575                 /*
 576                  * Temporarily forget the current slices struct and read
 577                  * the current one.
 578                  *
 579                  * NOTE:
 580                  *
 581                  * XXX should wait for current accesses on this disk to
 582                  * complete, then lock out future accesses and opens.
 583                  */
 584                 *sspp = NULL;
 585                 error = dsopen(dev, S_IFCHR, ssp->dss_oflags, sspp, info);
 586                 if (error != 0) {
 587                         *sspp = ssp;
 588                         return (error);
 589                 }
 590
 591                 /*
 592                  * Reopen everything.  This is a no-op except in the "force"
 593                  * case and when the raw bdev and cdev are both open.  Abort
 594                  * if anything fails.
 595                  */
 596                 for (slice = 0; slice < ssp->dss_nslices; slice++) {
 597                         for (part = 0; part < DKMAXPARTITIONS; ++part) {
 598                                 if (!dschkmask(&ssp->dss_slices[slice], part))
 599                                         continue;
 600                                 error = dsopen(dkmodslice(dkmodpart(dev, part),
 601                                                           slice),
 602                                                S_IFCHR, ssp->dss_oflags, sspp,
 603                                                info);
 604                                 if (error != 0) {
 605                                         *sspp = ssp;
 606                                         return (EBUSY);
 607                                 }
 608                         }
 609                 }
 610
 611                 dsgone(&ssp);
 612                 return (0);
 613
 614         case DIOCWDINFO32:
 615         case DIOCWDINFO64:
 616                 error = dsioctl(dev, ((cmd == DIOCWDINFO32) ?
 617                                         DIOCSDINFO32 : DIOCSDINFO64),
 618                                 data, flags, &ssp, info);
 619                 if (error == 0 && sp->ds_label.opaque == NULL)
 620                         error = EINVAL;
 621                 if (error != 0)
 622                         return (error);
 623
 624                 /*
 625                  * Allow the reserved area to be written, reload ops
 626                  * because the DIOCSDINFO op above may have installed
 627                  * a new label type.
 628                  */
 629                 ops = sp->ds_ops;
 630                 old_wlabel = sp->ds_wlabel;
 631                 set_ds_wlabel(ssp, slice, TRUE);
 632                 error = ops->op_writedisklabel(dev, ssp, sp, sp->ds_label);
 633                 set_ds_wlabel(ssp, slice, old_wlabel);
 634                 /* XXX should invalidate in-core label if write failed. */
 635                 return (error);
 636
 637         case DIOCWLABEL:
 638                 if (slice == WHOLE_DISK_SLICE)
 639                         return (ENODEV);
 640                 if (!(flags & FWRITE))
 641                         return (EBADF);
 642                 set_ds_wlabel(ssp, slice, *(int *)data != 0);
 643                 return (0);
 644
 645         default:
 646                 return (ENOIOCTL);
 647         }
 648 }
 649
 650 int
 651 dsisopen(struct diskslices *ssp)
 652 {
 653         int slice;
 654
 655         if (ssp == NULL)
 656                 return (0);
 657         for (slice = 0; slice < ssp->dss_nslices; slice++) {
 658                 if (dscountmask(&ssp->dss_slices[slice]))
 659                         return (1);
 660         }
 661         return (0);
 662 }
 663
 664 /*
 665  * Allocate a slices "struct" and initialize it to contain only an empty
 666  * compatibility slice (pointing to itself), a whole disk slice (covering
 667  * the disk as described by the label), and (nslices - BASE_SLICES) empty
 668  * slices beginning at BASE_SLICE.
 669  *
 670  * Note that the compatibility slice is no longer really a compatibility
 671  * slice.  It is slice 0 if a GPT label is present, and the dangerously
 672  * dedicated slice if no slice table otherwise exists.  Else it is 0-sized.
 673  */
 674 struct diskslices *
 675 dsmakeslicestruct(int nslices, struct disk_info *info)
 676 {
 677         struct diskslice *sp;
 678         struct diskslices *ssp;
 679
 680         ssp = kmalloc(offsetof(struct diskslices, dss_slices) +
 681                      nslices * sizeof *sp, M_DEVBUF, M_WAITOK);
 682         ssp->dss_first_bsd_slice = COMPATIBILITY_SLICE;
 683         ssp->dss_nslices = nslices;
 684         ssp->dss_oflags = 0;
 685
 686         /*
 687          * Figure out if we can use shifts or whether we have to
 688          * use mod/multply to translate byte offsets into sector numbers.
 689          */
 690         if ((info->d_media_blksize ^ (info->d_media_blksize - 1)) ==
 691              (info->d_media_blksize << 1) - 1) {
 692                 ssp->dss_secmult = info->d_media_blksize / DEV_BSIZE;
 693                 if (ssp->dss_secmult & (ssp->dss_secmult - 1))
 694                         ssp->dss_secshift = -1;
 695                 else
 696                         ssp->dss_secshift = ffs(ssp->dss_secmult) - 1;
 697         } else {
 698                 ssp->dss_secmult = 0;
 699                 ssp->dss_secshift = -1;
 700         }
 701         ssp->dss_secsize = info->d_media_blksize;
 702         sp = &ssp->dss_slices[0];
 703         bzero(sp, nslices * sizeof *sp);
 704         sp[WHOLE_DISK_SLICE].ds_size = info->d_media_blocks;
 705         return (ssp);
 706 }
 707
 708 char *
 709 dsname(cdev_t dev, int unit, int slice, int part, char *partname)
 710 {
 711         static char name[32];
 712         const char *dname;
 713         int used;
 714
 715         dname = dev_dname(dev);
 716         if (strlen(dname) > 16)
 717                 dname = "nametoolong";
 718         ksnprintf(name, sizeof(name), "%s%d", dname, unit);
 719         partname[0] = '\0';
 720         used = strlen(name);
 721
 722         if (slice != WHOLE_DISK_SLICE) {
 723                 /*
 724                  * slice or slice + partition.  BASE_SLICE is s1, but
 725                  * the compatibility slice (0) needs to be s0.
 726                  */
 727                 used += ksnprintf(name + used, sizeof(name) - used,
 728                                   "s%d", (slice ? slice - BASE_SLICE + 1 : 0));
 729                 if (part != WHOLE_SLICE_PART) {
 730                         used += ksnprintf(name + used, sizeof(name) - used,
 731                                           "%c", 'a' + part);
 732                         partname[0] = 'a' + part;
 733                         partname[1] = 0;
 734                 }
 735         } else if (part == WHOLE_SLICE_PART) {
 736                 /*
 737                  * whole-disk-device, raw access to disk
 738                  */
 739                 /* no string extension */
 740         } else if (part > 128) {
 741                 /*
 742                  * whole-disk-device, extended raw access partitions.
 743                  * (typically used to access CD audio tracks)
 744                  */
 745                 used += ksnprintf(name + used, sizeof(name) - used,
 746                                           "t%d", part - 128);
 747         } else {
 748                 /*
 749                  * whole-disk-device, illegal partition number
 750                  */
 751                 used += ksnprintf(name + used, sizeof(name) - used,
 752                                           "?%d", part);
 753         }
 754         return (name);
 755 }
 756
 757 /*
 758  * This should only be called when the unit is inactive and the strategy
 759  * routine should not allow it to become active unless we call it.  Our
 760  * strategy routine must be special to allow activity.
 761  */
 762 int
 763 dsopen(cdev_t dev, int mode, u_int flags,
 764         struct diskslices **sspp, struct disk_info *info)
 765 {
 766         cdev_t dev1;
 767         int error;
 768         int need_init;
 769         struct diskslice *sp;
 770         struct diskslices *ssp;
 771         int slice;
 772         int part;
 773
 774         dev->si_bsize_phys = info->d_media_blksize;
 775
 776         /*
 777          * Do not attempt to read the slice table or disk label when
 778          * accessing the whole-disk slice or a while-slice partition.
 779          */
 780         if (dkslice(dev) == WHOLE_DISK_SLICE)
 781                 flags |= DSO_ONESLICE | DSO_NOLABELS;
 782         if (dkpart(dev) == WHOLE_SLICE_PART)
 783                 flags |= DSO_NOLABELS;
 784
 785         /*
 786          * Reinitialize the slice table unless there is an open device
 787          * on the unit.
 788          *
 789          * It would be nice if we didn't have to do this but when a
 790          * user is slicing and partitioning up a disk it is a lot safer
 791          * to not take any chances.
 792          */
 793         ssp = *sspp;
 794         need_init = !dsisopen(ssp);
 795         if (ssp != NULL && need_init)
 796                 dsgone(sspp);
 797         if (need_init) {
 798                 /*
 799                  * Allocate a minimal slices "struct".  This will become
 800                  * the final slices "struct" if we don't want real slices
 801                  * or if we can't find any real slices.
 802                  *
 803                  * Then scan the disk
 804                  */
 805                 *sspp = dsmakeslicestruct(BASE_SLICE, info);
 806
 807                 if ((flags & DSO_ONESLICE) == 0) {
 808                         error = mbrinit(dev, info, sspp);
 809                         if (error != 0) {
 810                                 dsgone(sspp);
 811                                 return (error);
 812                         }
 813                 }
 814                 ssp = *sspp;
 815                 ssp->dss_oflags = flags;
 816
 817                 /*
 818                  * If there are no real slices, then make the compatiblity
 819                  * slice cover the whole disk.
 820                  */
 821                 if (ssp->dss_nslices == BASE_SLICE) {
 822                         sp = &ssp->dss_slices[COMPATIBILITY_SLICE];
 823
 824                         sp->ds_size = info->d_media_blocks;
 825                         sp->ds_reserved = 0;
 826                 }
 827
 828                 /*
 829                  * Set dss_first_bsd_slice to point at the first BSD
 830                  * slice, if any.
 831                  */
 832                 for (slice = BASE_SLICE; slice < ssp->dss_nslices; slice++) {
 833                         sp = &ssp->dss_slices[slice];
 834                         if (sp->ds_type == DOSPTYP_386BSD /* XXX */) {
 835 #if 0
 836                                 struct diskslice *csp;
 837 #endif
 838
 839                                 ssp->dss_first_bsd_slice = slice;
 840 #if 0
 841                                 /*
 842                                  * no longer supported, s0 is a real slice
 843                                  * for GPT
 844                                  */
 845                                 csp = &ssp->dss_slices[COMPATIBILITY_SLICE];
 846                                 csp->ds_offset = sp->ds_offset;
 847                                 csp->ds_size = sp->ds_size;
 848                                 csp->ds_type = sp->ds_type;
 849                                 csp->ds_reserved = sp->ds_reserved;
 850 #endif
 851                                 break;
 852                         }
 853                 }
 854
 855                 /*
 856                  * By definition accesses via the whole-disk device do not
 857                  * specify any reserved areas.  The whole disk may be read
 858                  * or written by the whole-disk device.
 859                  *
 860                  * The whole-disk slice does not ever have a label.
 861                  */
 862                 sp = &ssp->dss_slices[WHOLE_DISK_SLICE];
 863                 sp->ds_wlabel = TRUE;
 864                 sp->ds_reserved = 0;
 865         }
 866
 867         /*
 868          * Load the disklabel for the slice being accessed unless it is
 869          * a whole-disk-slice or a whole-slice-partition (as determined
 870          * by DSO_NOLABELS).
 871          *
 872          * We could scan all slices here and try to load up their
 873          * disklabels, but that would cause us to access slices that
 874          * the user may otherwise not intend us to access, or corrupted
 875          * slices, etc.
 876          *
 877          * XXX if there are no opens on the slice we may want to re-read
 878          * the disklabel anyway, even if we have one cached.
 879          */
 880         slice = dkslice(dev);
 881         if (slice >= ssp->dss_nslices)
 882                 return (ENXIO);
 883         sp = &ssp->dss_slices[slice];
 884         part = dkpart(dev);
 885
 886         if ((flags & DSO_NOLABELS) == 0 && sp->ds_label.opaque == NULL) {
 887                 dev1 = dkmodslice(dkmodpart(dev, WHOLE_SLICE_PART), slice);
 888
 889                 /*
 890                  * If opening a raw disk we do not try to
 891                  * read the disklabel now.  No interpretation of raw disks
 892                  * (e.g. like 'da0') ever occurs.  We will try to read the
 893                  * disklabel for a raw slice if asked to via DIOC* ioctls.
 894                  *
 895                  * Access to the label area is disallowed by default.  Note
 896                  * however that accesses via WHOLE_DISK_SLICE, and accesses
 897                  * via WHOLE_SLICE_PART for slices without valid disklabels,
 898                  * will allow writes and ignore the flag.
 899                  */
 900                 set_ds_wlabel(ssp, slice, FALSE);
 901                 dsreadandsetlabel(dev1, flags, ssp, sp, info);
 902         }
 903
 904         /*
 905          * If opening a particular partition the disklabel must exist and
 906          * the partition must be present in the label.
 907          *
 908          * If the partition is the special whole-disk-slice no partition
 909          * table need exist.
 910          */
 911         if (part != WHOLE_SLICE_PART && slice != WHOLE_DISK_SLICE) {
 912                 if (sp->ds_label.opaque == NULL ||
 913                     part >= sp->ds_ops->op_getnumparts(sp->ds_label)) {
 914                         return (EINVAL);
 915                 }
 916         }
 917
 918         /*
 919          * Do not allow special raw-extension partitions to be opened
 920          * if the device doesn't support them.  Raw-extension partitions
 921          * are typically used to handle CD tracks.
 922          */
 923         if (slice == WHOLE_DISK_SLICE && part >= 128 &&
 924             part != WHOLE_SLICE_PART) {
 925                 if ((info->d_dsflags & DSO_RAWEXTENSIONS) == 0)
 926                         return (EINVAL);
 927         }
 928
 929         /*
 930          * Ok, we are open
 931          */
 932         dssetmask(sp, part);
 933         return (0);
 934 }
 935
 936 /*
 937  * Attempt to read the disklabel.  If successful, store it in sp->ds_label.
 938  *
 939  * If we cannot read the disklabel and DSO_COMPATLABEL is set, we construct
 940  * a fake label covering the whole disk.
 941  */
 942 static
 943 int
 944 dsreadandsetlabel(cdev_t dev, u_int flags,
 945                   struct diskslices *ssp, struct diskslice *sp,
 946                   struct disk_info *info)
 947 {
 948         disklabel_t lp;
 949         disklabel_ops_t ops;
 950         const char *msg;
 951         const char *sname;
 952         char partname[2];
 953         int slice = dkslice(dev);
 954
 955         /*
 956          * Probe the disklabel
 957          */
 958         lp.opaque = NULL;
 959         sname = dsname(dev, dkunit(dev), slice, WHOLE_SLICE_PART, partname);
 960         ops = &disklabel32_ops;
 961         msg = ops->op_readdisklabel(dev, sp, &lp, info);
 962         if (msg && strcmp(msg, "no disk label") == 0) {
 963                 ops = &disklabel64_ops;
 964                 msg = disklabel64_ops.op_readdisklabel(dev, sp, &lp, info);
 965         }
 966
 967         /*
 968          * If we failed and COMPATLABEL is set, create a dummy disklabel.
 969          */
 970         if (msg != NULL && (flags & DSO_COMPATLABEL)) {
 971                 msg = NULL;
 972                 if (sp->ds_size >= 0x100000000ULL)
 973                         ops = &disklabel64_ops;
 974                 else
 975                         ops = &disklabel32_ops;
 976                 lp = ops->op_clone_label(info, sp);
 977         }
 978         if (msg != NULL) {
 979                 if (sp->ds_type == DOSPTYP_386BSD /* XXX */)
 980                         log(LOG_WARNING, "%s: cannot find label (%s)\n",
 981                             sname, msg);
 982                 if (lp.opaque)
 983                         kfree(lp.opaque, M_DEVBUF);
 984         } else {
 985                 set_ds_label(ssp, slice, lp, ops);
 986                 set_ds_wlabel(ssp, slice, FALSE);
 987         }
 988         return (msg ? EINVAL : 0);
 989 }
 990
 991 int64_t
 992 dssize(cdev_t dev, struct diskslices **sspp)
 993 {
 994         disklabel_t lp;
 995         disklabel_ops_t ops;
 996         int part;
 997         int slice;
 998         struct diskslices *ssp;
 999         u_int64_t start;
1000         u_int64_t blocks;
1001
1002         slice = dkslice(dev);
1003         part = dkpart(dev);
1004         ssp = *sspp;
1005         if (ssp == NULL || slice >= ssp->dss_nslices
1006             || !dschkmask(&ssp->dss_slices[slice], part)) {
1007                 if (dev_dopen(dev, FREAD, S_IFCHR, proc0.p_ucred) != 0)
1008                         return (-1);
1009                 dev_dclose(dev, FREAD, S_IFCHR);
1010                 ssp = *sspp;
1011         }
1012         lp = ssp->dss_slices[slice].ds_label;
1013         if (lp.opaque == NULL)
1014                 return (-1);
1015         ops = ssp->dss_slices[slice].ds_ops;
1016         if (ops->op_getpartbounds(ssp, lp, part, &start, &blocks))
1017                 return (-1);
1018         return ((int64_t)blocks);
1019 }
1020
1021 static void
1022 free_ds_label(struct diskslices *ssp, int slice)
1023 {
1024         struct diskslice *sp;
1025         disklabel_t lp;
1026
1027         sp = &ssp->dss_slices[slice];
1028         lp = sp->ds_label;
1029         if (lp.opaque != NULL) {
1030                 kfree(lp.opaque, M_DEVBUF);
1031                 lp.opaque = NULL;
1032                 set_ds_label(ssp, slice, lp, NULL);
1033         }
1034 }
1035
1036 static void
1037 set_ds_label(struct diskslices *ssp, int slice,
1038              disklabel_t lp, disklabel_ops_t ops)
1039 {
1040         struct diskslice *sp = &ssp->dss_slices[slice];
1041
1042         sp->ds_label = lp;
1043         sp->ds_ops = ops;
1044         if (lp.opaque && slice != WHOLE_DISK_SLICE)
1045                 ops->op_adjust_label_reserved(ssp, slice, sp);
1046         else
1047                 sp->ds_reserved = 0;
1048 }
1049
1050 static void
1051 set_ds_wlabel(struct diskslices *ssp, int slice, int wlabel)
1052 {
1053         ssp->dss_slices[slice].ds_wlabel = wlabel;
1054 }
1055