sys/vfs/hammer/hammer_ondisk.c

   1 /*
   2  * Copyright (c) 2007 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_ondisk.c,v 1.14 2007/12/31 05:33:12 dillon Exp $
  35  */
  36 /*
  37  * Manage HAMMER's on-disk structures.  These routines are primarily
  38  * responsible for interfacing with the kernel's I/O subsystem and for
  39  * managing in-memory structures.
  40  */
  41
  42 #include "hammer.h"
  43 #include <sys/fcntl.h>
  44 #include <sys/nlookup.h>
  45 #include <sys/buf.h>
  46 #include <sys/buf2.h>
  47
  48 static void hammer_free_volume(hammer_volume_t volume);
  49 static int hammer_load_volume(hammer_volume_t volume);
  50 static int hammer_load_supercl(hammer_supercl_t supercl, int isnew);
  51 static int hammer_load_cluster(hammer_cluster_t cluster, int isnew);
  52 static int hammer_load_buffer(hammer_buffer_t buffer, u_int64_t buf_type);
  53 static void hammer_remove_node_clist(hammer_buffer_t buffer,
  54                         hammer_node_t node);
  55 static void initbuffer(hammer_alist_t live, hammer_fsbuf_head_t head,
  56                         u_int64_t type);
  57 static void alloc_new_buffer(hammer_cluster_t cluster,
  58                         hammer_alist_t live, u_int64_t type, int32_t nelements,
  59                         int32_t start,
  60                         int *errorp, struct hammer_buffer **bufferp);
  61 #if 0
  62 static void readhammerbuf(hammer_volume_t vol, void *data,
  63                         int64_t offset);
  64 static void writehammerbuf(hammer_volume_t vol, const void *data,
  65                         int64_t offset);
  66 #endif
  67 static int64_t calculate_cluster_offset(hammer_volume_t vol, int32_t clu_no);
  68 static int64_t calculate_supercl_offset(hammer_volume_t vol, int32_t scl_no);
  69 static int32_t hammer_alloc_master(hammer_cluster_t cluster, int nblks,
  70                         int32_t start, int isfwd);
  71 static void hammer_adjust_stats(hammer_cluster_t cluster,
  72                         u_int64_t buf_type, int nblks);
  73
  74 struct hammer_alist_config Buf_alist_config;
  75 struct hammer_alist_config Vol_normal_alist_config;
  76 struct hammer_alist_config Vol_super_alist_config;
  77 struct hammer_alist_config Supercl_alist_config;
  78 struct hammer_alist_config Clu_master_alist_config;
  79 struct hammer_alist_config Clu_slave_alist_config;
  80
  81 /*
  82  * Red-Black tree support for various structures
  83  */
  84 static int
  85 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  86 {
  87         if (ip1->obj_id < ip2->obj_id)
  88                 return(-1);
  89         if (ip1->obj_id > ip2->obj_id)
  90                 return(1);
  91         if (ip1->obj_asof < ip2->obj_asof)
  92                 return(-1);
  93         if (ip1->obj_asof > ip2->obj_asof)
  94                 return(1);
  95         return(0);
  96 }
  97
  98 static int
  99 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
 100 {
 101         if (info->obj_id < ip->obj_id)
 102                 return(-1);
 103         if (info->obj_id > ip->obj_id)
 104                 return(1);
 105         if (info->obj_asof < ip->obj_asof)
 106                 return(-1);
 107         if (info->obj_asof > ip->obj_asof)
 108                 return(1);
 109         return(0);
 110 }
 111
 112 static int
 113 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
 114 {
 115         if (vol1->vol_no < vol2->vol_no)
 116                 return(-1);
 117         if (vol1->vol_no > vol2->vol_no)
 118                 return(1);
 119         return(0);
 120 }
 121
 122 static int
 123 hammer_scl_rb_compare(hammer_supercl_t cl1, hammer_supercl_t cl2)
 124 {
 125         if (cl1->scl_no < cl2->scl_no)
 126                 return(-1);
 127         if (cl1->scl_no > cl2->scl_no)
 128                 return(1);
 129         return(0);
 130 }
 131
 132 static int
 133 hammer_clu_rb_compare(hammer_cluster_t cl1, hammer_cluster_t cl2)
 134 {
 135         if (cl1->clu_no < cl2->clu_no)
 136                 return(-1);
 137         if (cl1->clu_no > cl2->clu_no)
 138                 return(1);
 139         return(0);
 140 }
 141
 142 static int
 143 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
 144 {
 145         if (buf1->buf_no < buf2->buf_no)
 146                 return(-1);
 147         if (buf1->buf_no > buf2->buf_no)
 148                 return(1);
 149         return(0);
 150 }
 151
 152 static int
 153 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
 154 {
 155         if (node1->node_offset < node2->node_offset)
 156                 return(-1);
 157         if (node1->node_offset > node2->node_offset)
 158                 return(1);
 159         return(0);
 160 }
 161
 162 /*
 163  * Note: The lookup function for hammer_ino_rb_tree winds up being named
 164  * hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).  The other lookup
 165  * functions are normal, e.g. hammer_clu_rb_tree_RB_LOOKUP(root, clu_no).
 166  */
 167 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 168 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 169                 hammer_inode_info_cmp, hammer_inode_info_t);
 170 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 171              hammer_vol_rb_compare, int32_t, vol_no);
 172 RB_GENERATE2(hammer_scl_rb_tree, hammer_supercl, rb_node,
 173              hammer_scl_rb_compare, int32_t, scl_no);
 174 RB_GENERATE2(hammer_clu_rb_tree, hammer_cluster, rb_node,
 175              hammer_clu_rb_compare, int32_t, clu_no);
 176 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 177              hammer_buf_rb_compare, int32_t, buf_no);
 178 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
 179              hammer_nod_rb_compare, int32_t, node_offset);
 180
 181 /************************************************************************
 182  *                              VOLUMES                                 *
 183  ************************************************************************
 184  *
 185  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 186  * code on failure.  Volumes must be loaded at mount time, get_volume() will
 187  * not load a new volume.
 188  *
 189  * Calls made to hammer_load_volume() or single-threaded
 190  */
 191 int
 192 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
 193 {
 194         struct mount *mp;
 195         hammer_volume_t volume;
 196         struct hammer_volume_ondisk *ondisk;
 197         struct nlookupdata nd;
 198         struct buf *bp = NULL;
 199         int error;
 200         int ronly;
 201
 202         mp = hmp->mp;
 203         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 204
 205         /*
 206          * Allocate a volume structure
 207          */
 208         ++hammer_count_volumes;
 209         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 210         volume->vol_name = kstrdup(volname, M_HAMMER);
 211         volume->hmp = hmp;
 212         volume->io.type = HAMMER_STRUCTURE_VOLUME;
 213         volume->io.offset = 0LL;
 214
 215         /*
 216          * Get the device vnode
 217          */
 218         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 219         if (error == 0)
 220                 error = nlookup(&nd);
 221         if (error == 0)
 222                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 223         nlookup_done(&nd);
 224         if (error == 0) {
 225                 vn_isdisk(volume->devvp, &error);
 226         }
 227         if (error == 0) {
 228                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 229                 error = VOP_OPEN(volume->devvp, (ronly ? FREAD : FREAD|FWRITE),
 230                                  FSCRED, NULL);
 231                 vn_unlock(volume->devvp);
 232         }
 233         if (error) {
 234                 hammer_free_volume(volume);
 235                 return(error);
 236         }
 237
 238         /*
 239          * Extract the volume number from the volume header and do various
 240          * sanity checks.
 241          */
 242         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 243         if (error)
 244                 goto late_failure;
 245         ondisk = (void *)bp->b_data;
 246         if (ondisk->head.buf_type != HAMMER_FSBUF_VOLUME) {
 247                 kprintf("hammer_mount: volume %s has an invalid header\n",
 248                         volume->vol_name);
 249                 error = EFTYPE;
 250                 goto late_failure;
 251         }
 252         volume->vol_no = ondisk->vol_no;
 253         volume->cluster_base = ondisk->vol_clo_beg;
 254         volume->vol_clsize = ondisk->vol_clsize;
 255         volume->vol_flags = ondisk->vol_flags;
 256         volume->nblocks = ondisk->vol_nblocks;
 257         RB_INIT(&volume->rb_clus_root);
 258         RB_INIT(&volume->rb_scls_root);
 259
 260         hmp->mp->mnt_stat.f_blocks += volume->nblocks;
 261
 262         if (RB_EMPTY(&hmp->rb_vols_root)) {
 263                 hmp->fsid = ondisk->vol_fsid;
 264         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 265                 kprintf("hammer_mount: volume %s's fsid does not match "
 266                         "other volumes\n", volume->vol_name);
 267                 error = EFTYPE;
 268                 goto late_failure;
 269         }
 270
 271         /*
 272          * Insert the volume structure into the red-black tree.
 273          */
 274         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 275                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 276                         volume->vol_name, volume->vol_no);
 277                 error = EEXIST;
 278         }
 279
 280         /*
 281          * Set the root volume and load the root cluster.  HAMMER special
 282          * cases rootvol and rootcl and will not deallocate the structures.
 283          * We do not hold a ref because this would prevent related I/O
 284          * from being flushed.
 285          */
 286         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 287                 hmp->rootvol = volume;
 288                 hmp->rootcl = hammer_get_cluster(volume,
 289                                                  ondisk->vol0_root_clu_no,
 290                                                  &error, 0);
 291                 hammer_rel_cluster(hmp->rootcl, 0);
 292                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 293         }
 294 late_failure:
 295         if (bp)
 296                 brelse(bp);
 297         if (error) {
 298                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 299                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 300                 hammer_free_volume(volume);
 301         }
 302         return (error);
 303 }
 304
 305 /*
 306  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 307  * so returns -1 on failure.
 308  */
 309 int
 310 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 311 {
 312         struct hammer_mount *hmp = volume->hmp;
 313         hammer_cluster_t rootcl;
 314         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 315
 316         /*
 317          * Sync clusters, sync volume
 318          */
 319
 320         hmp->mp->mnt_stat.f_blocks -= volume->nblocks;
 321
 322         /*
 323          * Clean up the root cluster, which is held unlocked in the root
 324          * volume.
 325          */
 326         if (hmp->rootvol == volume) {
 327                 if ((rootcl = hmp->rootcl) != NULL)
 328                         hmp->rootcl = NULL;
 329                 hmp->rootvol = NULL;
 330         }
 331
 332         /*
 333          * Unload clusters and super-clusters.  Unloading a super-cluster
 334          * also unloads related clusters, but the filesystem may not be
 335          * using super-clusters so unload clusters anyway.
 336          */
 337         RB_SCAN(hammer_clu_rb_tree, &volume->rb_clus_root, NULL,
 338                         hammer_unload_cluster, NULL);
 339         RB_SCAN(hammer_scl_rb_tree, &volume->rb_scls_root, NULL,
 340                         hammer_unload_supercl, NULL);
 341
 342         /*
 343          * Release our buffer and flush anything left in the buffer cache.
 344          */
 345         hammer_io_release(&volume->io, 1);
 346
 347         /*
 348          * There should be no references on the volume, no clusters, and
 349          * no super-clusters.
 350          */
 351         KKASSERT(volume->io.lock.refs == 0);
 352         KKASSERT(RB_EMPTY(&volume->rb_clus_root));
 353         KKASSERT(RB_EMPTY(&volume->rb_scls_root));
 354
 355         volume->ondisk = NULL;
 356         if (volume->devvp) {
 357                 if (ronly) {
 358                         vinvalbuf(volume->devvp, 0, 0, 0);
 359                         VOP_CLOSE(volume->devvp, FREAD);
 360                 } else {
 361                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 362                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 363                 }
 364         }
 365
 366         /*
 367          * Destroy the structure
 368          */
 369         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 370         hammer_free_volume(volume);
 371         return(0);
 372 }
 373
 374 static
 375 void
 376 hammer_free_volume(hammer_volume_t volume)
 377 {
 378         if (volume->vol_name) {
 379                 kfree(volume->vol_name, M_HAMMER);
 380                 volume->vol_name = NULL;
 381         }
 382         if (volume->devvp) {
 383                 vrele(volume->devvp);
 384                 volume->devvp = NULL;
 385         }
 386         --hammer_count_volumes;
 387         kfree(volume, M_HAMMER);
 388 }
 389
 390 /*
 391  * Get a HAMMER volume.  The volume must already exist.
 392  */
 393 hammer_volume_t
 394 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 395 {
 396         struct hammer_volume *volume;
 397
 398         /*
 399          * Locate the volume structure
 400          */
 401         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 402         if (volume == NULL) {
 403                 *errorp = ENOENT;
 404                 return(NULL);
 405         }
 406         hammer_ref(&volume->io.lock);
 407
 408         /*
 409          * Deal with on-disk info
 410          */
 411         if (volume->ondisk == NULL) {
 412                 *errorp = hammer_load_volume(volume);
 413                 if (*errorp) {
 414                         hammer_rel_volume(volume, 1);
 415                         volume = NULL;
 416                 }
 417         } else {
 418                 *errorp = 0;
 419         }
 420         return(volume);
 421 }
 422
 423 int
 424 hammer_ref_volume(hammer_volume_t volume)
 425 {
 426         int error;
 427
 428         hammer_ref(&volume->io.lock);
 429
 430         /*
 431          * Deal with on-disk info
 432          */
 433         if (volume->ondisk == NULL) {
 434                 error = hammer_load_volume(volume);
 435                 if (error)
 436                         hammer_rel_volume(volume, 1);
 437         } else {
 438                 error = 0;
 439         }
 440         return (error);
 441 }
 442
 443 hammer_volume_t
 444 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 445 {
 446         hammer_volume_t volume;
 447
 448         volume = hmp->rootvol;
 449         KKASSERT(volume != NULL);
 450         hammer_ref(&volume->io.lock);
 451
 452         /*
 453          * Deal with on-disk info
 454          */
 455         if (volume->ondisk == NULL) {
 456                 *errorp = hammer_load_volume(volume);
 457                 if (*errorp) {
 458                         hammer_rel_volume(volume, 1);
 459                         volume = NULL;
 460                 }
 461         } else {
 462                 *errorp = 0;
 463         }
 464         return (volume);
 465 }
 466
 467 /*
 468  * Load a volume's on-disk information.  The volume must be referenced and
 469  * not locked.  We temporarily acquire an exclusive lock to interlock
 470  * against releases or multiple get's.
 471  */
 472 static int
 473 hammer_load_volume(hammer_volume_t volume)
 474 {
 475         struct hammer_volume_ondisk *ondisk;
 476         int error;
 477
 478         hammer_lock_ex(&volume->io.lock);
 479         if (volume->ondisk == NULL) {
 480                 error = hammer_io_read(volume->devvp, &volume->io);
 481                 if (error) {
 482                         hammer_unlock(&volume->io.lock);
 483                         return (error);
 484                 }
 485                 volume->ondisk = ondisk = (void *)volume->io.bp->b_data;
 486
 487                 /*
 488                  * Configure the volume's A-lists.  These are used to
 489                  * allocate clusters.
 490                  */
 491                 if (volume->vol_flags & HAMMER_VOLF_USINGSUPERCL) {
 492                         volume->alist.config = &Vol_super_alist_config;
 493                         volume->alist.meta = ondisk->vol_almeta.super;
 494                         volume->alist.info = volume;
 495                 } else {
 496                         volume->alist.config = &Vol_normal_alist_config;
 497                         volume->alist.meta = ondisk->vol_almeta.normal;
 498                         volume->alist.info = NULL;
 499                 }
 500         } else {
 501                 error = 0;
 502         }
 503         hammer_unlock(&volume->io.lock);
 504         return(0);
 505 }
 506
 507 /*
 508  * Release a volume.  Call hammer_io_release on the last reference.  We have
 509  * to acquire an exclusive lock to interlock against volume->ondisk tests
 510  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 511  * lock to be held.
 512  *
 513  * Volumes are not unloaded from memory during normal operation.
 514  */
 515 void
 516 hammer_rel_volume(hammer_volume_t volume, int flush)
 517 {
 518         if (volume->io.lock.refs == 1) {
 519                 hammer_lock_ex(&volume->io.lock);
 520                 if (volume->io.lock.refs == 1) {
 521                         volume->ondisk = NULL;
 522                         hammer_io_release(&volume->io, flush);
 523                 }
 524                 hammer_unlock(&volume->io.lock);
 525         }
 526         hammer_unref(&volume->io.lock);
 527 }
 528
 529 /************************************************************************
 530  *                              SUPER-CLUSTERS                          *
 531  ************************************************************************
 532  *
 533  * Manage super-clusters.  Note that a supercl holds a reference to its
 534  * associated volume.
 535  */
 536 hammer_supercl_t
 537 hammer_get_supercl(hammer_volume_t volume, int32_t scl_no,
 538                    int *errorp, int isnew)
 539 {
 540         hammer_supercl_t supercl;
 541
 542         /*
 543          * Locate and lock the super-cluster structure, creating one
 544          * if necessary.
 545          */
 546 again:
 547         supercl = RB_LOOKUP(hammer_scl_rb_tree, &volume->rb_scls_root, scl_no);
 548         if (supercl == NULL) {
 549                 ++hammer_count_supercls;
 550                 supercl = kmalloc(sizeof(*supercl), M_HAMMER, M_WAITOK|M_ZERO);
 551                 supercl->scl_no = scl_no;
 552                 supercl->volume = volume;
 553                 supercl->io.offset = calculate_supercl_offset(volume, scl_no);
 554                 supercl->io.type = HAMMER_STRUCTURE_SUPERCL;
 555                 hammer_ref(&supercl->io.lock);
 556
 557                 /*
 558                  * Insert the cluster into the RB tree and handle late
 559                  * collisions.
 560                  */
 561                 if (RB_INSERT(hammer_scl_rb_tree, &volume->rb_scls_root, supercl)) {
 562                         hammer_unref(&supercl->io.lock);
 563                         --hammer_count_supercls;
 564                         kfree(supercl, M_HAMMER);
 565                         goto again;
 566                 }
 567                 hammer_ref(&volume->io.lock);
 568         } else {
 569                 hammer_ref(&supercl->io.lock);
 570         }
 571
 572         /*
 573          * Deal with on-disk info
 574          */
 575         if (supercl->ondisk == NULL || isnew) {
 576                 *errorp = hammer_load_supercl(supercl, isnew);
 577                 if (*errorp) {
 578                         hammer_rel_supercl(supercl, 1);
 579                         supercl = NULL;
 580                 }
 581         } else {
 582                 *errorp = 0;
 583         }
 584         return(supercl);
 585 }
 586
 587 static int
 588 hammer_load_supercl(hammer_supercl_t supercl, int isnew)
 589 {
 590         struct hammer_supercl_ondisk *ondisk;
 591         hammer_volume_t volume = supercl->volume;
 592         int error;
 593
 594         hammer_lock_ex(&supercl->io.lock);
 595         if (supercl->ondisk == NULL) {
 596                 if (isnew)
 597                         error = hammer_io_new(volume->devvp, &supercl->io);
 598                 else
 599                         error = hammer_io_read(volume->devvp, &supercl->io);
 600                 if (error) {
 601                         hammer_unlock(&supercl->io.lock);
 602                         return (error);
 603                 }
 604                 supercl->ondisk = ondisk = (void *)supercl->io.bp->b_data;
 605
 606                 supercl->alist.config = &Supercl_alist_config;
 607                 supercl->alist.meta = ondisk->scl_meta;
 608                 supercl->alist.info = NULL;
 609         } else if (isnew) {
 610                 error = hammer_io_new(volume->devvp, &supercl->io);
 611         } else {
 612                 error = 0;
 613         }
 614         if (error == 0 && isnew) {
 615                 /*
 616                  * If this is a new super-cluster we have to initialize
 617                  * various ondisk structural elements.  The caller is
 618                  * responsible for the remainder.
 619                  */
 620                 struct hammer_alist_live dummy;
 621
 622                 ondisk = supercl->ondisk;
 623                 dummy.config = &Buf_alist_config;
 624                 dummy.meta = ondisk->head.buf_almeta;
 625                 dummy.info = NULL;
 626                 initbuffer(&dummy, &ondisk->head, HAMMER_FSBUF_SUPERCL);
 627                 hammer_alist_init(&supercl->alist);
 628         }
 629         hammer_unlock(&supercl->io.lock);
 630         return (error);
 631 }
 632
 633 /*
 634  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 635  */
 636 int
 637 hammer_unload_supercl(hammer_supercl_t supercl, void *data __unused)
 638 {
 639         KKASSERT(supercl->io.lock.refs == 0);
 640         hammer_ref(&supercl->io.lock);
 641         hammer_rel_supercl(supercl, 1);
 642         return(0);
 643 }
 644
 645 /*
 646  * Release a super-cluster.  We have to deal with several places where
 647  * another thread can ref the super-cluster.
 648  *
 649  * Only destroy the structure itself if the related buffer cache buffer
 650  * was disassociated from it.  This ties the management of the structure
 651  * to the buffer cache subsystem.
 652  */
 653 void
 654 hammer_rel_supercl(hammer_supercl_t supercl, int flush)
 655 {
 656         hammer_volume_t volume;
 657
 658         if (supercl->io.lock.refs == 1) {
 659                 hammer_lock_ex(&supercl->io.lock);
 660                 if (supercl->io.lock.refs == 1) {
 661                         hammer_io_release(&supercl->io, flush);
 662                         if (supercl->io.bp == NULL &&
 663                             supercl->io.lock.refs == 1) {
 664                                 volume = supercl->volume;
 665                                 RB_REMOVE(hammer_scl_rb_tree,
 666                                           &volume->rb_scls_root, supercl);
 667                                 supercl->volume = NULL; /* sanity */
 668                                 --hammer_count_supercls;
 669                                 kfree(supercl, M_HAMMER);
 670                                 hammer_rel_volume(volume, 0);
 671                                 return;
 672                         }
 673                 }
 674                 hammer_unlock(&supercl->io.lock);
 675         }
 676         hammer_unref(&supercl->io.lock);
 677 }
 678
 679 /************************************************************************
 680  *                              CLUSTERS                                *
 681  ************************************************************************
 682  *
 683  */
 684 hammer_cluster_t
 685 hammer_get_cluster(hammer_volume_t volume, int32_t clu_no,
 686                    int *errorp, int isnew)
 687 {
 688         hammer_cluster_t cluster;
 689
 690 again:
 691         cluster = RB_LOOKUP(hammer_clu_rb_tree, &volume->rb_clus_root, clu_no);
 692         if (cluster == NULL) {
 693                 ++hammer_count_clusters;
 694                 cluster = kmalloc(sizeof(*cluster), M_HAMMER, M_WAITOK|M_ZERO);
 695                 cluster->clu_no = clu_no;
 696                 cluster->volume = volume;
 697                 cluster->io.offset = calculate_cluster_offset(volume, clu_no);
 698                 cluster->state = HAMMER_CLUSTER_IDLE;
 699                 RB_INIT(&cluster->rb_bufs_root);
 700                 RB_INIT(&cluster->rb_nods_root);
 701                 cluster->io.type = HAMMER_STRUCTURE_CLUSTER;
 702                 hammer_ref(&cluster->io.lock);
 703
 704                 /*
 705                  * Insert the cluster into the RB tree and handle late
 706                  * collisions.
 707                  */
 708                 if (RB_INSERT(hammer_clu_rb_tree, &volume->rb_clus_root, cluster)) {
 709                         hammer_unref(&cluster->io.lock);
 710                         --hammer_count_clusters;
 711                         kfree(cluster, M_HAMMER);
 712                         goto again;
 713                 }
 714                 hammer_ref(&volume->io.lock);
 715         } else {
 716                 hammer_ref(&cluster->io.lock);
 717         }
 718
 719         /*
 720          * Deal with on-disk info
 721          */
 722         if (cluster->ondisk == NULL || isnew) {
 723                 *errorp = hammer_load_cluster(cluster, isnew);
 724                 if (*errorp) {
 725                         hammer_rel_cluster(cluster, 1);
 726                         cluster = NULL;
 727                 }
 728         } else {
 729                 *errorp = 0;
 730         }
 731         return (cluster);
 732 }
 733
 734 hammer_cluster_t
 735 hammer_get_root_cluster(struct hammer_mount *hmp, int *errorp)
 736 {
 737         hammer_cluster_t cluster;
 738
 739         cluster = hmp->rootcl;
 740         KKASSERT(cluster != NULL);
 741         hammer_ref(&cluster->io.lock);
 742
 743         /*
 744          * Deal with on-disk info
 745          */
 746         if (cluster->ondisk == NULL) {
 747                 *errorp = hammer_load_cluster(cluster, 0);
 748                 if (*errorp) {
 749                         hammer_rel_cluster(cluster, 1);
 750                         cluster = NULL;
 751                 }
 752         } else {
 753                 *errorp = 0;
 754         }
 755         return (cluster);
 756 }
 757
 758 static
 759 int
 760 hammer_load_cluster(hammer_cluster_t cluster, int isnew)
 761 {
 762         hammer_volume_t volume = cluster->volume;
 763         struct hammer_cluster_ondisk *ondisk;
 764         int error;
 765
 766         /*
 767          * Load the cluster's on-disk info
 768          */
 769         hammer_lock_ex(&cluster->io.lock);
 770         if (cluster->ondisk == NULL) {
 771                 if (isnew)
 772                         error = hammer_io_new(volume->devvp, &cluster->io);
 773                 else
 774                         error = hammer_io_read(volume->devvp, &cluster->io);
 775                 if (error) {
 776                         hammer_unlock(&cluster->io.lock);
 777                         return (error);
 778                 }
 779                 cluster->ondisk = ondisk = (void *)cluster->io.bp->b_data;
 780
 781                 cluster->alist_master.config = &Clu_master_alist_config;
 782                 cluster->alist_master.meta = ondisk->clu_master_meta;
 783                 cluster->alist_btree.config = &Clu_slave_alist_config;
 784                 cluster->alist_btree.meta = ondisk->clu_btree_meta;
 785                 cluster->alist_btree.info = cluster;
 786                 cluster->alist_record.config = &Clu_slave_alist_config;
 787                 cluster->alist_record.meta = ondisk->clu_record_meta;
 788                 cluster->alist_record.info = cluster;
 789                 cluster->alist_mdata.config = &Clu_slave_alist_config;
 790                 cluster->alist_mdata.meta = ondisk->clu_mdata_meta;
 791                 cluster->alist_mdata.info = cluster;
 792
 793                 if (isnew == 0) {
 794                         cluster->clu_btree_beg = ondisk->clu_btree_beg;
 795                         cluster->clu_btree_end = ondisk->clu_btree_end;
 796                 }
 797         } else if (isnew) {
 798                 error = hammer_io_new(volume->devvp, &cluster->io);
 799         } else {
 800                 error = 0;
 801         }
 802         if (error == 0 && isnew) {
 803                 /*
 804                  * If this is a new cluster we have to initialize
 805                  * various ondisk structural elements.  The caller is
 806                  * responsible for the remainder.
 807                  */
 808                 struct hammer_alist_live dummy;
 809                 hammer_node_t croot;
 810                 hammer_volume_ondisk_t voldisk;
 811                 int32_t nbuffers;
 812
 813                 hammer_modify_cluster(cluster);
 814                 ondisk = cluster->ondisk;
 815                 voldisk = volume->ondisk;
 816
 817                 dummy.config = &Buf_alist_config;
 818                 dummy.meta = ondisk->head.buf_almeta;
 819                 dummy.info = NULL;
 820                 initbuffer(&dummy, &ondisk->head, HAMMER_FSBUF_CLUSTER);
 821
 822                 hammer_alist_init(&cluster->alist_master);
 823                 hammer_alist_init(&cluster->alist_btree);
 824                 hammer_alist_init(&cluster->alist_record);
 825                 hammer_alist_init(&cluster->alist_mdata);
 826
 827                 ondisk->vol_fsid = voldisk->vol_fsid;
 828                 ondisk->vol_fstype = voldisk->vol_fstype;
 829                 ondisk->clu_gen = 1;
 830                 ondisk->clu_id = 0;     /* XXX */
 831                 ondisk->clu_no = cluster->clu_no;
 832                 ondisk->clu_flags = 0;
 833                 ondisk->clu_start = HAMMER_BUFSIZE;
 834                 KKASSERT(voldisk->vol_clo_end > cluster->io.offset);
 835                 if (voldisk->vol_clo_end - cluster->io.offset >
 836                     voldisk->vol_clsize) {
 837                         ondisk->clu_limit = voldisk->vol_clsize;
 838                 } else {
 839                         ondisk->clu_limit = (int32_t)(voldisk->vol_clo_end -
 840                                                       cluster->io.offset);
 841                 }
 842                 nbuffers = ondisk->clu_limit / HAMMER_BUFSIZE;
 843                 hammer_alist_free(&cluster->alist_master, 1, nbuffers - 1);
 844                 ondisk->idx_data = 1 * HAMMER_FSBUF_MAXBLKS;
 845                 ondisk->idx_index = 0 * HAMMER_FSBUF_MAXBLKS;
 846                 ondisk->idx_record = nbuffers * HAMMER_FSBUF_MAXBLKS;
 847
 848                 /*
 849                  * Initialize the B-Tree.  We don't know what the caller
 850                  * intends to do with the cluster so make sure it causes
 851                  * an assertion if the caller makes no changes.
 852                  */
 853                 ondisk->clu_btree_parent_vol_no = -2;
 854                 ondisk->clu_btree_parent_clu_no = -2;
 855                 ondisk->clu_btree_parent_offset = -2;
 856                 ondisk->clu_btree_parent_clu_gen = -2;
 857                 hammer_modify_cluster_done(cluster);
 858
 859                 croot = hammer_alloc_btree(cluster, &error);
 860                 if (error == 0) {
 861                         hammer_modify_node(croot);
 862                         bzero(croot->ondisk, sizeof(*croot->ondisk));
 863                         croot->ondisk->count = 0;
 864                         croot->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
 865                         hammer_modify_node_done(croot);
 866                         hammer_modify_cluster(cluster);
 867                         ondisk->clu_btree_root = croot->node_offset;
 868                         hammer_modify_cluster_done(cluster);
 869                         hammer_rel_node(croot);
 870                 }
 871         }
 872         hammer_unlock(&cluster->io.lock);
 873         return (error);
 874 }
 875
 876 /*
 877  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 878  */
 879 int
 880 hammer_unload_cluster(hammer_cluster_t cluster, void *data __unused)
 881 {
 882         hammer_ref(&cluster->io.lock);
 883         RB_SCAN(hammer_buf_rb_tree, &cluster->rb_bufs_root, NULL,
 884                 hammer_unload_buffer, NULL);
 885         KKASSERT(cluster->io.lock.refs == 1);
 886         hammer_rel_cluster(cluster, 1);
 887         return(0);
 888 }
 889
 890 /*
 891  * Reference a cluster that is either already referenced or via a specially
 892  * handled pointer (aka rootcl).
 893  */
 894 int
 895 hammer_ref_cluster(hammer_cluster_t cluster)
 896 {
 897         int error;
 898
 899         KKASSERT(cluster != NULL);
 900         hammer_ref(&cluster->io.lock);
 901
 902         /*
 903          * Deal with on-disk info
 904          */
 905         if (cluster->ondisk == NULL) {
 906                 error = hammer_load_cluster(cluster, 0);
 907                 if (error)
 908                         hammer_rel_cluster(cluster, 1);
 909         } else {
 910                 error = 0;
 911         }
 912         return(error);
 913 }
 914
 915 /*
 916  * Release a cluster.  We have to deal with several places where
 917  * another thread can ref the cluster.
 918  *
 919  * Only destroy the structure itself if the related buffer cache buffer
 920  * was disassociated from it.  This ties the management of the structure
 921  * to the buffer cache subsystem.
 922  */
 923 void
 924 hammer_rel_cluster(hammer_cluster_t cluster, int flush)
 925 {
 926         hammer_node_t node;
 927         hammer_volume_t volume;
 928
 929         if (cluster->io.lock.refs == 1) {
 930                 hammer_lock_ex(&cluster->io.lock);
 931                 if (cluster->io.lock.refs == 1) {
 932                         /*
 933                          * Release the I/O.  If we or the kernel wants to
 934                          * flush, this will release the bp.  Otherwise the
 935                          * bp may be written and flushed passively by the
 936                          * kernel later on.
 937                          */
 938                         hammer_io_release(&cluster->io, flush);
 939
 940                         /*
 941                          * The B-Tree node cache is not counted in the
 942                          * cluster's reference count.  Clean out the
 943                          * cache.
 944                          *
 945                          * If the cluster acquires a new reference while we
 946                          * are trying to clean it out, abort the cleaning.
 947                          *
 948                          * Any actively referenced nodes will reference the
 949                          * related buffer and cluster, so a ref count check
 950                          * should be sufficient.
 951                          */
 952                         while (cluster->io.bp == NULL &&
 953                                cluster->io.lock.refs == 1 &&
 954                                (node = RB_ROOT(&cluster->rb_nods_root)) != NULL
 955                         ) {
 956                                 KKASSERT(node->lock.refs == 0);
 957                                 hammer_flush_node(node);
 958                         }
 959
 960                         /*
 961                          * Final cleanup
 962                          */
 963                         if (cluster != cluster->volume->hmp->rootcl &&
 964                             cluster->io.bp == NULL &&
 965                             cluster->io.lock.refs == 1 &&
 966                             RB_EMPTY(&cluster->rb_nods_root)) {
 967                                 KKASSERT(RB_EMPTY(&cluster->rb_bufs_root));
 968                                 volume = cluster->volume;
 969                                 RB_REMOVE(hammer_clu_rb_tree,
 970                                           &volume->rb_clus_root, cluster);
 971                                 cluster->volume = NULL; /* sanity */
 972                                 --hammer_count_clusters;
 973                                 kfree(cluster, M_HAMMER);
 974                                 hammer_rel_volume(volume, 0);
 975                                 return;
 976                         }
 977                 }
 978                 hammer_unlock(&cluster->io.lock);
 979         }
 980         hammer_unref(&cluster->io.lock);
 981 }
 982
 983 /************************************************************************
 984  *                              BUFFERS                                 *
 985  ************************************************************************
 986  *
 987  * Manage buffers.  Note that a buffer holds a reference to its associated
 988  * cluster, and its cluster will hold a reference to the cluster's volume.
 989  *
 990  * A non-zero buf_type indicates that a new buffer should be created and
 991  * zero'd.
 992  */
 993 hammer_buffer_t
 994 hammer_get_buffer(hammer_cluster_t cluster, int32_t buf_no,
 995                   u_int64_t buf_type, int *errorp)
 996 {
 997         hammer_buffer_t buffer;
 998
 999         /*
1000          * Find the buffer.  Note that buffer 0 corresponds to the cluster
1001          * header and should never be requested.
1002          */
1003         KKASSERT(buf_no >= cluster->ondisk->clu_start / HAMMER_BUFSIZE &&
1004                  buf_no < cluster->ondisk->clu_limit / HAMMER_BUFSIZE);
1005
1006         /*
1007          * Locate and lock the buffer structure, creating one if necessary.
1008          */
1009 again:
1010         buffer = RB_LOOKUP(hammer_buf_rb_tree, &cluster->rb_bufs_root, buf_no);
1011         if (buffer == NULL) {
1012                 ++hammer_count_buffers;
1013                 buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
1014                 buffer->buf_no = buf_no;
1015                 buffer->cluster = cluster;
1016                 buffer->volume = cluster->volume;
1017                 buffer->io.offset = cluster->io.offset +
1018                                     (buf_no * HAMMER_BUFSIZE);
1019                 buffer->io.type = HAMMER_STRUCTURE_BUFFER;
1020                 TAILQ_INIT(&buffer->clist);
1021                 hammer_ref(&buffer->io.lock);
1022
1023                 /*
1024                  * Insert the cluster into the RB tree and handle late
1025                  * collisions.
1026                  */
1027                 if (RB_INSERT(hammer_buf_rb_tree, &cluster->rb_bufs_root, buffer)) {
1028                         hammer_unref(&buffer->io.lock);
1029                         --hammer_count_buffers;
1030                         kfree(buffer, M_HAMMER);
1031                         goto again;
1032                 }
1033                 hammer_ref(&cluster->io.lock);
1034         } else {
1035                 hammer_ref(&buffer->io.lock);
1036         }
1037
1038         /*
1039          * Deal with on-disk info
1040          */
1041         if (buffer->ondisk == NULL || buf_type) {
1042                 *errorp = hammer_load_buffer(buffer, buf_type);
1043                 if (*errorp) {
1044                         hammer_rel_buffer(buffer, 1);
1045                         buffer = NULL;
1046                 }
1047         } else {
1048                 *errorp = 0;
1049         }
1050         return(buffer);
1051 }
1052
1053 static int
1054 hammer_load_buffer(hammer_buffer_t buffer, u_int64_t buf_type)
1055 {
1056         hammer_volume_t volume;
1057         hammer_fsbuf_ondisk_t ondisk;
1058         int error;
1059
1060         /*
1061          * Load the buffer's on-disk info
1062          */
1063         volume = buffer->volume;
1064         hammer_lock_ex(&buffer->io.lock);
1065         if (buffer->ondisk == NULL) {
1066                 if (buf_type) {
1067                         error = hammer_io_new(volume->devvp, &buffer->io);
1068                 } else {
1069                         error = hammer_io_read(volume->devvp, &buffer->io);
1070                 }
1071                 if (error) {
1072                         hammer_unlock(&buffer->io.lock);
1073                         return (error);
1074                 }
1075                 buffer->ondisk = ondisk = (void *)buffer->io.bp->b_data;
1076                 buffer->alist.config = &Buf_alist_config;
1077                 buffer->alist.meta = ondisk->head.buf_almeta;
1078                 buffer->buf_type = ondisk->head.buf_type;
1079         } else if (buf_type) {
1080                 error = hammer_io_new(volume->devvp, &buffer->io);
1081         } else {
1082                 error = 0;
1083         }
1084         if (error == 0 && buf_type) {
1085                 ondisk = buffer->ondisk;
1086                 initbuffer(&buffer->alist, &ondisk->head, buf_type);
1087                 buffer->buf_type = ondisk->head.buf_type;
1088         }
1089         hammer_unlock(&buffer->io.lock);
1090         return (error);
1091 }
1092
1093 /*
1094  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
1095  */
1096 int
1097 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
1098 {
1099         hammer_ref(&buffer->io.lock);
1100         hammer_flush_buffer_nodes(buffer);
1101         KKASSERT(buffer->io.lock.refs == 1);
1102         hammer_rel_buffer(buffer, 1);
1103         return(0);
1104 }
1105
1106 /*
1107  * Reference a buffer that is either already referenced or via a specially
1108  * handled pointer (aka cursor->buffer).
1109  */
1110 int
1111 hammer_ref_buffer(hammer_buffer_t buffer)
1112 {
1113         int error;
1114
1115         hammer_ref(&buffer->io.lock);
1116         if (buffer->ondisk == NULL) {
1117                 error = hammer_load_buffer(buffer, 0);
1118                 if (error) {
1119                         hammer_rel_buffer(buffer, 1);
1120                         /*
1121                          * NOTE: buffer pointer can become stale after
1122                          * the above release.
1123                          */
1124                 } else {
1125                         KKASSERT(buffer->buf_type ==
1126                                  buffer->ondisk->head.buf_type);
1127                 }
1128         } else {
1129                 error = 0;
1130         }
1131         return(error);
1132 }
1133
1134 /*
1135  * Release a buffer.  We have to deal with several places where
1136  * another thread can ref the buffer.
1137  *
1138  * Only destroy the structure itself if the related buffer cache buffer
1139  * was disassociated from it.  This ties the management of the structure
1140  * to the buffer cache subsystem.  buffer->ondisk determines whether the
1141  * embedded io is referenced or not.
1142  */
1143 void
1144 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
1145 {
1146         hammer_cluster_t cluster;
1147         hammer_node_t node;
1148
1149         if (buffer->io.lock.refs == 1) {
1150                 hammer_lock_ex(&buffer->io.lock);
1151                 if (buffer->io.lock.refs == 1) {
1152                         hammer_io_release(&buffer->io, flush);
1153
1154                         /*
1155                          * Clean out the B-Tree node cache, if any, then
1156                          * clean up the cluster ref and free the buffer.
1157                          *
1158                          * If the buffer acquires a new reference while we
1159                          * are trying to clean it out, abort the cleaning.
1160                          */
1161                         while (buffer->io.bp == NULL &&
1162                                buffer->io.lock.refs == 1 &&
1163                                (node = TAILQ_FIRST(&buffer->clist)) != NULL
1164                         ) {
1165                                 KKASSERT(node->lock.refs == 0);
1166                                 hammer_flush_node(node);
1167                         }
1168                         if (buffer->io.bp == NULL &&
1169                             hammer_islastref(&buffer->io.lock)) {
1170                                 cluster = buffer->cluster;
1171                                 RB_REMOVE(hammer_buf_rb_tree,
1172                                           &cluster->rb_bufs_root, buffer);
1173                                 buffer->cluster = NULL; /* sanity */
1174                                 --hammer_count_buffers;
1175                                 kfree(buffer, M_HAMMER);
1176                                 hammer_rel_cluster(cluster, 0);
1177                                 return;
1178                         }
1179                 }
1180                 hammer_unlock(&buffer->io.lock);
1181         }
1182         hammer_unref(&buffer->io.lock);
1183 }
1184
1185 /*
1186  * Flush passively cached B-Tree nodes associated with this buffer.
1187  *
1188  * NOTE: The buffer is referenced and locked.
1189  */
1190 void
1191 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1192 {
1193         hammer_node_t node;
1194
1195         node = TAILQ_FIRST(&buffer->clist);
1196         while (node) {
1197                 buffer->save_scan = TAILQ_NEXT(node, entry);
1198                 if (node->lock.refs == 0) {
1199                         hammer_ref(&node->lock);
1200                         node->flags |= HAMMER_NODE_FLUSH;
1201                         hammer_rel_node(node);
1202                 }
1203                 node = buffer->save_scan;
1204         }
1205 }
1206
1207 /************************************************************************
1208  *                              NODES                                   *
1209  ************************************************************************
1210  *
1211  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
1212  * method used by the HAMMER filesystem.
1213  *
1214  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1215  * associated with its buffer.  It can have an active buffer reference
1216  * even when the node itself has no references.  The node also passively
1217  * associates itself with its cluster without holding any cluster refs.
1218  * The cluster ref is indirectly maintained by the active buffer ref when
1219  * a node is acquired.
1220  *
1221  * A hammer_node can also be passively associated with other HAMMER
1222  * structures, such as inodes, while retaining 0 references.  These
1223  * associations can be cleared backwards using a pointer-to-pointer in
1224  * the hammer_node.
1225  *
1226  * This allows the HAMMER implementation to cache hammer_node's long-term
1227  * and short-cut a great deal of the infrastructure's complexity.  In
1228  * most cases a cached node can be reacquired without having to dip into
1229  * either the buffer or cluster management code.
1230  *
1231  * The caller must pass a referenced cluster on call and will retain
1232  * ownership of the reference on return.  The node will acquire its own
1233  * additional references, if necessary.
1234  */
1235 hammer_node_t
1236 hammer_get_node(hammer_cluster_t cluster, int32_t node_offset, int *errorp)
1237 {
1238         hammer_node_t node;
1239
1240         /*
1241          * Locate the structure, allocating one if necessary.
1242          */
1243 again:
1244         node = RB_LOOKUP(hammer_nod_rb_tree, &cluster->rb_nods_root,
1245                          node_offset);
1246         if (node == NULL) {
1247                 ++hammer_count_nodes;
1248                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
1249                 node->node_offset = node_offset;
1250                 node->cluster = cluster;
1251                 if (RB_INSERT(hammer_nod_rb_tree, &cluster->rb_nods_root,
1252                               node)) {
1253                         --hammer_count_nodes;
1254                         kfree(node, M_HAMMER);
1255                         goto again;
1256                 }
1257         }
1258         *errorp = hammer_ref_node(node);
1259         if (*errorp) {
1260                 /*
1261                  * NOTE: The node pointer may be stale on error return.
1262                  * In fact, its probably been destroyed.
1263                  */
1264                 node = NULL;
1265         }
1266         return(node);
1267 }
1268
1269 /*
1270  * Reference the node to prevent disassociations, then associate and
1271  * load the related buffer.  This routine can also be called to reference
1272  * a node from a cache pointer.
1273  *
1274  * NOTE: Because the caller does not have a ref on the node, the caller's
1275  * node pointer will be stale if an error is returned.  We may also wind
1276  * up clearing the related cache pointers.
1277  *
1278  * NOTE: The cluster is indirectly referenced by our buffer ref.
1279  */
1280 int
1281 hammer_ref_node(hammer_node_t node)
1282 {
1283         hammer_buffer_t buffer;
1284         int32_t buf_no;
1285         int error;
1286
1287         hammer_ref(&node->lock);
1288         error = 0;
1289         if (node->ondisk == NULL) {
1290                 hammer_lock_ex(&node->lock);
1291                 if (node->ondisk == NULL) {
1292                         /*
1293                          * This is a little confusing but the jist is that
1294                          * node->buffer determines whether the node is on
1295                          * the buffer's clist and node->ondisk determines
1296                          * whether the buffer is referenced.
1297                          */
1298                         if ((buffer = node->buffer) != NULL) {
1299                                 error = hammer_ref_buffer(buffer);
1300                         } else {
1301                                 buf_no = node->node_offset / HAMMER_BUFSIZE;
1302                                 buffer = hammer_get_buffer(node->cluster,
1303                                                            buf_no, 0, &error);
1304                                 if (buffer) {
1305                                         KKASSERT(error == 0);
1306                                         TAILQ_INSERT_TAIL(&buffer->clist,
1307                                                           node, entry);
1308                                         node->buffer = buffer;
1309                                 }
1310                         }
1311                         if (error == 0) {
1312                                 node->ondisk = (void *)((char *)buffer->ondisk +
1313                                        (node->node_offset & HAMMER_BUFMASK));
1314                         }
1315                 }
1316                 hammer_unlock(&node->lock);
1317         }
1318         if (error)
1319                 hammer_rel_node(node);
1320         return (error);
1321 }
1322
1323 /*
1324  * Release a hammer_node.  The node retains a passive association with
1325  * its cluster, buffer and caches.
1326  *
1327  * However, to avoid cluttering up kernel memory with tons of B-Tree
1328  * node cache structures we destroy the node if no passive cache or
1329  * (instantiated) buffer references exist.
1330  */
1331 void
1332 hammer_rel_node(hammer_node_t node)
1333 {
1334         hammer_cluster_t cluster;
1335         hammer_buffer_t buffer;
1336
1337         if (hammer_islastref(&node->lock)) {
1338                 cluster = node->cluster;
1339
1340                 /*
1341                  * Destroy the node if it is being deleted.  Free the node
1342                  * in the bitmap after we have unhooked it.
1343                  */
1344                 if (node->flags & (HAMMER_NODE_DELETED|HAMMER_NODE_FLUSH)) {
1345                         hammer_flush_node(node);
1346                         RB_REMOVE(hammer_nod_rb_tree, &cluster->rb_nods_root,
1347                                   node);
1348                         hammer_ref_cluster(cluster);
1349                         if ((buffer = node->buffer) != NULL) {
1350                                 node->buffer = NULL;
1351                                 hammer_remove_node_clist(buffer, node);
1352                                 if (node->ondisk) {
1353                                         node->ondisk = NULL;
1354                                         hammer_rel_buffer(buffer, 0);
1355                                 }
1356                         }
1357                         if (node->flags & HAMMER_NODE_DELETED) {
1358                                 hammer_free_btree(node->cluster,
1359                                                   node->node_offset);
1360                                 if (node->node_offset ==
1361                                     cluster->ondisk->clu_btree_root) {
1362                                         kprintf("FREE CLUSTER %d\n", cluster->clu_no);
1363                                         hammer_free_cluster(cluster);
1364                                         /*hammer_io_undirty(&cluster->io);*/
1365                                 }
1366                         }
1367                         hammer_rel_cluster(cluster, 0);
1368                         --hammer_count_nodes;
1369                         kfree(node, M_HAMMER);
1370                         return;
1371                 }
1372
1373                 /*
1374                  * node->ondisk determines whether we have a buffer reference
1375                  * to get rid of or not.  Only get rid of the reference if
1376                  * the kernel tried to flush the buffer.
1377                  *
1378                  * NOTE: Once unref'd the node can be physically destroyed,
1379                  * so our node is stale afterwords.
1380                  *
1381                  * This case occurs if the node still has cache references.
1382                  * We could remove the references and free the structure
1383                  * but for now we allow them (and the node structure) to
1384                  * remain intact.
1385                  */
1386                 if (node->ondisk && hammer_io_checkflush(&node->buffer->io)) {
1387                         hammer_flush_node(node);
1388                         buffer = node->buffer;
1389                         node->buffer = NULL;
1390                         node->ondisk = NULL;
1391                         hammer_remove_node_clist(buffer, node);
1392                         hammer_rel_buffer(buffer, 0);
1393                 }
1394
1395                 /*
1396                  * Clutter control, this case only occurs after a failed
1397                  * load since otherwise ondisk will be non-NULL.
1398                  */
1399                 if (node->cache1 == NULL && node->cache2 == NULL &&
1400                     node->ondisk == NULL) {
1401                         RB_REMOVE(hammer_nod_rb_tree, &cluster->rb_nods_root,
1402                                   node);
1403                         if ((buffer = node->buffer) != NULL) {
1404                                 node->buffer = NULL; /* sanity */
1405                                 node->ondisk = NULL; /* sanity */
1406                                 hammer_remove_node_clist(buffer, node);
1407                         }
1408                         --hammer_count_nodes;
1409                         kfree(node, M_HAMMER);
1410                         return;
1411                 }
1412
1413                 hammer_unref(&node->lock);
1414         } else {
1415                 hammer_unref(&node->lock);
1416         }
1417 }
1418
1419 /*
1420  * Cache-and-release a hammer_node.  Kinda like catching and releasing a
1421  * fish, but keeping an eye on him.  The node is passively cached in *cache.
1422  *
1423  * NOTE!  HAMMER may NULL *cache at any time, even after you have
1424  * referenced the node!
1425  */
1426 void
1427 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
1428 {
1429         hammer_node_t old;
1430
1431         /*
1432          * If the node is being deleted, don't cache it!
1433          */
1434         if (node->flags & HAMMER_NODE_DELETED)
1435                 return;
1436
1437         /*
1438          * Cache the node.  If we previously cached a different node we
1439          * have to give HAMMER a chance to destroy it.
1440          */
1441 again:
1442         if (node->cache1 != cache) {
1443                 if (node->cache2 == cache) {
1444                         struct hammer_node **tmp;
1445                         tmp = node->cache1;
1446                         node->cache1 = node->cache2;
1447                         node->cache2 = tmp;
1448                 } else {
1449                         if ((old = *cache) != NULL) {
1450                                 *cache = NULL;
1451                                 hammer_flush_node(old); /* can block */
1452                                 goto again;
1453                         }
1454                         if (node->cache2)
1455                                 *node->cache2 = NULL;
1456                         node->cache2 = node->cache1;
1457                         node->cache1 = cache;
1458                         *cache = node;
1459                 }
1460         }
1461 }
1462
1463 void
1464 hammer_uncache_node(struct hammer_node **cache)
1465 {
1466         hammer_node_t node;
1467
1468         if ((node = *cache) != NULL) {
1469                 *cache = NULL;
1470                 if (node->cache1 == cache) {
1471                         node->cache1 = node->cache2;
1472                         node->cache2 = NULL;
1473                 } else if (node->cache2 == cache) {
1474                         node->cache2 = NULL;
1475                 } else {
1476                         panic("hammer_uncache_node: missing cache linkage");
1477                 }
1478                 if (node->cache1 == NULL && node->cache2 == NULL)
1479                         hammer_flush_node(node);
1480         }
1481 }
1482
1483 /*
1484  * Remove a node's cache references and destroy the node if it has no
1485  * other references or backing store.
1486  */
1487 void
1488 hammer_flush_node(hammer_node_t node)
1489 {
1490         hammer_buffer_t buffer;
1491
1492         if (node->cache1)
1493                 *node->cache1 = NULL;
1494         if (node->cache2)
1495                 *node->cache2 = NULL;
1496         if (node->lock.refs == 0 && node->ondisk == NULL) {
1497                 RB_REMOVE(hammer_nod_rb_tree, &node->cluster->rb_nods_root,
1498                           node);
1499                 if ((buffer = node->buffer) != NULL) {
1500                         node->buffer = NULL;
1501                         hammer_remove_node_clist(buffer, node);
1502                         /* buffer is unreferenced because ondisk is NULL */
1503                 }
1504                 --hammer_count_nodes;
1505                 kfree(node, M_HAMMER);
1506         }
1507 }
1508
1509 /*
1510  * Remove a node from the buffer's clist.  Adjust save_scan as appropriate.
1511  * This is in its own little routine to properly handle interactions with
1512  * save_scan, so it is possible to block while scanning a buffer's node list.
1513  */
1514 static
1515 void
1516 hammer_remove_node_clist(hammer_buffer_t buffer, hammer_node_t node)
1517 {
1518         if (buffer->save_scan == node)
1519                 buffer->save_scan = TAILQ_NEXT(node, entry);
1520         TAILQ_REMOVE(&buffer->clist, node, entry);
1521 }
1522
1523 /************************************************************************
1524  *                              A-LIST ALLOCATORS                       *
1525  ************************************************************************/
1526
1527 /*
1528  * Allocate HAMMER clusters
1529  */
1530 hammer_cluster_t
1531 hammer_alloc_cluster(hammer_mount_t hmp, hammer_cluster_t cluster_hint,
1532                      int *errorp)
1533 {
1534         hammer_volume_t volume;
1535         hammer_cluster_t cluster;
1536         int32_t clu_no;
1537         int32_t clu_hint;
1538         int32_t vol_beg;
1539         int32_t vol_no;
1540
1541         /*
1542          * Figure out our starting volume and hint.
1543          */
1544         if (cluster_hint) {
1545                 vol_beg = cluster_hint->volume->vol_no;
1546                 clu_hint = cluster_hint->clu_no;
1547         } else {
1548                 vol_beg = hmp->volume_iterator;
1549                 clu_hint = -1;
1550         }
1551
1552         /*
1553          * Loop through volumes looking for a free cluster.  If allocating
1554          * a new cluster relative to an existing cluster try to find a free
1555          * cluster on either side (clu_hint >= 0), otherwise just do a
1556          * forwards iteration.
1557          */
1558         vol_no = vol_beg;
1559         do {
1560                 volume = hammer_get_volume(hmp, vol_no, errorp);
1561                 kprintf("VOLUME %p %d\n", volume, vol_no);
1562                 if (*errorp) {
1563                         clu_no = HAMMER_ALIST_BLOCK_NONE;
1564                         break;
1565                 }
1566                 hammer_modify_volume(volume);
1567                 if (clu_hint == -1) {
1568                         clu_hint = volume->clu_iterator;
1569                         clu_no = hammer_alist_alloc_fwd(&volume->alist, 1,
1570                                                         clu_hint);
1571                         if (clu_no == HAMMER_ALIST_BLOCK_NONE) {
1572                                 clu_no = hammer_alist_alloc_fwd(&volume->alist,
1573                                                                 1, 0);
1574                         }
1575                 } else {
1576                         clu_no = hammer_alist_alloc_fwd(&volume->alist, 1,
1577                                                         clu_hint);
1578                         if (clu_no == HAMMER_ALIST_BLOCK_NONE) {
1579                                 clu_no = hammer_alist_alloc_rev(&volume->alist,
1580                                                                 1, clu_hint);
1581                         }
1582                 }
1583                 hammer_modify_volume_done(volume);
1584                 if (clu_no != HAMMER_ALIST_BLOCK_NONE)
1585                         break;
1586                 hammer_rel_volume(volume, 0);
1587                 volume = NULL;
1588                 *errorp = ENOSPC;
1589                 vol_no = (vol_no + 1) % hmp->nvolumes;
1590                 clu_hint = -1;
1591         } while (vol_no != vol_beg);
1592
1593         /*
1594          * Acquire the cluster.  On success this will force *errorp to 0.
1595          */
1596         if (clu_no != HAMMER_ALIST_BLOCK_NONE) {
1597                 kprintf("ALLOC CLUSTER %d\n", clu_no);
1598                 cluster = hammer_get_cluster(volume, clu_no, errorp, 1);
1599                 volume->clu_iterator = clu_no;
1600                 hammer_rel_volume(volume, 0);
1601         } else {
1602                 cluster = NULL;
1603         }
1604         if (cluster)
1605                 hammer_lock_ex(&cluster->io.lock);
1606         return(cluster);
1607 }
1608
1609 void
1610 hammer_init_cluster(hammer_cluster_t cluster, hammer_base_elm_t left_bound,
1611                     hammer_base_elm_t right_bound)
1612 {
1613         hammer_cluster_ondisk_t ondisk = cluster->ondisk;
1614
1615         hammer_modify_cluster(cluster);
1616         ondisk->clu_btree_beg = *left_bound;
1617         ondisk->clu_btree_end = *right_bound;
1618         cluster->clu_btree_beg = ondisk->clu_btree_beg;
1619         cluster->clu_btree_end = ondisk->clu_btree_end;
1620         hammer_modify_cluster_done(cluster);
1621 }
1622
1623 /*
1624  * Deallocate a cluster
1625  */
1626 void
1627 hammer_free_cluster(hammer_cluster_t cluster)
1628 {
1629         hammer_modify_cluster(cluster);
1630         hammer_alist_free(&cluster->volume->alist, cluster->clu_no, 1);
1631         hammer_modify_cluster_done(cluster);
1632 }
1633
1634 /*
1635  * Allocate HAMMER elements - btree nodes, data storage, and record elements
1636  *
1637  * The passed *bufferp should be initialized to NULL.  On successive calls
1638  * *bufferp caches the most recent buffer used until put away by the caller.
1639  * Note that previously returned pointers using the cached buffer become
1640  * invalid on successive calls which reuse *bufferp.
1641  *
1642  * All allocations first attempt to use the block found at the specified
1643  * iterator.  If that fails the first available block is used.  If that
1644  * fails a new buffer is allocated and associated with the buffer type
1645  * A-list and the element is allocated out of the new buffer.
1646  */
1647
1648 hammer_node_t
1649 hammer_alloc_btree(hammer_cluster_t cluster, int *errorp)
1650 {
1651         hammer_buffer_t buffer;
1652         hammer_alist_t live;
1653         hammer_node_t node;
1654         int32_t elm_no;
1655         int32_t buf_no;
1656         int32_t node_offset;
1657
1658         /*
1659          * Allocate a B-Tree element
1660          */
1661         hammer_modify_cluster(cluster);
1662         buffer = NULL;
1663         live = &cluster->alist_btree;
1664         elm_no = hammer_alist_alloc_fwd(live, 1, cluster->ondisk->idx_index);
1665         if (elm_no == HAMMER_ALIST_BLOCK_NONE)
1666                 elm_no = hammer_alist_alloc_fwd(live, 1, 0);
1667         if (elm_no == HAMMER_ALIST_BLOCK_NONE) {
1668                 alloc_new_buffer(cluster, live,
1669                                  HAMMER_FSBUF_BTREE, HAMMER_BTREE_NODES,
1670                                  cluster->ondisk->idx_index, errorp, &buffer);
1671                 elm_no = hammer_alist_alloc(live, 1);
1672                 if (elm_no == HAMMER_ALIST_BLOCK_NONE) {
1673                         *errorp = ENOSPC;
1674                         if (buffer)
1675                                 hammer_rel_buffer(buffer, 0);
1676                         hammer_modify_cluster_done(cluster);
1677                         return(NULL);
1678                 }
1679         }
1680         cluster->ondisk->idx_index = elm_no;
1681         KKASSERT((elm_no & HAMMER_FSBUF_BLKMASK) < HAMMER_BTREE_NODES);
1682
1683         /*
1684          * Load and return the B-Tree element
1685          */
1686         buf_no = elm_no / HAMMER_FSBUF_MAXBLKS;
1687         node_offset = buf_no * HAMMER_BUFSIZE +
1688                       offsetof(union hammer_fsbuf_ondisk,
1689                                btree.nodes[elm_no & HAMMER_FSBUF_BLKMASK]);
1690         node = hammer_get_node(cluster, node_offset, errorp);
1691         if (node) {
1692                 hammer_modify_node(node);
1693                 bzero(node->ondisk, sizeof(*node->ondisk));
1694                 hammer_modify_node_done(node);
1695         } else {
1696                 hammer_alist_free(live, elm_no, 1);
1697                 hammer_rel_node(node);
1698                 node = NULL;
1699         }
1700         hammer_modify_cluster_done(cluster);
1701         if (buffer)
1702                 hammer_rel_buffer(buffer, 0);
1703         return(node);
1704 }
1705
1706 void *
1707 hammer_alloc_data(hammer_cluster_t cluster, int32_t bytes,
1708                   int *errorp, struct hammer_buffer **bufferp)
1709 {
1710         hammer_buffer_t buffer;
1711         hammer_alist_t live;
1712         int32_t elm_no;
1713         int32_t buf_no;
1714         int32_t nblks;
1715         void *item;
1716
1717         /*
1718          * Deal with large data blocks.  The blocksize is HAMMER_BUFSIZE
1719          * for these allocations.
1720          */
1721         hammer_modify_cluster(cluster);
1722         if ((bytes & HAMMER_BUFMASK) == 0) {
1723                 nblks = bytes / HAMMER_BUFSIZE;
1724                 /* only one block allowed for now (so buffer can hold it) */
1725                 KKASSERT(nblks == 1);
1726
1727                 buf_no = hammer_alloc_master(cluster, nblks,
1728                                              cluster->ondisk->idx_ldata, 1);
1729                 if (buf_no == HAMMER_ALIST_BLOCK_NONE) {
1730                         *errorp = ENOSPC;
1731                         hammer_modify_cluster_done(cluster);
1732                         return(NULL);
1733                 }
1734                 hammer_adjust_stats(cluster, HAMMER_FSBUF_DATA, nblks);
1735                 cluster->ondisk->idx_ldata = buf_no;
1736                 hammer_modify_cluster_done(cluster);
1737                 buffer = *bufferp;
1738                 *bufferp = hammer_get_buffer(cluster, buf_no, -1, errorp);
1739                 if (buffer)
1740                         hammer_rel_buffer(buffer, 0);
1741                 buffer = *bufferp;
1742                 return(buffer->ondisk);
1743         }
1744
1745         /*
1746          * Allocate a data element.  The block size is HAMMER_DATA_BLKSIZE
1747          * (64 bytes) for these allocations.
1748          */
1749         nblks = (bytes + HAMMER_DATA_BLKMASK) & ~HAMMER_DATA_BLKMASK;
1750         nblks /= HAMMER_DATA_BLKSIZE;
1751         live = &cluster->alist_mdata;
1752         elm_no = hammer_alist_alloc_fwd(live, nblks, cluster->ondisk->idx_data);
1753         if (elm_no == HAMMER_ALIST_BLOCK_NONE)
1754                 elm_no = hammer_alist_alloc_fwd(live, 1, 0);
1755         if (elm_no == HAMMER_ALIST_BLOCK_NONE) {
1756                 alloc_new_buffer(cluster, live,
1757                                  HAMMER_FSBUF_DATA, HAMMER_DATA_NODES,
1758                                  cluster->ondisk->idx_data, errorp, bufferp);
1759                 elm_no = hammer_alist_alloc(live, nblks);
1760                 if (elm_no == HAMMER_ALIST_BLOCK_NONE) {
1761                         *errorp = ENOSPC;
1762                         hammer_modify_cluster_done(cluster);
1763                         return(NULL);
1764                 }
1765         }
1766         cluster->ondisk->idx_index = elm_no;
1767         hammer_modify_cluster_done(cluster);
1768
1769         /*
1770          * Load and return the B-Tree element
1771          */
1772         buf_no = elm_no / HAMMER_FSBUF_MAXBLKS;
1773         buffer = *bufferp;
1774         if (buffer == NULL || buffer->cluster != cluster ||
1775             buffer->buf_no != buf_no) {
1776                 if (buffer)
1777                         hammer_rel_buffer(buffer, 0);
1778                 buffer = hammer_get_buffer(cluster, buf_no, 0, errorp);
1779                 *bufferp = buffer;
1780         }
1781         KKASSERT(buffer->ondisk->head.buf_type == HAMMER_FSBUF_DATA);
1782         KKASSERT((elm_no & HAMMER_FSBUF_BLKMASK) < HAMMER_DATA_NODES);
1783         hammer_modify_buffer(buffer);
1784         item = &buffer->ondisk->data.data[elm_no & HAMMER_FSBUF_BLKMASK];
1785         bzero(item, nblks * HAMMER_DATA_BLKSIZE);
1786         hammer_modify_buffer_done(buffer);
1787         *errorp = 0;
1788         return(item);
1789 }
1790
1791 void *
1792 hammer_alloc_record(hammer_cluster_t cluster,
1793                     int *errorp, struct hammer_buffer **bufferp)
1794 {
1795         hammer_buffer_t buffer;
1796         hammer_alist_t live;
1797         int32_t elm_no;
1798         int32_t buf_no;
1799         void *item;
1800
1801         /*
1802          * Allocate a record element
1803          */
1804         hammer_modify_cluster(cluster);
1805         live = &cluster->alist_record;
1806         elm_no = hammer_alist_alloc_rev(live, 1, cluster->ondisk->idx_record);
1807         if (elm_no == HAMMER_ALIST_BLOCK_NONE)
1808                 elm_no = hammer_alist_alloc_rev(live, 1,HAMMER_ALIST_BLOCK_MAX);
1809         if (elm_no == HAMMER_ALIST_BLOCK_NONE) {
1810                 alloc_new_buffer(cluster, live,
1811                                  HAMMER_FSBUF_RECORDS, HAMMER_RECORD_NODES,
1812                                  cluster->ondisk->idx_record, errorp, bufferp);
1813                 elm_no = hammer_alist_alloc_rev(live, 1,HAMMER_ALIST_BLOCK_MAX);
1814                 kprintf("hammer_alloc_record elm again %08x\n", elm_no);
1815                 if (elm_no == HAMMER_ALIST_BLOCK_NONE) {
1816                         *errorp = ENOSPC;
1817                         hammer_modify_cluster_done(cluster);
1818                         return(NULL);
1819                 }
1820         }
1821         cluster->ondisk->idx_record = elm_no;
1822         hammer_modify_cluster_done(cluster);
1823
1824         /*
1825          * Load and return the record element
1826          */
1827         buf_no = elm_no / HAMMER_FSBUF_MAXBLKS;
1828         buffer = *bufferp;
1829         if (buffer == NULL || buffer->cluster != cluster ||
1830             buffer->buf_no != buf_no) {
1831                 if (buffer)
1832                         hammer_rel_buffer(buffer, 0);
1833                 buffer = hammer_get_buffer(cluster, buf_no, 0, errorp);
1834                 *bufferp = buffer;
1835         }
1836         KKASSERT(buffer->ondisk->head.buf_type == HAMMER_FSBUF_RECORDS);
1837         KKASSERT((elm_no & HAMMER_FSBUF_BLKMASK) < HAMMER_RECORD_NODES);
1838         hammer_modify_buffer(buffer);
1839         item = &buffer->ondisk->record.recs[elm_no & HAMMER_FSBUF_BLKMASK];
1840         bzero(item, sizeof(union hammer_record_ondisk));
1841         hammer_modify_buffer_done(buffer);
1842         *errorp = 0;
1843         return(item);
1844 }
1845
1846 void
1847 hammer_free_data_ptr(hammer_buffer_t buffer, void *data, int bytes)
1848 {
1849         int32_t elm_no;
1850         int32_t nblks;
1851         hammer_alist_t live;
1852
1853         hammer_modify_cluster(buffer->cluster);
1854         if ((bytes & HAMMER_BUFMASK) == 0) {
1855                 nblks = bytes / HAMMER_BUFSIZE;
1856                 KKASSERT(nblks == 1 && data == (void *)buffer->ondisk);
1857                 hammer_alist_free(&buffer->cluster->alist_master,
1858                                   buffer->buf_no, nblks);
1859                 hammer_adjust_stats(buffer->cluster, HAMMER_FSBUF_DATA, -nblks);
1860                 hammer_modify_cluster_done(buffer->cluster);
1861                 return;
1862         }
1863
1864         elm_no = ((char *)data - (char *)buffer->ondisk->data.data) /
1865                  HAMMER_DATA_BLKSIZE;
1866         KKASSERT(elm_no >= 0 && elm_no < HAMMER_DATA_NODES);
1867         elm_no += buffer->buf_no * HAMMER_FSBUF_MAXBLKS;
1868         nblks = (bytes + HAMMER_DATA_BLKMASK) & ~HAMMER_DATA_BLKMASK;
1869         nblks /= HAMMER_DATA_BLKSIZE;
1870         live = &buffer->cluster->alist_mdata;
1871         hammer_alist_free(live, elm_no, nblks);
1872         hammer_modify_cluster_done(buffer->cluster);
1873 }
1874
1875 void
1876 hammer_free_record_ptr(hammer_buffer_t buffer, union hammer_record_ondisk *rec)
1877 {
1878         int32_t elm_no;
1879         hammer_alist_t live;
1880
1881         hammer_modify_cluster(buffer->cluster);
1882         elm_no = rec - &buffer->ondisk->record.recs[0];
1883         KKASSERT(elm_no >= 0 && elm_no < HAMMER_BTREE_NODES);
1884         elm_no += buffer->buf_no * HAMMER_FSBUF_MAXBLKS;
1885         live = &buffer->cluster->alist_record;
1886         hammer_alist_free(live, elm_no, 1);
1887         hammer_modify_cluster_done(buffer->cluster);
1888 }
1889
1890 void
1891 hammer_free_btree(hammer_cluster_t cluster, int32_t bclu_offset)
1892 {
1893         const int32_t blksize = sizeof(struct hammer_node_ondisk);
1894         int32_t fsbuf_offset = bclu_offset & HAMMER_BUFMASK;
1895         hammer_alist_t live;
1896         int32_t elm_no;
1897
1898         hammer_modify_cluster(cluster);
1899         elm_no = bclu_offset / HAMMER_BUFSIZE * HAMMER_FSBUF_MAXBLKS;
1900         fsbuf_offset -= offsetof(union hammer_fsbuf_ondisk, btree.nodes[0]);
1901         live = &cluster->alist_btree;
1902         KKASSERT(fsbuf_offset >= 0 && fsbuf_offset % blksize == 0);
1903         elm_no += fsbuf_offset / blksize;
1904         hammer_alist_free(live, elm_no, 1);
1905         hammer_modify_cluster_done(cluster);
1906 }
1907
1908 void
1909 hammer_free_data(hammer_cluster_t cluster, int32_t bclu_offset, int32_t bytes)
1910 {
1911         const int32_t blksize = HAMMER_DATA_BLKSIZE;
1912         int32_t fsbuf_offset = bclu_offset & HAMMER_BUFMASK;
1913         hammer_alist_t live;
1914         int32_t elm_no;
1915         int32_t buf_no;
1916         int32_t nblks;
1917
1918         hammer_modify_cluster(cluster);
1919         if ((bytes & HAMMER_BUFMASK) == 0) {
1920                 nblks = bytes / HAMMER_BUFSIZE;
1921                 KKASSERT(nblks == 1 && (bclu_offset & HAMMER_BUFMASK) == 0);
1922                 buf_no = bclu_offset / HAMMER_BUFSIZE;
1923                 hammer_alist_free(&cluster->alist_master, buf_no, nblks);
1924                 hammer_adjust_stats(cluster, HAMMER_FSBUF_DATA, -nblks);
1925                 hammer_modify_cluster_done(cluster);
1926                 return;
1927         }
1928
1929         elm_no = bclu_offset / HAMMER_BUFSIZE * HAMMER_FSBUF_MAXBLKS;
1930         fsbuf_offset -= offsetof(union hammer_fsbuf_ondisk, data.data[0][0]);
1931         live = &cluster->alist_mdata;
1932         nblks = (bytes + HAMMER_DATA_BLKMASK) & ~HAMMER_DATA_BLKMASK;
1933         nblks /= HAMMER_DATA_BLKSIZE;
1934         KKASSERT(fsbuf_offset >= 0 && fsbuf_offset % blksize == 0);
1935         elm_no += fsbuf_offset / blksize;
1936         hammer_alist_free(live, elm_no, nblks);
1937         hammer_modify_cluster_done(cluster);
1938 }
1939
1940 void
1941 hammer_free_record(hammer_cluster_t cluster, int32_t bclu_offset)
1942 {
1943         const int32_t blksize = sizeof(union hammer_record_ondisk);
1944         int32_t fsbuf_offset = bclu_offset & HAMMER_BUFMASK;
1945         hammer_alist_t live;
1946         int32_t elm_no;
1947
1948         hammer_modify_cluster(cluster);
1949         elm_no = bclu_offset / HAMMER_BUFSIZE * HAMMER_FSBUF_MAXBLKS;
1950         fsbuf_offset -= offsetof(union hammer_fsbuf_ondisk, record.recs[0]);
1951         live = &cluster->alist_record;
1952         KKASSERT(fsbuf_offset >= 0 && fsbuf_offset % blksize == 0);
1953         elm_no += fsbuf_offset / blksize;
1954         hammer_alist_free(live, elm_no, 1);
1955         hammer_modify_cluster_done(cluster);
1956 }
1957
1958
1959 /*
1960  * Allocate a new filesystem buffer and assign it to the specified
1961  * filesystem buffer type.  The new buffer will be added to the
1962  * type-specific A-list and initialized.
1963  */
1964 static void
1965 alloc_new_buffer(hammer_cluster_t cluster, hammer_alist_t live,
1966                  u_int64_t type, int32_t nelements,
1967                  int start, int *errorp, struct hammer_buffer **bufferp)
1968 {
1969         hammer_buffer_t buffer;
1970         int32_t buf_no;
1971         int isfwd;
1972
1973         if (*bufferp)
1974                 hammer_rel_buffer(*bufferp, 0);
1975         *bufferp = NULL;
1976
1977         start = start / HAMMER_FSBUF_MAXBLKS;   /* convert to buf_no */
1978         isfwd = (type != HAMMER_FSBUF_RECORDS);
1979         buf_no = hammer_alloc_master(cluster, 1, start, isfwd);
1980         if (buf_no == HAMMER_ALIST_BLOCK_NONE) {
1981                 *errorp = ENOSPC;
1982                 return;
1983         }
1984
1985         /*
1986          * The new buffer must be initialized (type != 0) regardless of
1987          * whether we already have it cached or not, so don't try to
1988          * optimize the cached buffer check.  Just call hammer_get_buffer().
1989          */
1990         buffer = hammer_get_buffer(cluster, buf_no, type, errorp);
1991         *bufferp = buffer;
1992
1993         /*
1994          * Finally, do a meta-free of the buffer's elements into the
1995          * type-specific A-list and update our statistics to reflect
1996          * the allocation.
1997          */
1998         if (buffer) {
1999                 kprintf("alloc_new_buffer buf_no %d type %016llx nelms %d\n",
2000                         buf_no, type, nelements);
2001                 hammer_modify_buffer(buffer);  /*XXX*/
2002                 hammer_alist_free(live, buf_no * HAMMER_FSBUF_MAXBLKS,
2003                                   nelements);
2004                 hammer_modify_buffer_done(buffer);  /*XXX*/
2005                 hammer_adjust_stats(cluster, type, 1);
2006         }
2007 }
2008
2009 /*
2010  * Sync dirty buffers to the media
2011  */
2012
2013 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
2014 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
2015
2016 int
2017 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
2018 {
2019         struct hammer_sync_info info;
2020
2021         info.error = 0;
2022         info.waitfor = waitfor;
2023
2024         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
2025                       hammer_sync_scan1, hammer_sync_scan2, &info);
2026
2027         RB_SCAN(hammer_vol_rb_tree, &hmp->rb_vols_root, NULL,
2028                 hammer_sync_volume, &info);
2029         return(info.error);
2030 }
2031
2032 static int
2033 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
2034 {
2035         struct hammer_inode *ip;
2036
2037         ip = VTOI(vp);
2038         if (vp->v_type == VNON || ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2039             RB_EMPTY(&vp->v_rbdirty_tree))) {
2040                 return(-1);
2041         }
2042         return(0);
2043 }
2044
2045 static int
2046 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2047 {
2048         struct hammer_sync_info *info = data;
2049         struct hammer_inode *ip;
2050         int error;
2051
2052         ip = VTOI(vp);
2053         if (vp->v_type == VNON || vp->v_type == VBAD ||
2054             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2055              RB_EMPTY(&vp->v_rbdirty_tree))) {
2056                 return(0);
2057         }
2058         if (vp->v_type != VCHR) {
2059                 error = VOP_FSYNC(vp, info->waitfor);
2060                 if (error)
2061                         info->error = error;
2062         }
2063         return(0);
2064 }
2065
2066 int
2067 hammer_sync_volume(hammer_volume_t volume, void *data)
2068 {
2069         struct hammer_sync_info *info = data;
2070
2071         RB_SCAN(hammer_clu_rb_tree, &volume->rb_clus_root, NULL,
2072                 hammer_sync_cluster, info);
2073         if (hammer_ref_volume(volume) == 0) {
2074                 hammer_io_flush(&volume->io, info);
2075                 hammer_rel_volume(volume, 0);
2076         }
2077         return(0);
2078 }
2079
2080 int
2081 hammer_sync_cluster(hammer_cluster_t cluster, void *data)
2082 {
2083         struct hammer_sync_info *info = data;
2084
2085         RB_SCAN(hammer_buf_rb_tree, &cluster->rb_bufs_root, NULL,
2086                 hammer_sync_buffer, info);
2087         switch(cluster->state) {
2088         case HAMMER_CLUSTER_OPEN:
2089         case HAMMER_CLUSTER_IDLE:
2090                 if (hammer_ref_cluster(cluster) == 0) {
2091                         hammer_io_flush(&cluster->io, info);
2092                         hammer_rel_cluster(cluster, 0);
2093                 }
2094                 break;
2095         default:
2096                 break;
2097         }
2098         return(0);
2099 }
2100
2101 int
2102 hammer_sync_buffer(hammer_buffer_t buffer, void *data)
2103 {
2104         struct hammer_sync_info *info = data;
2105
2106         if (hammer_ref_buffer(buffer) == 0) {
2107                 hammer_lock_ex(&buffer->io.lock);
2108                 hammer_flush_buffer_nodes(buffer);
2109                 hammer_unlock(&buffer->io.lock);
2110                 hammer_io_flush(&buffer->io, info);
2111                 hammer_rel_buffer(buffer, 0);
2112         }
2113         return(0);
2114 }
2115
2116 /*
2117  * Generic buffer initialization
2118  */
2119 static void
2120 initbuffer(hammer_alist_t live, hammer_fsbuf_head_t head, u_int64_t type)
2121 {
2122         head->buf_type = type;
2123         hammer_alist_init(live);
2124 }
2125
2126 /*
2127  * Calculate the cluster's offset in the volume.  This calculation is
2128  * slightly more complex when using superclusters because superclusters
2129  * are grouped in blocks of 16, followed by 16 x N clusters where N
2130  * is the number of clusters a supercluster can manage.
2131  */
2132 static int64_t
2133 calculate_cluster_offset(hammer_volume_t volume, int32_t clu_no)
2134 {
2135         int32_t scl_group;
2136         int64_t scl_group_size;
2137         int64_t off;
2138
2139         if (volume->vol_flags & HAMMER_VOLF_USINGSUPERCL) {
2140                 scl_group = clu_no / HAMMER_VOL_SUPERCLUSTER_GROUP /
2141                             HAMMER_SCL_MAXCLUSTERS;
2142                 scl_group_size =
2143                             ((int64_t)HAMMER_BUFSIZE *
2144                              HAMMER_VOL_SUPERCLUSTER_GROUP) +
2145                             ((int64_t)HAMMER_VOL_SUPERCLUSTER_GROUP *
2146                              volume->vol_clsize * HAMMER_SCL_MAXCLUSTERS);
2147                 scl_group_size +=
2148                             HAMMER_VOL_SUPERCLUSTER_GROUP * HAMMER_BUFSIZE;
2149
2150                 off = volume->cluster_base +
2151                       scl_group * scl_group_size +
2152                       (HAMMER_BUFSIZE * HAMMER_VOL_SUPERCLUSTER_GROUP) +
2153                       ((int64_t)clu_no % ((int64_t)HAMMER_SCL_MAXCLUSTERS *
2154                        HAMMER_VOL_SUPERCLUSTER_GROUP))
2155                       * volume->vol_clsize;
2156         } else {
2157                 off = volume->cluster_base +
2158                       (int64_t)clu_no * volume->vol_clsize;
2159         }
2160         return(off);
2161 }
2162
2163 /*
2164  * Calculate a super-cluster's offset in the volume.
2165  */
2166 static int64_t
2167 calculate_supercl_offset(hammer_volume_t volume, int32_t scl_no)
2168 {
2169         int64_t off;
2170         int32_t scl_group;
2171         int64_t scl_group_size;
2172
2173         KKASSERT (volume->vol_flags & HAMMER_VOLF_USINGSUPERCL);
2174         scl_group = scl_no / HAMMER_VOL_SUPERCLUSTER_GROUP;
2175         if (scl_group) {
2176                 scl_group_size =
2177                             ((int64_t)HAMMER_BUFSIZE *
2178                              HAMMER_VOL_SUPERCLUSTER_GROUP) +
2179                             ((int64_t)HAMMER_VOL_SUPERCLUSTER_GROUP *
2180                              volume->vol_clsize * HAMMER_SCL_MAXCLUSTERS);
2181                 scl_group_size +=
2182                             HAMMER_VOL_SUPERCLUSTER_GROUP * HAMMER_BUFSIZE;
2183                 off = volume->cluster_base + (scl_group * scl_group_size) +
2184                       (scl_no % HAMMER_VOL_SUPERCLUSTER_GROUP) * HAMMER_BUFSIZE;
2185         } else {
2186                 off = volume->cluster_base + (scl_no * HAMMER_BUFSIZE);
2187         }
2188         return(off);
2189 }
2190
2191 /*
2192  *
2193  *
2194  */
2195 static int32_t
2196 hammer_alloc_master(hammer_cluster_t cluster, int nblks,
2197                     int32_t start, int isfwd)
2198 {
2199         int32_t buf_no;
2200
2201         hammer_modify_cluster(cluster);
2202         if (isfwd) {
2203                 buf_no = hammer_alist_alloc_fwd(&cluster->alist_master,
2204                                                 nblks, start);
2205                 if (buf_no == HAMMER_ALIST_BLOCK_NONE) {
2206                         buf_no = hammer_alist_alloc_fwd(&cluster->alist_master,
2207                                                 nblks, 0);
2208                 }
2209         } else {
2210                 buf_no = hammer_alist_alloc_rev(&cluster->alist_master,
2211                                                 nblks, start);
2212                 if (buf_no == HAMMER_ALIST_BLOCK_NONE) {
2213                         buf_no = hammer_alist_alloc_rev(&cluster->alist_master,
2214                                                 nblks, HAMMER_ALIST_BLOCK_MAX);
2215                 }
2216         }
2217         hammer_modify_cluster_done(cluster);
2218
2219         /*
2220          * Recover space from empty record, b-tree, and data a-lists.
2221          */
2222
2223         return(buf_no);
2224 }
2225
2226 /*
2227  * Adjust allocation statistics
2228  */
2229 static void
2230 hammer_adjust_stats(hammer_cluster_t cluster, u_int64_t buf_type, int nblks)
2231 {
2232         hammer_modify_cluster(cluster);
2233         hammer_modify_volume(cluster->volume);
2234         hammer_modify_volume(cluster->volume->hmp->rootvol);
2235
2236         switch(buf_type) {
2237         case HAMMER_FSBUF_BTREE:
2238                 cluster->ondisk->stat_idx_bufs += nblks;
2239                 cluster->volume->ondisk->vol_stat_idx_bufs += nblks;
2240                 cluster->volume->hmp->rootvol->ondisk->vol0_stat_idx_bufs += nblks;
2241                 break;
2242         case HAMMER_FSBUF_DATA:
2243                 cluster->ondisk->stat_data_bufs += nblks;
2244                 cluster->volume->ondisk->vol_stat_data_bufs += nblks;
2245                 cluster->volume->hmp->rootvol->ondisk->vol0_stat_data_bufs += nblks;
2246                 break;
2247         case HAMMER_FSBUF_RECORDS:
2248                 cluster->ondisk->stat_rec_bufs += nblks;
2249                 cluster->volume->ondisk->vol_stat_rec_bufs += nblks;
2250                 cluster->volume->hmp->rootvol->ondisk->vol0_stat_rec_bufs += nblks;
2251                 break;
2252         }
2253         hammer_modify_cluster_done(cluster);
2254         hammer_modify_volume_done(cluster->volume);
2255         hammer_modify_volume_done(cluster->volume->hmp->rootvol);
2256 }
2257
2258 /*
2259  * A-LIST SUPPORT
2260  *
2261  * Setup the parameters for the various A-lists we use in hammer.  The
2262  * supercluster A-list must be chained to the cluster A-list and cluster
2263  * slave A-lists are chained to buffer A-lists.
2264  *
2265  * See hammer_init_alist_config() below.
2266  */
2267
2268 /*
2269  * A-LIST - cluster recursion into a filesystem buffer
2270  */
2271 static int
2272 buffer_alist_init(void *info, int32_t blk, int32_t radix)
2273 {
2274         return(0);
2275 #if 0
2276         hammer_cluster_t cluster = info;
2277         hammer_buffer_t buffer;
2278         int32_t buf_no;
2279         int error = 0;
2280
2281         /*
2282          * Calculate the buffer number, initialize based on the buffer type.
2283          * The buffer has already been allocated so assert that it has been
2284          * initialized.
2285          */
2286         buf_no = blk / HAMMER_FSBUF_MAXBLKS;
2287         buffer = hammer_get_buffer(cluster, buf_no, 0, &error);
2288         if (buffer) {
2289                 hammer_adjust_stats(cluster, buffer->ondisk->head.buf_type, 1);
2290                 hammer_rel_buffer(buffer, 0);
2291         }
2292         return (error);
2293 #endif
2294 }
2295
2296 static int
2297 buffer_alist_destroy(void *info, int32_t blk, int32_t radix)
2298 {
2299         return(0);
2300 #if 0
2301         hammer_cluster_t cluster = info;
2302         hammer_buffer_t buffer;
2303         int32_t buf_no;
2304         int error = 0;
2305
2306         /*
2307          * Calculate the buffer number, initialize based on the buffer type.
2308          * The buffer has already been allocated so assert that it has been
2309          * initialized.
2310          */
2311         buf_no = blk / HAMMER_FSBUF_MAXBLKS;
2312         buffer = hammer_get_buffer(cluster, buf_no, 0, &error);
2313         if (buffer) {
2314                 hammer_adjust_stats(cluster, buffer->ondisk->head.buf_type, -1);
2315                 hammer_rel_buffer(buffer, 0);
2316         }
2317         return (error);
2318 #endif
2319 }
2320
2321 /*
2322  * Note: atblk can be negative and atblk - blk can go negative.
2323  */
2324 static int
2325 buffer_alist_alloc_fwd(void *info, int32_t blk, int32_t radix,
2326                       int32_t count, int32_t atblk, int32_t *fullp)
2327 {
2328         hammer_cluster_t cluster = info;
2329         hammer_buffer_t buffer;
2330         int32_t buf_no;
2331         int32_t r;
2332         int error = 0;
2333
2334         buf_no = blk / HAMMER_FSBUF_MAXBLKS;
2335         buffer = hammer_get_buffer(cluster, buf_no, 0, &error);
2336         if (buffer) {
2337                 KKASSERT(buffer->ondisk->head.buf_type != 0);
2338
2339                 hammer_modify_buffer(buffer);
2340                 r = hammer_alist_alloc_fwd(&buffer->alist, count, atblk - blk);
2341                 if (r != HAMMER_ALIST_BLOCK_NONE)
2342                         r += blk;
2343                 hammer_modify_buffer_done(buffer);
2344                 *fullp = hammer_alist_isfull(&buffer->alist);
2345                 hammer_rel_buffer(buffer, 0);
2346         } else {
2347                 r = HAMMER_ALIST_BLOCK_NONE;
2348         }
2349         return(r);
2350 }
2351
2352 static int
2353 buffer_alist_alloc_rev(void *info, int32_t blk, int32_t radix,
2354                       int32_t count, int32_t atblk, int32_t *fullp)
2355 {
2356         hammer_cluster_t cluster = info;
2357         hammer_buffer_t buffer;
2358         int32_t buf_no;
2359         int32_t r;
2360         int error = 0;
2361
2362         buf_no = blk / HAMMER_FSBUF_MAXBLKS;
2363         buffer = hammer_get_buffer(cluster, buf_no, 0, &error);
2364         if (buffer) {
2365                 KKASSERT(buffer->ondisk->head.buf_type != 0);
2366                 hammer_modify_buffer(buffer);
2367                 r = hammer_alist_alloc_rev(&buffer->alist, count, atblk - blk);
2368                 if (r != HAMMER_ALIST_BLOCK_NONE)
2369                         r += blk;
2370                 hammer_modify_buffer_done(buffer);
2371                 *fullp = hammer_alist_isfull(&buffer->alist);
2372                 hammer_rel_buffer(buffer, 0);
2373         } else {
2374                 r = HAMMER_ALIST_BLOCK_NONE;
2375                 *fullp = 0;
2376         }
2377         return(r);
2378 }
2379
2380 static void
2381 buffer_alist_free(void *info, int32_t blk, int32_t radix,
2382                  int32_t base_blk, int32_t count, int32_t *emptyp)
2383 {
2384         hammer_cluster_t cluster = info;
2385         hammer_buffer_t buffer;
2386         int32_t buf_no;
2387         int error = 0;
2388
2389         buf_no = blk / HAMMER_FSBUF_MAXBLKS;
2390         buffer = hammer_get_buffer(cluster, buf_no, 0, &error);
2391         if (buffer) {
2392                 KKASSERT(buffer->ondisk->head.buf_type != 0);
2393                 hammer_modify_buffer(buffer);
2394                 hammer_alist_free(&buffer->alist, base_blk, count);
2395                 hammer_modify_buffer_done(buffer);
2396                 *emptyp = hammer_alist_isempty(&buffer->alist);
2397                 /* XXX don't bother updating the buffer is completely empty? */
2398                 hammer_rel_buffer(buffer, 0);
2399         } else {
2400                 *emptyp = 0;
2401         }
2402 }
2403
2404 static void
2405 buffer_alist_print(void *info, int32_t blk, int32_t radix, int tab)
2406 {
2407 }
2408
2409 /*
2410  * A-LIST - super-cluster recursion into a cluster and cluster recursion
2411  * into a filesystem buffer.  A-List's are mostly self-contained entities,
2412  * but callbacks must be installed to recurse from one A-List to another.
2413  *
2414  * Implementing these callbacks allows us to operate a multi-layered A-List
2415  * as a single entity.
2416  */
2417 static int
2418 super_alist_init(void *info, int32_t blk, int32_t radix)
2419 {
2420         hammer_volume_t volume = info;
2421         hammer_supercl_t supercl;
2422         int32_t scl_no;
2423         int error = 0;
2424
2425         /*
2426          * Calculate the super-cluster number containing the cluster (blk)
2427          * and obtain the super-cluster buffer.
2428          */
2429         scl_no = blk / HAMMER_SCL_MAXCLUSTERS;
2430         supercl = hammer_get_supercl(volume, scl_no, &error, 1);
2431         if (supercl)
2432                 hammer_rel_supercl(supercl, 0);
2433         return (error);
2434 }
2435
2436 static int
2437 super_alist_destroy(void *info, int32_t blk, int32_t radix)
2438 {
2439         return(0);
2440 }
2441
2442 static int
2443 super_alist_alloc_fwd(void *info, int32_t blk, int32_t radix,
2444                       int32_t count, int32_t atblk, int32_t *fullp)
2445 {
2446         hammer_volume_t volume = info;
2447         hammer_supercl_t supercl;
2448         int32_t scl_no;
2449         int32_t r;
2450         int error = 0;
2451
2452         scl_no = blk / HAMMER_SCL_MAXCLUSTERS;
2453         supercl = hammer_get_supercl(volume, scl_no, &error, 0);
2454         if (supercl) {
2455                 hammer_modify_supercl(supercl);
2456                 r = hammer_alist_alloc_fwd(&supercl->alist, count, atblk - blk);
2457                 if (r != HAMMER_ALIST_BLOCK_NONE)
2458                         r += blk;
2459                 hammer_modify_supercl_done(supercl);
2460                 *fullp = hammer_alist_isfull(&supercl->alist);
2461                 hammer_rel_supercl(supercl, 0);
2462         } else {
2463                 r = HAMMER_ALIST_BLOCK_NONE;
2464                 *fullp = 0;
2465         }
2466         return(r);
2467 }
2468
2469 static int
2470 super_alist_alloc_rev(void *info, int32_t blk, int32_t radix,
2471                       int32_t count, int32_t atblk, int32_t *fullp)
2472 {
2473         hammer_volume_t volume = info;
2474         hammer_supercl_t supercl;
2475         int32_t scl_no;
2476         int32_t r;
2477         int error = 0;
2478
2479         scl_no = blk / HAMMER_SCL_MAXCLUSTERS;
2480         supercl = hammer_get_supercl(volume, scl_no, &error, 0);
2481         if (supercl) {
2482                 hammer_modify_supercl(supercl);
2483                 r = hammer_alist_alloc_rev(&supercl->alist, count, atblk - blk);
2484                 if (r != HAMMER_ALIST_BLOCK_NONE)
2485                         r += blk;
2486                 hammer_modify_supercl_done(supercl);
2487                 *fullp = hammer_alist_isfull(&supercl->alist);
2488                 hammer_rel_supercl(supercl, 0);
2489         } else {
2490                 r = HAMMER_ALIST_BLOCK_NONE;
2491                 *fullp = 0;
2492         }
2493         return(r);
2494 }
2495
2496 static void
2497 super_alist_free(void *info, int32_t blk, int32_t radix,
2498                  int32_t base_blk, int32_t count, int32_t *emptyp)
2499 {
2500         hammer_volume_t volume = info;
2501         hammer_supercl_t supercl;
2502         int32_t scl_no;
2503         int error = 0;
2504
2505         scl_no = blk / HAMMER_SCL_MAXCLUSTERS;
2506         supercl = hammer_get_supercl(volume, scl_no, &error, 0);
2507         if (supercl) {
2508                 hammer_modify_supercl(supercl);
2509                 hammer_alist_free(&supercl->alist, base_blk, count);
2510                 hammer_modify_supercl_done(supercl);
2511                 *emptyp = hammer_alist_isempty(&supercl->alist);
2512                 hammer_rel_supercl(supercl, 0);
2513         } else {
2514                 *emptyp = 0;
2515         }
2516 }
2517
2518 static void
2519 super_alist_print(void *info, int32_t blk, int32_t radix, int tab)
2520 {
2521 }
2522
2523 void
2524 hammer_init_alist_config(void)
2525 {
2526         hammer_alist_config_t config;
2527
2528         hammer_alist_template(&Buf_alist_config, HAMMER_FSBUF_MAXBLKS,
2529                               1, HAMMER_FSBUF_METAELMS);
2530         hammer_alist_template(&Vol_normal_alist_config, HAMMER_VOL_MAXCLUSTERS,
2531                               1, HAMMER_VOL_METAELMS_1LYR);
2532         hammer_alist_template(&Vol_super_alist_config,
2533                           HAMMER_VOL_MAXSUPERCLUSTERS * HAMMER_SCL_MAXCLUSTERS,
2534                               HAMMER_SCL_MAXCLUSTERS, HAMMER_VOL_METAELMS_2LYR);
2535         hammer_alist_template(&Supercl_alist_config, HAMMER_VOL_MAXCLUSTERS,
2536                               1, HAMMER_SUPERCL_METAELMS);
2537         hammer_alist_template(&Clu_master_alist_config, HAMMER_CLU_MAXBUFFERS,
2538                               1, HAMMER_CLU_MASTER_METAELMS);
2539         hammer_alist_template(&Clu_slave_alist_config,
2540                               HAMMER_CLU_MAXBUFFERS * HAMMER_FSBUF_MAXBLKS,
2541                               HAMMER_FSBUF_MAXBLKS, HAMMER_CLU_SLAVE_METAELMS);
2542
2543         config = &Vol_super_alist_config;
2544         config->bl_radix_init = super_alist_init;
2545         config->bl_radix_destroy = super_alist_destroy;
2546         config->bl_radix_alloc_fwd = super_alist_alloc_fwd;
2547         config->bl_radix_alloc_rev = super_alist_alloc_rev;
2548         config->bl_radix_free = super_alist_free;
2549         config->bl_radix_print = super_alist_print;
2550
2551         config = &Clu_slave_alist_config;
2552         config->bl_radix_init = buffer_alist_init;
2553         config->bl_radix_destroy = buffer_alist_destroy;
2554         config->bl_radix_alloc_fwd = buffer_alist_alloc_fwd;
2555         config->bl_radix_alloc_rev = buffer_alist_alloc_rev;
2556         config->bl_radix_free = buffer_alist_free;
2557         config->bl_radix_print = buffer_alist_print;
2558 }
2559