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 1.76 2008/06/17 04:02:38 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_inode_t ip);
  46 static int      hammer_setup_parent_inodes_helper(hammer_record_t record);
  47 static void     hammer_inode_wakereclaims(hammer_inode_t ip);
  48
  49 #ifdef DEBUG_TRUNCATE
  50 extern struct hammer_inode *HammerTruncIp;
  51 #endif
  52
  53 /*
  54  * The kernel is not actively referencing this vnode but is still holding
  55  * it cached.
  56  *
  57  * This is called from the frontend.
  58  */
  59 int
  60 hammer_vop_inactive(struct vop_inactive_args *ap)
  61 {
  62         struct hammer_inode *ip = VTOI(ap->a_vp);
  63
  64         /*
  65          * Degenerate case
  66          */
  67         if (ip == NULL) {
  68                 vrecycle(ap->a_vp);
  69                 return(0);
  70         }
  71
  72         /*
  73          * If the inode no longer has visibility in the filesystem and is
  74          * fairly clean, try to recycle it immediately.  This can deadlock
  75          * in vfsync() if we aren't careful.
  76          *
  77          * Do not queue the inode to the flusher if we still have visibility,
  78          * otherwise namespace calls such as chmod will unnecessarily generate
  79          * multiple inode updates.
  80          */
  81         hammer_inode_unloadable_check(ip, 0);
  82         if (ip->ino_data.nlinks == 0) {
  83                 if (ip->flags & HAMMER_INODE_MODMASK)
  84                         hammer_flush_inode(ip, 0);
  85                 else
  86                         vrecycle(ap->a_vp);
  87         }
  88         return(0);
  89 }
  90
  91 /*
  92  * Release the vnode association.  This is typically (but not always)
  93  * the last reference on the inode.
  94  *
  95  * Once the association is lost we are on our own with regards to
  96  * flushing the inode.
  97  */
  98 int
  99 hammer_vop_reclaim(struct vop_reclaim_args *ap)
 100 {
 101         struct hammer_reclaim reclaim;
 102         struct hammer_inode *ip;
 103         hammer_mount_t hmp;
 104         struct vnode *vp;
 105         int delay;
 106
 107         vp = ap->a_vp;
 108
 109         if ((ip = vp->v_data) != NULL) {
 110                 hmp = ip->hmp;
 111                 vp->v_data = NULL;
 112                 ip->vp = NULL;
 113
 114                 /*
 115                  * Setup our reclaim pipeline.  We only let so many detached
 116                  * (and dirty) inodes build up before we start blocking.  Do
 117                  * not bother tracking the immediate increment/decrement if
 118                  * the inode is not actually dirty.
 119                  *
 120                  * When we block we don't care *which* inode has finished
 121                  * reclaiming, as lone as one does.
 122                  */
 123                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
 124                     ((ip->flags|ip->sync_flags) & HAMMER_INODE_MODMASK)) {
 125                         ++hammer_count_reclaiming;
 126                         ++hmp->inode_reclaims;
 127                         ip->flags |= HAMMER_INODE_RECLAIM;
 128                         if (hmp->inode_reclaims > HAMMER_RECLAIM_PIPESIZE) {
 129                                 reclaim.okydoky = 0;
 130                                 TAILQ_INSERT_TAIL(&hmp->reclaim_list,
 131                                                   &reclaim, entry);
 132                         } else {
 133                                 reclaim.okydoky = 1;
 134                         }
 135                 } else {
 136                         reclaim.okydoky = 1;
 137                 }
 138                 hammer_rel_inode(ip, 1);
 139
 140                 /*
 141                  * Reclaim pipeline.  We can't let too many reclaimed inodes
 142                  * build-up in the flusher or the flusher loses its locality
 143                  * of reference, or worse blows out our memory.  Once we have
 144                  * exceeded the reclaim pipe size start slowing down.  Our
 145                  * imposed delay can be cut short if the flusher catches up
 146                  * to us.
 147                  */
 148                 if (reclaim.okydoky == 0) {
 149                         delay = (hmp->inode_reclaims -
 150                                  HAMMER_RECLAIM_PIPESIZE) * hz /
 151                                 HAMMER_RECLAIM_PIPESIZE;
 152                         if (delay <= 0)
 153                                 delay = 1;
 154                         hammer_flusher_async(hmp);
 155                         if (reclaim.okydoky == 0) {
 156                                 tsleep(&reclaim, 0, "hmrrcm", delay);
 157                         }
 158                         if (reclaim.okydoky == 0) {
 159                                 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim,
 160                                              entry);
 161                         }
 162                 }
 163         }
 164         return(0);
 165 }
 166
 167 /*
 168  * Return a locked vnode for the specified inode.  The inode must be
 169  * referenced but NOT LOCKED on entry and will remain referenced on
 170  * return.
 171  *
 172  * Called from the frontend.
 173  */
 174 int
 175 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 176 {
 177         hammer_mount_t hmp;
 178         struct vnode *vp;
 179         int error = 0;
 180
 181         hmp = ip->hmp;
 182
 183         for (;;) {
 184                 if ((vp = ip->vp) == NULL) {
 185                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
 186                         if (error)
 187                                 break;
 188                         hammer_lock_ex(&ip->lock);
 189                         if (ip->vp != NULL) {
 190                                 hammer_unlock(&ip->lock);
 191                                 vp->v_type = VBAD;
 192                                 vx_put(vp);
 193                                 continue;
 194                         }
 195                         hammer_ref(&ip->lock);
 196                         vp = *vpp;
 197                         ip->vp = vp;
 198                         vp->v_type =
 199                                 hammer_get_vnode_type(ip->ino_data.obj_type);
 200
 201                         hammer_inode_wakereclaims(ip);
 202
 203                         switch(ip->ino_data.obj_type) {
 204                         case HAMMER_OBJTYPE_CDEV:
 205                         case HAMMER_OBJTYPE_BDEV:
 206                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
 207                                 addaliasu(vp, ip->ino_data.rmajor,
 208                                           ip->ino_data.rminor);
 209                                 break;
 210                         case HAMMER_OBJTYPE_FIFO:
 211                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
 212                                 break;
 213                         default:
 214                                 break;
 215                         }
 216
 217                         /*
 218                          * Only mark as the root vnode if the ip is not
 219                          * historical, otherwise the VFS cache will get
 220                          * confused.  The other half of the special handling
 221                          * is in hammer_vop_nlookupdotdot().
 222                          */
 223                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 224                             ip->obj_asof == hmp->asof) {
 225                                 vp->v_flag |= VROOT;
 226                         }
 227
 228                         vp->v_data = (void *)ip;
 229                         /* vnode locked by getnewvnode() */
 230                         /* make related vnode dirty if inode dirty? */
 231                         hammer_unlock(&ip->lock);
 232                         if (vp->v_type == VREG)
 233                                 vinitvmio(vp, ip->ino_data.size);
 234                         break;
 235                 }
 236
 237                 /*
 238                  * loop if the vget fails (aka races), or if the vp
 239                  * no longer matches ip->vp.
 240                  */
 241                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 242                         if (vp == ip->vp)
 243                                 break;
 244                         vput(vp);
 245                 }
 246         }
 247         *vpp = vp;
 248         return(error);
 249 }
 250
 251 /*
 252  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 253  * do not attach or detach the related vnode (use hammer_get_vnode() for
 254  * that).
 255  *
 256  * The flags argument is only applied for newly created inodes, and only
 257  * certain flags are inherited.
 258  *
 259  * Called from the frontend.
 260  */
 261 struct hammer_inode *
 262 hammer_get_inode(hammer_transaction_t trans, struct hammer_node **cache,
 263                  u_int64_t obj_id, hammer_tid_t asof, int flags, int *errorp)
 264 {
 265         hammer_mount_t hmp = trans->hmp;
 266         struct hammer_inode_info iinfo;
 267         struct hammer_cursor cursor;
 268         struct hammer_inode *ip;
 269
 270         /*
 271          * Determine if we already have an inode cached.  If we do then
 272          * we are golden.
 273          */
 274         iinfo.obj_id = obj_id;
 275         iinfo.obj_asof = asof;
 276 loop:
 277         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 278         if (ip) {
 279                 hammer_ref(&ip->lock);
 280                 *errorp = 0;
 281                 return(ip);
 282         }
 283
 284         /*
 285          * Allocate a new inode structure and deal with races later.
 286          */
 287         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 288         ++hammer_count_inodes;
 289         ++hmp->count_inodes;
 290         ip->obj_id = obj_id;
 291         ip->obj_asof = iinfo.obj_asof;
 292         ip->hmp = hmp;
 293         ip->flags = flags & HAMMER_INODE_RO;
 294         if (hmp->ronly)
 295                 ip->flags |= HAMMER_INODE_RO;
 296         ip->sync_trunc_off = ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 297         RB_INIT(&ip->rec_tree);
 298         TAILQ_INIT(&ip->target_list);
 299
 300         /*
 301          * Locate the on-disk inode.
 302          */
 303 retry:
 304         hammer_init_cursor(trans, &cursor, cache, NULL);
 305         cursor.key_beg.localization = HAMMER_LOCALIZE_INODE;
 306         cursor.key_beg.obj_id = ip->obj_id;
 307         cursor.key_beg.key = 0;
 308         cursor.key_beg.create_tid = 0;
 309         cursor.key_beg.delete_tid = 0;
 310         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 311         cursor.key_beg.obj_type = 0;
 312         cursor.asof = iinfo.obj_asof;
 313         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 314                        HAMMER_CURSOR_ASOF;
 315
 316         *errorp = hammer_btree_lookup(&cursor);
 317         if (*errorp == EDEADLK) {
 318                 hammer_done_cursor(&cursor);
 319                 goto retry;
 320         }
 321
 322         /*
 323          * On success the B-Tree lookup will hold the appropriate
 324          * buffer cache buffers and provide a pointer to the requested
 325          * information.  Copy the information to the in-memory inode
 326          * and cache the B-Tree node to improve future operations.
 327          */
 328         if (*errorp == 0) {
 329                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 330                 ip->ino_data = cursor.data->inode;
 331                 hammer_cache_node(cursor.node, &ip->cache[0]);
 332                 if (cache)
 333                         hammer_cache_node(cursor.node, cache);
 334
 335                 /*
 336                  * The file should not contain any data past the file size
 337                  * stored in the inode.  Setting sync_trunc_off to the
 338                  * file size instead of max reduces B-Tree lookup overheads
 339                  * on append by allowing the flusher to avoid checking for
 340                  * record overwrites.
 341                  */
 342                 ip->sync_trunc_off = ip->ino_data.size;
 343         }
 344
 345         /*
 346          * The inode is placed on the red-black tree and will be synced to
 347          * the media when flushed or by the filesystem sync.  If this races
 348          * another instantiation/lookup the insertion will fail.
 349          */
 350         if (*errorp == 0) {
 351                 hammer_ref(&ip->lock);
 352                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 353                         hammer_uncache_node(&ip->cache[0]);
 354                         hammer_uncache_node(&ip->cache[1]);
 355                         KKASSERT(ip->lock.refs == 1);
 356                         --hammer_count_inodes;
 357                         --hmp->count_inodes;
 358                         kfree(ip, M_HAMMER);
 359                         hammer_done_cursor(&cursor);
 360                         goto loop;
 361                 }
 362                 ip->flags |= HAMMER_INODE_ONDISK;
 363         } else {
 364                 /*
 365                  * Do not panic on read-only accesses which fail, particularly
 366                  * historical accesses where the snapshot might not have
 367                  * complete connectivity.
 368                  */
 369                 if ((flags & HAMMER_INODE_RO) == 0) {
 370                         kprintf("hammer_get_inode: failed ip %p obj_id %016llx cursor %p error %d\n",
 371                                 ip, ip->obj_id, &cursor, *errorp);
 372                         Debugger("x");
 373                 }
 374                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 375                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 376                         --hmp->rsv_inodes;
 377                 }
 378                 hmp->rsv_databufs -= ip->rsv_databufs;
 379                 ip->rsv_databufs = 0;                          /* sanity */
 380
 381                 --hammer_count_inodes;
 382                 --hmp->count_inodes;
 383                 kfree(ip, M_HAMMER);
 384                 ip = NULL;
 385         }
 386         hammer_done_cursor(&cursor);
 387         return (ip);
 388 }
 389
 390 /*
 391  * Create a new filesystem object, returning the inode in *ipp.  The
 392  * returned inode will be referenced.
 393  *
 394  * The inode is created in-memory.
 395  */
 396 int
 397 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 398                     struct ucred *cred, hammer_inode_t dip,
 399                     struct hammer_inode **ipp)
 400 {
 401         hammer_mount_t hmp;
 402         hammer_inode_t ip;
 403         uid_t xuid;
 404
 405         hmp = trans->hmp;
 406         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 407         ++hammer_count_inodes;
 408         ++hmp->count_inodes;
 409         ip->obj_id = hammer_alloc_objid(trans, dip);
 410         KKASSERT(ip->obj_id != 0);
 411         ip->obj_asof = hmp->asof;
 412         ip->hmp = hmp;
 413         ip->flush_state = HAMMER_FST_IDLE;
 414         ip->flags = HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES;
 415
 416         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 417         RB_INIT(&ip->rec_tree);
 418         TAILQ_INIT(&ip->target_list);
 419
 420         ip->ino_leaf.atime = trans->time;
 421         ip->ino_data.mtime = trans->time;
 422         ip->ino_data.size = 0;
 423         ip->ino_data.nlinks = 0;
 424
 425         /*
 426          * A nohistory designator on the parent directory is inherited by
 427          * the child.
 428          */
 429         ip->ino_data.uflags = dip->ino_data.uflags &
 430                               (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 431
 432         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 433         ip->ino_leaf.base.localization = HAMMER_LOCALIZE_INODE;
 434         ip->ino_leaf.base.obj_id = ip->obj_id;
 435         ip->ino_leaf.base.key = 0;
 436         ip->ino_leaf.base.create_tid = 0;
 437         ip->ino_leaf.base.delete_tid = 0;
 438         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 439         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 440
 441         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 442         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 443         ip->ino_data.mode = vap->va_mode;
 444         ip->ino_data.ctime = trans->time;
 445         ip->ino_data.parent_obj_id = (dip) ? dip->ino_leaf.base.obj_id : 0;
 446
 447         switch(ip->ino_leaf.base.obj_type) {
 448         case HAMMER_OBJTYPE_CDEV:
 449         case HAMMER_OBJTYPE_BDEV:
 450                 ip->ino_data.rmajor = vap->va_rmajor;
 451                 ip->ino_data.rminor = vap->va_rminor;
 452                 break;
 453         default:
 454                 break;
 455         }
 456
 457         /*
 458          * Calculate default uid/gid and overwrite with information from
 459          * the vap.
 460          */
 461         xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 462         xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode, xuid, cred,
 463                                      &vap->va_mode);
 464         ip->ino_data.mode = vap->va_mode;
 465
 466         if (vap->va_vaflags & VA_UID_UUID_VALID)
 467                 ip->ino_data.uid = vap->va_uid_uuid;
 468         else if (vap->va_uid != (uid_t)VNOVAL)
 469                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 470         else
 471                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 472
 473         if (vap->va_vaflags & VA_GID_UUID_VALID)
 474                 ip->ino_data.gid = vap->va_gid_uuid;
 475         else if (vap->va_gid != (gid_t)VNOVAL)
 476                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 477         else
 478                 ip->ino_data.gid = dip->ino_data.gid;
 479
 480         hammer_ref(&ip->lock);
 481         if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 482                 hammer_unref(&ip->lock);
 483                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 484         }
 485         *ipp = ip;
 486         return(0);
 487 }
 488
 489 /*
 490  * Called by hammer_sync_inode().
 491  */
 492 static int
 493 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 494 {
 495         hammer_transaction_t trans = cursor->trans;
 496         hammer_record_t record;
 497         int error;
 498
 499 retry:
 500         error = 0;
 501
 502         /*
 503          * If the inode has a presence on-disk then locate it and mark
 504          * it deleted, setting DELONDISK.
 505          *
 506          * The record may or may not be physically deleted, depending on
 507          * the retention policy.
 508          */
 509         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 510             HAMMER_INODE_ONDISK) {
 511                 hammer_normalize_cursor(cursor);
 512                 cursor->key_beg.localization = HAMMER_LOCALIZE_INODE;
 513                 cursor->key_beg.obj_id = ip->obj_id;
 514                 cursor->key_beg.key = 0;
 515                 cursor->key_beg.create_tid = 0;
 516                 cursor->key_beg.delete_tid = 0;
 517                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 518                 cursor->key_beg.obj_type = 0;
 519                 cursor->asof = ip->obj_asof;
 520                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 521                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 522                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 523
 524                 error = hammer_btree_lookup(cursor);
 525                 if (hammer_debug_inode)
 526                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 527                 if (error) {
 528                         kprintf("error %d\n", error);
 529                         Debugger("hammer_update_inode");
 530                 }
 531
 532                 if (error == 0) {
 533                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
 534                         if (hammer_debug_inode)
 535                                 kprintf(" error %d\n", error);
 536                         if (error && error != EDEADLK) {
 537                                 kprintf("error %d\n", error);
 538                                 Debugger("hammer_update_inode2");
 539                         }
 540                         if (error == 0) {
 541                                 ip->flags |= HAMMER_INODE_DELONDISK;
 542                         }
 543                         if (cursor->node)
 544                                 hammer_cache_node(cursor->node, &ip->cache[0]);
 545                 }
 546                 if (error == EDEADLK) {
 547                         hammer_done_cursor(cursor);
 548                         error = hammer_init_cursor(trans, cursor,
 549                                                    &ip->cache[0], ip);
 550                         if (hammer_debug_inode)
 551                                 kprintf("IPDED %p %d\n", ip, error);
 552                         if (error == 0)
 553                                 goto retry;
 554                 }
 555         }
 556
 557         /*
 558          * Ok, write out the initial record or a new record (after deleting
 559          * the old one), unless the DELETED flag is set.  This routine will
 560          * clear DELONDISK if it writes out a record.
 561          *
 562          * Update our inode statistics if this is the first application of
 563          * the inode on-disk.
 564          */
 565         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 566                 /*
 567                  * Generate a record and write it to the media
 568                  */
 569                 record = hammer_alloc_mem_record(ip, 0);
 570                 record->type = HAMMER_MEM_RECORD_INODE;
 571                 record->flush_state = HAMMER_FST_FLUSH;
 572                 record->leaf = ip->sync_ino_leaf;
 573                 record->leaf.base.create_tid = trans->tid;
 574                 record->leaf.data_len = sizeof(ip->sync_ino_data);
 575                 record->data = (void *)&ip->sync_ino_data;
 576                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
 577                 for (;;) {
 578                         error = hammer_ip_sync_record_cursor(cursor, record);
 579                         if (hammer_debug_inode)
 580                                 kprintf("GENREC %p rec %08x %d\n",
 581                                         ip, record->flags, error);
 582                         if (error != EDEADLK)
 583                                 break;
 584                         hammer_done_cursor(cursor);
 585                         error = hammer_init_cursor(trans, cursor,
 586                                                    &ip->cache[0], ip);
 587                         if (hammer_debug_inode)
 588                                 kprintf("GENREC reinit %d\n", error);
 589                         if (error)
 590                                 break;
 591                 }
 592                 if (error) {
 593                         kprintf("error %d\n", error);
 594                         Debugger("hammer_update_inode3");
 595                 }
 596
 597                 /*
 598                  * The record isn't managed by the inode's record tree,
 599                  * destroy it whether we succeed or fail.
 600                  */
 601                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 602                 record->flags |= HAMMER_RECF_DELETED_FE;
 603                 record->flush_state = HAMMER_FST_IDLE;
 604                 hammer_rel_mem_record(record);
 605
 606                 /*
 607                  * Finish up.
 608                  */
 609                 if (error == 0) {
 610                         if (hammer_debug_inode)
 611                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
 612                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
 613                                             HAMMER_INODE_ITIMES);
 614                         ip->flags &= ~HAMMER_INODE_DELONDISK;
 615
 616                         /*
 617                          * Root volume count of inodes
 618                          */
 619                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
 620                                 hammer_modify_volume_field(trans,
 621                                                            trans->rootvol,
 622                                                            vol0_stat_inodes);
 623                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
 624                                 hammer_modify_volume_done(trans->rootvol);
 625                                 ip->flags |= HAMMER_INODE_ONDISK;
 626                                 if (hammer_debug_inode)
 627                                         kprintf("NOWONDISK %p\n", ip);
 628                         }
 629                 }
 630         }
 631
 632         /*
 633          * If the inode has been destroyed, clean out any left-over flags
 634          * that may have been set by the frontend.
 635          */
 636         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
 637                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
 638                                     HAMMER_INODE_ITIMES);
 639         }
 640         return(error);
 641 }
 642
 643 /*
 644  * Update only the itimes fields.  This is done no-historically.  The
 645  * record is updated in-place on the disk.
 646  */
 647 static int
 648 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
 649 {
 650         hammer_transaction_t trans = cursor->trans;
 651         struct hammer_btree_leaf_elm *leaf;
 652         int error;
 653
 654 retry:
 655         error = 0;
 656         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 657             HAMMER_INODE_ONDISK) {
 658                 hammer_normalize_cursor(cursor);
 659                 cursor->key_beg.localization = HAMMER_LOCALIZE_INODE;
 660                 cursor->key_beg.obj_id = ip->obj_id;
 661                 cursor->key_beg.key = 0;
 662                 cursor->key_beg.create_tid = 0;
 663                 cursor->key_beg.delete_tid = 0;
 664                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 665                 cursor->key_beg.obj_type = 0;
 666                 cursor->asof = ip->obj_asof;
 667                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 668                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 669                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 670
 671                 error = hammer_btree_lookup(cursor);
 672                 if (error) {
 673                         kprintf("error %d\n", error);
 674                         Debugger("hammer_update_itimes1");
 675                 }
 676                 if (error == 0) {
 677                         /*
 678                          * Do not generate UNDO records for atime updates.
 679                          */
 680                         leaf = cursor->leaf;
 681                         hammer_modify_node(trans, cursor->node,
 682                                            &leaf->atime, sizeof(leaf->atime));
 683                         leaf->atime = ip->sync_ino_leaf.atime;
 684                         hammer_modify_node_done(cursor->node);
 685                         /*rec->ino_mtime = ip->sync_ino_rec.ino_mtime;*/
 686                         ip->sync_flags &= ~HAMMER_INODE_ITIMES;
 687                         /* XXX recalculate crc */
 688                         hammer_cache_node(cursor->node, &ip->cache[0]);
 689                 }
 690                 if (error == EDEADLK) {
 691                         hammer_done_cursor(cursor);
 692                         error = hammer_init_cursor(trans, cursor,
 693                                                    &ip->cache[0], ip);
 694                         if (error == 0)
 695                                 goto retry;
 696                 }
 697         }
 698         return(error);
 699 }
 700
 701 /*
 702  * Release a reference on an inode, flush as requested.
 703  *
 704  * On the last reference we queue the inode to the flusher for its final
 705  * disposition.
 706  */
 707 void
 708 hammer_rel_inode(struct hammer_inode *ip, int flush)
 709 {
 710         hammer_mount_t hmp = ip->hmp;
 711
 712         /*
 713          * Handle disposition when dropping the last ref.
 714          */
 715         for (;;) {
 716                 if (ip->lock.refs == 1) {
 717                         /*
 718                          * Determine whether on-disk action is needed for
 719                          * the inode's final disposition.
 720                          */
 721                         KKASSERT(ip->vp == NULL);
 722                         hammer_inode_unloadable_check(ip, 0);
 723                         if (ip->flags & HAMMER_INODE_MODMASK) {
 724                                 if (hmp->rsv_inodes > desiredvnodes) {
 725                                         hammer_flush_inode(ip,
 726                                                            HAMMER_FLUSH_SIGNAL);
 727                                 } else {
 728                                         hammer_flush_inode(ip, 0);
 729                                 }
 730                         } else if (ip->lock.refs == 1) {
 731                                 hammer_unload_inode(ip);
 732                                 break;
 733                         }
 734                 } else {
 735                         if (flush)
 736                                 hammer_flush_inode(ip, 0);
 737
 738                         /*
 739                          * The inode still has multiple refs, try to drop
 740                          * one ref.
 741                          */
 742                         KKASSERT(ip->lock.refs >= 1);
 743                         if (ip->lock.refs > 1) {
 744                                 hammer_unref(&ip->lock);
 745                                 break;
 746                         }
 747                 }
 748         }
 749 }
 750
 751 /*
 752  * Unload and destroy the specified inode.  Must be called with one remaining
 753  * reference.  The reference is disposed of.
 754  *
 755  * This can only be called in the context of the flusher.
 756  */
 757 static int
 758 hammer_unload_inode(struct hammer_inode *ip)
 759 {
 760         hammer_mount_t hmp = ip->hmp;
 761
 762         KASSERT(ip->lock.refs == 1,
 763                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
 764         KKASSERT(ip->vp == NULL);
 765         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
 766         KKASSERT(ip->cursor_ip_refs == 0);
 767         KKASSERT(ip->lock.lockcount == 0);
 768         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
 769
 770         KKASSERT(RB_EMPTY(&ip->rec_tree));
 771         KKASSERT(TAILQ_EMPTY(&ip->target_list));
 772
 773         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
 774
 775         hammer_uncache_node(&ip->cache[0]);
 776         hammer_uncache_node(&ip->cache[1]);
 777         if (ip->objid_cache)
 778                 hammer_clear_objid(ip);
 779         --hammer_count_inodes;
 780         --hmp->count_inodes;
 781
 782         hammer_inode_wakereclaims(ip);
 783         kfree(ip, M_HAMMER);
 784
 785         return(0);
 786 }
 787
 788 /*
 789  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
 790  * the read-only flag for cached inodes.
 791  *
 792  * This routine is called from a RB_SCAN().
 793  */
 794 int
 795 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
 796 {
 797         hammer_mount_t hmp = ip->hmp;
 798
 799         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
 800                 ip->flags |= HAMMER_INODE_RO;
 801         else
 802                 ip->flags &= ~HAMMER_INODE_RO;
 803         return(0);
 804 }
 805
 806 /*
 807  * A transaction has modified an inode, requiring updates as specified by
 808  * the passed flags.
 809  *
 810  * HAMMER_INODE_DDIRTY: Inode data has been updated
 811  * HAMMER_INODE_XDIRTY: Dirty in-memory records
 812  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
 813  * HAMMER_INODE_DELETED: Inode record/data must be deleted
 814  * HAMMER_INODE_ITIMES: mtime/atime has been updated
 815  */
 816 void
 817 hammer_modify_inode(hammer_inode_t ip, int flags)
 818 {
 819         KKASSERT ((ip->flags & HAMMER_INODE_RO) == 0 ||
 820                   (flags & (HAMMER_INODE_DDIRTY |
 821                             HAMMER_INODE_XDIRTY | HAMMER_INODE_BUFS |
 822                             HAMMER_INODE_DELETED | HAMMER_INODE_ITIMES)) == 0);
 823         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
 824                 ip->flags |= HAMMER_INODE_RSV_INODES;
 825                 ++ip->hmp->rsv_inodes;
 826         }
 827
 828         ip->flags |= flags;
 829 }
 830
 831 /*
 832  * Request that an inode be flushed.  This whole mess cannot block and may
 833  * recurse (if not synchronous).  Once requested HAMMER will attempt to
 834  * actively flush the inode until the flush can be done.
 835  *
 836  * The inode may already be flushing, or may be in a setup state.  We can
 837  * place the inode in a flushing state if it is currently idle and flag it
 838  * to reflush if it is currently flushing.
 839  *
 840  * If the HAMMER_FLUSH_SYNCHRONOUS flag is specified we will attempt to
 841  * flush the indoe synchronously using the caller's context.
 842  */
 843 void
 844 hammer_flush_inode(hammer_inode_t ip, int flags)
 845 {
 846         int good;
 847
 848         /*
 849          * Trivial 'nothing to flush' case.  If the inode is ina SETUP
 850          * state we have to put it back into an IDLE state so we can
 851          * drop the extra ref.
 852          */
 853         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
 854                 if (ip->flush_state == HAMMER_FST_SETUP) {
 855                         ip->flush_state = HAMMER_FST_IDLE;
 856                         hammer_rel_inode(ip, 0);
 857                 }
 858                 return;
 859         }
 860
 861         /*
 862          * Our flush action will depend on the current state.
 863          */
 864         switch(ip->flush_state) {
 865         case HAMMER_FST_IDLE:
 866                 /*
 867                  * We have no dependancies and can flush immediately.  Some
 868                  * our children may not be flushable so we have to re-test
 869                  * with that additional knowledge.
 870                  */
 871                 hammer_flush_inode_core(ip, flags);
 872                 break;
 873         case HAMMER_FST_SETUP:
 874                 /*
 875                  * Recurse upwards through dependancies via target_list
 876                  * and start their flusher actions going if possible.
 877                  *
 878                  * 'good' is our connectivity.  -1 means we have none and
 879                  * can't flush, 0 means there weren't any dependancies, and
 880                  * 1 means we have good connectivity.
 881                  */
 882                 good = hammer_setup_parent_inodes(ip);
 883
 884                 /*
 885                  * We can continue if good >= 0.  Determine how many records
 886                  * under our inode can be flushed (and mark them).
 887                  */
 888                 if (good >= 0) {
 889                         hammer_flush_inode_core(ip, flags);
 890                 } else {
 891                         ip->flags |= HAMMER_INODE_REFLUSH;
 892                         if (flags & HAMMER_FLUSH_SIGNAL) {
 893                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 894                                 hammer_flusher_async(ip->hmp);
 895                         }
 896                 }
 897                 break;
 898         default:
 899                 /*
 900                  * We are already flushing, flag the inode to reflush
 901                  * if needed after it completes its current flush.
 902                  */
 903                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
 904                         ip->flags |= HAMMER_INODE_REFLUSH;
 905                 if (flags & HAMMER_FLUSH_SIGNAL) {
 906                         ip->flags |= HAMMER_INODE_RESIGNAL;
 907                         hammer_flusher_async(ip->hmp);
 908                 }
 909                 break;
 910         }
 911 }
 912
 913 /*
 914  * Scan ip->target_list, which is a list of records owned by PARENTS to our
 915  * ip which reference our ip.
 916  *
 917  * XXX This is a huge mess of recursive code, but not one bit of it blocks
 918  *     so for now do not ref/deref the structures.  Note that if we use the
 919  *     ref/rel code later, the rel CAN block.
 920  */
 921 static int
 922 hammer_setup_parent_inodes(hammer_inode_t ip)
 923 {
 924         hammer_record_t depend;
 925 #if 0
 926         hammer_record_t next;
 927         hammer_inode_t  pip;
 928 #endif
 929         int good;
 930         int r;
 931
 932         good = 0;
 933         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 934                 r = hammer_setup_parent_inodes_helper(depend);
 935                 KKASSERT(depend->target_ip == ip);
 936                 if (r < 0 && good == 0)
 937                         good = -1;
 938                 if (r > 0)
 939                         good = 1;
 940         }
 941         return(good);
 942
 943 #if 0
 944 retry:
 945         good = 0;
 946         next = TAILQ_FIRST(&ip->target_list);
 947         if (next) {
 948                 hammer_ref(&next->lock);
 949                 hammer_ref(&next->ip->lock);
 950         }
 951         while ((depend = next) != NULL) {
 952                 if (depend->target_ip == NULL) {
 953                         pip = depend->ip;
 954                         hammer_rel_mem_record(depend);
 955                         hammer_rel_inode(pip, 0);
 956                         goto retry;
 957                 }
 958                 KKASSERT(depend->target_ip == ip);
 959                 next = TAILQ_NEXT(depend, target_entry);
 960                 if (next) {
 961                         hammer_ref(&next->lock);
 962                         hammer_ref(&next->ip->lock);
 963                 }
 964                 r = hammer_setup_parent_inodes_helper(depend);
 965                 if (r < 0 && good == 0)
 966                         good = -1;
 967                 if (r > 0)
 968                         good = 1;
 969                 pip = depend->ip;
 970                 hammer_rel_mem_record(depend);
 971                 hammer_rel_inode(pip, 0);
 972         }
 973         return(good);
 974 #endif
 975 }
 976
 977 /*
 978  * This helper function takes a record representing the dependancy between
 979  * the parent inode and child inode.
 980  *
 981  * record->ip           = parent inode
 982  * record->target_ip    = child inode
 983  *
 984  * We are asked to recurse upwards and convert the record from SETUP
 985  * to FLUSH if possible.
 986  *
 987  * Return 1 if the record gives us connectivity
 988  *
 989  * Return 0 if the record is not relevant
 990  *
 991  * Return -1 if we can't resolve the dependancy and there is no connectivity.
 992  */
 993 static int
 994 hammer_setup_parent_inodes_helper(hammer_record_t record)
 995 {
 996         hammer_mount_t hmp;
 997         hammer_inode_t pip;
 998         int good;
 999
1000         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1001         pip = record->ip;
1002         hmp = pip->hmp;
1003
1004         /*
1005          * If the record is already flushing, is it in our flush group?
1006          *
1007          * If it is in our flush group but it is a general record or a
1008          * delete-on-disk, it does not improve our connectivity (return 0),
1009          * and if the target inode is not trying to destroy itself we can't
1010          * allow the operation yet anyway (the second return -1).
1011          */
1012         if (record->flush_state == HAMMER_FST_FLUSH) {
1013                 if (record->flush_group != hmp->flusher.next) {
1014                         pip->flags |= HAMMER_INODE_REFLUSH;
1015                         return(-1);
1016                 }
1017                 if (record->type == HAMMER_MEM_RECORD_ADD)
1018                         return(1);
1019                 /* GENERAL or DEL */
1020                 return(0);
1021         }
1022
1023         /*
1024          * It must be a setup record.  Try to resolve the setup dependancies
1025          * by recursing upwards so we can place ip on the flush list.
1026          */
1027         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1028
1029         good = hammer_setup_parent_inodes(pip);
1030
1031         /*
1032          * We can't flush ip because it has no connectivity (XXX also check
1033          * nlinks for pre-existing connectivity!).  Flag it so any resolution
1034          * recurses back down.
1035          */
1036         if (good < 0) {
1037                 pip->flags |= HAMMER_INODE_REFLUSH;
1038                 return(good);
1039         }
1040
1041         /*
1042          * We are go, place the parent inode in a flushing state so we can
1043          * place its record in a flushing state.  Note that the parent
1044          * may already be flushing.  The record must be in the same flush
1045          * group as the parent.
1046          */
1047         if (pip->flush_state != HAMMER_FST_FLUSH)
1048                 hammer_flush_inode_core(pip, HAMMER_FLUSH_RECURSION);
1049         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1050         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1051
1052 #if 0
1053         if (record->type == HAMMER_MEM_RECORD_DEL &&
1054             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1055                 /*
1056                  * Regardless of flushing state we cannot sync this path if the
1057                  * record represents a delete-on-disk but the target inode
1058                  * is not ready to sync its own deletion.
1059                  *
1060                  * XXX need to count effective nlinks to determine whether
1061                  * the flush is ok, otherwise removing a hardlink will
1062                  * just leave the DEL record to rot.
1063                  */
1064                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1065                 return(-1);
1066         } else
1067 #endif
1068         if (pip->flush_group == pip->hmp->flusher.next) {
1069                 /*
1070                  * This is the record we wanted to synchronize.  If the
1071                  * record went into a flush state while we blocked it
1072                  * had better be in the correct flush group.
1073                  */
1074                 if (record->flush_state != HAMMER_FST_FLUSH) {
1075                         record->flush_state = HAMMER_FST_FLUSH;
1076                         record->flush_group = pip->flush_group;
1077                         hammer_ref(&record->lock);
1078                 } else {
1079                         KKASSERT(record->flush_group == pip->flush_group);
1080                 }
1081                 if (record->type == HAMMER_MEM_RECORD_ADD)
1082                         return(1);
1083
1084                 /*
1085                  * A general or delete-on-disk record does not contribute
1086                  * to our visibility.  We can still flush it, however.
1087                  */
1088                 return(0);
1089         } else {
1090                 /*
1091                  * We couldn't resolve the dependancies, request that the
1092                  * inode be flushed when the dependancies can be resolved.
1093                  */
1094                 pip->flags |= HAMMER_INODE_REFLUSH;
1095                 return(-1);
1096         }
1097 }
1098
1099 /*
1100  * This is the core routine placing an inode into the FST_FLUSH state.
1101  */
1102 static void
1103 hammer_flush_inode_core(hammer_inode_t ip, int flags)
1104 {
1105         int go_count;
1106
1107         /*
1108          * Set flush state and prevent the flusher from cycling into
1109          * the next flush group.  Do not place the ip on the list yet.
1110          * Inodes not in the idle state get an extra reference.
1111          */
1112         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1113         if (ip->flush_state == HAMMER_FST_IDLE)
1114                 hammer_ref(&ip->lock);
1115         ip->flush_state = HAMMER_FST_FLUSH;
1116         ip->flush_group = ip->hmp->flusher.next;
1117         ++ip->hmp->flusher.group_lock;
1118         ++ip->hmp->count_iqueued;
1119         ++hammer_count_iqueued;
1120
1121         /*
1122          * We need to be able to vfsync/truncate from the backend.
1123          */
1124         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1125         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1126                 ip->flags |= HAMMER_INODE_VHELD;
1127                 vref(ip->vp);
1128         }
1129
1130         /*
1131          * Figure out how many in-memory records we can actually flush
1132          * (not including inode meta-data, buffers, etc).
1133          */
1134         if (flags & HAMMER_FLUSH_RECURSION) {
1135                 go_count = 1;
1136         } else {
1137                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1138                                    hammer_setup_child_callback, NULL);
1139         }
1140
1141         /*
1142          * This is a more involved test that includes go_count.  If we
1143          * can't flush, flag the inode and return.  If go_count is 0 we
1144          * were are unable to flush any records in our rec_tree and
1145          * must ignore the XDIRTY flag.
1146          */
1147         if (go_count == 0) {
1148                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1149                         ip->flags |= HAMMER_INODE_REFLUSH;
1150
1151                         --ip->hmp->count_iqueued;
1152                         --hammer_count_iqueued;
1153
1154                         ip->flush_state = HAMMER_FST_SETUP;
1155                         if (ip->flags & HAMMER_INODE_VHELD) {
1156                                 ip->flags &= ~HAMMER_INODE_VHELD;
1157                                 vrele(ip->vp);
1158                         }
1159                         if (flags & HAMMER_FLUSH_SIGNAL) {
1160                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1161                                 hammer_flusher_async(ip->hmp);
1162                         }
1163                         if (--ip->hmp->flusher.group_lock == 0)
1164                                 wakeup(&ip->hmp->flusher.group_lock);
1165                         return;
1166                 }
1167         }
1168
1169         /*
1170          * Snapshot the state of the inode for the backend flusher.
1171          *
1172          * The truncation must be retained in the frontend until after
1173          * we've actually performed the record deletion.
1174          *
1175          * We continue to retain sync_trunc_off even when all truncations
1176          * have been resolved as an optimization to determine if we can
1177          * skip the B-Tree lookup for overwrite deletions.
1178          *
1179          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1180          * and stays in ip->flags.  Once set, it stays set until the
1181          * inode is destroyed.
1182          */
1183         ip->sync_flags = (ip->flags & HAMMER_INODE_MODMASK);
1184         if (ip->sync_flags & HAMMER_INODE_TRUNCATED)
1185                 ip->sync_trunc_off = ip->trunc_off;
1186         ip->sync_ino_leaf = ip->ino_leaf;
1187         ip->sync_ino_data = ip->ino_data;
1188         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1189         ip->flags &= ~HAMMER_INODE_MODMASK;
1190 #ifdef DEBUG_TRUNCATE
1191         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1192                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1193 #endif
1194
1195         /*
1196          * The flusher list inherits our inode and reference.
1197          */
1198         TAILQ_INSERT_TAIL(&ip->hmp->flush_list, ip, flush_entry);
1199         if (--ip->hmp->flusher.group_lock == 0)
1200                 wakeup(&ip->hmp->flusher.group_lock);
1201
1202         if (flags & HAMMER_FLUSH_SIGNAL) {
1203                 hammer_flusher_async(ip->hmp);
1204         }
1205 }
1206
1207 /*
1208  * Callback for scan of ip->rec_tree.  Try to include each record in our
1209  * flush.  ip->flush_group has been set but the inode has not yet been
1210  * moved into a flushing state.
1211  *
1212  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1213  * both inodes.
1214  *
1215  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1216  * the caller from shortcutting the flush.
1217  */
1218 static int
1219 hammer_setup_child_callback(hammer_record_t rec, void *data)
1220 {
1221         hammer_inode_t target_ip;
1222         hammer_inode_t ip;
1223         int r;
1224
1225         /*
1226          * Deleted records are ignored.  Note that the flush detects deleted
1227          * front-end records at multiple points to deal with races.  This is
1228          * just the first line of defense.  The only time DELETED_FE cannot
1229          * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1230          *
1231          * Don't get confused between record deletion and, say, directory
1232          * entry deletion.  The deletion of a directory entry that is on
1233          * the media has nothing to do with the record deletion flags.
1234          */
1235         if (rec->flags & (HAMMER_RECF_DELETED_FE|HAMMER_RECF_DELETED_BE))
1236                 return(0);
1237
1238         /*
1239          * If the record is in an idle state it has no dependancies and
1240          * can be flushed.
1241          */
1242         ip = rec->ip;
1243         r = 0;
1244
1245         switch(rec->flush_state) {
1246         case HAMMER_FST_IDLE:
1247                 /*
1248                  * Record has no setup dependancy, we can flush it.
1249                  */
1250                 KKASSERT(rec->target_ip == NULL);
1251                 rec->flush_state = HAMMER_FST_FLUSH;
1252                 rec->flush_group = ip->flush_group;
1253                 hammer_ref(&rec->lock);
1254                 r = 1;
1255                 break;
1256         case HAMMER_FST_SETUP:
1257                 /*
1258                  * Record has a setup dependancy.  Try to include the
1259                  * target ip in the flush.
1260                  *
1261                  * We have to be careful here, if we do not do the right
1262                  * thing we can lose track of dirty inodes and the system
1263                  * will lockup trying to allocate buffers.
1264                  */
1265                 target_ip = rec->target_ip;
1266                 KKASSERT(target_ip != NULL);
1267                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1268                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1269                         /*
1270                          * If the target IP is already flushing in our group
1271                          * we are golden, otherwise make sure the target
1272                          * reflushes.
1273                          */
1274                         if (target_ip->flush_group == ip->flush_group) {
1275                                 rec->flush_state = HAMMER_FST_FLUSH;
1276                                 rec->flush_group = ip->flush_group;
1277                                 hammer_ref(&rec->lock);
1278                                 r = 1;
1279                         } else {
1280                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1281                         }
1282                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1283                         /*
1284                          * If the target IP is not flushing we can force
1285                          * it to flush, even if it is unable to write out
1286                          * any of its own records we have at least one in
1287                          * hand that we CAN deal with.
1288                          */
1289                         rec->flush_state = HAMMER_FST_FLUSH;
1290                         rec->flush_group = ip->flush_group;
1291                         hammer_ref(&rec->lock);
1292                         hammer_flush_inode_core(target_ip,
1293                                                 HAMMER_FLUSH_RECURSION);
1294                         r = 1;
1295                 } else {
1296                         /*
1297                          * General or delete-on-disk record.
1298                          *
1299                          * XXX this needs help.  If a delete-on-disk we could
1300                          * disconnect the target.  If the target has its own
1301                          * dependancies they really need to be flushed.
1302                          *
1303                          * XXX
1304                          */
1305                         rec->flush_state = HAMMER_FST_FLUSH;
1306                         rec->flush_group = ip->flush_group;
1307                         hammer_ref(&rec->lock);
1308                         hammer_flush_inode_core(target_ip,
1309                                                 HAMMER_FLUSH_RECURSION);
1310                         r = 1;
1311                 }
1312                 break;
1313         case HAMMER_FST_FLUSH:
1314                 /*
1315                  * Record already associated with a flush group.  It had
1316                  * better be ours.
1317                  */
1318                 KKASSERT(rec->flush_group == ip->flush_group);
1319                 r = 1;
1320                 break;
1321         }
1322         return(r);
1323 }
1324
1325 /*
1326  * Wait for a previously queued flush to complete
1327  */
1328 void
1329 hammer_wait_inode(hammer_inode_t ip)
1330 {
1331         while (ip->flush_state != HAMMER_FST_IDLE) {
1332                 if (ip->flush_state == HAMMER_FST_SETUP) {
1333                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1334                 } else {
1335                         ip->flags |= HAMMER_INODE_FLUSHW;
1336                         tsleep(&ip->flags, 0, "hmrwin", 0);
1337                 }
1338         }
1339 }
1340
1341 /*
1342  * Wait for records to drain
1343  */
1344 void
1345 hammer_wait_inode_recs(hammer_inode_t ip)
1346 {
1347         while (ip->rsv_recs > hammer_limit_irecs) {
1348                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1349                 if (ip->rsv_recs > hammer_limit_irecs) {
1350                         ip->flags |= HAMMER_INODE_PARTIALW;
1351                         tsleep(&ip->flags, 0, "hmrwpp", 0);
1352                 }
1353         }
1354 }
1355
1356 /*
1357  * Called by the backend code when a flush has been completed.
1358  * The inode has already been removed from the flush list.
1359  *
1360  * A pipelined flush can occur, in which case we must re-enter the
1361  * inode on the list and re-copy its fields.
1362  */
1363 void
1364 hammer_flush_inode_done(hammer_inode_t ip)
1365 {
1366         hammer_mount_t hmp;
1367         int dorel;
1368
1369         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1370
1371         hmp = ip->hmp;
1372
1373         /*
1374          * Merge left-over flags back into the frontend and fix the state.
1375          */
1376         ip->flags |= ip->sync_flags;
1377
1378         /*
1379          * The backend may have adjusted nlinks, so if the adjusted nlinks
1380          * does not match the fronttend set the frontend's RDIRTY flag again.
1381          */
1382         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
1383                 ip->flags |= HAMMER_INODE_DDIRTY;
1384
1385         /*
1386          * Fix up the dirty buffer status.  IO completions will also
1387          * try to clean up rsv_databufs.
1388          */
1389         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
1390                 ip->flags |= HAMMER_INODE_BUFS;
1391         } else {
1392                 hmp->rsv_databufs -= ip->rsv_databufs;
1393                 ip->rsv_databufs = 0;
1394         }
1395
1396         /*
1397          * Re-set the XDIRTY flag if some of the inode's in-memory records
1398          * could not be flushed.
1399          */
1400         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
1401                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
1402                  (!RB_EMPTY(&ip->rec_tree) &&
1403                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
1404
1405         /*
1406          * Do not lose track of inodes which no longer have vnode
1407          * assocations, otherwise they may never get flushed again.
1408          */
1409         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1410                 ip->flags |= HAMMER_INODE_REFLUSH;
1411
1412         /*
1413          * Adjust flush_state.  The target state (idle or setup) shouldn't
1414          * be terribly important since we will reflush if we really need
1415          * to do anything. XXX
1416          */
1417         if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
1418                 ip->flush_state = HAMMER_FST_IDLE;
1419                 dorel = 1;
1420         } else {
1421                 ip->flush_state = HAMMER_FST_SETUP;
1422                 dorel = 0;
1423         }
1424
1425         --hmp->count_iqueued;
1426         --hammer_count_iqueued;
1427
1428         /*
1429          * Clean up the vnode ref
1430          */
1431         if (ip->flags & HAMMER_INODE_VHELD) {
1432                 ip->flags &= ~HAMMER_INODE_VHELD;
1433                 vrele(ip->vp);
1434         }
1435
1436         /*
1437          * If the frontend made more changes and requested another flush,
1438          * then try to get it running.
1439          */
1440         if (ip->flags & HAMMER_INODE_REFLUSH) {
1441                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1442                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1443                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1444                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1445                 } else {
1446                         hammer_flush_inode(ip, 0);
1447                 }
1448         }
1449
1450         /*
1451          * If the inode is now clean drop the space reservation.
1452          */
1453         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1454             (ip->flags & HAMMER_INODE_RSV_INODES)) {
1455                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
1456                 --hmp->rsv_inodes;
1457         }
1458
1459         /*
1460          * Finally, if the frontend is waiting for a flush to complete,
1461          * wake it up.
1462          */
1463         if (ip->flush_state != HAMMER_FST_FLUSH) {
1464                 if (ip->flags & HAMMER_INODE_FLUSHW) {
1465                         ip->flags &= ~HAMMER_INODE_FLUSHW;
1466                         wakeup(&ip->flags);
1467                 }
1468         }
1469         if (dorel)
1470                 hammer_rel_inode(ip, 0);
1471 }
1472
1473 /*
1474  * Called from hammer_sync_inode() to synchronize in-memory records
1475  * to the media.
1476  */
1477 static int
1478 hammer_sync_record_callback(hammer_record_t record, void *data)
1479 {
1480         hammer_cursor_t cursor = data;
1481         hammer_transaction_t trans = cursor->trans;
1482         int error;
1483
1484         /*
1485          * Skip records that do not belong to the current flush.
1486          */
1487         ++hammer_stats_record_iterations;
1488         if (record->flush_state != HAMMER_FST_FLUSH)
1489                 return(0);
1490
1491 #if 1
1492         if (record->flush_group != record->ip->flush_group) {
1493                 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1494                 Debugger("blah2");
1495                 return(0);
1496         }
1497 #endif
1498         KKASSERT(record->flush_group == record->ip->flush_group);
1499
1500         /*
1501          * Interlock the record using the BE flag.  Once BE is set the
1502          * frontend cannot change the state of FE.
1503          *
1504          * NOTE: If FE is set prior to us setting BE we still sync the
1505          * record out, but the flush completion code converts it to
1506          * a delete-on-disk record instead of destroying it.
1507          */
1508         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1509         record->flags |= HAMMER_RECF_INTERLOCK_BE;
1510
1511         /*
1512          * The backend may have already disposed of the record.
1513          */
1514         if (record->flags & HAMMER_RECF_DELETED_BE) {
1515                 error = 0;
1516                 goto done;
1517         }
1518
1519         /*
1520          * If the whole inode is being deleting all on-disk records will
1521          * be deleted very soon, we can't sync any new records to disk
1522          * because they will be deleted in the same transaction they were
1523          * created in (delete_tid == create_tid), which will assert.
1524          *
1525          * XXX There may be a case with RECORD_ADD with DELETED_FE set
1526          * that we currently panic on.
1527          */
1528         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
1529                 switch(record->type) {
1530                 case HAMMER_MEM_RECORD_DATA:
1531                         /*
1532                          * We don't have to do anything, if the record was
1533                          * committed the space will have been accounted for
1534                          * in the blockmap.
1535                          */
1536                         /* fall through */
1537                 case HAMMER_MEM_RECORD_GENERAL:
1538                         record->flags |= HAMMER_RECF_DELETED_FE;
1539                         record->flags |= HAMMER_RECF_DELETED_BE;
1540                         error = 0;
1541                         goto done;
1542                 case HAMMER_MEM_RECORD_ADD:
1543                         panic("hammer_sync_record_callback: illegal add "
1544                               "during inode deletion record %p", record);
1545                         break; /* NOT REACHED */
1546                 case HAMMER_MEM_RECORD_INODE:
1547                         panic("hammer_sync_record_callback: attempt to "
1548                               "sync inode record %p?", record);
1549                         break; /* NOT REACHED */
1550                 case HAMMER_MEM_RECORD_DEL:
1551                         /*
1552                          * Follow through and issue the on-disk deletion
1553                          */
1554                         break;
1555                 }
1556         }
1557
1558         /*
1559          * If DELETED_FE is set special handling is needed for directory
1560          * entries.  Dependant pieces related to the directory entry may
1561          * have already been synced to disk.  If this occurs we have to
1562          * sync the directory entry and then change the in-memory record
1563          * from an ADD to a DELETE to cover the fact that it's been
1564          * deleted by the frontend.
1565          *
1566          * A directory delete covering record (MEM_RECORD_DEL) can never
1567          * be deleted by the frontend.
1568          *
1569          * Any other record type (aka DATA) can be deleted by the frontend.
1570          * XXX At the moment the flusher must skip it because there may
1571          * be another data record in the flush group for the same block,
1572          * meaning that some frontend data changes can leak into the backend's
1573          * synchronization point.
1574          */
1575         if (record->flags & HAMMER_RECF_DELETED_FE) {
1576                 if (record->type == HAMMER_MEM_RECORD_ADD) {
1577                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
1578                 } else {
1579                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
1580                         record->flags |= HAMMER_RECF_DELETED_BE;
1581                         error = 0;
1582                         goto done;
1583                 }
1584         }
1585
1586         /*
1587          * Assign the create_tid for new records.  Deletions already
1588          * have the record's entire key properly set up.
1589          */
1590         if (record->type != HAMMER_MEM_RECORD_DEL)
1591                 record->leaf.base.create_tid = trans->tid;
1592         for (;;) {
1593                 error = hammer_ip_sync_record_cursor(cursor, record);
1594                 if (error != EDEADLK)
1595                         break;
1596                 hammer_done_cursor(cursor);
1597                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
1598                                            record->ip);
1599                 if (error)
1600                         break;
1601         }
1602         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1603
1604         if (error) {
1605                 error = -error;
1606                 if (error != -ENOSPC) {
1607                         kprintf("hammer_sync_record_callback: sync failed rec "
1608                                 "%p, error %d\n", record, error);
1609                         Debugger("sync failed rec");
1610                 }
1611         }
1612 done:
1613         hammer_flush_record_done(record, error);
1614         return(error);
1615 }
1616
1617 /*
1618  * XXX error handling
1619  */
1620 int
1621 hammer_sync_inode(hammer_inode_t ip)
1622 {
1623         struct hammer_transaction trans;
1624         struct hammer_cursor cursor;
1625         hammer_node_t tmp_node;
1626         hammer_record_t depend;
1627         hammer_record_t next;
1628         int error, tmp_error;
1629         u_int64_t nlinks;
1630
1631         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
1632                 return(0);
1633
1634         hammer_start_transaction_fls(&trans, ip->hmp);
1635         error = hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1636         if (error)
1637                 goto done;
1638
1639         /*
1640          * Any directory records referencing this inode which are not in
1641          * our current flush group must adjust our nlink count for the
1642          * purposes of synchronization to disk.
1643          *
1644          * Records which are in our flush group can be unlinked from our
1645          * inode now, potentially allowing the inode to be physically
1646          * deleted.
1647          *
1648          * This cannot block.
1649          */
1650         nlinks = ip->ino_data.nlinks;
1651         next = TAILQ_FIRST(&ip->target_list);
1652         while ((depend = next) != NULL) {
1653                 next = TAILQ_NEXT(depend, target_entry);
1654                 if (depend->flush_state == HAMMER_FST_FLUSH &&
1655                     depend->flush_group == ip->hmp->flusher.act) {
1656                         /*
1657                          * If this is an ADD that was deleted by the frontend
1658                          * the frontend nlinks count will have already been
1659                          * decremented, but the backend is going to sync its
1660                          * directory entry and must account for it.  The
1661                          * record will be converted to a delete-on-disk when
1662                          * it gets synced.
1663                          *
1664                          * If the ADD was not deleted by the frontend we
1665                          * can remove the dependancy from our target_list.
1666                          */
1667                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
1668                                 ++nlinks;
1669                         } else {
1670                                 TAILQ_REMOVE(&ip->target_list, depend,
1671                                              target_entry);
1672                                 depend->target_ip = NULL;
1673                         }
1674                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
1675                         /*
1676                          * Not part of our flush group
1677                          */
1678                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
1679                         switch(depend->type) {
1680                         case HAMMER_MEM_RECORD_ADD:
1681                                 --nlinks;
1682                                 break;
1683                         case HAMMER_MEM_RECORD_DEL:
1684                                 ++nlinks;
1685                                 break;
1686                         default:
1687                                 break;
1688                         }
1689                 }
1690         }
1691
1692         /*
1693          * Set dirty if we had to modify the link count.
1694          */
1695         if (ip->sync_ino_data.nlinks != nlinks) {
1696                 KKASSERT((int64_t)nlinks >= 0);
1697                 ip->sync_ino_data.nlinks = nlinks;
1698                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1699         }
1700
1701         /*
1702          * If there is a trunction queued destroy any data past the (aligned)
1703          * truncation point.  Userland will have dealt with the buffer
1704          * containing the truncation point for us.
1705          *
1706          * We don't flush pending frontend data buffers until after we've
1707          * dealt with the truncation.
1708          */
1709         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
1710                 /*
1711                  * Interlock trunc_off.  The VOP front-end may continue to
1712                  * make adjustments to it while we are blocked.
1713                  */
1714                 off_t trunc_off;
1715                 off_t aligned_trunc_off;
1716
1717                 trunc_off = ip->sync_trunc_off;
1718                 aligned_trunc_off = (trunc_off + HAMMER_BUFMASK) &
1719                                     ~HAMMER_BUFMASK64;
1720
1721                 /*
1722                  * Delete any whole blocks on-media.  The front-end has
1723                  * already cleaned out any partial block and made it
1724                  * pending.  The front-end may have updated trunc_off
1725                  * while we were blocked so we only use sync_trunc_off.
1726                  */
1727                 error = hammer_ip_delete_range(&cursor, ip,
1728                                                 aligned_trunc_off,
1729                                                 0x7FFFFFFFFFFFFFFFLL, 1);
1730                 if (error)
1731                         Debugger("hammer_ip_delete_range errored");
1732
1733                 /*
1734                  * Clear the truncation flag on the backend after we have
1735                  * complete the deletions.  Backend data is now good again
1736                  * (including new records we are about to sync, below).
1737                  *
1738                  * Leave sync_trunc_off intact.  As we write additional
1739                  * records the backend will update sync_trunc_off.  This
1740                  * tells the backend whether it can skip the overwrite
1741                  * test.  This should work properly even when the backend
1742                  * writes full blocks where the truncation point straddles
1743                  * the block because the comparison is against the base
1744                  * offset of the record.
1745                  */
1746                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1747                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
1748         } else {
1749                 error = 0;
1750         }
1751
1752         /*
1753          * Now sync related records.  These will typically be directory
1754          * entries or delete-on-disk records.
1755          *
1756          * Not all records will be flushed, but clear XDIRTY anyway.  We
1757          * will set it again in the frontend hammer_flush_inode_done()
1758          * if records remain.
1759          */
1760         if (error == 0) {
1761                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1762                                     hammer_sync_record_callback, &cursor);
1763                 if (tmp_error < 0)
1764                         tmp_error = -error;
1765                 if (tmp_error)
1766                         error = tmp_error;
1767         }
1768         hammer_cache_node(cursor.node, &ip->cache[1]);
1769
1770         /*
1771          * Re-seek for inode update.
1772          */
1773         if (error == 0) {
1774                 tmp_node = hammer_ref_node_safe(ip->hmp, &ip->cache[0], &error);
1775                 if (tmp_node) {
1776                         hammer_cursor_seek(&cursor, tmp_node, 0);
1777                         hammer_rel_node(tmp_node);
1778                 }
1779                 error = 0;
1780         }
1781
1782         /*
1783          * If we are deleting the inode the frontend had better not have
1784          * any active references on elements making up the inode.
1785          */
1786         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
1787                 RB_EMPTY(&ip->rec_tree)  &&
1788             (ip->sync_flags & HAMMER_INODE_DELETING) &&
1789             (ip->flags & HAMMER_INODE_DELETED) == 0) {
1790                 int count1 = 0;
1791
1792                 ip->flags |= HAMMER_INODE_DELETED;
1793                 error = hammer_ip_delete_range_all(&cursor, ip, &count1);
1794                 if (error == 0) {
1795                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
1796                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1797                         KKASSERT(RB_EMPTY(&ip->rec_tree));
1798
1799                         /*
1800                          * Set delete_tid in both the frontend and backend
1801                          * copy of the inode record.  The DELETED flag handles
1802                          * this, do not set RDIRTY.
1803                          */
1804                         ip->ino_leaf.base.delete_tid = trans.tid;
1805                         ip->sync_ino_leaf.base.delete_tid = trans.tid;
1806
1807                         /*
1808                          * Adjust the inode count in the volume header
1809                          */
1810                         if (ip->flags & HAMMER_INODE_ONDISK) {
1811                                 hammer_modify_volume_field(&trans,
1812                                                            trans.rootvol,
1813                                                            vol0_stat_inodes);
1814                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1815                                 hammer_modify_volume_done(trans.rootvol);
1816                         }
1817                 } else {
1818                         ip->flags &= ~HAMMER_INODE_DELETED;
1819                         Debugger("hammer_ip_delete_range_all errored");
1820                 }
1821         }
1822
1823         ip->sync_flags &= ~HAMMER_INODE_BUFS;
1824
1825         if (error)
1826                 Debugger("RB_SCAN errored");
1827
1828         /*
1829          * Now update the inode's on-disk inode-data and/or on-disk record.
1830          * DELETED and ONDISK are managed only in ip->flags.
1831          */
1832         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
1833         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
1834                 /*
1835                  * If deleted and on-disk, don't set any additional flags.
1836                  * the delete flag takes care of things.
1837                  *
1838                  * Clear flags which may have been set by the frontend.
1839                  */
1840                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY|
1841                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1842                                     HAMMER_INODE_DELETING);
1843                 break;
1844         case HAMMER_INODE_DELETED:
1845                 /*
1846                  * Take care of the case where a deleted inode was never
1847                  * flushed to the disk in the first place.
1848                  *
1849                  * Clear flags which may have been set by the frontend.
1850                  */
1851                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY|
1852                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1853                                     HAMMER_INODE_DELETING);
1854                 while (RB_ROOT(&ip->rec_tree)) {
1855                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
1856                         hammer_ref(&record->lock);
1857                         KKASSERT(record->lock.refs == 1);
1858                         record->flags |= HAMMER_RECF_DELETED_FE;
1859                         record->flags |= HAMMER_RECF_DELETED_BE;
1860                         hammer_rel_mem_record(record);
1861                 }
1862                 break;
1863         case HAMMER_INODE_ONDISK:
1864                 /*
1865                  * If already on-disk, do not set any additional flags.
1866                  */
1867                 break;
1868         default:
1869                 /*
1870                  * If not on-disk and not deleted, set both dirty flags
1871                  * to force an initial record to be written.  Also set
1872                  * the create_tid for the inode.
1873                  *
1874                  * Set create_tid in both the frontend and backend
1875                  * copy of the inode record.
1876                  */
1877                 ip->ino_leaf.base.create_tid = trans.tid;
1878                 ip->sync_ino_leaf.base.create_tid = trans.tid;
1879                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1880                 break;
1881         }
1882
1883         /*
1884          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
1885          * is already on-disk the old record is marked as deleted.
1886          *
1887          * If DELETED is set hammer_update_inode() will delete the existing
1888          * record without writing out a new one.
1889          *
1890          * If *ONLY* the ITIMES flag is set we can update the record in-place.
1891          */
1892         if (ip->flags & HAMMER_INODE_DELETED) {
1893                 error = hammer_update_inode(&cursor, ip);
1894         } else
1895         if ((ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES)) ==
1896             HAMMER_INODE_ITIMES) {
1897                 error = hammer_update_itimes(&cursor, ip);
1898         } else
1899         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES)) {
1900                 error = hammer_update_inode(&cursor, ip);
1901         }
1902         if (error)
1903                 Debugger("hammer_update_itimes/inode errored");
1904 done:
1905         /*
1906          * Save the TID we used to sync the inode with to make sure we
1907          * do not improperly reuse it.
1908          */
1909         hammer_done_cursor(&cursor);
1910         hammer_done_transaction(&trans);
1911         return(error);
1912 }
1913
1914 /*
1915  * This routine is called when the OS is no longer actively referencing
1916  * the inode (but might still be keeping it cached), or when releasing
1917  * the last reference to an inode.
1918  *
1919  * At this point if the inode's nlinks count is zero we want to destroy
1920  * it, which may mean destroying it on-media too.
1921  */
1922 void
1923 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
1924 {
1925         struct vnode *vp;
1926
1927         /*
1928          * Set the DELETING flag when the link count drops to 0 and the
1929          * OS no longer has any opens on the inode.
1930          *
1931          * The backend will clear DELETING (a mod flag) and set DELETED
1932          * (a state flag) when it is actually able to perform the
1933          * operation.
1934          */
1935         if (ip->ino_data.nlinks == 0 &&
1936             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
1937                 ip->flags |= HAMMER_INODE_DELETING;
1938                 ip->flags |= HAMMER_INODE_TRUNCATED;
1939                 ip->trunc_off = 0;
1940                 vp = NULL;
1941                 if (getvp) {
1942                         if (hammer_get_vnode(ip, &vp) != 0)
1943                                 return;
1944                 }
1945
1946                 /*
1947                  * Final cleanup
1948                  */
1949                 if (ip->vp) {
1950                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
1951                         vnode_pager_setsize(ip->vp, 0);
1952                 }
1953                 if (getvp) {
1954                         vput(vp);
1955                 }
1956         }
1957 }
1958
1959 /*
1960  * Re-test an inode when a dependancy had gone away to see if we
1961  * can chain flush it.
1962  */
1963 void
1964 hammer_test_inode(hammer_inode_t ip)
1965 {
1966         if (ip->flags & HAMMER_INODE_REFLUSH) {
1967                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1968                 hammer_ref(&ip->lock);
1969                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1970                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1971                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1972                 } else {
1973                         hammer_flush_inode(ip, 0);
1974                 }
1975                 hammer_rel_inode(ip, 0);
1976         }
1977 }
1978
1979 /*
1980  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
1981  * reassociated with a vp or just before it gets freed.
1982  *
1983  * Wakeup one thread blocked waiting on reclaims to complete.  Note that
1984  * the inode the thread is waiting on behalf of is a different inode then
1985  * the inode we are called with.  This is to create a pipeline.
1986  */
1987 static void
1988 hammer_inode_wakereclaims(hammer_inode_t ip)
1989 {
1990         struct hammer_reclaim *reclaim;
1991         hammer_mount_t hmp = ip->hmp;
1992
1993         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
1994                 return;
1995
1996         --hammer_count_reclaiming;
1997         --hmp->inode_reclaims;
1998         ip->flags &= ~HAMMER_INODE_RECLAIM;
1999
2000         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
2001                 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2002                 reclaim->okydoky = 1;
2003                 wakeup(reclaim);
2004         }
2005 }
2006