sys/vfs/hammer/hammer_inode.c

   1 /*
   2  * Copyright (c) 2007-2008 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_inode.c,v 1.59 2008/05/13 20:46:55 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38 #include <vm/vm_extern.h>
  39 #include <sys/buf.h>
  40 #include <sys/buf2.h>
  41
  42 static int hammer_unload_inode(struct hammer_inode *ip);
  43 static void hammer_flush_inode_core(hammer_inode_t ip, int flags);
  44 static int hammer_setup_child_callback(hammer_record_t rec, void *data);
  45 static int hammer_setup_parent_inodes(hammer_record_t record);
  46
  47 /*
  48  * The kernel is not actively referencing this vnode but is still holding
  49  * it cached.
  50  *
  51  * This is called from the frontend.
  52  */
  53 int
  54 hammer_vop_inactive(struct vop_inactive_args *ap)
  55 {
  56         struct hammer_inode *ip = VTOI(ap->a_vp);
  57
  58         /*
  59          * Degenerate case
  60          */
  61         if (ip == NULL) {
  62                 vrecycle(ap->a_vp);
  63                 return(0);
  64         }
  65
  66         /*
  67          * If the inode no longer has visibility in the filesystem and is
  68          * fairly clean, try to recycle it immediately.  This can deadlock
  69          * in vfsync() if we aren't careful.
  70          *
  71          * Do not queue the inode to the flusher if we still have visibility,
  72          * otherwise namespace calls such as chmod will unnecessarily generate
  73          * multiple inode updates.
  74          */
  75         hammer_inode_unloadable_check(ip, 0);
  76         if (ip->ino_data.nlinks == 0) {
  77                 if (ip->flags & HAMMER_INODE_MODMASK)
  78                         hammer_flush_inode(ip, 0);
  79                 else
  80                         vrecycle(ap->a_vp);
  81         }
  82         return(0);
  83 }
  84
  85 /*
  86  * Release the vnode association.  This is typically (but not always)
  87  * the last reference on the inode.
  88  *
  89  * Once the association is lost we are on our own with regards to
  90  * flushing the inode.
  91  */
  92 int
  93 hammer_vop_reclaim(struct vop_reclaim_args *ap)
  94 {
  95         struct hammer_inode *ip;
  96         struct vnode *vp;
  97
  98         vp = ap->a_vp;
  99
 100         if ((ip = vp->v_data) != NULL) {
 101                 vp->v_data = NULL;
 102                 ip->vp = NULL;
 103                 hammer_rel_inode(ip, 1);
 104         }
 105         return(0);
 106 }
 107
 108 /*
 109  * Return a locked vnode for the specified inode.  The inode must be
 110  * referenced but NOT LOCKED on entry and will remain referenced on
 111  * return.
 112  *
 113  * Called from the frontend.
 114  */
 115 int
 116 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 117 {
 118         struct vnode *vp;
 119         int error = 0;
 120
 121         for (;;) {
 122                 if ((vp = ip->vp) == NULL) {
 123                         error = getnewvnode(VT_HAMMER, ip->hmp->mp, vpp, 0, 0);
 124                         if (error)
 125                                 break;
 126                         hammer_lock_ex(&ip->lock);
 127                         if (ip->vp != NULL) {
 128                                 hammer_unlock(&ip->lock);
 129                                 vp->v_type = VBAD;
 130                                 vx_put(vp);
 131                                 continue;
 132                         }
 133                         hammer_ref(&ip->lock);
 134                         vp = *vpp;
 135                         ip->vp = vp;
 136                         vp->v_type =
 137                                 hammer_get_vnode_type(ip->ino_data.obj_type);
 138
 139                         switch(ip->ino_data.obj_type) {
 140                         case HAMMER_OBJTYPE_CDEV:
 141                         case HAMMER_OBJTYPE_BDEV:
 142                                 vp->v_ops = &ip->hmp->mp->mnt_vn_spec_ops;
 143                                 addaliasu(vp, ip->ino_data.rmajor,
 144                                           ip->ino_data.rminor);
 145                                 break;
 146                         case HAMMER_OBJTYPE_FIFO:
 147                                 vp->v_ops = &ip->hmp->mp->mnt_vn_fifo_ops;
 148                                 break;
 149                         default:
 150                                 break;
 151                         }
 152
 153                         /*
 154                          * Only mark as the root vnode if the ip is not
 155                          * historical, otherwise the VFS cache will get
 156                          * confused.  The other half of the special handling
 157                          * is in hammer_vop_nlookupdotdot().
 158                          */
 159                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 160                             ip->obj_asof == ip->hmp->asof) {
 161                                 vp->v_flag |= VROOT;
 162                         }
 163
 164                         vp->v_data = (void *)ip;
 165                         /* vnode locked by getnewvnode() */
 166                         /* make related vnode dirty if inode dirty? */
 167                         hammer_unlock(&ip->lock);
 168                         if (vp->v_type == VREG)
 169                                 vinitvmio(vp, ip->ino_data.size);
 170                         break;
 171                 }
 172
 173                 /*
 174                  * loop if the vget fails (aka races), or if the vp
 175                  * no longer matches ip->vp.
 176                  */
 177                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 178                         if (vp == ip->vp)
 179                                 break;
 180                         vput(vp);
 181                 }
 182         }
 183         *vpp = vp;
 184         return(error);
 185 }
 186
 187 /*
 188  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 189  * do not attach or detach the related vnode (use hammer_get_vnode() for
 190  * that).
 191  *
 192  * The flags argument is only applied for newly created inodes, and only
 193  * certain flags are inherited.
 194  *
 195  * Called from the frontend.
 196  */
 197 struct hammer_inode *
 198 hammer_get_inode(hammer_transaction_t trans, struct hammer_node **cache,
 199                  u_int64_t obj_id, hammer_tid_t asof, int flags, int *errorp)
 200 {
 201         hammer_mount_t hmp = trans->hmp;
 202         struct hammer_inode_info iinfo;
 203         struct hammer_cursor cursor;
 204         struct hammer_inode *ip;
 205
 206         /*
 207          * Determine if we already have an inode cached.  If we do then
 208          * we are golden.
 209          */
 210         iinfo.obj_id = obj_id;
 211         iinfo.obj_asof = asof;
 212 loop:
 213         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 214         if (ip) {
 215                 hammer_ref(&ip->lock);
 216                 *errorp = 0;
 217                 return(ip);
 218         }
 219
 220         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 221         ++hammer_count_inodes;
 222         ip->obj_id = obj_id;
 223         ip->obj_asof = iinfo.obj_asof;
 224         ip->hmp = hmp;
 225         ip->flags = flags & HAMMER_INODE_RO;
 226         if (hmp->ronly)
 227                 ip->flags |= HAMMER_INODE_RO;
 228         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 229         RB_INIT(&ip->rec_tree);
 230         TAILQ_INIT(&ip->bio_list);
 231         TAILQ_INIT(&ip->bio_alt_list);
 232         TAILQ_INIT(&ip->target_list);
 233
 234         /*
 235          * Locate the on-disk inode.
 236          */
 237 retry:
 238         hammer_init_cursor(trans, &cursor, cache, NULL);
 239         cursor.key_beg.obj_id = ip->obj_id;
 240         cursor.key_beg.key = 0;
 241         cursor.key_beg.create_tid = 0;
 242         cursor.key_beg.delete_tid = 0;
 243         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 244         cursor.key_beg.obj_type = 0;
 245         cursor.asof = iinfo.obj_asof;
 246         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 247                        HAMMER_CURSOR_ASOF;
 248
 249         *errorp = hammer_btree_lookup(&cursor);
 250         if (*errorp == EDEADLK) {
 251                 hammer_done_cursor(&cursor);
 252                 goto retry;
 253         }
 254
 255         /*
 256          * On success the B-Tree lookup will hold the appropriate
 257          * buffer cache buffers and provide a pointer to the requested
 258          * information.  Copy the information to the in-memory inode
 259          * and cache the B-Tree node to improve future operations.
 260          */
 261         if (*errorp == 0) {
 262                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 263                 ip->ino_data = cursor.data->inode;
 264                 hammer_cache_node(cursor.node, &ip->cache[0]);
 265                 if (cache)
 266                         hammer_cache_node(cursor.node, cache);
 267         }
 268
 269         /*
 270          * On success load the inode's record and data and insert the
 271          * inode into the B-Tree.  It is possible to race another lookup
 272          * insertion of the same inode so deal with that condition too.
 273          *
 274          * The cursor's locked node interlocks against others creating and
 275          * destroying ip while we were blocked.
 276          */
 277         if (*errorp == 0) {
 278                 hammer_ref(&ip->lock);
 279                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 280                         hammer_uncache_node(&ip->cache[0]);
 281                         hammer_uncache_node(&ip->cache[1]);
 282                         KKASSERT(ip->lock.refs == 1);
 283                         --hammer_count_inodes;
 284                         kfree(ip, M_HAMMER);
 285                         hammer_done_cursor(&cursor);
 286                         goto loop;
 287                 }
 288                 ip->flags |= HAMMER_INODE_ONDISK;
 289         } else {
 290                 /*
 291                  * Do not panic on read-only accesses which fail, particularly
 292                  * historical accesses where the snapshot might not have
 293                  * complete connectivity.
 294                  */
 295                 if ((flags & HAMMER_INODE_RO) == 0) {
 296                         kprintf("hammer_get_inode: failed ip %p obj_id %016llx cursor %p error %d\n",
 297                                 ip, ip->obj_id, &cursor, *errorp);
 298                         Debugger("x");
 299                 }
 300                 --hammer_count_inodes;
 301                 kfree(ip, M_HAMMER);
 302                 ip = NULL;
 303         }
 304         hammer_done_cursor(&cursor);
 305         return (ip);
 306 }
 307
 308 /*
 309  * Create a new filesystem object, returning the inode in *ipp.  The
 310  * returned inode will be referenced.
 311  *
 312  * The inode is created in-memory.
 313  */
 314 int
 315 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 316                     struct ucred *cred, hammer_inode_t dip,
 317                     struct hammer_inode **ipp)
 318 {
 319         hammer_mount_t hmp;
 320         hammer_inode_t ip;
 321         uid_t xuid;
 322
 323         hmp = trans->hmp;
 324         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 325         ++hammer_count_inodes;
 326         ip->obj_id = hammer_alloc_objid(trans, dip);
 327         KKASSERT(ip->obj_id != 0);
 328         ip->obj_asof = hmp->asof;
 329         ip->hmp = hmp;
 330         ip->flush_state = HAMMER_FST_IDLE;
 331         ip->flags = HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES;
 332
 333         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 334         RB_INIT(&ip->rec_tree);
 335         TAILQ_INIT(&ip->bio_list);
 336         TAILQ_INIT(&ip->bio_alt_list);
 337         TAILQ_INIT(&ip->target_list);
 338
 339         ip->ino_leaf.atime = trans->time;
 340         ip->ino_data.mtime = trans->time;
 341         ip->ino_data.size = 0;
 342         ip->ino_data.nlinks = 0;
 343         /* XXX */
 344         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 345         ip->ino_leaf.base.obj_id = ip->obj_id;
 346         ip->ino_leaf.base.key = 0;
 347         ip->ino_leaf.base.create_tid = 0;
 348         ip->ino_leaf.base.delete_tid = 0;
 349         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 350         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 351
 352         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 353         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 354         ip->ino_data.mode = vap->va_mode;
 355         ip->ino_data.ctime = trans->time;
 356         ip->ino_data.parent_obj_id = (dip) ? dip->ino_leaf.base.obj_id : 0;
 357
 358         switch(ip->ino_leaf.base.obj_type) {
 359         case HAMMER_OBJTYPE_CDEV:
 360         case HAMMER_OBJTYPE_BDEV:
 361                 ip->ino_data.rmajor = vap->va_rmajor;
 362                 ip->ino_data.rminor = vap->va_rminor;
 363                 break;
 364         default:
 365                 break;
 366         }
 367
 368         /*
 369          * Calculate default uid/gid and overwrite with information from
 370          * the vap.
 371          */
 372         xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 373         ip->ino_data.gid = dip->ino_data.gid;
 374         xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode, xuid, cred,
 375                                      &vap->va_mode);
 376         ip->ino_data.mode = vap->va_mode;
 377
 378         if (vap->va_vaflags & VA_UID_UUID_VALID)
 379                 ip->ino_data.uid = vap->va_uid_uuid;
 380         else if (vap->va_uid != (uid_t)VNOVAL)
 381                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 382         if (vap->va_vaflags & VA_GID_UUID_VALID)
 383                 ip->ino_data.gid = vap->va_gid_uuid;
 384         else if (vap->va_gid != (gid_t)VNOVAL)
 385                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 386
 387         hammer_ref(&ip->lock);
 388         if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 389                 hammer_unref(&ip->lock);
 390                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 391         }
 392         *ipp = ip;
 393         return(0);
 394 }
 395
 396 /*
 397  * Called by hammer_sync_inode().
 398  */
 399 static int
 400 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 401 {
 402         hammer_transaction_t trans = cursor->trans;
 403         hammer_record_t record;
 404         int error;
 405
 406 retry:
 407         error = 0;
 408
 409         /*
 410          * If the inode has a presence on-disk then locate it and mark
 411          * it deleted, setting DELONDISK.
 412          *
 413          * The record may or may not be physically deleted, depending on
 414          * the retention policy.
 415          */
 416         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 417             HAMMER_INODE_ONDISK) {
 418                 hammer_normalize_cursor(cursor);
 419                 cursor->key_beg.obj_id = ip->obj_id;
 420                 cursor->key_beg.key = 0;
 421                 cursor->key_beg.create_tid = 0;
 422                 cursor->key_beg.delete_tid = 0;
 423                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 424                 cursor->key_beg.obj_type = 0;
 425                 cursor->asof = ip->obj_asof;
 426                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 427                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 428                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 429
 430                 error = hammer_btree_lookup(cursor);
 431                 if (hammer_debug_inode)
 432                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 433                 if (error) {
 434                         kprintf("error %d\n", error);
 435                         Debugger("hammer_update_inode");
 436                 }
 437
 438                 if (error == 0) {
 439                         error = hammer_ip_delete_record(cursor, trans->tid);
 440                         if (hammer_debug_inode)
 441                                 kprintf(" error %d\n", error);
 442                         if (error && error != EDEADLK) {
 443                                 kprintf("error %d\n", error);
 444                                 Debugger("hammer_update_inode2");
 445                         }
 446                         if (error == 0) {
 447                                 ip->flags |= HAMMER_INODE_DELONDISK;
 448                         }
 449                         if (cursor->node)
 450                                 hammer_cache_node(cursor->node, &ip->cache[0]);
 451                 }
 452                 if (error == EDEADLK) {
 453                         hammer_done_cursor(cursor);
 454                         error = hammer_init_cursor(trans, cursor,
 455                                                    &ip->cache[0], ip);
 456                         if (hammer_debug_inode)
 457                                 kprintf("IPDED %p %d\n", ip, error);
 458                         if (error == 0)
 459                                 goto retry;
 460                 }
 461         }
 462
 463         /*
 464          * Ok, write out the initial record or a new record (after deleting
 465          * the old one), unless the DELETED flag is set.  This routine will
 466          * clear DELONDISK if it writes out a record.
 467          *
 468          * Update our inode statistics if this is the first application of
 469          * the inode on-disk.
 470          */
 471         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 472                 /*
 473                  * Generate a record and write it to the media
 474                  */
 475                 record = hammer_alloc_mem_record(ip, 0);
 476                 record->type = HAMMER_MEM_RECORD_INODE;
 477                 record->flush_state = HAMMER_FST_FLUSH;
 478                 record->leaf = ip->sync_ino_leaf;
 479                 record->leaf.base.create_tid = trans->tid;
 480                 record->leaf.data_len = sizeof(ip->sync_ino_data);
 481                 record->data = (void *)&ip->sync_ino_data;
 482                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
 483                 for (;;) {
 484                         error = hammer_ip_sync_record_cursor(cursor, record);
 485                         if (hammer_debug_inode)
 486                                 kprintf("GENREC %p rec %08x %d\n",
 487                                         ip, record->flags, error);
 488                         if (error != EDEADLK)
 489                                 break;
 490                         hammer_done_cursor(cursor);
 491                         error = hammer_init_cursor(trans, cursor,
 492                                                    &ip->cache[0], ip);
 493                         if (hammer_debug_inode)
 494                                 kprintf("GENREC reinit %d\n", error);
 495                         if (error)
 496                                 break;
 497                 }
 498                 if (error) {
 499                         kprintf("error %d\n", error);
 500                         Debugger("hammer_update_inode3");
 501                 }
 502
 503                 /*
 504                  * The record isn't managed by the inode's record tree,
 505                  * destroy it whether we succeed or fail.
 506                  */
 507                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 508                 record->flags |= HAMMER_RECF_DELETED_FE;
 509                 record->flush_state = HAMMER_FST_IDLE;
 510                 hammer_rel_mem_record(record);
 511
 512                 /*
 513                  * Finish up.
 514                  */
 515                 if (error == 0) {
 516                         if (hammer_debug_inode)
 517                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
 518                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
 519                                             HAMMER_INODE_ITIMES);
 520                         ip->flags &= ~HAMMER_INODE_DELONDISK;
 521
 522                         /*
 523                          * Root volume count of inodes
 524                          */
 525                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
 526                                 hammer_modify_volume_field(trans,
 527                                                            trans->rootvol,
 528                                                            vol0_stat_inodes);
 529                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
 530                                 hammer_modify_volume_done(trans->rootvol);
 531                                 ip->flags |= HAMMER_INODE_ONDISK;
 532                                 if (hammer_debug_inode)
 533                                         kprintf("NOWONDISK %p\n", ip);
 534                         }
 535                 }
 536         }
 537
 538         /*
 539          * If the inode has been destroyed, clean out any left-over flags
 540          * that may have been set by the frontend.
 541          */
 542         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
 543                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
 544                                     HAMMER_INODE_ITIMES);
 545         }
 546         return(error);
 547 }
 548
 549 /*
 550  * Update only the itimes fields.  This is done no-historically.  The
 551  * record is updated in-place on the disk.
 552  */
 553 static int
 554 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
 555 {
 556         hammer_transaction_t trans = cursor->trans;
 557         struct hammer_btree_leaf_elm *leaf;
 558         int error;
 559
 560 retry:
 561         error = 0;
 562         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 563             HAMMER_INODE_ONDISK) {
 564                 hammer_normalize_cursor(cursor);
 565                 cursor->key_beg.obj_id = ip->obj_id;
 566                 cursor->key_beg.key = 0;
 567                 cursor->key_beg.create_tid = 0;
 568                 cursor->key_beg.delete_tid = 0;
 569                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 570                 cursor->key_beg.obj_type = 0;
 571                 cursor->asof = ip->obj_asof;
 572                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 573                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 574                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 575
 576                 error = hammer_btree_lookup(cursor);
 577                 if (error) {
 578                         kprintf("error %d\n", error);
 579                         Debugger("hammer_update_itimes1");
 580                 }
 581                 if (error == 0) {
 582                         /*
 583                          * Do not generate UNDO records for atime updates.
 584                          */
 585                         leaf = cursor->leaf;
 586                         hammer_modify_node(trans, cursor->node,
 587                                            &leaf->atime, sizeof(leaf->atime));
 588                         leaf->atime = ip->sync_ino_leaf.atime;
 589                         hammer_modify_node_done(cursor->node);
 590                         /*rec->ino_mtime = ip->sync_ino_rec.ino_mtime;*/
 591                         ip->sync_flags &= ~HAMMER_INODE_ITIMES;
 592                         /* XXX recalculate crc */
 593                         hammer_cache_node(cursor->node, &ip->cache[0]);
 594                 }
 595                 if (error == EDEADLK) {
 596                         hammer_done_cursor(cursor);
 597                         error = hammer_init_cursor(trans, cursor,
 598                                                    &ip->cache[0], ip);
 599                         if (error == 0)
 600                                 goto retry;
 601                 }
 602         }
 603         return(error);
 604 }
 605
 606 /*
 607  * Release a reference on an inode, flush as requested.
 608  *
 609  * On the last reference we queue the inode to the flusher for its final
 610  * disposition.
 611  */
 612 void
 613 hammer_rel_inode(struct hammer_inode *ip, int flush)
 614 {
 615         hammer_mount_t hmp = ip->hmp;
 616
 617         /*
 618          * Handle disposition when dropping the last ref.
 619          */
 620         for (;;) {
 621                 if (ip->lock.refs == 1) {
 622                         /*
 623                          * Determine whether on-disk action is needed for
 624                          * the inode's final disposition.
 625                          */
 626                         KKASSERT(ip->vp == NULL);
 627                         hammer_inode_unloadable_check(ip, 0);
 628                         if (ip->flags & HAMMER_INODE_MODMASK) {
 629                                 hammer_flush_inode(ip, 0);
 630                         } else if (ip->lock.refs == 1) {
 631                                 hammer_unload_inode(ip);
 632                                 break;
 633                         }
 634                 } else {
 635                         if (flush)
 636                                 hammer_flush_inode(ip, 0);
 637
 638                         /*
 639                          * The inode still has multiple refs, try to drop
 640                          * one ref.
 641                          */
 642                         KKASSERT(ip->lock.refs >= 1);
 643                         if (ip->lock.refs > 1) {
 644                                 hammer_unref(&ip->lock);
 645                                 break;
 646                         }
 647                 }
 648         }
 649
 650         /*
 651          * XXX bad hack until I add code to track inodes in SETUP.  We
 652          * can queue a lot of inodes to the syncer but if we don't wake
 653          * it up the undo sets will be too large or too many unflushed
 654          * records will build up and blow our malloc limit.
 655          */
 656         if (++hmp->reclaim_count > 256) {
 657                 hmp->reclaim_count = 0;
 658                 hammer_flusher_async(hmp);
 659         }
 660 }
 661
 662 /*
 663  * Unload and destroy the specified inode.  Must be called with one remaining
 664  * reference.  The reference is disposed of.
 665  *
 666  * This can only be called in the context of the flusher.
 667  */
 668 static int
 669 hammer_unload_inode(struct hammer_inode *ip)
 670 {
 671         KASSERT(ip->lock.refs == 1,
 672                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
 673         KKASSERT(ip->vp == NULL);
 674         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
 675         KKASSERT(ip->cursor_ip_refs == 0);
 676         KKASSERT(ip->lock.lockcount == 0);
 677         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
 678
 679         KKASSERT(RB_EMPTY(&ip->rec_tree));
 680         KKASSERT(TAILQ_EMPTY(&ip->target_list));
 681         KKASSERT(TAILQ_EMPTY(&ip->bio_list));
 682         KKASSERT(TAILQ_EMPTY(&ip->bio_alt_list));
 683
 684         RB_REMOVE(hammer_ino_rb_tree, &ip->hmp->rb_inos_root, ip);
 685
 686         hammer_uncache_node(&ip->cache[0]);
 687         hammer_uncache_node(&ip->cache[1]);
 688         if (ip->objid_cache)
 689                 hammer_clear_objid(ip);
 690         --hammer_count_inodes;
 691         kfree(ip, M_HAMMER);
 692
 693         return(0);
 694 }
 695
 696 /*
 697  * A transaction has modified an inode, requiring updates as specified by
 698  * the passed flags.
 699  *
 700  * HAMMER_INODE_DDIRTY: Inode data has been updated
 701  * HAMMER_INODE_XDIRTY: Dirty in-memory records
 702  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
 703  * HAMMER_INODE_DELETED: Inode record/data must be deleted
 704  * HAMMER_INODE_ITIMES: mtime/atime has been updated
 705  */
 706 void
 707 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
 708 {
 709         KKASSERT ((ip->flags & HAMMER_INODE_RO) == 0 ||
 710                   (flags & (HAMMER_INODE_DDIRTY |
 711                             HAMMER_INODE_XDIRTY | HAMMER_INODE_BUFS |
 712                             HAMMER_INODE_DELETED | HAMMER_INODE_ITIMES)) == 0);
 713
 714         ip->flags |= flags;
 715 }
 716
 717 /*
 718  * Request that an inode be flushed.  This whole mess cannot block and may
 719  * recurse.  Once requested HAMMER will attempt to actively flush it until
 720  * the flush can be done.
 721  *
 722  * The inode may already be flushing, or may be in a setup state.  We can
 723  * place the inode in a flushing state if it is currently idle and flag it
 724  * to reflush if it is currently flushing.
 725  */
 726 void
 727 hammer_flush_inode(hammer_inode_t ip, int flags)
 728 {
 729         hammer_record_t depend;
 730         int r, good;
 731
 732         /*
 733          * Trivial 'nothing to flush' case.  If the inode is ina SETUP
 734          * state we have to put it back into an IDLE state so we can
 735          * drop the extra ref.
 736          */
 737         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
 738                 if (ip->flush_state == HAMMER_FST_SETUP) {
 739                         ip->flush_state = HAMMER_FST_IDLE;
 740                         hammer_rel_inode(ip, 0);
 741                 }
 742                 return;
 743         }
 744
 745         /*
 746          * Our flush action will depend on the current state.
 747          */
 748         switch(ip->flush_state) {
 749         case HAMMER_FST_IDLE:
 750                 /*
 751                  * We have no dependancies and can flush immediately.  Some
 752                  * our children may not be flushable so we have to re-test
 753                  * with that additional knowledge.
 754                  */
 755                 hammer_flush_inode_core(ip, flags);
 756                 break;
 757         case HAMMER_FST_SETUP:
 758                 /*
 759                  * Recurse upwards through dependancies via target_list
 760                  * and start their flusher actions going if possible.
 761                  *
 762                  * 'good' is our connectivity.  -1 means we have none and
 763                  * can't flush, 0 means there weren't any dependancies, and
 764                  * 1 means we have good connectivity.
 765                  */
 766                 good = 0;
 767                 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 768                         r = hammer_setup_parent_inodes(depend);
 769                         if (r < 0 && good == 0)
 770                                 good = -1;
 771                         if (r > 0)
 772                                 good = 1;
 773                 }
 774
 775                 /*
 776                  * We can continue if good >= 0.  Determine how many records
 777                  * under our inode can be flushed (and mark them).
 778                  */
 779                 if (good >= 0) {
 780                         hammer_flush_inode_core(ip, flags);
 781                 } else {
 782                         ip->flags |= HAMMER_INODE_REFLUSH;
 783                         if (flags & HAMMER_FLUSH_SIGNAL) {
 784                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 785                                 hammer_flusher_async(ip->hmp);
 786                         }
 787                 }
 788                 break;
 789         default:
 790                 /*
 791                  * We are already flushing, flag the inode to reflush
 792                  * if needed after it completes its current flush.
 793                  */
 794                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
 795                         ip->flags |= HAMMER_INODE_REFLUSH;
 796                 if (flags & HAMMER_FLUSH_SIGNAL) {
 797                         ip->flags |= HAMMER_INODE_RESIGNAL;
 798                         hammer_flusher_async(ip->hmp);
 799                 }
 800                 break;
 801         }
 802 }
 803
 804 /*
 805  * We are asked to recurse upwards and convert the record from SETUP
 806  * to FLUSH if possible.  record->ip is a parent of the caller's inode,
 807  * and record->target_ip is the caller's inode.
 808  *
 809  * Return 1 if the record gives us connectivity
 810  *
 811  * Return 0 if the record is not relevant
 812  *
 813  * Return -1 if we can't resolve the dependancy and there is no connectivity.
 814  */
 815 static int
 816 hammer_setup_parent_inodes(hammer_record_t record)
 817 {
 818         hammer_mount_t hmp = record->ip->hmp;
 819         hammer_record_t depend;
 820         hammer_inode_t ip;
 821         int r, good;
 822
 823         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
 824         ip = record->ip;
 825
 826         /*
 827          * If the record is already flushing, is it in our flush group?
 828          *
 829          * If it is in our flush group but it is a general record or a
 830          * delete-on-disk, it does not improve our connectivity (return 0),
 831          * and if the target inode is not trying to destroy itself we can't
 832          * allow the operation yet anyway (the second return -1).
 833          */
 834         if (record->flush_state == HAMMER_FST_FLUSH) {
 835                 if (record->flush_group != hmp->flusher_next) {
 836                         ip->flags |= HAMMER_INODE_REFLUSH;
 837                         return(-1);
 838                 }
 839                 if (record->type == HAMMER_MEM_RECORD_ADD)
 840                         return(1);
 841                 /* GENERAL or DEL */
 842                 return(0);
 843         }
 844
 845         /*
 846          * It must be a setup record.  Try to resolve the setup dependancies
 847          * by recursing upwards so we can place ip on the flush list.
 848          */
 849         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
 850
 851         good = 0;
 852         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 853                 r = hammer_setup_parent_inodes(depend);
 854                 if (r < 0 && good == 0)
 855                         good = -1;
 856                 if (r > 0)
 857                         good = 1;
 858         }
 859
 860         /*
 861          * We can't flush ip because it has no connectivity (XXX also check
 862          * nlinks for pre-existing connectivity!).  Flag it so any resolution
 863          * recurses back down.
 864          */
 865         if (good < 0) {
 866                 ip->flags |= HAMMER_INODE_REFLUSH;
 867                 return(good);
 868         }
 869
 870         /*
 871          * We are go, place the parent inode in a flushing state so we can
 872          * place its record in a flushing state.  Note that the parent
 873          * may already be flushing.  The record must be in the same flush
 874          * group as the parent.
 875          */
 876         if (ip->flush_state != HAMMER_FST_FLUSH)
 877                 hammer_flush_inode_core(ip, HAMMER_FLUSH_RECURSION);
 878         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
 879         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
 880
 881 #if 0
 882         if (record->type == HAMMER_MEM_RECORD_DEL &&
 883             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
 884                 /*
 885                  * Regardless of flushing state we cannot sync this path if the
 886                  * record represents a delete-on-disk but the target inode
 887                  * is not ready to sync its own deletion.
 888                  *
 889                  * XXX need to count effective nlinks to determine whether
 890                  * the flush is ok, otherwise removing a hardlink will
 891                  * just leave the DEL record to rot.
 892                  */
 893                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
 894                 return(-1);
 895         } else
 896 #endif
 897         if (ip->flush_group == ip->hmp->flusher_next) {
 898                 /*
 899                  * This is the record we wanted to synchronize.
 900                  */
 901                 record->flush_state = HAMMER_FST_FLUSH;
 902                 record->flush_group = ip->flush_group;
 903                 hammer_ref(&record->lock);
 904                 if (record->type == HAMMER_MEM_RECORD_ADD)
 905                         return(1);
 906
 907                 /*
 908                  * A general or delete-on-disk record does not contribute
 909                  * to our visibility.  We can still flush it, however.
 910                  */
 911                 return(0);
 912         } else {
 913                 /*
 914                  * We couldn't resolve the dependancies, request that the
 915                  * inode be flushed when the dependancies can be resolved.
 916                  */
 917                 ip->flags |= HAMMER_INODE_REFLUSH;
 918                 return(-1);
 919         }
 920 }
 921
 922 /*
 923  * This is the core routine placing an inode into the FST_FLUSH state.
 924  */
 925 static void
 926 hammer_flush_inode_core(hammer_inode_t ip, int flags)
 927 {
 928         int go_count;
 929
 930         /*
 931          * Set flush state and prevent the flusher from cycling into
 932          * the next flush group.  Do not place the ip on the list yet.
 933          * Inodes not in the idle state get an extra reference.
 934          */
 935         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
 936         if (ip->flush_state == HAMMER_FST_IDLE)
 937                 hammer_ref(&ip->lock);
 938         ip->flush_state = HAMMER_FST_FLUSH;
 939         ip->flush_group = ip->hmp->flusher_next;
 940         ++ip->hmp->flusher_lock;
 941
 942         /*
 943          * We need to be able to vfsync/truncate from the backend.
 944          */
 945         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
 946         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
 947                 ip->flags |= HAMMER_INODE_VHELD;
 948                 vref(ip->vp);
 949         }
 950
 951         /*
 952          * Figure out how many in-memory records we can actually flush
 953          * (not including inode meta-data, buffers, etc).
 954          */
 955         if (flags & HAMMER_FLUSH_RECURSION) {
 956                 go_count = 1;
 957         } else {
 958                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
 959                                    hammer_setup_child_callback, NULL);
 960         }
 961
 962         /*
 963          * This is a more involved test that includes go_count.  If we
 964          * can't flush, flag the inode and return.  If go_count is 0 we
 965          * were are unable to flush any records in our rec_tree and
 966          * must ignore the XDIRTY flag.
 967          */
 968         if (go_count == 0) {
 969                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
 970                         ip->flags |= HAMMER_INODE_REFLUSH;
 971                         ip->flush_state = HAMMER_FST_SETUP;
 972                         if (ip->flags & HAMMER_INODE_VHELD) {
 973                                 ip->flags &= ~HAMMER_INODE_VHELD;
 974                                 vrele(ip->vp);
 975                         }
 976                         if (flags & HAMMER_FLUSH_SIGNAL) {
 977                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 978                                 hammer_flusher_async(ip->hmp);
 979                         }
 980                         if (--ip->hmp->flusher_lock == 0)
 981                                 wakeup(&ip->hmp->flusher_lock);
 982                         return;
 983                 }
 984         }
 985
 986         /*
 987          * Snapshot the state of the inode for the backend flusher.
 988          *
 989          * The truncation must be retained in the frontend until after
 990          * we've actually performed the record deletion.
 991          *
 992          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
 993          * and stays in ip->flags.  Once set, it stays set until the
 994          * inode is destroyed.
 995          */
 996         ip->sync_flags = (ip->flags & HAMMER_INODE_MODMASK);
 997         ip->sync_trunc_off = ip->trunc_off;
 998         ip->sync_ino_leaf = ip->ino_leaf;
 999         ip->sync_ino_data = ip->ino_data;
1000         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1001
1002         /*
1003          * The flusher list inherits our inode and reference.
1004          */
1005         TAILQ_INSERT_TAIL(&ip->hmp->flush_list, ip, flush_entry);
1006         if (--ip->hmp->flusher_lock == 0)
1007                 wakeup(&ip->hmp->flusher_lock);
1008
1009         if (flags & HAMMER_FLUSH_SIGNAL)
1010                 hammer_flusher_async(ip->hmp);
1011 }
1012
1013 /*
1014  * Callback for scan of ip->rec_tree.  Try to include each record in our
1015  * flush.  ip->flush_group has been set but the inode has not yet been
1016  * moved into a flushing state.
1017  *
1018  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1019  * both inodes.
1020  *
1021  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1022  * the caller from shortcutting the flush.
1023  */
1024 static int
1025 hammer_setup_child_callback(hammer_record_t rec, void *data)
1026 {
1027         hammer_inode_t target_ip;
1028         hammer_inode_t ip;
1029         int r;
1030
1031         /*
1032          * If the record has been deleted by the backend (it's being held
1033          * by the frontend in a race), just ignore it.
1034          */
1035         if (rec->flags & HAMMER_RECF_DELETED_BE)
1036                 return(0);
1037
1038         /*
1039          * If the record is in an idle state it has no dependancies and
1040          * can be flushed.
1041          */
1042         ip = rec->ip;
1043         r = 0;
1044
1045         switch(rec->flush_state) {
1046         case HAMMER_FST_IDLE:
1047                 /*
1048                  * Record has no setup dependancy, we can flush it.
1049                  */
1050                 KKASSERT(rec->target_ip == NULL);
1051                 rec->flush_state = HAMMER_FST_FLUSH;
1052                 rec->flush_group = ip->flush_group;
1053                 hammer_ref(&rec->lock);
1054                 r = 1;
1055                 break;
1056         case HAMMER_FST_SETUP:
1057                 /*
1058                  * Record has a setup dependancy.  Try to include the
1059                  * target ip in the flush.
1060                  *
1061                  * We have to be careful here, if we do not do the right
1062                  * thing we can lose track of dirty inodes and the system
1063                  * will lockup trying to allocate buffers.
1064                  */
1065                 target_ip = rec->target_ip;
1066                 KKASSERT(target_ip != NULL);
1067                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1068                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1069                         /*
1070                          * If the target IP is already flushing in our group
1071                          * we are golden, otherwise make sure the target
1072                          * reflushes.
1073                          */
1074                         if (target_ip->flush_group == ip->flush_group) {
1075                                 rec->flush_state = HAMMER_FST_FLUSH;
1076                                 rec->flush_group = ip->flush_group;
1077                                 hammer_ref(&rec->lock);
1078                                 r = 1;
1079                         } else {
1080                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1081                         }
1082                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1083                         /*
1084                          * If the target IP is not flushing we can force
1085                          * it to flush, even if it is unable to write out
1086                          * any of its own records we have at least one in
1087                          * hand that we CAN deal with.
1088                          */
1089                         rec->flush_state = HAMMER_FST_FLUSH;
1090                         rec->flush_group = ip->flush_group;
1091                         hammer_ref(&rec->lock);
1092                         hammer_flush_inode_core(target_ip,
1093                                                 HAMMER_FLUSH_RECURSION);
1094                         r = 1;
1095                 } else {
1096                         /*
1097                          * General or delete-on-disk record.
1098                          *
1099                          * XXX this needs help.  If a delete-on-disk we could
1100                          * disconnect the target.  If the target has its own
1101                          * dependancies they really need to be flushed.
1102                          *
1103                          * XXX
1104                          */
1105                         rec->flush_state = HAMMER_FST_FLUSH;
1106                         rec->flush_group = ip->flush_group;
1107                         hammer_ref(&rec->lock);
1108                         hammer_flush_inode_core(target_ip,
1109                                                 HAMMER_FLUSH_RECURSION);
1110                         r = 1;
1111                 }
1112                 break;
1113         case HAMMER_FST_FLUSH:
1114                 /*
1115                  * Record already associated with a flush group.  It had
1116                  * better be ours.
1117                  */
1118                 KKASSERT(rec->flush_group == ip->flush_group);
1119                 r = 1;
1120                 break;
1121         }
1122         return(r);
1123 }
1124
1125 /*
1126  * Wait for a previously queued flush to complete
1127  */
1128 void
1129 hammer_wait_inode(hammer_inode_t ip)
1130 {
1131         while (ip->flush_state != HAMMER_FST_IDLE) {
1132                 ip->flags |= HAMMER_INODE_FLUSHW;
1133                 tsleep(&ip->flags, 0, "hmrwin", 0);
1134         }
1135 }
1136
1137 /*
1138  * Called by the backend code when a flush has been completed.
1139  * The inode has already been removed from the flush list.
1140  *
1141  * A pipelined flush can occur, in which case we must re-enter the
1142  * inode on the list and re-copy its fields.
1143  */
1144 void
1145 hammer_flush_inode_done(hammer_inode_t ip)
1146 {
1147         struct bio *bio;
1148         int dorel = 0;
1149
1150         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1151
1152         /*
1153          * Allow BIOs to queue to the inode's primary bioq again.
1154          */
1155         ip->flags &= ~HAMMER_INODE_WRITE_ALT;
1156
1157         /*
1158          * Merge left-over flags back into the frontend and fix the state.
1159          */
1160         ip->flags |= ip->sync_flags;
1161
1162         /*
1163          * The backend may have adjusted nlinks, so if the adjusted nlinks
1164          * does not match the fronttend set the frontend's RDIRTY flag again.
1165          */
1166         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
1167                 ip->flags |= HAMMER_INODE_DDIRTY;
1168
1169         /*
1170          * Reflush any BIOs that wound up in the alt list.  Our inode will
1171          * also wind up at the end of the flusher's list.
1172          */
1173         while ((bio = TAILQ_FIRST(&ip->bio_alt_list)) != NULL) {
1174                 TAILQ_REMOVE(&ip->bio_alt_list, bio, bio_act);
1175                 TAILQ_INSERT_TAIL(&ip->bio_list, bio, bio_act);
1176         }
1177         /*
1178          * Fix up the dirty buffer status.
1179          */
1180         if (TAILQ_FIRST(&ip->bio_list) ||
1181             (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree))) {
1182                 ip->flags |= HAMMER_INODE_BUFS;
1183         }
1184
1185         /*
1186          * Re-set the XDIRTY flag if some of the inode's in-memory records
1187          * could not be flushed.
1188          */
1189         if (RB_ROOT(&ip->rec_tree))
1190                 ip->flags |= HAMMER_INODE_XDIRTY;
1191
1192         /*
1193          * Do not lose track of inodes which no longer have vnode
1194          * assocations, otherwise they may never get flushed again.
1195          */
1196         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1197                 ip->flags |= HAMMER_INODE_REFLUSH;
1198
1199         /*
1200          * Adjust flush_state.  The target state (idle or setup) shouldn't
1201          * be terribly important since we will reflush if we really need
1202          * to do anything. XXX
1203          */
1204         if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
1205                 ip->flush_state = HAMMER_FST_IDLE;
1206                 dorel = 1;
1207         } else {
1208                 ip->flush_state = HAMMER_FST_SETUP;
1209         }
1210
1211         /*
1212          * Clean up the vnode ref
1213          */
1214         if (ip->flags & HAMMER_INODE_VHELD) {
1215                 ip->flags &= ~HAMMER_INODE_VHELD;
1216                 vrele(ip->vp);
1217         }
1218
1219         /*
1220          * If the frontend made more changes and requested another flush,
1221          * then try to get it running.
1222          */
1223         if (ip->flags & HAMMER_INODE_REFLUSH) {
1224                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1225                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1226                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1227                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1228                 } else {
1229                         hammer_flush_inode(ip, 0);
1230                 }
1231         }
1232
1233         /*
1234          * Finally, if the frontend is waiting for a flush to complete,
1235          * wake it up.
1236          */
1237         if (ip->flush_state != HAMMER_FST_FLUSH) {
1238                 if (ip->flags & HAMMER_INODE_FLUSHW) {
1239                         ip->flags &= ~HAMMER_INODE_FLUSHW;
1240                         wakeup(&ip->flags);
1241                 }
1242         }
1243         if (dorel)
1244                 hammer_rel_inode(ip, 0);
1245 }
1246
1247 /*
1248  * Called from hammer_sync_inode() to synchronize in-memory records
1249  * to the media.
1250  */
1251 static int
1252 hammer_sync_record_callback(hammer_record_t record, void *data)
1253 {
1254         hammer_cursor_t cursor = data;
1255         hammer_transaction_t trans = cursor->trans;
1256         int error;
1257
1258         /*
1259          * Skip records that do not belong to the current flush.
1260          */
1261         if (record->flush_state != HAMMER_FST_FLUSH)
1262                 return(0);
1263         KKASSERT((record->flags & HAMMER_RECF_DELETED_BE) == 0);
1264 #if 1
1265         if (record->flush_group != record->ip->flush_group) {
1266                 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1267                 Debugger("blah2");
1268                 return(0);
1269         }
1270 #endif
1271         KKASSERT(record->flush_group == record->ip->flush_group);
1272
1273         /*
1274          * Interlock the record using the BE flag.  Once BE is set the
1275          * frontend cannot change the state of FE.
1276          *
1277          * NOTE: If FE is set prior to us setting BE we still sync the
1278          * record out, but the flush completion code converts it to
1279          * a delete-on-disk record instead of destroying it.
1280          */
1281         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1282         record->flags |= HAMMER_RECF_INTERLOCK_BE;
1283
1284         /*
1285          * If the whole inode is being deleting all on-disk records will
1286          * be deleted very soon, we can't sync any new records to disk
1287          * because they will be deleted in the same transaction they were
1288          * created in (delete_tid == create_tid), which will assert.
1289          *
1290          * XXX There may be a case with RECORD_ADD with DELETED_FE set
1291          * that we currently panic on.
1292          */
1293         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
1294                 switch(record->type) {
1295                 case HAMMER_MEM_RECORD_GENERAL:
1296                         record->flags |= HAMMER_RECF_DELETED_FE;
1297                         record->flags |= HAMMER_RECF_DELETED_BE;
1298                         error = 0;
1299                         goto done;
1300                 case HAMMER_MEM_RECORD_ADD:
1301                         panic("hammer_sync_record_callback: illegal add "
1302                               "during inode deletion record %p", record);
1303                         break; /* NOT REACHED */
1304                 case HAMMER_MEM_RECORD_INODE:
1305                         panic("hammer_sync_record_callback: attempt to "
1306                               "sync inode record %p?", record);
1307                         break; /* NOT REACHED */
1308                 case HAMMER_MEM_RECORD_DEL:
1309                         /*
1310                          * Follow through and issue the on-disk deletion
1311                          */
1312                         break;
1313                 }
1314         }
1315
1316         /*
1317          * If DELETED_FE is set we may have already sent dependant pieces
1318          * to the disk and we must flush the record as if it hadn't been
1319          * deleted.  This creates a bit of a mess because we have to
1320          * have ip_sync_record convert the record to MEM_RECORD_DEL before
1321          * it inserts the B-Tree record.  Otherwise the media sync might
1322          * be visible to the frontend.
1323          */
1324         if (record->flags & HAMMER_RECF_DELETED_FE) {
1325                 if (record->type == HAMMER_MEM_RECORD_ADD) {
1326                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
1327                 } else {
1328                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
1329                         return(0);
1330                 }
1331         }
1332
1333         /*
1334          * Assign the create_tid for new records.  Deletions already
1335          * have the record's entire key properly set up.
1336          */
1337         if (record->type != HAMMER_MEM_RECORD_DEL)
1338                 record->leaf.base.create_tid = trans->tid;
1339         for (;;) {
1340                 error = hammer_ip_sync_record_cursor(cursor, record);
1341                 if (error != EDEADLK)
1342                         break;
1343                 hammer_done_cursor(cursor);
1344                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
1345                                            record->ip);
1346                 if (error)
1347                         break;
1348         }
1349         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1350
1351         if (error) {
1352                 error = -error;
1353                 if (error != -ENOSPC) {
1354                         kprintf("hammer_sync_record_callback: sync failed rec "
1355                                 "%p, error %d\n", record, error);
1356                         Debugger("sync failed rec");
1357                 }
1358         }
1359 done:
1360         hammer_flush_record_done(record, error);
1361         return(error);
1362 }
1363
1364 /*
1365  * XXX error handling
1366  */
1367 int
1368 hammer_sync_inode(hammer_inode_t ip)
1369 {
1370         struct hammer_transaction trans;
1371         struct hammer_cursor cursor;
1372         struct bio *bio;
1373         hammer_record_t depend;
1374         hammer_record_t next;
1375         int error, tmp_error;
1376         u_int64_t nlinks;
1377
1378         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
1379                 return(0);
1380
1381         hammer_start_transaction_fls(&trans, ip->hmp);
1382         error = hammer_init_cursor(&trans, &cursor, &ip->cache[0], ip);
1383         if (error)
1384                 goto done;
1385
1386         /*
1387          * Any directory records referencing this inode which are not in
1388          * our current flush group must adjust our nlink count for the
1389          * purposes of synchronization to disk.
1390          *
1391          * Records which are in our flush group can be unlinked from our
1392          * inode now, potentially allowing the inode to be physically
1393          * deleted.
1394          */
1395         nlinks = ip->ino_data.nlinks;
1396         next = TAILQ_FIRST(&ip->target_list);
1397         while ((depend = next) != NULL) {
1398                 next = TAILQ_NEXT(depend, target_entry);
1399                 if (depend->flush_state == HAMMER_FST_FLUSH &&
1400                     depend->flush_group == ip->hmp->flusher_act) {
1401                         /*
1402                          * If this is an ADD that was deleted by the frontend
1403                          * the frontend nlinks count will have already been
1404                          * decremented, but the backend is going to sync its
1405                          * directory entry and must account for it.  The
1406                          * record will be converted to a delete-on-disk when
1407                          * it gets synced.
1408                          *
1409                          * If the ADD was not deleted by the frontend we
1410                          * can remove the dependancy from our target_list.
1411                          */
1412                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
1413                                 ++nlinks;
1414                         } else {
1415                                 TAILQ_REMOVE(&ip->target_list, depend,
1416                                              target_entry);
1417                                 depend->target_ip = NULL;
1418                         }
1419                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
1420                         /*
1421                          * Not part of our flush group
1422                          */
1423                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
1424                         switch(depend->type) {
1425                         case HAMMER_MEM_RECORD_ADD:
1426                                 --nlinks;
1427                                 break;
1428                         case HAMMER_MEM_RECORD_DEL:
1429                                 ++nlinks;
1430                                 break;
1431                         default:
1432                                 break;
1433                         }
1434                 }
1435         }
1436
1437         /*
1438          * Set dirty if we had to modify the link count.
1439          */
1440         if (ip->sync_ino_data.nlinks != nlinks) {
1441                 KKASSERT((int64_t)nlinks >= 0);
1442                 ip->sync_ino_data.nlinks = nlinks;
1443                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1444         }
1445
1446         /*
1447          * Queue up as many dirty buffers as we can then set a flag to
1448          * cause any further BIOs to go to the alternative queue.
1449          */
1450         if (ip->flags & HAMMER_INODE_VHELD)
1451                 error = vfsync(ip->vp, MNT_NOWAIT, 1, NULL, NULL);
1452         ip->flags |= HAMMER_INODE_WRITE_ALT;
1453
1454         /*
1455          * The buffer cache may contain dirty buffers beyond the inode
1456          * state we copied from the frontend to the backend.  Because
1457          * we are syncing our buffer cache on the backend, resync
1458          * the truncation point and the file size so we don't wipe out
1459          * any data.
1460          *
1461          * Syncing the buffer cache on the frontend has serious problems
1462          * because it prevents us from passively queueing dirty inodes
1463          * to the backend (the BIO's could stall indefinitely).
1464          */
1465         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1466                 ip->sync_trunc_off = ip->trunc_off;
1467                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1468         }
1469         if (ip->sync_ino_data.size != ip->ino_data.size) {
1470                 ip->sync_ino_data.size = ip->ino_data.size;
1471                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1472         }
1473
1474         /*
1475          * If there is a trunction queued destroy any data past the (aligned)
1476          * truncation point.  Userland will have dealt with the buffer
1477          * containing the truncation point for us.
1478          *
1479          * We don't flush pending frontend data buffers until after we've
1480          * dealth with the truncation.
1481          *
1482          * Don't bother if the inode is or has been deleted.
1483          */
1484         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
1485                 /*
1486                  * Interlock trunc_off.  The VOP front-end may continue to
1487                  * make adjustments to it while we are blocked.
1488                  */
1489                 off_t trunc_off;
1490                 off_t aligned_trunc_off;
1491
1492                 trunc_off = ip->sync_trunc_off;
1493                 aligned_trunc_off = (trunc_off + HAMMER_BUFMASK) &
1494                                     ~HAMMER_BUFMASK64;
1495
1496                 /*
1497                  * Delete any whole blocks on-media.  The front-end has
1498                  * already cleaned out any partial block and made it
1499                  * pending.  The front-end may have updated trunc_off
1500                  * while we were blocked so do not just unconditionally
1501                  * set it to the maximum offset.
1502                  */
1503                 error = hammer_ip_delete_range(&cursor, ip,
1504                                                 aligned_trunc_off,
1505                                                 0x7FFFFFFFFFFFFFFFLL);
1506                 if (error)
1507                         Debugger("hammer_ip_delete_range errored");
1508                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1509                 if (ip->trunc_off >= trunc_off) {
1510                         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1511                         ip->flags &= ~HAMMER_INODE_TRUNCATED;
1512                 }
1513         } else {
1514                 error = 0;
1515         }
1516
1517         /*
1518          * Now sync related records.  These will typically be directory
1519          * entries or delete-on-disk records.
1520          *
1521          * Not all records will be flushed, but clear XDIRTY anyway.  We
1522          * will set it again in the frontend hammer_flush_inode_done()
1523          * if records remain.
1524          */
1525         if (error == 0) {
1526                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1527                                     hammer_sync_record_callback, &cursor);
1528                 if (tmp_error < 0)
1529                         tmp_error = -error;
1530                 if (tmp_error)
1531                         error = tmp_error;
1532                 if (RB_EMPTY(&ip->rec_tree))
1533                         ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
1534         }
1535
1536         /*
1537          * If we are deleting the inode the frontend had better not have
1538          * any active references on elements making up the inode.
1539          */
1540         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
1541                 RB_EMPTY(&ip->rec_tree)  &&
1542             (ip->sync_flags & HAMMER_INODE_DELETING) &&
1543             (ip->flags & HAMMER_INODE_DELETED) == 0) {
1544                 int count1 = 0;
1545
1546                 hkprintf("Y");
1547                 ip->flags |= HAMMER_INODE_DELETED;
1548                 error = hammer_ip_delete_range_all(&cursor, ip, &count1);
1549                 if (error == 0) {
1550                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
1551                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1552                         KKASSERT(RB_EMPTY(&ip->rec_tree));
1553
1554                         /*
1555                          * Set delete_tid in both the frontend and backend
1556                          * copy of the inode record.  The DELETED flag handles
1557                          * this, do not set RDIRTY.
1558                          */
1559                         ip->ino_leaf.base.delete_tid = trans.tid;
1560                         ip->sync_ino_leaf.base.delete_tid = trans.tid;
1561
1562                         /*
1563                          * Adjust the inode count in the volume header
1564                          */
1565                         if (ip->flags & HAMMER_INODE_ONDISK) {
1566                                 hammer_modify_volume_field(&trans,
1567                                                            trans.rootvol,
1568                                                            vol0_stat_inodes);
1569                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1570                                 hammer_modify_volume_done(trans.rootvol);
1571                         }
1572                 } else {
1573                         ip->flags &= ~HAMMER_INODE_DELETED;
1574                         Debugger("hammer_ip_delete_range_all errored");
1575                 }
1576         }
1577
1578         /*
1579          * Flush any queued BIOs.  These will just biodone() the IO's if
1580          * the inode has been deleted.
1581          */
1582         while ((bio = TAILQ_FIRST(&ip->bio_list)) != NULL) {
1583                 TAILQ_REMOVE(&ip->bio_list, bio, bio_act);
1584                 tmp_error = hammer_dowrite(&cursor, ip, bio);
1585                 if (tmp_error)
1586                         error = tmp_error;
1587         }
1588         ip->sync_flags &= ~HAMMER_INODE_BUFS;
1589
1590         if (error)
1591                 Debugger("RB_SCAN errored");
1592
1593         /*
1594          * Now update the inode's on-disk inode-data and/or on-disk record.
1595          * DELETED and ONDISK are managed only in ip->flags.
1596          */
1597         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
1598         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
1599                 /*
1600                  * If deleted and on-disk, don't set any additional flags.
1601                  * the delete flag takes care of things.
1602                  *
1603                  * Clear flags which may have been set by the frontend.
1604                  */
1605                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY|
1606                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1607                                     HAMMER_INODE_DELETING);
1608                 break;
1609         case HAMMER_INODE_DELETED:
1610                 /*
1611                  * Take care of the case where a deleted inode was never
1612                  * flushed to the disk in the first place.
1613                  *
1614                  * Clear flags which may have been set by the frontend.
1615                  */
1616                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY|
1617                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1618                                     HAMMER_INODE_DELETING);
1619                 while (RB_ROOT(&ip->rec_tree)) {
1620                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
1621                         hammer_ref(&record->lock);
1622                         KKASSERT(record->lock.refs == 1);
1623                         record->flags |= HAMMER_RECF_DELETED_FE;
1624                         record->flags |= HAMMER_RECF_DELETED_BE;
1625                         hammer_rel_mem_record(record);
1626                 }
1627                 break;
1628         case HAMMER_INODE_ONDISK:
1629                 /*
1630                  * If already on-disk, do not set any additional flags.
1631                  */
1632                 break;
1633         default:
1634                 /*
1635                  * If not on-disk and not deleted, set both dirty flags
1636                  * to force an initial record to be written.  Also set
1637                  * the create_tid for the inode.
1638                  *
1639                  * Set create_tid in both the frontend and backend
1640                  * copy of the inode record.
1641                  */
1642                 ip->ino_leaf.base.create_tid = trans.tid;
1643                 ip->sync_ino_leaf.base.create_tid = trans.tid;
1644                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1645                 break;
1646         }
1647
1648         /*
1649          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
1650          * is already on-disk the old record is marked as deleted.
1651          *
1652          * If DELETED is set hammer_update_inode() will delete the existing
1653          * record without writing out a new one.
1654          *
1655          * If *ONLY* the ITIMES flag is set we can update the record in-place.
1656          */
1657         if (ip->flags & HAMMER_INODE_DELETED) {
1658                 error = hammer_update_inode(&cursor, ip);
1659         } else
1660         if ((ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES)) ==
1661             HAMMER_INODE_ITIMES) {
1662                 error = hammer_update_itimes(&cursor, ip);
1663         } else
1664         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES)) {
1665                 error = hammer_update_inode(&cursor, ip);
1666         }
1667         if (error)
1668                 Debugger("hammer_update_itimes/inode errored");
1669 done:
1670         /*
1671          * Save the TID we used to sync the inode with to make sure we
1672          * do not improperly reuse it.
1673          */
1674         hammer_done_cursor(&cursor);
1675         hammer_done_transaction(&trans);
1676         return(error);
1677 }
1678
1679 /*
1680  * This routine is called when the OS is no longer actively referencing
1681  * the inode (but might still be keeping it cached), or when releasing
1682  * the last reference to an inode.
1683  *
1684  * At this point if the inode's nlinks count is zero we want to destroy
1685  * it, which may mean destroying it on-media too.
1686  */
1687 void
1688 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
1689 {
1690         struct vnode *vp;
1691
1692         /*
1693          * Set the DELETING flag when the link count drops to 0 and the
1694          * OS no longer has any opens on the inode.
1695          *
1696          * The backend will clear DELETING (a mod flag) and set DELETED
1697          * (a state flag) when it is actually able to perform the
1698          * operation.
1699          */
1700         if (ip->ino_data.nlinks == 0 &&
1701             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
1702                 ip->flags |= HAMMER_INODE_DELETING;
1703                 ip->flags |= HAMMER_INODE_TRUNCATED;
1704                 ip->trunc_off = 0;
1705                 vp = NULL;
1706                 if (getvp) {
1707                         if (hammer_get_vnode(ip, &vp) != 0)
1708                                 return;
1709                 }
1710                 if (ip->vp) {
1711                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
1712                         vnode_pager_setsize(ip->vp, 0);
1713                 }
1714                 if (getvp) {
1715                         vput(vp);
1716                 }
1717         }
1718 }
1719
1720 /*
1721  * Re-test an inode when a dependancy had gone away to see if we
1722  * can chain flush it.
1723  */
1724 void
1725 hammer_test_inode(hammer_inode_t ip)
1726 {
1727         if (ip->flags & HAMMER_INODE_REFLUSH) {
1728                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1729                 hammer_ref(&ip->lock);
1730                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1731                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1732                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1733                 } else {
1734                         hammer_flush_inode(ip, 0);
1735                 }
1736                 hammer_rel_inode(ip, 0);
1737         }
1738 }
1739