src/vol/vg_cache.c

   1 /*
   2  * Copyright 2009-2010, Sine Nomine Associates and others.
   3  * All Rights Reserved.
   4  *
   5  * This software has been released under the terms of the IBM Public
   6  * License.  For details, see the LICENSE file in the top-level source
   7  * directory or online at http://www.openafs.org/dl/license10.html
   8  */
   9
  10 /*
  11  * demand attach fs
  12  * volume group membership cache
  13  */
  14
  15 #include <afsconfig.h>
  16 #include <afs/param.h>
  17
  18 #include <roken.h>
  19
  20 #ifdef HAVE_SYS_FILE_H
  21 #include <sys/file.h>
  22 #endif
  23
  24 #ifdef AFS_DEMAND_ATTACH_FS
  25
  26 #include <afs/opr.h>
  27 #include <rx/rx_queue.h>
  28 #include <opr/lock.h>
  29 #include <lock.h>
  30 #include <afs/afsutil.h>
  31 #include "nfs.h"
  32 #include <afs/afsint.h>
  33 #include "ihandle.h"
  34 #include "vnode.h"
  35 #include "volume.h"
  36 #include "viceinode.h"
  37 #include "voldefs.h"
  38 #include "partition.h"
  39 #include <afs/errors.h>
  40
  41 #define __VOL_VG_CACHE_IMPL 1
  42
  43 #include "vg_cache.h"
  44 #include "vg_cache_impl.h"
  45
  46 static int _VVGC_lookup(struct DiskPartition64 *,
  47                         VolumeId volid,
  48                         VVGCache_entry_t ** entry,
  49                         VVGCache_hash_entry_t ** hentry);
  50 static int _VVGC_entry_alloc(VVGCache_entry_t ** entry);
  51 static int _VVGC_entry_free(VVGCache_entry_t * entry);
  52 static int _VVGC_entry_get(VVGCache_entry_t * entry);
  53 static int _VVGC_entry_put(struct DiskPartition64 *,
  54                            VVGCache_entry_t * entry);
  55 static int _VVGC_entry_add(struct DiskPartition64 *,
  56                            VolumeId volid,
  57                            VVGCache_entry_t **,
  58                            VVGCache_hash_entry_t **);
  59 static int _VVGC_entry_cl_add(VVGCache_entry_t *, VolumeId);
  60 static int _VVGC_entry_cl_del(struct DiskPartition64 *, VVGCache_entry_t *,
  61                               VolumeId);
  62 static int _VVGC_entry_export(VVGCache_entry_t *, VVGCache_query_t *);
  63 static int _VVGC_hash_entry_alloc(VVGCache_hash_entry_t ** entry);
  64 static int _VVGC_hash_entry_free(VVGCache_hash_entry_t * entry);
  65 static int _VVGC_hash_entry_add(struct DiskPartition64 *,
  66                                 VolumeId,
  67                                 VVGCache_entry_t *,
  68                                 VVGCache_hash_entry_t **);
  69 static int _VVGC_hash_entry_del(VVGCache_hash_entry_t * entry);
  70 static int _VVGC_hash_entry_unlink(VVGCache_hash_entry_t * entry);
  71
  72 VVGCache_hash_table_t VVGCache_hash_table;
  73 VVGCache_t VVGCache;
  74
  75 /**
  76  * initialize volume group cache subsystem.
  77  *
  78  * @return operation status
  79  *    @retval 0 success
  80  */
  81 int
  82 VVGCache_PkgInit(void)
  83 {
  84     int code = 0;
  85     int i;
  86
  87     /* allocate hash table */
  88     VVGCache_hash_table.hash_buckets =
  89         malloc(VolumeHashTable.Size * sizeof(struct rx_queue));
  90     if (VVGCache_hash_table.hash_buckets == NULL) {
  91         code = ENOMEM;
  92         goto error;
  93     }
  94
  95     /* setup hash chain heads */
  96     for (i = 0; i < VolumeHashTable.Size; i++) {
  97         queue_Init(&VVGCache_hash_table.hash_buckets[i]);
  98     }
  99
 100     /* initialize per-partition VVGC state */
 101     for (i = 0; i <= VOLMAXPARTS; i++) {
 102         VVGCache.part[i].state = VVGC_PART_STATE_INVALID;
 103         VVGCache.part[i].dlist_hash_buckets = NULL;
 104         CV_INIT(&VVGCache.part[i].cv, "cache part", CV_DEFAULT, 0);
 105         if (code) {
 106             goto error;
 107         }
 108     }
 109
 110  error:
 111     return code;
 112 }
 113
 114 /**
 115  * shut down volume group cache subsystem.
 116  *
 117  * @return operation status
 118  *    @retval 0 success
 119  *
 120  * @todo implement
 121  */
 122 int
 123 VVGCache_PkgShutdown(void)
 124 {
 125     int i;
 126
 127     /* fix it later */
 128
 129     /* free hash table */
 130     free(VVGCache_hash_table.hash_buckets);
 131     VVGCache_hash_table.hash_buckets = NULL;
 132
 133     /* destroy per-partition VVGC state */
 134     for (i = 0; i <= VOLMAXPARTS; i++) {
 135         VVGCache.part[i].state = VVGC_PART_STATE_INVALID;
 136         CV_DESTROY(&VVGCache.part[i].cv);
 137     }
 138
 139     return EOPNOTSUPP;
 140 }
 141
 142 /**
 143  * allocate a cache entry.
 144  *
 145  * @param[out] entry_out  pointer to newly allocated entry
 146  *
 147  * @return operation status
 148  *    @retval 0 success
 149  *
 150  * @internal
 151  */
 152 static int
 153 _VVGC_entry_alloc(VVGCache_entry_t ** entry_out)
 154 {
 155     *entry_out = calloc(1, sizeof(VVGCache_entry_t));
 156
 157     if (*entry_out == NULL)
 158         return ENOMEM;
 159
 160     return 0;
 161 }
 162
 163 /**
 164  * free a cache entry.
 165  *
 166  * @param[in] entry   cache entry
 167  *
 168  * @return operation status
 169  *    @retval 0 success
 170  *
 171  * @internal
 172  */
 173 static int
 174 _VVGC_entry_free(VVGCache_entry_t * entry)
 175 {
 176     int code = 0;
 177
 178     opr_Assert(entry->refcnt == 0);
 179     free(entry);
 180
 181     return code;
 182 }
 183
 184 /**
 185  * allocate and register an entry for a volume group.
 186  *
 187  * @param[in]  dp         disk partition object
 188  * @param[in]  volid      volume id
 189  * @param[out] entry_out  vg cache object pointer
 190  * @param[out] hash_out   vg cache hash entry object pointer
 191  *
 192  * @pre - VOL_LOCK held
 193  *      - no such entry exists in hash table
 194  *
 195  * @return operation status
 196  *    @retval 0 success
 197  *
 198  * @internal
 199  */
 200 static int
 201 _VVGC_entry_add(struct DiskPartition64 * dp,
 202                 VolumeId volid,
 203                 VVGCache_entry_t ** entry_out,
 204                 VVGCache_hash_entry_t ** hash_out)
 205 {
 206     int code = 0;
 207     VVGCache_entry_t * ent;
 208
 209     code = _VVGC_entry_alloc(&ent);
 210     if (code) {
 211         goto error;
 212     }
 213
 214     ent->rw = volid;
 215     /* refcnt will be inc'd when a child is added */
 216     ent->refcnt = 0;
 217
 218     code = _VVGC_hash_entry_add(dp, volid, ent, hash_out);
 219     if (code) {
 220         goto error;
 221     }
 222
 223     if (entry_out) {
 224         *entry_out = ent;
 225     }
 226     return code;
 227
 228  error:
 229     if (ent) {
 230         _VVGC_entry_free(ent);
 231         ent = NULL;
 232     }
 233     return code;
 234 }
 235
 236 /**
 237  * add a volid to the entry's child list.
 238  *
 239  * @param[in]   ent     volume group object
 240  * @param[in]   volid   volume id
 241  *
 242  * @return operation status
 243  *    @retval 0 success
 244  *    @retval -1 child table is full
 245  *
 246  * @internal
 247  */
 248 static int
 249 _VVGC_entry_cl_add(VVGCache_entry_t * ent,
 250                    VolumeId volid)
 251 {
 252     int code = 0, i;
 253     int empty_idx = -1;
 254
 255     /* search table to avoid duplicates */
 256     for (i = 0; i < VOL_VG_MAX_VOLS; i++) {
 257         if (ent->children[i] == volid) {
 258             ViceLog(1, ("VVGC_entry_cl_add: tried to add duplicate vol "
 259                        "%lu to VG %lu\n",
 260                        afs_printable_uint32_lu(volid),
 261                        afs_printable_uint32_lu(ent->rw)));
 262             goto done;
 263         }
 264         if (empty_idx == -1 && !ent->children[i]) {
 265             empty_idx = i;
 266             /* don't break; make sure we go through all children so we don't
 267              * add a duplicate entry */
 268         }
 269     }
 270
 271     /* verify table isn't full */
 272     if (empty_idx == -1) {
 273         code = -1;
 274         ViceLog(0, ("VVGC_entry_cl_add: tried to add vol %lu to VG %lu, but VG "
 275             "is full\n", afs_printable_uint32_lu(volid),
 276             afs_printable_uint32_lu(ent->rw)));
 277         goto done;
 278     }
 279
 280     /* add entry */
 281     ent->children[empty_idx] = volid;
 282
 283     /* inc refcount */
 284     code = _VVGC_entry_get(ent);
 285
 286  done:
 287     return code;
 288 }
 289
 290 /**
 291  * delete a volid from the entry's child list.
 292  *
 293  * @param[in]   dp      disk partition object
 294  * @param[in]   ent     volume group object
 295  * @param[in]   volid   volume id
 296  *
 297  * @return operation status
 298  *    @retval 0 success
 299  *    @retval -1 no such entry found
 300  *
 301  * @internal
 302  */
 303 static int
 304 _VVGC_entry_cl_del(struct DiskPartition64 *dp,
 305                    VVGCache_entry_t * ent,
 306                    VolumeId volid)
 307 {
 308     int code = -1, i;
 309
 310     for (i = 0; i < VOL_VG_MAX_VOLS; i++) {
 311         if (ent->children[i] == volid) {
 312             ent->children[i] = 0;
 313             code = 0;
 314             goto done;
 315         }
 316     }
 317
 318  done:
 319     if (!code) {
 320         code = _VVGC_entry_put(dp, ent);
 321     }
 322
 323     return code;
 324 }
 325
 326 /**
 327  * add a refcount to an entry.
 328  *
 329  * @param[in] entry   cache entry
 330  *
 331  * @pre VOL_LOCK held
 332  *
 333  * @return operation status
 334  *    @retval 0 success
 335  *
 336  * @internal
 337  */
 338 static int _VVGC_entry_get(VVGCache_entry_t * entry)
 339 {
 340     entry->refcnt++;
 341     return 0;
 342 }
 343
 344 /**
 345  * put back a reference to an entry.
 346  *
 347  * @param[in] dp      disk partition object
 348  * @param[in] entry   cache entry
 349  *
 350  * @pre VOL_LOCK held
 351  *
 352  * @warning do not attempt to deref pointer after calling this interface
 353  *
 354  * @return operation status
 355  *    @retval 0 success
 356  *
 357  * @note dp is needed to lookup the RW hash entry to unlink, if we are
 358  *       putting back the final reference and freeing
 359  *
 360  * @internal
 361  */
 362 static int
 363 _VVGC_entry_put(struct DiskPartition64 * dp, VVGCache_entry_t * entry)
 364 {
 365     int code = 0;
 366
 367     opr_Assert(entry->refcnt > 0);
 368
 369     if (--entry->refcnt == 0) {
 370         VVGCache_entry_t *nentry;
 371         VVGCache_hash_entry_t *hentry;
 372
 373         /* first, try to delete the RW id hash entry pointing to this
 374          * entry */
 375         code = _VVGC_lookup(dp, entry->rw, &nentry, &hentry);
 376         if (!code) {
 377             if (nentry != entry) {
 378                 /* looking up the rw of this entry points to a different
 379                  * entry; should not happen */
 380                 ViceLog(0, ("VVGC_entry_put: error: entry lookup for entry %lu "
 381                     "found different entry than was passed",
 382                     afs_printable_uint32_lu(entry->rw)));
 383                 code = -1;
 384             } else {
 385                 code = _VVGC_hash_entry_unlink(hentry);
 386                 hentry = NULL;
 387             }
 388         } else if (code == ENOENT) {
 389             /* ignore ENOENT; this shouldn't happen, since the RW hash
 390              * entry should always exist if the entry does... but we
 391              * were going to delete it anyway, so try to continue */
 392             ViceLog(0, ("VVGC_entry_put: warning: tried to unlink entry for "
 393                 "vol %lu, but RW hash entry doesn't exist; continuing "
 394                 "anyway...\n", afs_printable_uint32_lu(entry->rw)));
 395
 396             code = 0;
 397         }
 398
 399         /* now, just free the entry itself */
 400         if (!code) {
 401             code = _VVGC_entry_free(entry);
 402         }
 403     }
 404
 405     return code;
 406 }
 407
 408 /**
 409  * export a volume group entry in the external object format.
 410  *
 411  * @param[in]  ent   internal-format volume group object
 412  * @param[out] qry   external-format volume group object
 413  *
 414  * @pre VOL_LOCK held
 415  *
 416  * @return operation status
 417  *    @retval 0 success
 418  *
 419  * @internal
 420  */
 421 static int
 422 _VVGC_entry_export(VVGCache_entry_t * ent, VVGCache_query_t * qry)
 423 {
 424     int i;
 425
 426     qry->rw = ent->rw;
 427     for (i = 0; i < VOL_VG_MAX_VOLS; i++) {
 428         qry->children[i] = ent->children[i];
 429     }
 430
 431     return 0;
 432 }
 433
 434 /**
 435  * allocate a hash table entry structure.
 436  *
 437  * @param[out] entry_out  address in which to store newly allocated hash entry struct
 438  *
 439  * @return operation status
 440  *    @retval 0 success
 441  *
 442  * @internal
 443  */
 444 static int
 445 _VVGC_hash_entry_alloc(VVGCache_hash_entry_t ** entry_out)
 446 {
 447     int code = 0;
 448     VVGCache_hash_entry_t * ent;
 449
 450     *entry_out = ent = malloc(sizeof(VVGCache_hash_entry_t));
 451     if (ent == NULL) {
 452         code = ENOMEM;
 453     }
 454
 455     return code;
 456 }
 457
 458 /**
 459  * free a hash table entry structure.
 460  *
 461  * @param[in] entry   hash table entry structure to be freed
 462  *
 463  * @return operation status
 464  *    @retval 0 success
 465  *
 466  * @internal
 467  */
 468 static int
 469 _VVGC_hash_entry_free(VVGCache_hash_entry_t * entry)
 470 {
 471     int code = 0;
 472
 473     free(entry);
 474
 475     return code;
 476 }
 477
 478 /**
 479  * add an entry to the hash table.
 480  *
 481  * @param[in]  dp        disk partition object
 482  * @param[in]  volid     volume id
 483  * @param[in]  ent       volume group object
 484  * @param[out] hash_out  address in which to store pointer to hash entry
 485  *
 486  * @pre VOL_LOCK held
 487  *
 488  * @return operation status
 489  *    @retval 0 success
 490  *    @retval EEXIST hash entry for volid already exists, and it points to
 491  *                   a different VG entry
 492  *
 493  * @internal
 494  */
 495 static int
 496 _VVGC_hash_entry_add(struct DiskPartition64 * dp,
 497                      VolumeId volid,
 498                      VVGCache_entry_t * ent,
 499                      VVGCache_hash_entry_t ** hash_out)
 500 {
 501     int code = 0;
 502     VVGCache_hash_entry_t * hent;
 503     int hash = VVGC_HASH(volid);
 504     VVGCache_entry_t *nent;
 505
 506     code = _VVGC_lookup(dp, volid, &nent, hash_out);
 507     if (!code) {
 508         if (ent != nent) {
 509             ViceLog(0, ("_VVGC_hash_entry_add: tried to add a duplicate "
 510                         " nonmatching entry for vol %lu: original "
 511                         "(%"AFS_PTR_FMT",%lu) new (%"AFS_PTR_FMT",%lu)\n",
 512                         afs_printable_uint32_lu(volid),
 513                         nent, afs_printable_uint32_lu(nent->rw),
 514                         ent, afs_printable_uint32_lu(ent->rw)));
 515             return EEXIST;
 516         }
 517         ViceLog(1, ("_VVGC_hash_entry_add: tried to add duplicate "
 518                       "hash entry for vol %lu, VG %lu",
 519                       afs_printable_uint32_lu(volid),
 520                       afs_printable_uint32_lu(ent->rw)));
 521         /* accept attempts to add matching duplicate entries; just
 522          * pretend we added it */
 523         return 0;
 524     }
 525
 526     code = _VVGC_hash_entry_alloc(&hent);
 527     if (code) {
 528         goto done;
 529     }
 530
 531     hent->entry = ent;
 532     hent->dp    = dp;
 533     hent->volid = volid;
 534     queue_Append(&VVGCache_hash_table.hash_buckets[hash],
 535                  hent);
 536
 537  done:
 538     if (hash_out) {
 539         *hash_out = hent;
 540     }
 541     return code;
 542 }
 543
 544 /**
 545  * remove an entry from the hash table.
 546  *
 547  * @param[in] hent   hash table entry
 548  *
 549  * @pre VOL_LOCK held
 550  *
 551  * @return operation status
 552  *    @retval 0 success
 553  *
 554  * @internal
 555  */
 556 static int
 557 _VVGC_hash_entry_del(VVGCache_hash_entry_t * hent)
 558 {
 559     int code = 0, res;
 560     int rw = 0;
 561
 562     if (hent->entry->rw == hent->volid) {
 563         rw = 1;
 564     }
 565
 566     code = _VVGC_entry_cl_del(hent->dp, hent->entry, hent->volid);
 567     /* note: hent->entry is possibly NULL after _VVGC_entry_cl_del, and
 568      * if hent->entry->rw == hent->volid, it is possible for hent to
 569      * have been freed */
 570
 571     if (!rw) {
 572         /* If we are the RW id, don't unlink, since we still need the
 573          * hash entry to exist, so when we lookup children, they can
 574          * look up the RW id hash chain, and they will all go to the
 575          * same object.
 576          *
 577          * If we are the last entry and the entry should be deleted,
 578          * _VVGC_entry_cl_del will take care of unlinking the RW hash entry.
 579          */
 580         res = _VVGC_hash_entry_unlink(hent);
 581         if (res) {
 582             code = res;
 583         }
 584     }
 585
 586     return code;
 587 }
 588
 589 /**
 590  * low-level interface to remove an entry from the hash table.
 591  *
 592  * Does not alter the refcount or worry about the children lists or
 593  * anything like that; just removes the hash table entry, frees it, and
 594  * that's all. You probably want @see _VVGC_hash_entry_del instead.
 595  *
 596  * @param[in] hent   hash table entry
 597  *
 598  * @pre VOL_LOCK held
 599  *
 600  * @return operation status
 601  *    @retval 0 success
 602  *
 603  * @internal
 604  */
 605 static int
 606 _VVGC_hash_entry_unlink(VVGCache_hash_entry_t * hent)
 607 {
 608     int code;
 609
 610     queue_Remove(hent);
 611     hent->entry = NULL;
 612     hent->volid = 0;
 613     code = _VVGC_hash_entry_free(hent);
 614
 615     return code;
 616 }
 617
 618 /**
 619  * lookup a vg cache entry given any member volume id.
 620  *
 621  * @param[in]  dp           disk partition object
 622  * @param[in]  volid        vg member volume id
 623  * @param[out] entry_out    address in which to store volume group entry structure pointer
 624  * @param[out] hash_out     address in which to store hash entry pointer
 625  *
 626  * @pre VOL_LOCK held
 627  *
 628  * @warning - it is up to the caller to get a ref to entry_out, if needed
 629  *          - hash_out must not be referenced after dropping VOL_LOCK
 630  *
 631  * @return operation status
 632  *    @retval 0 success
 633  *    @retval ENOENT volume id not found
 634  *    @retval EINVAL partition's VGC is invalid
 635  *
 636  * @internal
 637  */
 638 static int
 639 _VVGC_lookup(struct DiskPartition64 * dp,
 640                VolumeId volid,
 641                VVGCache_entry_t ** entry_out,
 642                VVGCache_hash_entry_t ** hash_out)
 643 {
 644     int code = ENOENT;
 645     int bucket = VVGC_HASH(volid);
 646     struct VVGCache_hash_entry * ent, * nent;
 647
 648     if (VVGCache.part[dp->index].state == VVGC_PART_STATE_INVALID) {
 649         return EINVAL;
 650     }
 651
 652     *entry_out = NULL;
 653
 654     for (queue_Scan(&VVGCache_hash_table.hash_buckets[bucket],
 655                     ent,
 656                     nent,
 657                     VVGCache_hash_entry)) {
 658         if (ent->volid == volid && ent->dp == dp) {
 659             code = 0;
 660             *entry_out = ent->entry;
 661             if (hash_out) {
 662                 *hash_out = ent;
 663             }
 664             break;
 665         }
 666     }
 667
 668     return code;
 669 }
 670
 671 /**
 672  * add an entry to the volume group cache.
 673  *
 674  * @param[in] dp       disk partition object
 675  * @param[in] parent   parent volume id
 676  * @param[in] child    child volume id
 677  * @param[out] newvg   if non-NULL, *newvg is 1 if adding this added a
 678  *                     new VG, 0 if we added to an existing VG
 679  *
 680  * @pre VOL_LOCK held
 681  *
 682  * @return operation status
 683  *    @retval 0 success
 684  *    @retval -1 parent and child are already registered in
 685  *               different VGs
 686  */
 687 int
 688 VVGCache_entry_add_r(struct DiskPartition64 * dp,
 689                      VolumeId parent,
 690                      VolumeId child,
 691                      afs_int32 *newvg)
 692 {
 693     int code = 0, res;
 694     VVGCache_entry_t * child_ent, * parent_ent;
 695
 696     if (newvg) {
 697         *newvg = 0;
 698     }
 699
 700     /* check for existing entries */
 701     res = _VVGC_lookup(dp, child, &child_ent, NULL);
 702     if (res && res != ENOENT) {
 703         code = res;
 704         goto done;
 705     }
 706
 707     res = _VVGC_lookup(dp, parent, &parent_ent, NULL);
 708     if (res && res != ENOENT) {
 709         code = res;
 710         goto done;
 711     }
 712
 713     /*
 714      * branch based upon existence of parent and child nodes
 715      */
 716     if (parent_ent && child_ent) {
 717         /* both exist.  we're done.
 718          * if they point different places, then report the error. */
 719         if (child_ent != parent_ent) {
 720             code = -1;
 721         }
 722         if (parent == child) {
 723             /* if we're adding the RW entry as a child, the RW id may
 724              * not be in the child array yet, so make sure not to skip
 725              * over that */
 726             goto cladd;
 727         }
 728         goto done;
 729     } else if (!parent_ent && child_ent) {
 730         /* child exists.
 731          * update vg root volid, and add hash entry. */
 732         parent_ent = child_ent;
 733         parent_ent->rw = parent;
 734
 735         code = _VVGC_hash_entry_add(dp,
 736                                     parent,
 737                                     parent_ent,
 738                                     NULL);
 739         goto done;
 740     } else if (!child_ent && !parent_ent) {
 741         code = _VVGC_entry_add(dp,
 742                                parent,
 743                                &parent_ent,
 744                                NULL);
 745         if (code) {
 746             goto done;
 747         }
 748         if (newvg) {
 749             *newvg = 1;
 750         }
 751         if (child == parent) {
 752             /* if we're the RW, skip over adding the child hash entry;
 753              * we already added the hash entry when creating the entry */
 754             child_ent = parent_ent;
 755             goto cladd;
 756         }
 757     }
 758
 759     opr_Assert(!child_ent);
 760     child_ent = parent_ent;
 761     code = _VVGC_hash_entry_add(dp,
 762                                 child,
 763                                 child_ent,
 764                                 NULL);
 765     if (code) {
 766         goto done;
 767     }
 768
 769  cladd:
 770     code = _VVGC_entry_cl_add(child_ent, child);
 771
 772  done:
 773     if (code && code != EINVAL) {
 774         ViceLog(0, ("VVGCache_entry_add: error %d trying to add vol %lu to VG"
 775             " %lu on partition %s", code, afs_printable_uint32_lu(child),
 776             afs_printable_uint32_lu(parent), VPartitionPath(dp)));
 777     }
 778
 779     if (code == 0 && VVGCache.part[dp->index].state == VVGC_PART_STATE_UPDATING) {
 780         /* we successfully added the entry; make sure it's not on the
 781          * to-delete list, so it doesn't get deleted later */
 782         code = _VVGC_dlist_del_r(dp, parent, child);
 783         if (code && code != ENOENT) {
 784             ViceLog(0, ("VVGCache_entry_add: error %d trying to remove vol "
 785                         "%lu (parent %lu) from the to-delete list for part "
 786                         "%s.\n", code, afs_printable_uint32_lu(child),
 787                         afs_printable_uint32_lu(parent),
 788                         VPartitionPath(dp)));
 789         } else {
 790             code = 0;
 791         }
 792     }
 793
 794     return code;
 795 }
 796
 797 /**
 798  * add an entry to the volume group cache.
 799  *
 800  * @param[in] dp       disk partition object
 801  * @param[in] parent   parent volume id
 802  * @param[in] child    child volume id
 803  * @param[out] newvg   if non-NULL, *newvg is 1 if adding this added a
 804  *                     new VG, 0 if we added to an existing VG
 805  *
 806  * @return operation status
 807  *    @retval 0 success
 808  */
 809 int
 810 VVGCache_entry_add(struct DiskPartition64 * dp,
 811                    VolumeId parent,
 812                    VolumeId child,
 813                    afs_int32 *newvg)
 814 {
 815     int code = 0;
 816
 817     VOL_LOCK;
 818     VVGCache_entry_add_r(dp, parent, child, newvg);
 819     VOL_UNLOCK;
 820
 821     return code;
 822 }
 823
 824 /**
 825  * delete an entry from the volume group cache.
 826  *
 827  * If partition is scanning, actually puts the entry on a list of entries
 828  * to delete when the scan is done.
 829  *
 830  * @param[in] dp       disk partition object
 831  * @param[in] parent   parent volume id
 832  * @param[in] child    child volume id
 833  *
 834  * @pre VOL_LOCK held
 835  *
 836  * @return operation status
 837  *    @retval 0 success
 838  */
 839 int
 840 VVGCache_entry_del_r(struct DiskPartition64 * dp,
 841                      VolumeId parent, VolumeId child)
 842 {
 843     if (VVGCache.part[dp->index].state == VVGC_PART_STATE_UPDATING) {
 844         int code;
 845         code = _VVGC_dlist_add_r(dp, parent, child);
 846         if (code) {
 847             return code;
 848         }
 849     }
 850     return _VVGC_entry_purge_r(dp, parent, child);
 851 }
 852
 853 /**
 854  * delete an entry from the volume group cache.
 855  *
 856  * @param[in] dp       disk partition object
 857  * @param[in] parent   parent volume id
 858  * @param[in] child    child volume id
 859  *
 860  * @pre VOL_LOCK held
 861  *
 862  * @internal
 863  *
 864  * @return operation status
 865  *    @retval 0 success
 866  */
 867 int
 868 _VVGC_entry_purge_r(struct DiskPartition64 * dp,
 869                     VolumeId parent, VolumeId child)
 870 {
 871     int code = 0, res;
 872     VVGCache_entry_t * parent_ent, * child_ent;
 873     VVGCache_hash_entry_t * child_hent;
 874
 875     /* check mappings for each volid */
 876     res = _VVGC_lookup(dp, parent, &parent_ent, NULL);
 877     if (res) {
 878         code = res;
 879         goto done;
 880     }
 881     res = _VVGC_lookup(dp, child, &child_ent, &child_hent);
 882     if (res) {
 883         code = res;
 884         goto done;
 885     }
 886
 887     /* if the mappings don't match, we have a serious error */
 888     if (parent_ent != child_ent) {
 889         ViceLog(0, ("VVGCache_entry_del: trying to delete vol %lu from VG %lu, "
 890             "but vol %lu points to VGC entry %"AFS_PTR_FMT" and VG %lu "
 891             "points to VGC entry %"AFS_PTR_FMT"\n",
 892             afs_printable_uint32_lu(child),
 893             afs_printable_uint32_lu(parent),
 894             afs_printable_uint32_lu(child),
 895             child_ent, afs_printable_uint32_lu(parent), parent_ent));
 896         code = -1;
 897         goto done;
 898     }
 899
 900     code = _VVGC_hash_entry_del(child_hent);
 901
 902  done:
 903     return code;
 904 }
 905
 906 /**
 907  * delete an entry from the volume group cache.
 908  *
 909  * @param[in] dp       disk partition object
 910  * @param[in] parent   parent volume id
 911  * @param[in] child    child volume id
 912  *
 913  * @return operation status
 914  *    @retval 0 success
 915  */
 916 int
 917 VVGCache_entry_del(struct DiskPartition64 * dp,
 918                    VolumeId parent, VolumeId child)
 919 {
 920     int code;
 921
 922     VOL_LOCK;
 923     code = VVGCache_entry_del_r(dp, parent, child);
 924     VOL_UNLOCK;
 925
 926     return code;
 927 }
 928
 929 /**
 930  * query a volume group by any member volume id.
 931  *
 932  * @param[in]  dp        disk partition object
 933  * @param[in]  volume    volume id of a member of VG
 934  * @param[out] res       vg membership data
 935  *
 936  * @pre VOL_LOCK held
 937  *
 938  * @return operation status
 939  *    @retval 0 success
 940  *    @retval EAGAIN partition needs to finish scanning
 941  */
 942 int
 943 VVGCache_query_r(struct DiskPartition64 * dp,
 944                  VolumeId volume,
 945                  VVGCache_query_t * res)
 946 {
 947     int code = 0;
 948     VVGCache_entry_t * ent;
 949
 950     /* If cache for this partition doesn't exist; start a scan */
 951     if (VVGCache.part[dp->index].state == VVGC_PART_STATE_INVALID) {
 952         code = VVGCache_scanStart_r(dp);
 953         if (code == 0 || code == -3) {
 954             /* -3 means another thread already started scanning */
 955             return EAGAIN;
 956         }
 957         return code;
 958     }
 959     if (VVGCache.part[dp->index].state == VVGC_PART_STATE_UPDATING) {
 960         return EAGAIN;
 961     }
 962
 963     code = _VVGC_lookup(dp, volume, &ent, NULL);
 964     if (!code) {
 965         code = _VVGC_entry_export(ent, res);
 966     }
 967
 968     return code;
 969 }
 970
 971 /**
 972  * query a volume group by any member volume id.
 973  *
 974  * @param[in]  dp        disk partition object
 975  * @param[in]  volume    volume id of a member of VG
 976  * @param[out] res       vg membership data
 977  *
 978  * @return operation status
 979  *    @retval 0 success
 980  */
 981 int
 982 VVGCache_query(struct DiskPartition64 * dp,
 983                VolumeId volume, VVGCache_query_t * res)
 984 {
 985     int code;
 986
 987     VOL_LOCK;
 988     code = VVGCache_query_r(dp, volume, res);
 989     VOL_UNLOCK;
 990
 991     return code;
 992 }
 993
 994 /**
 995  * begin asynchronous scan of on-disk volume group metadata.
 996  *
 997  * @param[in] dp       disk partition object
 998  *
 999  * @pre VOL_LOCK held
1000  *
1001  * @return operation status
1002  *    @retval 0 success
1003  */
1004 int
1005 VVGCache_scanStart_r(struct DiskPartition64 * dp)
1006 {
1007     int code = 0, res;
1008
1009     if (dp) {
1010         code = _VVGC_scan_start(dp);
1011     } else {
1012         /* start a scanner thread on each partition */
1013         for (dp = DiskPartitionList; dp; dp = dp->next) {
1014             res = _VVGC_scan_start(dp);
1015             if (res) {
1016                 code = res;
1017             }
1018         }
1019     }
1020
1021     return code;
1022 }
1023
1024 /**
1025  * begin asynchronous scan of on-disk volume group metadata.
1026  *
1027  * @param[in] dp       disk partition object
1028  *
1029  * @return operation status
1030  *    @retval 0 success
1031  */
1032 int
1033 VVGCache_scanStart(struct DiskPartition64 * dp)
1034 {
1035     int code;
1036
1037     VOL_LOCK;
1038     code = VVGCache_scanStart_r(dp);
1039     VOL_UNLOCK;
1040
1041     return code;
1042 }
1043
1044 /**
1045  * wait for async on-disk VG metadata scan to complete.
1046  *
1047  * @param[in] dp       disk partition object
1048  *
1049  * @pre VOL_LOCK held
1050  *
1051  * @warning this routine must drop VOL_LOCK internally
1052  *
1053  * @return operation status
1054  *    @retval 0 success
1055  */
1056 int
1057 VVGCache_scanWait_r(struct DiskPartition64 * dp)
1058 {
1059     int code = 0;
1060
1061     while (VVGCache.part[dp->index].state == VVGC_PART_STATE_UPDATING) {
1062         VOL_CV_WAIT(&VVGCache.part[dp->index].cv);
1063     }
1064
1065     return code;
1066 }
1067
1068 /**
1069  * wait for async on-disk VG metadata scan to complete.
1070  *
1071  * @param[in] dp       disk partition object
1072  *
1073  * @return operation status
1074  *    @retval 0 success
1075  */
1076 int
1077 VVGCache_scanWait(struct DiskPartition64 * dp)
1078 {
1079     int code;
1080
1081     VOL_LOCK;
1082     code = VVGCache_scanWait_r(dp);
1083     VOL_UNLOCK;
1084
1085     return code;
1086 }
1087
1088 /**
1089  * flush all cache entries for a given disk partition.
1090  *
1091  * @param[in] part  disk partition object
1092  *
1093  * @pre VOL_LOCK held
1094  *
1095  * @return operation status
1096  *    @retval 0 success
1097  *
1098  * @internal
1099  */
1100 int
1101 _VVGC_flush_part_r(struct DiskPartition64 * part)
1102 {
1103     int code = 0, res;
1104     int i;
1105     VVGCache_hash_entry_t * ent, * nent;
1106
1107     for (i = 0; i < VolumeHashTable.Size; i++) {
1108         for (queue_Scan(&VVGCache_hash_table.hash_buckets[i],
1109                         ent,
1110                         nent,
1111                         VVGCache_hash_entry)) {
1112             if (ent->dp == part) {
1113                 VolumeId volid = ent->volid;
1114                 res = _VVGC_hash_entry_del(ent);
1115                 if (res) {
1116                     ViceLog(0, ("_VVGC_flush_part_r: error %d deleting hash entry for %lu\n",
1117                         res, afs_printable_uint32_lu(volid)));
1118                     code = res;
1119                 }
1120             }
1121         }
1122     }
1123
1124     return code;
1125 }
1126
1127 /**
1128  * flush all cache entries for a given disk partition.
1129  *
1130  * @param[in] part  disk partition object
1131  *
1132  * @return operation status
1133  *    @retval 0 success
1134  *
1135  * @internal
1136  */
1137 int
1138 _VVGC_flush_part(struct DiskPartition64 * part)
1139 {
1140     int code;
1141
1142     VOL_LOCK;
1143     code = _VVGC_flush_part_r(part);
1144     VOL_UNLOCK;
1145
1146     return code;
1147 }
1148
1149
1150 /**
1151  * change VVGC partition state.
1152  *
1153  * @param[in]  part    disk partition object
1154  * @param[in]  state   new state
1155  *
1156  * @pre VOL_LOCK is held
1157  *
1158  * @return old state
1159  *
1160  * @internal
1161  */
1162 int
1163 _VVGC_state_change(struct DiskPartition64 * part,
1164                    VVGCache_part_state_t state)
1165 {
1166     VVGCache_part_state_t old_state;
1167
1168     old_state = VVGCache.part[part->index].state;
1169     VVGCache.part[part->index].state = state;
1170
1171     if (old_state != state) {
1172         CV_BROADCAST(&VVGCache.part[part->index].cv);
1173     }
1174
1175     return old_state;
1176 }
1177
1178 #endif /* AFS_DEMAND_ATTACH_FS */