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 1.110 2008/08/09 07:04:16 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_free_inode(hammer_inode_t ip);
  44 static void     hammer_flush_inode_core(hammer_inode_t ip,
  45                                         hammer_flush_group_t flg, int flags);
  46 static int      hammer_setup_child_callback(hammer_record_t rec, void *data);
  47 #if 0
  48 static int      hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
  49 #endif
  50 static int      hammer_setup_parent_inodes(hammer_inode_t ip,
  51                                         hammer_flush_group_t flg);
  52 static int      hammer_setup_parent_inodes_helper(hammer_record_t record,
  53                                         hammer_flush_group_t flg);
  54 static void     hammer_inode_wakereclaims(hammer_inode_t ip);
  55
  56 #ifdef DEBUG_TRUNCATE
  57 extern struct hammer_inode *HammerTruncIp;
  58 #endif
  59
  60 /*
  61  * RB-Tree support for inode structures
  62  */
  63 int
  64 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  65 {
  66         if (ip1->obj_localization < ip2->obj_localization)
  67                 return(-1);
  68         if (ip1->obj_localization > ip2->obj_localization)
  69                 return(1);
  70         if (ip1->obj_id < ip2->obj_id)
  71                 return(-1);
  72         if (ip1->obj_id > ip2->obj_id)
  73                 return(1);
  74         if (ip1->obj_asof < ip2->obj_asof)
  75                 return(-1);
  76         if (ip1->obj_asof > ip2->obj_asof)
  77                 return(1);
  78         return(0);
  79 }
  80
  81 /*
  82  * RB-Tree support for inode structures / special LOOKUP_INFO
  83  */
  84 static int
  85 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  86 {
  87         if (info->obj_localization < ip->obj_localization)
  88                 return(-1);
  89         if (info->obj_localization > ip->obj_localization)
  90                 return(1);
  91         if (info->obj_id < ip->obj_id)
  92                 return(-1);
  93         if (info->obj_id > ip->obj_id)
  94                 return(1);
  95         if (info->obj_asof < ip->obj_asof)
  96                 return(-1);
  97         if (info->obj_asof > ip->obj_asof)
  98                 return(1);
  99         return(0);
 100 }
 101
 102 /*
 103  * Used by hammer_scan_inode_snapshots() to locate all of an object's
 104  * snapshots.  Note that the asof field is not tested, which we can get
 105  * away with because it is the lowest-priority field.
 106  */
 107 static int
 108 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
 109 {
 110         hammer_inode_info_t info = data;
 111
 112         if (ip->obj_localization > info->obj_localization)
 113                 return(1);
 114         if (ip->obj_localization < info->obj_localization)
 115                 return(-1);
 116         if (ip->obj_id > info->obj_id)
 117                 return(1);
 118         if (ip->obj_id < info->obj_id)
 119                 return(-1);
 120         return(0);
 121 }
 122
 123 /*
 124  * Used by hammer_unload_pseudofs() to locate all inodes associated with
 125  * a particular PFS.
 126  */
 127 static int
 128 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
 129 {
 130         u_int32_t localization = *(u_int32_t *)data;
 131         if (ip->obj_localization > localization)
 132                 return(1);
 133         if (ip->obj_localization < localization)
 134                 return(-1);
 135         return(0);
 136 }
 137
 138 /*
 139  * RB-Tree support for pseudofs structures
 140  */
 141 static int
 142 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
 143 {
 144         if (p1->localization < p2->localization)
 145                 return(-1);
 146         if (p1->localization > p2->localization)
 147                 return(1);
 148         return(0);
 149 }
 150
 151
 152 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 153 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 154                 hammer_inode_info_cmp, hammer_inode_info_t);
 155 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 156              hammer_pfs_rb_compare, u_int32_t, localization);
 157
 158 /*
 159  * The kernel is not actively referencing this vnode but is still holding
 160  * it cached.
 161  *
 162  * This is called from the frontend.
 163  */
 164 int
 165 hammer_vop_inactive(struct vop_inactive_args *ap)
 166 {
 167         struct hammer_inode *ip = VTOI(ap->a_vp);
 168
 169         /*
 170          * Degenerate case
 171          */
 172         if (ip == NULL) {
 173                 vrecycle(ap->a_vp);
 174                 return(0);
 175         }
 176
 177         /*
 178          * If the inode no longer has visibility in the filesystem try to
 179          * recycle it immediately, even if the inode is dirty.  Recycling
 180          * it quickly allows the system to reclaim buffer cache and VM
 181          * resources which can matter a lot in a heavily loaded system.
 182          *
 183          * This can deadlock in vfsync() if we aren't careful.
 184          *
 185          * Do not queue the inode to the flusher if we still have visibility,
 186          * otherwise namespace calls such as chmod will unnecessarily generate
 187          * multiple inode updates.
 188          */
 189         hammer_inode_unloadable_check(ip, 0);
 190         if (ip->ino_data.nlinks == 0) {
 191                 if (ip->flags & HAMMER_INODE_MODMASK)
 192                         hammer_flush_inode(ip, 0);
 193                 vrecycle(ap->a_vp);
 194         }
 195         return(0);
 196 }
 197
 198 /*
 199  * Release the vnode association.  This is typically (but not always)
 200  * the last reference on the inode.
 201  *
 202  * Once the association is lost we are on our own with regards to
 203  * flushing the inode.
 204  */
 205 int
 206 hammer_vop_reclaim(struct vop_reclaim_args *ap)
 207 {
 208         struct hammer_inode *ip;
 209         hammer_mount_t hmp;
 210         struct vnode *vp;
 211
 212         vp = ap->a_vp;
 213
 214         if ((ip = vp->v_data) != NULL) {
 215                 hmp = ip->hmp;
 216                 vp->v_data = NULL;
 217                 ip->vp = NULL;
 218
 219                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
 220                         ++hammer_count_reclaiming;
 221                         ++hmp->inode_reclaims;
 222                         ip->flags |= HAMMER_INODE_RECLAIM;
 223
 224                         /*
 225                          * Poke the flusher.  If we don't do this programs
 226                          * will start to stall on the reclaiming count.
 227                          */
 228                         if (hmp->inode_reclaims > HAMMER_RECLAIM_FLUSH &&
 229                            (hmp->inode_reclaims & 255) == 0) {
 230                                hammer_flusher_async(hmp, NULL);
 231                         }
 232                 }
 233                 hammer_rel_inode(ip, 1);
 234         }
 235         return(0);
 236 }
 237
 238 /*
 239  * Return a locked vnode for the specified inode.  The inode must be
 240  * referenced but NOT LOCKED on entry and will remain referenced on
 241  * return.
 242  *
 243  * Called from the frontend.
 244  */
 245 int
 246 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 247 {
 248         hammer_mount_t hmp;
 249         struct vnode *vp;
 250         int error = 0;
 251         u_int8_t obj_type;
 252
 253         hmp = ip->hmp;
 254
 255         for (;;) {
 256                 if ((vp = ip->vp) == NULL) {
 257                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
 258                         if (error)
 259                                 break;
 260                         hammer_lock_ex(&ip->lock);
 261                         if (ip->vp != NULL) {
 262                                 hammer_unlock(&ip->lock);
 263                                 vp->v_type = VBAD;
 264                                 vx_put(vp);
 265                                 continue;
 266                         }
 267                         hammer_ref(&ip->lock);
 268                         vp = *vpp;
 269                         ip->vp = vp;
 270
 271                         obj_type = ip->ino_data.obj_type;
 272                         vp->v_type = hammer_get_vnode_type(obj_type);
 273
 274                         hammer_inode_wakereclaims(ip);
 275
 276                         switch(ip->ino_data.obj_type) {
 277                         case HAMMER_OBJTYPE_CDEV:
 278                         case HAMMER_OBJTYPE_BDEV:
 279                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
 280                                 addaliasu(vp, ip->ino_data.rmajor,
 281                                           ip->ino_data.rminor);
 282                                 break;
 283                         case HAMMER_OBJTYPE_FIFO:
 284                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
 285                                 break;
 286                         default:
 287                                 break;
 288                         }
 289
 290                         /*
 291                          * Only mark as the root vnode if the ip is not
 292                          * historical, otherwise the VFS cache will get
 293                          * confused.  The other half of the special handling
 294                          * is in hammer_vop_nlookupdotdot().
 295                          *
 296                          * Pseudo-filesystem roots also do not count.
 297                          */
 298                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 299                             ip->obj_asof == hmp->asof &&
 300                             ip->obj_localization == 0) {
 301                                 vp->v_flag |= VROOT;
 302                         }
 303
 304                         vp->v_data = (void *)ip;
 305                         /* vnode locked by getnewvnode() */
 306                         /* make related vnode dirty if inode dirty? */
 307                         hammer_unlock(&ip->lock);
 308                         if (vp->v_type == VREG)
 309                                 vinitvmio(vp, ip->ino_data.size);
 310                         break;
 311                 }
 312
 313                 /*
 314                  * loop if the vget fails (aka races), or if the vp
 315                  * no longer matches ip->vp.
 316                  */
 317                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 318                         if (vp == ip->vp)
 319                                 break;
 320                         vput(vp);
 321                 }
 322         }
 323         *vpp = vp;
 324         return(error);
 325 }
 326
 327 /*
 328  * Locate all copies of the inode for obj_id compatible with the specified
 329  * asof, reference, and issue the related call-back.  This routine is used
 330  * for direct-io invalidation and does not create any new inodes.
 331  */
 332 void
 333 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
 334                             int (*callback)(hammer_inode_t ip, void *data),
 335                             void *data)
 336 {
 337         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
 338                                    hammer_inode_info_cmp_all_history,
 339                                    callback, iinfo);
 340 }
 341
 342 /*
 343  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 344  * do not attach or detach the related vnode (use hammer_get_vnode() for
 345  * that).
 346  *
 347  * The flags argument is only applied for newly created inodes, and only
 348  * certain flags are inherited.
 349  *
 350  * Called from the frontend.
 351  */
 352 struct hammer_inode *
 353 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
 354                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 355                  int flags, int *errorp)
 356 {
 357         hammer_mount_t hmp = trans->hmp;
 358         struct hammer_inode_info iinfo;
 359         struct hammer_cursor cursor;
 360         struct hammer_inode *ip;
 361
 362
 363         /*
 364          * Determine if we already have an inode cached.  If we do then
 365          * we are golden.
 366          */
 367         iinfo.obj_id = obj_id;
 368         iinfo.obj_asof = asof;
 369         iinfo.obj_localization = localization;
 370 loop:
 371         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 372         if (ip) {
 373                 hammer_ref(&ip->lock);
 374                 *errorp = 0;
 375                 return(ip);
 376         }
 377
 378         /*
 379          * Allocate a new inode structure and deal with races later.
 380          */
 381         ip = kmalloc(sizeof(*ip), M_HAMMER_INO, M_WAITOK|M_ZERO);
 382         ++hammer_count_inodes;
 383         ++hmp->count_inodes;
 384         ip->obj_id = obj_id;
 385         ip->obj_asof = iinfo.obj_asof;
 386         ip->obj_localization = localization;
 387         ip->hmp = hmp;
 388         ip->flags = flags & HAMMER_INODE_RO;
 389         ip->cache[0].ip = ip;
 390         ip->cache[1].ip = ip;
 391         if (hmp->ronly)
 392                 ip->flags |= HAMMER_INODE_RO;
 393         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 394                 0x7FFFFFFFFFFFFFFFLL;
 395         RB_INIT(&ip->rec_tree);
 396         TAILQ_INIT(&ip->target_list);
 397         hammer_ref(&ip->lock);
 398
 399         /*
 400          * Locate the on-disk inode.  If this is a PFS root we always
 401          * access the current version of the root inode and (if it is not
 402          * a master) always access information under it with a snapshot
 403          * TID.
 404          */
 405 retry:
 406         hammer_init_cursor(trans, &cursor, (dip ? &dip->cache[0] : NULL), NULL);
 407         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
 408         cursor.key_beg.obj_id = ip->obj_id;
 409         cursor.key_beg.key = 0;
 410         cursor.key_beg.create_tid = 0;
 411         cursor.key_beg.delete_tid = 0;
 412         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 413         cursor.key_beg.obj_type = 0;
 414
 415         cursor.asof = iinfo.obj_asof;
 416         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 417                        HAMMER_CURSOR_ASOF;
 418
 419         *errorp = hammer_btree_lookup(&cursor);
 420         if (*errorp == EDEADLK) {
 421                 hammer_done_cursor(&cursor);
 422                 goto retry;
 423         }
 424
 425         /*
 426          * On success the B-Tree lookup will hold the appropriate
 427          * buffer cache buffers and provide a pointer to the requested
 428          * information.  Copy the information to the in-memory inode
 429          * and cache the B-Tree node to improve future operations.
 430          */
 431         if (*errorp == 0) {
 432                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 433                 ip->ino_data = cursor.data->inode;
 434
 435                 /*
 436                  * cache[0] tries to cache the location of the object inode.
 437                  * The assumption is that it is near the directory inode.
 438                  *
 439                  * cache[1] tries to cache the location of the object data.
 440                  * The assumption is that it is near the directory data.
 441                  */
 442                 hammer_cache_node(&ip->cache[0], cursor.node);
 443                 if (dip && dip->cache[1].node)
 444                         hammer_cache_node(&ip->cache[1], dip->cache[1].node);
 445
 446                 /*
 447                  * The file should not contain any data past the file size
 448                  * stored in the inode.  Setting save_trunc_off to the
 449                  * file size instead of max reduces B-Tree lookup overheads
 450                  * on append by allowing the flusher to avoid checking for
 451                  * record overwrites.
 452                  */
 453                 ip->save_trunc_off = ip->ino_data.size;
 454
 455                 /*
 456                  * Locate and assign the pseudofs management structure to
 457                  * the inode.
 458                  */
 459                 if (dip && dip->obj_localization == ip->obj_localization) {
 460                         ip->pfsm = dip->pfsm;
 461                         hammer_ref(&ip->pfsm->lock);
 462                 } else {
 463                         ip->pfsm = hammer_load_pseudofs(trans,
 464                                                         ip->obj_localization,
 465                                                         errorp);
 466                         *errorp = 0;    /* ignore ENOENT */
 467                 }
 468         }
 469
 470         /*
 471          * The inode is placed on the red-black tree and will be synced to
 472          * the media when flushed or by the filesystem sync.  If this races
 473          * another instantiation/lookup the insertion will fail.
 474          */
 475         if (*errorp == 0) {
 476                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 477                         hammer_free_inode(ip);
 478                         hammer_done_cursor(&cursor);
 479                         goto loop;
 480                 }
 481                 ip->flags |= HAMMER_INODE_ONDISK;
 482         } else {
 483                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 484                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 485                         --hmp->rsv_inodes;
 486                 }
 487
 488                 hammer_free_inode(ip);
 489                 ip = NULL;
 490         }
 491         hammer_done_cursor(&cursor);
 492         return (ip);
 493 }
 494
 495 /*
 496  * Create a new filesystem object, returning the inode in *ipp.  The
 497  * returned inode will be referenced.  The inode is created in-memory.
 498  *
 499  * If pfsm is non-NULL the caller wishes to create the root inode for
 500  * a master PFS.
 501  */
 502 int
 503 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 504                     struct ucred *cred, hammer_inode_t dip,
 505                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
 506 {
 507         hammer_mount_t hmp;
 508         hammer_inode_t ip;
 509         uid_t xuid;
 510         int error;
 511
 512         hmp = trans->hmp;
 513
 514         ip = kmalloc(sizeof(*ip), M_HAMMER_INO, M_WAITOK|M_ZERO);
 515         ++hammer_count_inodes;
 516         ++hmp->count_inodes;
 517
 518         if (pfsm) {
 519                 KKASSERT(pfsm->localization != 0);
 520                 ip->obj_id = HAMMER_OBJID_ROOT;
 521                 ip->obj_localization = pfsm->localization;
 522         } else {
 523                 KKASSERT(dip != NULL);
 524                 ip->obj_id = hammer_alloc_objid(hmp, dip);
 525                 ip->obj_localization = dip->obj_localization;
 526         }
 527
 528         KKASSERT(ip->obj_id != 0);
 529         ip->obj_asof = hmp->asof;
 530         ip->hmp = hmp;
 531         ip->flush_state = HAMMER_FST_IDLE;
 532         ip->flags = HAMMER_INODE_DDIRTY |
 533                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
 534         ip->cache[0].ip = ip;
 535         ip->cache[1].ip = ip;
 536
 537         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 538         /* ip->save_trunc_off = 0; (already zero) */
 539         RB_INIT(&ip->rec_tree);
 540         TAILQ_INIT(&ip->target_list);
 541
 542         ip->ino_data.atime = trans->time;
 543         ip->ino_data.mtime = trans->time;
 544         ip->ino_data.size = 0;
 545         ip->ino_data.nlinks = 0;
 546
 547         /*
 548          * A nohistory designator on the parent directory is inherited by
 549          * the child.  We will do this even for pseudo-fs creation... the
 550          * sysad can turn it off.
 551          */
 552         if (dip) {
 553                 ip->ino_data.uflags = dip->ino_data.uflags &
 554                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 555         }
 556
 557         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 558         ip->ino_leaf.base.localization = ip->obj_localization +
 559                                          HAMMER_LOCALIZE_INODE;
 560         ip->ino_leaf.base.obj_id = ip->obj_id;
 561         ip->ino_leaf.base.key = 0;
 562         ip->ino_leaf.base.create_tid = 0;
 563         ip->ino_leaf.base.delete_tid = 0;
 564         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 565         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 566
 567         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 568         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 569         ip->ino_data.mode = vap->va_mode;
 570         ip->ino_data.ctime = trans->time;
 571
 572         /*
 573          * Setup the ".." pointer.  This only needs to be done for directories
 574          * but we do it for all objects as a recovery aid.
 575          */
 576         if (dip)
 577                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
 578 #if 0
 579         /*
 580          * The parent_obj_localization field only applies to pseudo-fs roots.
 581          * XXX this is no longer applicable, PFSs are no longer directly
 582          * tied into the parent's directory structure.
 583          */
 584         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
 585             ip->obj_id == HAMMER_OBJID_ROOT) {
 586                 ip->ino_data.ext.obj.parent_obj_localization =
 587                                                 dip->obj_localization;
 588         }
 589 #endif
 590
 591         switch(ip->ino_leaf.base.obj_type) {
 592         case HAMMER_OBJTYPE_CDEV:
 593         case HAMMER_OBJTYPE_BDEV:
 594                 ip->ino_data.rmajor = vap->va_rmajor;
 595                 ip->ino_data.rminor = vap->va_rminor;
 596                 break;
 597         default:
 598                 break;
 599         }
 600
 601         /*
 602          * Calculate default uid/gid and overwrite with information from
 603          * the vap.
 604          */
 605         if (dip) {
 606                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 607                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
 608                                              xuid, cred, &vap->va_mode);
 609         } else {
 610                 xuid = 0;
 611         }
 612         ip->ino_data.mode = vap->va_mode;
 613
 614         if (vap->va_vaflags & VA_UID_UUID_VALID)
 615                 ip->ino_data.uid = vap->va_uid_uuid;
 616         else if (vap->va_uid != (uid_t)VNOVAL)
 617                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 618         else
 619                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 620
 621         if (vap->va_vaflags & VA_GID_UUID_VALID)
 622                 ip->ino_data.gid = vap->va_gid_uuid;
 623         else if (vap->va_gid != (gid_t)VNOVAL)
 624                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 625         else if (dip)
 626                 ip->ino_data.gid = dip->ino_data.gid;
 627
 628         hammer_ref(&ip->lock);
 629
 630         if (pfsm) {
 631                 ip->pfsm = pfsm;
 632                 hammer_ref(&pfsm->lock);
 633                 error = 0;
 634         } else if (dip->obj_localization == ip->obj_localization) {
 635                 ip->pfsm = dip->pfsm;
 636                 hammer_ref(&ip->pfsm->lock);
 637                 error = 0;
 638         } else {
 639                 ip->pfsm = hammer_load_pseudofs(trans,
 640                                                 ip->obj_localization,
 641                                                 &error);
 642                 error = 0;      /* ignore ENOENT */
 643         }
 644
 645         if (error) {
 646                 hammer_free_inode(ip);
 647                 ip = NULL;
 648         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 649                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 650                 /* not reached */
 651                 hammer_free_inode(ip);
 652         }
 653         *ipp = ip;
 654         return(error);
 655 }
 656
 657 /*
 658  * Final cleanup / freeing of an inode structure
 659  */
 660 static void
 661 hammer_free_inode(hammer_inode_t ip)
 662 {
 663         KKASSERT(ip->lock.refs == 1);
 664         hammer_uncache_node(&ip->cache[0]);
 665         hammer_uncache_node(&ip->cache[1]);
 666         hammer_inode_wakereclaims(ip);
 667         if (ip->objid_cache)
 668                 hammer_clear_objid(ip);
 669         --hammer_count_inodes;
 670         --ip->hmp->count_inodes;
 671         if (ip->pfsm) {
 672                 hammer_rel_pseudofs(ip->hmp, ip->pfsm);
 673                 ip->pfsm = NULL;
 674         }
 675         kfree(ip, M_HAMMER_INO);
 676         ip = NULL;
 677 }
 678
 679 /*
 680  * Retrieve pseudo-fs data.  NULL will never be returned.
 681  *
 682  * If an error occurs *errorp will be set and a default template is returned,
 683  * otherwise *errorp is set to 0.  Typically when an error occurs it will
 684  * be ENOENT.
 685  */
 686 hammer_pseudofs_inmem_t
 687 hammer_load_pseudofs(hammer_transaction_t trans,
 688                      u_int32_t localization, int *errorp)
 689 {
 690         hammer_mount_t hmp = trans->hmp;
 691         hammer_inode_t ip;
 692         hammer_pseudofs_inmem_t pfsm;
 693         struct hammer_cursor cursor;
 694         int bytes;
 695
 696 retry:
 697         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
 698         if (pfsm) {
 699                 hammer_ref(&pfsm->lock);
 700                 *errorp = 0;
 701                 return(pfsm);
 702         }
 703
 704         /*
 705          * PFS records are stored in the root inode (not the PFS root inode,
 706          * but the real root).  Avoid an infinite recursion if loading
 707          * the PFS for the real root.
 708          */
 709         if (localization) {
 710                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
 711                                       HAMMER_MAX_TID,
 712                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
 713         } else {
 714                 ip = NULL;
 715         }
 716
 717         pfsm = kmalloc(sizeof(*pfsm), M_HAMMER, M_WAITOK | M_ZERO);
 718         pfsm->localization = localization;
 719         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
 720         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
 721
 722         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
 723         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
 724                                       HAMMER_LOCALIZE_MISC;
 725         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 726         cursor.key_beg.create_tid = 0;
 727         cursor.key_beg.delete_tid = 0;
 728         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 729         cursor.key_beg.obj_type = 0;
 730         cursor.key_beg.key = localization;
 731         cursor.asof = HAMMER_MAX_TID;
 732         cursor.flags |= HAMMER_CURSOR_ASOF;
 733
 734         if (ip)
 735                 *errorp = hammer_ip_lookup(&cursor);
 736         else
 737                 *errorp = hammer_btree_lookup(&cursor);
 738         if (*errorp == 0) {
 739                 *errorp = hammer_ip_resolve_data(&cursor);
 740                 if (*errorp == 0) {
 741                         if (cursor.data->pfsd.mirror_flags &
 742                             HAMMER_PFSD_DELETED) {
 743                                 *errorp = ENOENT;
 744                         } else {
 745                                 bytes = cursor.leaf->data_len;
 746                                 if (bytes > sizeof(pfsm->pfsd))
 747                                         bytes = sizeof(pfsm->pfsd);
 748                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
 749                         }
 750                 }
 751         }
 752         hammer_done_cursor(&cursor);
 753
 754         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 755         hammer_ref(&pfsm->lock);
 756         if (ip)
 757                 hammer_rel_inode(ip, 0);
 758         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
 759                 kfree(pfsm, M_HAMMER);
 760                 goto retry;
 761         }
 762         return(pfsm);
 763 }
 764
 765 /*
 766  * Store pseudo-fs data.  The backend will automatically delete any prior
 767  * on-disk pseudo-fs data but we have to delete in-memory versions.
 768  */
 769 int
 770 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
 771 {
 772         struct hammer_cursor cursor;
 773         hammer_record_t record;
 774         hammer_inode_t ip;
 775         int error;
 776
 777         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 778                               HAMMER_DEF_LOCALIZATION, 0, &error);
 779 retry:
 780         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 781         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
 782         cursor.key_beg.localization = ip->obj_localization +
 783                                       HAMMER_LOCALIZE_MISC;
 784         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 785         cursor.key_beg.create_tid = 0;
 786         cursor.key_beg.delete_tid = 0;
 787         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 788         cursor.key_beg.obj_type = 0;
 789         cursor.key_beg.key = pfsm->localization;
 790         cursor.asof = HAMMER_MAX_TID;
 791         cursor.flags |= HAMMER_CURSOR_ASOF;
 792
 793         error = hammer_ip_lookup(&cursor);
 794         if (error == 0 && hammer_cursor_inmem(&cursor)) {
 795                 record = cursor.iprec;
 796                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 797                         KKASSERT(cursor.deadlk_rec == NULL);
 798                         hammer_ref(&record->lock);
 799                         cursor.deadlk_rec = record;
 800                         error = EDEADLK;
 801                 } else {
 802                         record->flags |= HAMMER_RECF_DELETED_FE;
 803                         error = 0;
 804                 }
 805         }
 806         if (error == 0 || error == ENOENT) {
 807                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
 808                 record->type = HAMMER_MEM_RECORD_GENERAL;
 809
 810                 record->leaf.base.localization = ip->obj_localization +
 811                                                  HAMMER_LOCALIZE_MISC;
 812                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
 813                 record->leaf.base.key = pfsm->localization;
 814                 record->leaf.data_len = sizeof(pfsm->pfsd);
 815                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
 816                 error = hammer_ip_add_record(trans, record);
 817         }
 818         hammer_done_cursor(&cursor);
 819         if (error == EDEADLK)
 820                 goto retry;
 821         hammer_rel_inode(ip, 0);
 822         return(error);
 823 }
 824
 825 /*
 826  * Create a root directory for a PFS if one does not alredy exist.
 827  *
 828  * The PFS root stands alone so we must also bump the nlinks count
 829  * to prevent it from being destroyed on release.
 830  */
 831 int
 832 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
 833                        hammer_pseudofs_inmem_t pfsm)
 834 {
 835         hammer_inode_t ip;
 836         struct vattr vap;
 837         int error;
 838
 839         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 840                               pfsm->localization, 0, &error);
 841         if (ip == NULL) {
 842                 vattr_null(&vap);
 843                 vap.va_mode = 0755;
 844                 vap.va_type = VDIR;
 845                 error = hammer_create_inode(trans, &vap, cred, NULL, pfsm, &ip);
 846                 if (error == 0) {
 847                         ++ip->ino_data.nlinks;
 848                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 849                 }
 850         }
 851         if (ip)
 852                 hammer_rel_inode(ip, 0);
 853         return(error);
 854 }
 855
 856 /*
 857  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
 858  * if we are unable to disassociate all the inodes.
 859  */
 860 static
 861 int
 862 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
 863 {
 864         int res;
 865
 866         hammer_ref(&ip->lock);
 867         if (ip->lock.refs == 2 && ip->vp)
 868                 vclean_unlocked(ip->vp);
 869         if (ip->lock.refs == 1 && ip->vp == NULL)
 870                 res = 0;
 871         else
 872                 res = -1;       /* stop, someone is using the inode */
 873         hammer_rel_inode(ip, 0);
 874         return(res);
 875 }
 876
 877 int
 878 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
 879 {
 880         int res;
 881         int try;
 882
 883         for (try = res = 0; try < 4; ++try) {
 884                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
 885                                            hammer_inode_pfs_cmp,
 886                                            hammer_unload_pseudofs_callback,
 887                                            &localization);
 888                 if (res == 0 && try > 1)
 889                         break;
 890                 hammer_flusher_sync(trans->hmp);
 891         }
 892         if (res != 0)
 893                 res = ENOTEMPTY;
 894         return(res);
 895 }
 896
 897
 898 /*
 899  * Release a reference on a PFS
 900  */
 901 void
 902 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
 903 {
 904         hammer_unref(&pfsm->lock);
 905         if (pfsm->lock.refs == 0) {
 906                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
 907                 kfree(pfsm, M_HAMMER);
 908         }
 909 }
 910
 911 /*
 912  * Called by hammer_sync_inode().
 913  */
 914 static int
 915 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 916 {
 917         hammer_transaction_t trans = cursor->trans;
 918         hammer_record_t record;
 919         int error;
 920         int redirty;
 921
 922 retry:
 923         error = 0;
 924
 925         /*
 926          * If the inode has a presence on-disk then locate it and mark
 927          * it deleted, setting DELONDISK.
 928          *
 929          * The record may or may not be physically deleted, depending on
 930          * the retention policy.
 931          */
 932         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 933             HAMMER_INODE_ONDISK) {
 934                 hammer_normalize_cursor(cursor);
 935                 cursor->key_beg.localization = ip->obj_localization +
 936                                                HAMMER_LOCALIZE_INODE;
 937                 cursor->key_beg.obj_id = ip->obj_id;
 938                 cursor->key_beg.key = 0;
 939                 cursor->key_beg.create_tid = 0;
 940                 cursor->key_beg.delete_tid = 0;
 941                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 942                 cursor->key_beg.obj_type = 0;
 943                 cursor->asof = ip->obj_asof;
 944                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 945                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 946                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 947
 948                 error = hammer_btree_lookup(cursor);
 949                 if (hammer_debug_inode)
 950                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 951
 952                 if (error == 0) {
 953                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
 954                         if (hammer_debug_inode)
 955                                 kprintf(" error %d\n", error);
 956                         if (error == 0) {
 957                                 ip->flags |= HAMMER_INODE_DELONDISK;
 958                         }
 959                         if (cursor->node)
 960                                 hammer_cache_node(&ip->cache[0], cursor->node);
 961                 }
 962                 if (error == EDEADLK) {
 963                         hammer_done_cursor(cursor);
 964                         error = hammer_init_cursor(trans, cursor,
 965                                                    &ip->cache[0], ip);
 966                         if (hammer_debug_inode)
 967                                 kprintf("IPDED %p %d\n", ip, error);
 968                         if (error == 0)
 969                                 goto retry;
 970                 }
 971         }
 972
 973         /*
 974          * Ok, write out the initial record or a new record (after deleting
 975          * the old one), unless the DELETED flag is set.  This routine will
 976          * clear DELONDISK if it writes out a record.
 977          *
 978          * Update our inode statistics if this is the first application of
 979          * the inode on-disk.
 980          */
 981         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 982                 /*
 983                  * Generate a record and write it to the media.  We clean-up
 984                  * the state before releasing so we do not have to set-up
 985                  * a flush_group.
 986                  */
 987                 record = hammer_alloc_mem_record(ip, 0);
 988                 record->type = HAMMER_MEM_RECORD_INODE;
 989                 record->flush_state = HAMMER_FST_FLUSH;
 990                 record->leaf = ip->sync_ino_leaf;
 991                 record->leaf.base.create_tid = trans->tid;
 992                 record->leaf.data_len = sizeof(ip->sync_ino_data);
 993                 record->leaf.create_ts = trans->time32;
 994                 record->data = (void *)&ip->sync_ino_data;
 995                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
 996
 997                 /*
 998                  * If this flag is set we cannot sync the new file size
 999                  * because we haven't finished related truncations.  The
1000                  * inode will be flushed in another flush group to finish
1001                  * the job.
1002                  */
1003                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1004                     ip->sync_ino_data.size != ip->ino_data.size) {
1005                         redirty = 1;
1006                         ip->sync_ino_data.size = ip->ino_data.size;
1007                 } else {
1008                         redirty = 0;
1009                 }
1010
1011                 for (;;) {
1012                         error = hammer_ip_sync_record_cursor(cursor, record);
1013                         if (hammer_debug_inode)
1014                                 kprintf("GENREC %p rec %08x %d\n",
1015                                         ip, record->flags, error);
1016                         if (error != EDEADLK)
1017                                 break;
1018                         hammer_done_cursor(cursor);
1019                         error = hammer_init_cursor(trans, cursor,
1020                                                    &ip->cache[0], ip);
1021                         if (hammer_debug_inode)
1022                                 kprintf("GENREC reinit %d\n", error);
1023                         if (error)
1024                                 break;
1025                 }
1026
1027                 /*
1028                  * The record isn't managed by the inode's record tree,
1029                  * destroy it whether we succeed or fail.
1030                  */
1031                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1032                 record->flags |= HAMMER_RECF_DELETED_FE | HAMMER_RECF_COMMITTED;
1033                 record->flush_state = HAMMER_FST_IDLE;
1034                 hammer_rel_mem_record(record);
1035
1036                 /*
1037                  * Finish up.
1038                  */
1039                 if (error == 0) {
1040                         if (hammer_debug_inode)
1041                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1042                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1043                                             HAMMER_INODE_ATIME |
1044                                             HAMMER_INODE_MTIME);
1045                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1046                         if (redirty)
1047                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1048
1049                         /*
1050                          * Root volume count of inodes
1051                          */
1052                         hammer_sync_lock_sh(trans);
1053                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1054                                 hammer_modify_volume_field(trans,
1055                                                            trans->rootvol,
1056                                                            vol0_stat_inodes);
1057                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1058                                 hammer_modify_volume_done(trans->rootvol);
1059                                 ip->flags |= HAMMER_INODE_ONDISK;
1060                                 if (hammer_debug_inode)
1061                                         kprintf("NOWONDISK %p\n", ip);
1062                         }
1063                         hammer_sync_unlock(trans);
1064                 }
1065         }
1066
1067         /*
1068          * If the inode has been destroyed, clean out any left-over flags
1069          * that may have been set by the frontend.
1070          */
1071         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1072                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1073                                     HAMMER_INODE_ATIME |
1074                                     HAMMER_INODE_MTIME);
1075         }
1076         return(error);
1077 }
1078
1079 /*
1080  * Update only the itimes fields.
1081  *
1082  * ATIME can be updated without generating any UNDO.  MTIME is updated
1083  * with UNDO so it is guaranteed to be synchronized properly in case of
1084  * a crash.
1085  *
1086  * Neither field is included in the B-Tree leaf element's CRC, which is how
1087  * we can get away with updating ATIME the way we do.
1088  */
1089 static int
1090 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1091 {
1092         hammer_transaction_t trans = cursor->trans;
1093         int error;
1094
1095 retry:
1096         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1097             HAMMER_INODE_ONDISK) {
1098                 return(0);
1099         }
1100
1101         hammer_normalize_cursor(cursor);
1102         cursor->key_beg.localization = ip->obj_localization +
1103                                        HAMMER_LOCALIZE_INODE;
1104         cursor->key_beg.obj_id = ip->obj_id;
1105         cursor->key_beg.key = 0;
1106         cursor->key_beg.create_tid = 0;
1107         cursor->key_beg.delete_tid = 0;
1108         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1109         cursor->key_beg.obj_type = 0;
1110         cursor->asof = ip->obj_asof;
1111         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1112         cursor->flags |= HAMMER_CURSOR_ASOF;
1113         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1114         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1115         cursor->flags |= HAMMER_CURSOR_BACKEND;
1116
1117         error = hammer_btree_lookup(cursor);
1118         if (error == 0) {
1119                 hammer_cache_node(&ip->cache[0], cursor->node);
1120                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1121                         /*
1122                          * Updating MTIME requires an UNDO.  Just cover
1123                          * both atime and mtime.
1124                          */
1125                         hammer_sync_lock_sh(trans);
1126                         hammer_modify_buffer(trans, cursor->data_buffer,
1127                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1128                                      HAMMER_ITIMES_BYTES);
1129                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1130                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1131                         hammer_modify_buffer_done(cursor->data_buffer);
1132                         hammer_sync_unlock(trans);
1133                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1134                         /*
1135                          * Updating atime only can be done in-place with
1136                          * no UNDO.
1137                          */
1138                         hammer_sync_lock_sh(trans);
1139                         hammer_modify_buffer(trans, cursor->data_buffer,
1140                                              NULL, 0);
1141                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1142                         hammer_modify_buffer_done(cursor->data_buffer);
1143                         hammer_sync_unlock(trans);
1144                 }
1145                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1146         }
1147         if (error == EDEADLK) {
1148                 hammer_done_cursor(cursor);
1149                 error = hammer_init_cursor(trans, cursor,
1150                                            &ip->cache[0], ip);
1151                 if (error == 0)
1152                         goto retry;
1153         }
1154         return(error);
1155 }
1156
1157 /*
1158  * Release a reference on an inode, flush as requested.
1159  *
1160  * On the last reference we queue the inode to the flusher for its final
1161  * disposition.
1162  */
1163 void
1164 hammer_rel_inode(struct hammer_inode *ip, int flush)
1165 {
1166         hammer_mount_t hmp = ip->hmp;
1167
1168         /*
1169          * Handle disposition when dropping the last ref.
1170          */
1171         for (;;) {
1172                 if (ip->lock.refs == 1) {
1173                         /*
1174                          * Determine whether on-disk action is needed for
1175                          * the inode's final disposition.
1176                          */
1177                         KKASSERT(ip->vp == NULL);
1178                         hammer_inode_unloadable_check(ip, 0);
1179                         if (ip->flags & HAMMER_INODE_MODMASK) {
1180                                 if (hmp->rsv_inodes > desiredvnodes) {
1181                                         hammer_flush_inode(ip,
1182                                                            HAMMER_FLUSH_SIGNAL);
1183                                 } else {
1184                                         hammer_flush_inode(ip, 0);
1185                                 }
1186                         } else if (ip->lock.refs == 1) {
1187                                 hammer_unload_inode(ip);
1188                                 break;
1189                         }
1190                 } else {
1191                         if (flush)
1192                                 hammer_flush_inode(ip, 0);
1193
1194                         /*
1195                          * The inode still has multiple refs, try to drop
1196                          * one ref.
1197                          */
1198                         KKASSERT(ip->lock.refs >= 1);
1199                         if (ip->lock.refs > 1) {
1200                                 hammer_unref(&ip->lock);
1201                                 break;
1202                         }
1203                 }
1204         }
1205 }
1206
1207 /*
1208  * Unload and destroy the specified inode.  Must be called with one remaining
1209  * reference.  The reference is disposed of.
1210  *
1211  * The inode must be completely clean.
1212  */
1213 static int
1214 hammer_unload_inode(struct hammer_inode *ip)
1215 {
1216         hammer_mount_t hmp = ip->hmp;
1217
1218         KASSERT(ip->lock.refs == 1,
1219                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1220         KKASSERT(ip->vp == NULL);
1221         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1222         KKASSERT(ip->cursor_ip_refs == 0);
1223         KKASSERT(ip->lock.lockcount == 0);
1224         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1225
1226         KKASSERT(RB_EMPTY(&ip->rec_tree));
1227         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1228
1229         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1230
1231         hammer_free_inode(ip);
1232         return(0);
1233 }
1234
1235 /*
1236  * Called during unmounting if a critical error occured.  The in-memory
1237  * inode and all related structures are destroyed.
1238  *
1239  * If a critical error did not occur the unmount code calls the standard
1240  * release and asserts that the inode is gone.
1241  */
1242 int
1243 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1244 {
1245         hammer_record_t rec;
1246
1247         /*
1248          * Get rid of the inodes in-memory records, regardless of their
1249          * state, and clear the mod-mask.
1250          */
1251         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1252                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1253                 rec->target_ip = NULL;
1254                 if (rec->flush_state == HAMMER_FST_SETUP)
1255                         rec->flush_state = HAMMER_FST_IDLE;
1256         }
1257         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1258                 if (rec->flush_state == HAMMER_FST_FLUSH)
1259                         --rec->flush_group->refs;
1260                 else
1261                         hammer_ref(&rec->lock);
1262                 KKASSERT(rec->lock.refs == 1);
1263                 rec->flush_state = HAMMER_FST_IDLE;
1264                 rec->flush_group = NULL;
1265                 rec->flags |= HAMMER_RECF_DELETED_FE;
1266                 rec->flags |= HAMMER_RECF_DELETED_BE;
1267                 hammer_rel_mem_record(rec);
1268         }
1269         ip->flags &= ~HAMMER_INODE_MODMASK;
1270         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1271         KKASSERT(ip->vp == NULL);
1272
1273         /*
1274          * Remove the inode from any flush group, force it idle.  FLUSH
1275          * and SETUP states have an inode ref.
1276          */
1277         switch(ip->flush_state) {
1278         case HAMMER_FST_FLUSH:
1279                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1280                 --ip->flush_group->refs;
1281                 ip->flush_group = NULL;
1282                 /* fall through */
1283         case HAMMER_FST_SETUP:
1284                 hammer_unref(&ip->lock);
1285                 ip->flush_state = HAMMER_FST_IDLE;
1286                 /* fall through */
1287         case HAMMER_FST_IDLE:
1288                 break;
1289         }
1290
1291         /*
1292          * There shouldn't be any associated vnode.  The unload needs at
1293          * least one ref, if we do have a vp steal its ip ref.
1294          */
1295         if (ip->vp) {
1296                 kprintf("hammer_destroy_inode_callback: Unexpected "
1297                         "vnode association ip %p vp %p\n", ip, ip->vp);
1298                 ip->vp->v_data = NULL;
1299                 ip->vp = NULL;
1300         } else {
1301                 hammer_ref(&ip->lock);
1302         }
1303         hammer_unload_inode(ip);
1304         return(0);
1305 }
1306
1307 /*
1308  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1309  * the read-only flag for cached inodes.
1310  *
1311  * This routine is called from a RB_SCAN().
1312  */
1313 int
1314 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1315 {
1316         hammer_mount_t hmp = ip->hmp;
1317
1318         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1319                 ip->flags |= HAMMER_INODE_RO;
1320         else
1321                 ip->flags &= ~HAMMER_INODE_RO;
1322         return(0);
1323 }
1324
1325 /*
1326  * A transaction has modified an inode, requiring updates as specified by
1327  * the passed flags.
1328  *
1329  * HAMMER_INODE_DDIRTY: Inode data has been updated
1330  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1331  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1332  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1333  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1334  */
1335 void
1336 hammer_modify_inode(hammer_inode_t ip, int flags)
1337 {
1338         /*
1339          * ronly of 0 or 2 does not trigger assertion.
1340          * 2 is a special error state
1341          */
1342         KKASSERT(ip->hmp->ronly != 1 ||
1343                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1344                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1345                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1346         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1347                 ip->flags |= HAMMER_INODE_RSV_INODES;
1348                 ++ip->hmp->rsv_inodes;
1349         }
1350
1351         ip->flags |= flags;
1352 }
1353
1354 /*
1355  * Request that an inode be flushed.  This whole mess cannot block and may
1356  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1357  * actively flush the inode until the flush can be done.
1358  *
1359  * The inode may already be flushing, or may be in a setup state.  We can
1360  * place the inode in a flushing state if it is currently idle and flag it
1361  * to reflush if it is currently flushing.
1362  *
1363  * Upon return if the inode could not be flushed due to a setup
1364  * dependancy, then it will be automatically flushed when the dependancy
1365  * is satisfied.
1366  */
1367 void
1368 hammer_flush_inode(hammer_inode_t ip, int flags)
1369 {
1370         hammer_mount_t hmp;
1371         hammer_flush_group_t flg;
1372         int good;
1373
1374         /*
1375          * next_flush_group is the first flush group we can place the inode
1376          * in.  It may be NULL.  If it becomes full we append a new flush
1377          * group and make that the next_flush_group.
1378          */
1379         hmp = ip->hmp;
1380         while ((flg = hmp->next_flush_group) != NULL) {
1381                 KKASSERT(flg->running == 0);
1382                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1383                         break;
1384                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1385                 hammer_flusher_async(ip->hmp, flg);
1386         }
1387         if (flg == NULL) {
1388                 flg = kmalloc(sizeof(*flg), M_HAMMER, M_WAITOK|M_ZERO);
1389                 hmp->next_flush_group = flg;
1390                 TAILQ_INIT(&flg->flush_list);
1391                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1392         }
1393
1394         /*
1395          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1396          * state we have to put it back into an IDLE state so we can
1397          * drop the extra ref.
1398          *
1399          * If we have a parent dependancy we must still fall through
1400          * so we can run it.
1401          */
1402         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1403                 if (ip->flush_state == HAMMER_FST_SETUP &&
1404                     TAILQ_EMPTY(&ip->target_list)) {
1405                         ip->flush_state = HAMMER_FST_IDLE;
1406                         hammer_rel_inode(ip, 0);
1407                 }
1408                 if (ip->flush_state == HAMMER_FST_IDLE)
1409                         return;
1410         }
1411
1412         /*
1413          * Our flush action will depend on the current state.
1414          */
1415         switch(ip->flush_state) {
1416         case HAMMER_FST_IDLE:
1417                 /*
1418                  * We have no dependancies and can flush immediately.  Some
1419                  * our children may not be flushable so we have to re-test
1420                  * with that additional knowledge.
1421                  */
1422                 hammer_flush_inode_core(ip, flg, flags);
1423                 break;
1424         case HAMMER_FST_SETUP:
1425                 /*
1426                  * Recurse upwards through dependancies via target_list
1427                  * and start their flusher actions going if possible.
1428                  *
1429                  * 'good' is our connectivity.  -1 means we have none and
1430                  * can't flush, 0 means there weren't any dependancies, and
1431                  * 1 means we have good connectivity.
1432                  */
1433                 good = hammer_setup_parent_inodes(ip, flg);
1434
1435                 if (good >= 0) {
1436                         /*
1437                          * We can continue if good >= 0.  Determine how
1438                          * many records under our inode can be flushed (and
1439                          * mark them).
1440                          */
1441                         hammer_flush_inode_core(ip, flg, flags);
1442                 } else {
1443                         /*
1444                          * Parent has no connectivity, tell it to flush
1445                          * us as soon as it does.
1446                          *
1447                          * The REFLUSH flag is also needed to trigger
1448                          * dependancy wakeups.
1449                          */
1450                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1451                                      HAMMER_INODE_REFLUSH;
1452                         if (flags & HAMMER_FLUSH_SIGNAL) {
1453                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1454                                 hammer_flusher_async(ip->hmp, flg);
1455                         }
1456                 }
1457                 break;
1458         case HAMMER_FST_FLUSH:
1459                 /*
1460                  * We are already flushing, flag the inode to reflush
1461                  * if needed after it completes its current flush.
1462                  *
1463                  * The REFLUSH flag is also needed to trigger
1464                  * dependancy wakeups.
1465                  */
1466                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1467                         ip->flags |= HAMMER_INODE_REFLUSH;
1468                 if (flags & HAMMER_FLUSH_SIGNAL) {
1469                         ip->flags |= HAMMER_INODE_RESIGNAL;
1470                         hammer_flusher_async(ip->hmp, flg);
1471                 }
1472                 break;
1473         }
1474 }
1475
1476 /*
1477  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1478  * ip which reference our ip.
1479  *
1480  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1481  *     so for now do not ref/deref the structures.  Note that if we use the
1482  *     ref/rel code later, the rel CAN block.
1483  */
1484 static int
1485 hammer_setup_parent_inodes(hammer_inode_t ip, hammer_flush_group_t flg)
1486 {
1487         hammer_record_t depend;
1488         int good;
1489         int r;
1490
1491         good = 0;
1492         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1493                 r = hammer_setup_parent_inodes_helper(depend, flg);
1494                 KKASSERT(depend->target_ip == ip);
1495                 if (r < 0 && good == 0)
1496                         good = -1;
1497                 if (r > 0)
1498                         good = 1;
1499         }
1500         return(good);
1501 }
1502
1503 /*
1504  * This helper function takes a record representing the dependancy between
1505  * the parent inode and child inode.
1506  *
1507  * record->ip           = parent inode
1508  * record->target_ip    = child inode
1509  *
1510  * We are asked to recurse upwards and convert the record from SETUP
1511  * to FLUSH if possible.
1512  *
1513  * Return 1 if the record gives us connectivity
1514  *
1515  * Return 0 if the record is not relevant
1516  *
1517  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1518  */
1519 static int
1520 hammer_setup_parent_inodes_helper(hammer_record_t record,
1521                                   hammer_flush_group_t flg)
1522 {
1523         hammer_mount_t hmp;
1524         hammer_inode_t pip;
1525         int good;
1526
1527         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1528         pip = record->ip;
1529         hmp = pip->hmp;
1530
1531         /*
1532          * If the record is already flushing, is it in our flush group?
1533          *
1534          * If it is in our flush group but it is a general record or a
1535          * delete-on-disk, it does not improve our connectivity (return 0),
1536          * and if the target inode is not trying to destroy itself we can't
1537          * allow the operation yet anyway (the second return -1).
1538          */
1539         if (record->flush_state == HAMMER_FST_FLUSH) {
1540                 /*
1541                  * If not in our flush group ask the parent to reflush
1542                  * us as soon as possible.
1543                  */
1544                 if (record->flush_group != flg) {
1545                         pip->flags |= HAMMER_INODE_REFLUSH;
1546                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1547                         return(-1);
1548                 }
1549
1550                 /*
1551                  * If in our flush group everything is already set up,
1552                  * just return whether the record will improve our
1553                  * visibility or not.
1554                  */
1555                 if (record->type == HAMMER_MEM_RECORD_ADD)
1556                         return(1);
1557                 return(0);
1558         }
1559
1560         /*
1561          * It must be a setup record.  Try to resolve the setup dependancies
1562          * by recursing upwards so we can place ip on the flush list.
1563          */
1564         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1565
1566         good = hammer_setup_parent_inodes(pip, flg);
1567
1568         /*
1569          * If good < 0 the parent has no connectivity and we cannot safely
1570          * flush the directory entry, which also means we can't flush our
1571          * ip.  Flag the parent and us for downward recursion once the
1572          * parent's connectivity is resolved.
1573          */
1574         if (good < 0) {
1575                 /* pip->flags |= HAMMER_INODE_CONN_DOWN; set by recursion */
1576                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1577                 return(good);
1578         }
1579
1580         /*
1581          * We are go, place the parent inode in a flushing state so we can
1582          * place its record in a flushing state.  Note that the parent
1583          * may already be flushing.  The record must be in the same flush
1584          * group as the parent.
1585          */
1586         if (pip->flush_state != HAMMER_FST_FLUSH)
1587                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1588         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1589         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1590
1591 #if 0
1592         if (record->type == HAMMER_MEM_RECORD_DEL &&
1593             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1594                 /*
1595                  * Regardless of flushing state we cannot sync this path if the
1596                  * record represents a delete-on-disk but the target inode
1597                  * is not ready to sync its own deletion.
1598                  *
1599                  * XXX need to count effective nlinks to determine whether
1600                  * the flush is ok, otherwise removing a hardlink will
1601                  * just leave the DEL record to rot.
1602                  */
1603                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1604                 return(-1);
1605         } else
1606 #endif
1607         if (pip->flush_group == flg) {
1608                 /*
1609                  * Because we have not calculated nlinks yet we can just
1610                  * set records to the flush state if the parent is in
1611                  * the same flush group as we are.
1612                  */
1613                 record->flush_state = HAMMER_FST_FLUSH;
1614                 record->flush_group = flg;
1615                 ++record->flush_group->refs;
1616                 hammer_ref(&record->lock);
1617
1618                 /*
1619                  * A general directory-add contributes to our visibility.
1620                  *
1621                  * Otherwise it is probably a directory-delete or
1622                  * delete-on-disk record and does not contribute to our
1623                  * visbility (but we can still flush it).
1624                  */
1625                 if (record->type == HAMMER_MEM_RECORD_ADD)
1626                         return(1);
1627                 return(0);
1628         } else {
1629                 /*
1630                  * If the parent is not in our flush group we cannot
1631                  * flush this record yet, there is no visibility.
1632                  * We tell the parent to reflush and mark ourselves
1633                  * so the parent knows it should flush us too.
1634                  */
1635                 pip->flags |= HAMMER_INODE_REFLUSH;
1636                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1637                 return(-1);
1638         }
1639 }
1640
1641 /*
1642  * This is the core routine placing an inode into the FST_FLUSH state.
1643  */
1644 static void
1645 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1646 {
1647         int go_count;
1648
1649         /*
1650          * Set flush state and prevent the flusher from cycling into
1651          * the next flush group.  Do not place the ip on the list yet.
1652          * Inodes not in the idle state get an extra reference.
1653          */
1654         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1655         if (ip->flush_state == HAMMER_FST_IDLE)
1656                 hammer_ref(&ip->lock);
1657         ip->flush_state = HAMMER_FST_FLUSH;
1658         ip->flush_group = flg;
1659         ++ip->hmp->flusher.group_lock;
1660         ++ip->hmp->count_iqueued;
1661         ++hammer_count_iqueued;
1662         ++flg->total_count;
1663
1664         /*
1665          * We need to be able to vfsync/truncate from the backend.
1666          */
1667         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1668         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1669                 ip->flags |= HAMMER_INODE_VHELD;
1670                 vref(ip->vp);
1671         }
1672
1673         /*
1674          * Figure out how many in-memory records we can actually flush
1675          * (not including inode meta-data, buffers, etc).
1676          */
1677         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1678         if (flags & HAMMER_FLUSH_RECURSION) {
1679                 /*
1680                  * If this is a upwards recursion we do not want to
1681                  * recurse down again!
1682                  */
1683                 go_count = 1;
1684 #if 0
1685         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1686                 /*
1687                  * No new records are added if we must complete a flush
1688                  * from a previous cycle, but we do have to move the records
1689                  * from the previous cycle to the current one.
1690                  */
1691 #if 0
1692                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1693                                    hammer_syncgrp_child_callback, NULL);
1694 #endif
1695                 go_count = 1;
1696 #endif
1697         } else {
1698                 /*
1699                  * Normal flush, scan records and bring them into the flush.
1700                  * Directory adds and deletes are usually skipped (they are
1701                  * grouped with the related inode rather then with the
1702                  * directory).
1703                  *
1704                  * go_count can be negative, which means the scan aborted
1705                  * due to the flush group being over-full and we should
1706                  * flush what we have.
1707                  */
1708                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1709                                    hammer_setup_child_callback, NULL);
1710         }
1711
1712         /*
1713          * This is a more involved test that includes go_count.  If we
1714          * can't flush, flag the inode and return.  If go_count is 0 we
1715          * were are unable to flush any records in our rec_tree and
1716          * must ignore the XDIRTY flag.
1717          */
1718         if (go_count == 0) {
1719                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1720                         --ip->hmp->count_iqueued;
1721                         --hammer_count_iqueued;
1722
1723                         --flg->total_count;
1724                         ip->flush_state = HAMMER_FST_SETUP;
1725                         ip->flush_group = NULL;
1726                         if (ip->flags & HAMMER_INODE_VHELD) {
1727                                 ip->flags &= ~HAMMER_INODE_VHELD;
1728                                 vrele(ip->vp);
1729                         }
1730
1731                         /*
1732                          * REFLUSH is needed to trigger dependancy wakeups
1733                          * when an inode is in SETUP.
1734                          */
1735                         ip->flags |= HAMMER_INODE_REFLUSH;
1736                         if (flags & HAMMER_FLUSH_SIGNAL) {
1737                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1738                                 hammer_flusher_async(ip->hmp, flg);
1739                         }
1740                         if (--ip->hmp->flusher.group_lock == 0)
1741                                 wakeup(&ip->hmp->flusher.group_lock);
1742                         return;
1743                 }
1744         }
1745
1746         /*
1747          * Snapshot the state of the inode for the backend flusher.
1748          *
1749          * We continue to retain save_trunc_off even when all truncations
1750          * have been resolved as an optimization to determine if we can
1751          * skip the B-Tree lookup for overwrite deletions.
1752          *
1753          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1754          * and stays in ip->flags.  Once set, it stays set until the
1755          * inode is destroyed.
1756          */
1757         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1758                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1759                 ip->sync_trunc_off = ip->trunc_off;
1760                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1761                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1762                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1763
1764                 /*
1765                  * The save_trunc_off used to cache whether the B-Tree
1766                  * holds any records past that point is not used until
1767                  * after the truncation has succeeded, so we can safely
1768                  * set it now.
1769                  */
1770                 if (ip->save_trunc_off > ip->sync_trunc_off)
1771                         ip->save_trunc_off = ip->sync_trunc_off;
1772         }
1773         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1774                            ~HAMMER_INODE_TRUNCATED);
1775         ip->sync_ino_leaf = ip->ino_leaf;
1776         ip->sync_ino_data = ip->ino_data;
1777         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1778 #ifdef DEBUG_TRUNCATE
1779         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1780                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1781 #endif
1782
1783         /*
1784          * The flusher list inherits our inode and reference.
1785          */
1786         KKASSERT(flg->running == 0);
1787         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
1788         if (--ip->hmp->flusher.group_lock == 0)
1789                 wakeup(&ip->hmp->flusher.group_lock);
1790
1791         if (flags & HAMMER_FLUSH_SIGNAL) {
1792                 hammer_flusher_async(ip->hmp, flg);
1793         }
1794 }
1795
1796 /*
1797  * Callback for scan of ip->rec_tree.  Try to include each record in our
1798  * flush.  ip->flush_group has been set but the inode has not yet been
1799  * moved into a flushing state.
1800  *
1801  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1802  * both inodes.
1803  *
1804  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1805  * the caller from shortcutting the flush.
1806  */
1807 static int
1808 hammer_setup_child_callback(hammer_record_t rec, void *data)
1809 {
1810         hammer_flush_group_t flg;
1811         hammer_inode_t target_ip;
1812         hammer_inode_t ip;
1813         int r;
1814
1815         /*
1816          * Deleted records are ignored.  Note that the flush detects deleted
1817          * front-end records at multiple points to deal with races.  This is
1818          * just the first line of defense.  The only time DELETED_FE cannot
1819          * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1820          *
1821          * Don't get confused between record deletion and, say, directory
1822          * entry deletion.  The deletion of a directory entry that is on
1823          * the media has nothing to do with the record deletion flags.
1824          */
1825         if (rec->flags & (HAMMER_RECF_DELETED_FE|HAMMER_RECF_DELETED_BE)) {
1826                 if (rec->flush_state == HAMMER_FST_FLUSH) {
1827                         KKASSERT(rec->flush_group == rec->ip->flush_group);
1828                         r = 1;
1829                 } else {
1830                         r = 0;
1831                 }
1832                 return(r);
1833         }
1834
1835         /*
1836          * If the record is in an idle state it has no dependancies and
1837          * can be flushed.
1838          */
1839         ip = rec->ip;
1840         flg = ip->flush_group;
1841         r = 0;
1842
1843         switch(rec->flush_state) {
1844         case HAMMER_FST_IDLE:
1845                 /*
1846                  * The record has no setup dependancy, we can flush it.
1847                  */
1848                 KKASSERT(rec->target_ip == NULL);
1849                 rec->flush_state = HAMMER_FST_FLUSH;
1850                 rec->flush_group = flg;
1851                 ++flg->refs;
1852                 hammer_ref(&rec->lock);
1853                 r = 1;
1854                 break;
1855         case HAMMER_FST_SETUP:
1856                 /*
1857                  * The record has a setup dependancy.  These are typically
1858                  * directory entry adds and deletes.  Such entries will be
1859                  * flushed when their inodes are flushed so we do not
1860                  * usually have to add them to the flush here.  However,
1861                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
1862                  * it is asking us to flush this record (and it).
1863                  */
1864                 target_ip = rec->target_ip;
1865                 KKASSERT(target_ip != NULL);
1866                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1867
1868                 /*
1869                  * If the target IP is already flushing in our group
1870                  * we could associate the record, but target_ip has
1871                  * already synced ino_data to sync_ino_data and we
1872                  * would also have to adjust nlinks.   Plus there are
1873                  * ordering issues for adds and deletes.
1874                  *
1875                  * Reflush downward if this is an ADD, and upward if
1876                  * this is a DEL.
1877                  */
1878                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1879                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
1880                                 ip->flags |= HAMMER_INODE_REFLUSH;
1881                         else
1882                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1883                         break;
1884                 }
1885
1886                 /*
1887                  * Target IP is not yet flushing.  This can get complex
1888                  * because we have to be careful about the recursion.
1889                  *
1890                  * Directories create an issue for us in that if a flush
1891                  * of a directory is requested the expectation is to flush
1892                  * any pending directory entries, but this will cause the
1893                  * related inodes to recursively flush as well.  We can't
1894                  * really defer the operation so just get as many as we
1895                  * can and
1896                  */
1897 #if 0
1898                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
1899                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
1900                         /*
1901                          * We aren't reclaiming and the target ip was not
1902                          * previously prevented from flushing due to this
1903                          * record dependancy.  Do not flush this record.
1904                          */
1905                         /*r = 0;*/
1906                 } else
1907 #endif
1908                 if (flg->total_count + flg->refs >
1909                            ip->hmp->undo_rec_limit) {
1910                         /*
1911                          * Our flush group is over-full and we risk blowing
1912                          * out the UNDO FIFO.  Stop the scan, flush what we
1913                          * have, then reflush the directory.
1914                          *
1915                          * The directory may be forced through multiple
1916                          * flush groups before it can be completely
1917                          * flushed.
1918                          */
1919                         ip->flags |= HAMMER_INODE_RESIGNAL |
1920                                      HAMMER_INODE_REFLUSH;
1921                         r = -1;
1922                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1923                         /*
1924                          * If the target IP is not flushing we can force
1925                          * it to flush, even if it is unable to write out
1926                          * any of its own records we have at least one in
1927                          * hand that we CAN deal with.
1928                          */
1929                         rec->flush_state = HAMMER_FST_FLUSH;
1930                         rec->flush_group = flg;
1931                         ++flg->refs;
1932                         hammer_ref(&rec->lock);
1933                         hammer_flush_inode_core(target_ip, flg,
1934                                                 HAMMER_FLUSH_RECURSION);
1935                         r = 1;
1936                 } else {
1937                         /*
1938                          * General or delete-on-disk record.
1939                          *
1940                          * XXX this needs help.  If a delete-on-disk we could
1941                          * disconnect the target.  If the target has its own
1942                          * dependancies they really need to be flushed.
1943                          *
1944                          * XXX
1945                          */
1946                         rec->flush_state = HAMMER_FST_FLUSH;
1947                         rec->flush_group = flg;
1948                         ++flg->refs;
1949                         hammer_ref(&rec->lock);
1950                         hammer_flush_inode_core(target_ip, flg,
1951                                                 HAMMER_FLUSH_RECURSION);
1952                         r = 1;
1953                 }
1954                 break;
1955         case HAMMER_FST_FLUSH:
1956                 /*
1957                  * The flush_group should already match.
1958                  */
1959                 KKASSERT(rec->flush_group == flg);
1960                 r = 1;
1961                 break;
1962         }
1963         return(r);
1964 }
1965
1966 #if 0
1967 /*
1968  * This version just moves records already in a flush state to the new
1969  * flush group and that is it.
1970  */
1971 static int
1972 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
1973 {
1974         hammer_inode_t ip = rec->ip;
1975
1976         switch(rec->flush_state) {
1977         case HAMMER_FST_FLUSH:
1978                 KKASSERT(rec->flush_group == ip->flush_group);
1979                 break;
1980         default:
1981                 break;
1982         }
1983         return(0);
1984 }
1985 #endif
1986
1987 /*
1988  * Wait for a previously queued flush to complete.
1989  *
1990  * If a critical error occured we don't try to wait.
1991  */
1992 void
1993 hammer_wait_inode(hammer_inode_t ip)
1994 {
1995         hammer_flush_group_t flg;
1996
1997         flg = NULL;
1998         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
1999                 while (ip->flush_state != HAMMER_FST_IDLE &&
2000                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2001                         if (ip->flush_state == HAMMER_FST_SETUP)
2002                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2003                         if (ip->flush_state != HAMMER_FST_IDLE) {
2004                                 ip->flags |= HAMMER_INODE_FLUSHW;
2005                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2006                         }
2007                 }
2008         }
2009 }
2010
2011 /*
2012  * Called by the backend code when a flush has been completed.
2013  * The inode has already been removed from the flush list.
2014  *
2015  * A pipelined flush can occur, in which case we must re-enter the
2016  * inode on the list and re-copy its fields.
2017  */
2018 void
2019 hammer_flush_inode_done(hammer_inode_t ip, int error)
2020 {
2021         hammer_mount_t hmp;
2022         int dorel;
2023
2024         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2025
2026         hmp = ip->hmp;
2027
2028         /*
2029          * Merge left-over flags back into the frontend and fix the state.
2030          * Incomplete truncations are retained by the backend.
2031          */
2032         ip->error = error;
2033         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2034         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2035
2036         /*
2037          * The backend may have adjusted nlinks, so if the adjusted nlinks
2038          * does not match the fronttend set the frontend's RDIRTY flag again.
2039          */
2040         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2041                 ip->flags |= HAMMER_INODE_DDIRTY;
2042
2043         /*
2044          * Fix up the dirty buffer status.
2045          */
2046         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2047                 ip->flags |= HAMMER_INODE_BUFS;
2048         }
2049
2050         /*
2051          * Re-set the XDIRTY flag if some of the inode's in-memory records
2052          * could not be flushed.
2053          */
2054         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2055                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2056                  (!RB_EMPTY(&ip->rec_tree) &&
2057                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2058
2059         /*
2060          * Do not lose track of inodes which no longer have vnode
2061          * assocations, otherwise they may never get flushed again.
2062          */
2063         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
2064                 ip->flags |= HAMMER_INODE_REFLUSH;
2065
2066         /*
2067          * Adjust the flush state.
2068          */
2069         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2070                 /*
2071                  * We were unable to flush out all our records, leave the
2072                  * inode in a flush state and in the current flush group.
2073                  * The flush group will be re-run.
2074                  *
2075                  * This occurs if the UNDO block gets too full or there is
2076                  * too much dirty meta-data and allows the flusher to
2077                  * finalize the UNDO block and then re-flush.
2078                  */
2079                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2080                 dorel = 0;
2081         } else {
2082                 /*
2083                  * Remove from the flush_group
2084                  */
2085                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2086                 ip->flush_group = NULL;
2087
2088                 /*
2089                  * Clean up the vnode ref and tracking counts.
2090                  */
2091                 if (ip->flags & HAMMER_INODE_VHELD) {
2092                         ip->flags &= ~HAMMER_INODE_VHELD;
2093                         vrele(ip->vp);
2094                 }
2095                 --hmp->count_iqueued;
2096                 --hammer_count_iqueued;
2097
2098                 /*
2099                  * And adjust the state.
2100                  */
2101                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2102                         ip->flush_state = HAMMER_FST_IDLE;
2103                         dorel = 1;
2104                 } else {
2105                         ip->flush_state = HAMMER_FST_SETUP;
2106                         dorel = 0;
2107                 }
2108
2109                 /*
2110                  * If the frontend is waiting for a flush to complete,
2111                  * wake it up.
2112                  */
2113                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2114                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2115                         wakeup(&ip->flags);
2116                 }
2117
2118                 /*
2119                  * If the frontend made more changes and requested another
2120                  * flush, then try to get it running.
2121                  *
2122                  * Reflushes are aborted when the inode is errored out.
2123                  */
2124                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2125                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2126                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2127                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2128                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2129                         } else {
2130                                 hammer_flush_inode(ip, 0);
2131                         }
2132                 }
2133         }
2134
2135         /*
2136          * If we have no parent dependancies we can clear CONN_DOWN
2137          */
2138         if (TAILQ_EMPTY(&ip->target_list))
2139                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2140
2141         /*
2142          * If the inode is now clean drop the space reservation.
2143          */
2144         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2145             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2146                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2147                 --hmp->rsv_inodes;
2148         }
2149
2150         if (dorel)
2151                 hammer_rel_inode(ip, 0);
2152 }
2153
2154 /*
2155  * Called from hammer_sync_inode() to synchronize in-memory records
2156  * to the media.
2157  */
2158 static int
2159 hammer_sync_record_callback(hammer_record_t record, void *data)
2160 {
2161         hammer_cursor_t cursor = data;
2162         hammer_transaction_t trans = cursor->trans;
2163         hammer_mount_t hmp = trans->hmp;
2164         int error;
2165
2166         /*
2167          * Skip records that do not belong to the current flush.
2168          */
2169         ++hammer_stats_record_iterations;
2170         if (record->flush_state != HAMMER_FST_FLUSH)
2171                 return(0);
2172
2173 #if 1
2174         if (record->flush_group != record->ip->flush_group) {
2175                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2176                 Debugger("blah2");
2177                 return(0);
2178         }
2179 #endif
2180         KKASSERT(record->flush_group == record->ip->flush_group);
2181
2182         /*
2183          * Interlock the record using the BE flag.  Once BE is set the
2184          * frontend cannot change the state of FE.
2185          *
2186          * NOTE: If FE is set prior to us setting BE we still sync the
2187          * record out, but the flush completion code converts it to
2188          * a delete-on-disk record instead of destroying it.
2189          */
2190         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2191         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2192
2193         /*
2194          * The backend may have already disposed of the record.
2195          */
2196         if (record->flags & HAMMER_RECF_DELETED_BE) {
2197                 error = 0;
2198                 goto done;
2199         }
2200
2201         /*
2202          * If the whole inode is being deleting all on-disk records will
2203          * be deleted very soon, we can't sync any new records to disk
2204          * because they will be deleted in the same transaction they were
2205          * created in (delete_tid == create_tid), which will assert.
2206          *
2207          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2208          * that we currently panic on.
2209          */
2210         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2211                 switch(record->type) {
2212                 case HAMMER_MEM_RECORD_DATA:
2213                         /*
2214                          * We don't have to do anything, if the record was
2215                          * committed the space will have been accounted for
2216                          * in the blockmap.
2217                          */
2218                         /* fall through */
2219                 case HAMMER_MEM_RECORD_GENERAL:
2220                         record->flags |= HAMMER_RECF_DELETED_FE;
2221                         record->flags |= HAMMER_RECF_DELETED_BE;
2222                         error = 0;
2223                         goto done;
2224                 case HAMMER_MEM_RECORD_ADD:
2225                         panic("hammer_sync_record_callback: illegal add "
2226                               "during inode deletion record %p", record);
2227                         break; /* NOT REACHED */
2228                 case HAMMER_MEM_RECORD_INODE:
2229                         panic("hammer_sync_record_callback: attempt to "
2230                               "sync inode record %p?", record);
2231                         break; /* NOT REACHED */
2232                 case HAMMER_MEM_RECORD_DEL:
2233                         /*
2234                          * Follow through and issue the on-disk deletion
2235                          */
2236                         break;
2237                 }
2238         }
2239
2240         /*
2241          * If DELETED_FE is set special handling is needed for directory
2242          * entries.  Dependant pieces related to the directory entry may
2243          * have already been synced to disk.  If this occurs we have to
2244          * sync the directory entry and then change the in-memory record
2245          * from an ADD to a DELETE to cover the fact that it's been
2246          * deleted by the frontend.
2247          *
2248          * A directory delete covering record (MEM_RECORD_DEL) can never
2249          * be deleted by the frontend.
2250          *
2251          * Any other record type (aka DATA) can be deleted by the frontend.
2252          * XXX At the moment the flusher must skip it because there may
2253          * be another data record in the flush group for the same block,
2254          * meaning that some frontend data changes can leak into the backend's
2255          * synchronization point.
2256          */
2257         if (record->flags & HAMMER_RECF_DELETED_FE) {
2258                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2259                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2260                 } else {
2261                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2262                         record->flags |= HAMMER_RECF_DELETED_BE;
2263                         error = 0;
2264                         goto done;
2265                 }
2266         }
2267
2268         /*
2269          * Assign the create_tid for new records.  Deletions already
2270          * have the record's entire key properly set up.
2271          */
2272         if (record->type != HAMMER_MEM_RECORD_DEL)
2273                 record->leaf.base.create_tid = trans->tid;
2274                 record->leaf.create_ts = trans->time32;
2275         for (;;) {
2276                 error = hammer_ip_sync_record_cursor(cursor, record);
2277                 if (error != EDEADLK)
2278                         break;
2279                 hammer_done_cursor(cursor);
2280                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2281                                            record->ip);
2282                 if (error)
2283                         break;
2284         }
2285         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2286
2287         if (error)
2288                 error = -error;
2289 done:
2290         hammer_flush_record_done(record, error);
2291
2292         /*
2293          * Do partial finalization if we have built up too many dirty
2294          * buffers.  Otherwise a buffer cache deadlock can occur when
2295          * doing things like creating tens of thousands of tiny files.
2296          *
2297          * We must release our cursor lock to avoid a 3-way deadlock
2298          * due to the exclusive sync lock the finalizer must get.
2299          */
2300         if (hammer_flusher_meta_limit(hmp)) {
2301                 hammer_unlock_cursor(cursor, 0);
2302                 hammer_flusher_finalize(trans, 0);
2303                 hammer_lock_cursor(cursor, 0);
2304         }
2305
2306         return(error);
2307 }
2308
2309 /*
2310  * Backend function called by the flusher to sync an inode to media.
2311  */
2312 int
2313 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2314 {
2315         struct hammer_cursor cursor;
2316         hammer_node_t tmp_node;
2317         hammer_record_t depend;
2318         hammer_record_t next;
2319         int error, tmp_error;
2320         u_int64_t nlinks;
2321
2322         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2323                 return(0);
2324
2325         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2326         if (error)
2327                 goto done;
2328
2329         /*
2330          * Any directory records referencing this inode which are not in
2331          * our current flush group must adjust our nlink count for the
2332          * purposes of synchronization to disk.
2333          *
2334          * Records which are in our flush group can be unlinked from our
2335          * inode now, potentially allowing the inode to be physically
2336          * deleted.
2337          *
2338          * This cannot block.
2339          */
2340         nlinks = ip->ino_data.nlinks;
2341         next = TAILQ_FIRST(&ip->target_list);
2342         while ((depend = next) != NULL) {
2343                 next = TAILQ_NEXT(depend, target_entry);
2344                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2345                     depend->flush_group == ip->flush_group) {
2346                         /*
2347                          * If this is an ADD that was deleted by the frontend
2348                          * the frontend nlinks count will have already been
2349                          * decremented, but the backend is going to sync its
2350                          * directory entry and must account for it.  The
2351                          * record will be converted to a delete-on-disk when
2352                          * it gets synced.
2353                          *
2354                          * If the ADD was not deleted by the frontend we
2355                          * can remove the dependancy from our target_list.
2356                          */
2357                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2358                                 ++nlinks;
2359                         } else {
2360                                 TAILQ_REMOVE(&ip->target_list, depend,
2361                                              target_entry);
2362                                 depend->target_ip = NULL;
2363                         }
2364                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2365                         /*
2366                          * Not part of our flush group
2367                          */
2368                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2369                         switch(depend->type) {
2370                         case HAMMER_MEM_RECORD_ADD:
2371                                 --nlinks;
2372                                 break;
2373                         case HAMMER_MEM_RECORD_DEL:
2374                                 ++nlinks;
2375                                 break;
2376                         default:
2377                                 break;
2378                         }
2379                 }
2380         }
2381
2382         /*
2383          * Set dirty if we had to modify the link count.
2384          */
2385         if (ip->sync_ino_data.nlinks != nlinks) {
2386                 KKASSERT((int64_t)nlinks >= 0);
2387                 ip->sync_ino_data.nlinks = nlinks;
2388                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2389         }
2390
2391         /*
2392          * If there is a trunction queued destroy any data past the (aligned)
2393          * truncation point.  Userland will have dealt with the buffer
2394          * containing the truncation point for us.
2395          *
2396          * We don't flush pending frontend data buffers until after we've
2397          * dealt with the truncation.
2398          */
2399         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2400                 /*
2401                  * Interlock trunc_off.  The VOP front-end may continue to
2402                  * make adjustments to it while we are blocked.
2403                  */
2404                 off_t trunc_off;
2405                 off_t aligned_trunc_off;
2406                 int blkmask;
2407
2408                 trunc_off = ip->sync_trunc_off;
2409                 blkmask = hammer_blocksize(trunc_off) - 1;
2410                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2411
2412                 /*
2413                  * Delete any whole blocks on-media.  The front-end has
2414                  * already cleaned out any partial block and made it
2415                  * pending.  The front-end may have updated trunc_off
2416                  * while we were blocked so we only use sync_trunc_off.
2417                  *
2418                  * This operation can blow out the buffer cache, EWOULDBLOCK
2419                  * means we were unable to complete the deletion.  The
2420                  * deletion will update sync_trunc_off in that case.
2421                  */
2422                 error = hammer_ip_delete_range(&cursor, ip,
2423                                                 aligned_trunc_off,
2424                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2425                 if (error == EWOULDBLOCK) {
2426                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2427                         error = 0;
2428                         goto defer_buffer_flush;
2429                 }
2430
2431                 if (error)
2432                         goto done;
2433
2434                 /*
2435                  * Clear the truncation flag on the backend after we have
2436                  * complete the deletions.  Backend data is now good again
2437                  * (including new records we are about to sync, below).
2438                  *
2439                  * Leave sync_trunc_off intact.  As we write additional
2440                  * records the backend will update sync_trunc_off.  This
2441                  * tells the backend whether it can skip the overwrite
2442                  * test.  This should work properly even when the backend
2443                  * writes full blocks where the truncation point straddles
2444                  * the block because the comparison is against the base
2445                  * offset of the record.
2446                  */
2447                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2448                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2449         } else {
2450                 error = 0;
2451         }
2452
2453         /*
2454          * Now sync related records.  These will typically be directory
2455          * entries, records tracking direct-writes, or delete-on-disk records.
2456          */
2457         if (error == 0) {
2458                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2459                                     hammer_sync_record_callback, &cursor);
2460                 if (tmp_error < 0)
2461                         tmp_error = -error;
2462                 if (tmp_error)
2463                         error = tmp_error;
2464         }
2465         hammer_cache_node(&ip->cache[1], cursor.node);
2466
2467         /*
2468          * Re-seek for inode update, assuming our cache hasn't been ripped
2469          * out from under us.
2470          */
2471         if (error == 0) {
2472                 tmp_node = hammer_ref_node_safe(ip->hmp, &ip->cache[0], &error);
2473                 if (tmp_node) {
2474                         hammer_cursor_downgrade(&cursor);
2475                         hammer_lock_sh(&tmp_node->lock);
2476                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2477                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2478                         hammer_unlock(&tmp_node->lock);
2479                         hammer_rel_node(tmp_node);
2480                 }
2481                 error = 0;
2482         }
2483
2484         /*
2485          * If we are deleting the inode the frontend had better not have
2486          * any active references on elements making up the inode.
2487          *
2488          * The call to hammer_ip_delete_clean() cleans up auxillary records
2489          * but not DB or DATA records.  Those must have already been deleted
2490          * by the normal truncation mechanic.
2491          */
2492         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2493                 RB_EMPTY(&ip->rec_tree)  &&
2494             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2495             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2496                 int count1 = 0;
2497
2498                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2499                 if (error == 0) {
2500                         ip->flags |= HAMMER_INODE_DELETED;
2501                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2502                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2503                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2504
2505                         /*
2506                          * Set delete_tid in both the frontend and backend
2507                          * copy of the inode record.  The DELETED flag handles
2508                          * this, do not set RDIRTY.
2509                          */
2510                         ip->ino_leaf.base.delete_tid = trans->tid;
2511                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2512                         ip->ino_leaf.delete_ts = trans->time32;
2513                         ip->sync_ino_leaf.delete_ts = trans->time32;
2514
2515
2516                         /*
2517                          * Adjust the inode count in the volume header
2518                          */
2519                         hammer_sync_lock_sh(trans);
2520                         if (ip->flags & HAMMER_INODE_ONDISK) {
2521                                 hammer_modify_volume_field(trans,
2522                                                            trans->rootvol,
2523                                                            vol0_stat_inodes);
2524                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2525                                 hammer_modify_volume_done(trans->rootvol);
2526                         }
2527                         hammer_sync_unlock(trans);
2528                 }
2529         }
2530
2531         if (error)
2532                 goto done;
2533         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2534
2535 defer_buffer_flush:
2536         /*
2537          * Now update the inode's on-disk inode-data and/or on-disk record.
2538          * DELETED and ONDISK are managed only in ip->flags.
2539          *
2540          * In the case of a defered buffer flush we still update the on-disk
2541          * inode to satisfy visibility requirements if there happen to be
2542          * directory dependancies.
2543          */
2544         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2545         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2546                 /*
2547                  * If deleted and on-disk, don't set any additional flags.
2548                  * the delete flag takes care of things.
2549                  *
2550                  * Clear flags which may have been set by the frontend.
2551                  */
2552                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2553                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2554                                     HAMMER_INODE_DELETING);
2555                 break;
2556         case HAMMER_INODE_DELETED:
2557                 /*
2558                  * Take care of the case where a deleted inode was never
2559                  * flushed to the disk in the first place.
2560                  *
2561                  * Clear flags which may have been set by the frontend.
2562                  */
2563                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2564                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2565                                     HAMMER_INODE_DELETING);
2566                 while (RB_ROOT(&ip->rec_tree)) {
2567                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2568                         hammer_ref(&record->lock);
2569                         KKASSERT(record->lock.refs == 1);
2570                         record->flags |= HAMMER_RECF_DELETED_FE;
2571                         record->flags |= HAMMER_RECF_DELETED_BE;
2572                         hammer_rel_mem_record(record);
2573                 }
2574                 break;
2575         case HAMMER_INODE_ONDISK:
2576                 /*
2577                  * If already on-disk, do not set any additional flags.
2578                  */
2579                 break;
2580         default:
2581                 /*
2582                  * If not on-disk and not deleted, set DDIRTY to force
2583                  * an initial record to be written.
2584                  *
2585                  * Also set the create_tid in both the frontend and backend
2586                  * copy of the inode record.
2587                  */
2588                 ip->ino_leaf.base.create_tid = trans->tid;
2589                 ip->ino_leaf.create_ts = trans->time32;
2590                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2591                 ip->sync_ino_leaf.create_ts = trans->time32;
2592                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2593                 break;
2594         }
2595
2596         /*
2597          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2598          * is already on-disk the old record is marked as deleted.
2599          *
2600          * If DELETED is set hammer_update_inode() will delete the existing
2601          * record without writing out a new one.
2602          *
2603          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2604          */
2605         if (ip->flags & HAMMER_INODE_DELETED) {
2606                 error = hammer_update_inode(&cursor, ip);
2607         } else
2608         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2609             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2610                 error = hammer_update_itimes(&cursor, ip);
2611         } else
2612         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2613                 error = hammer_update_inode(&cursor, ip);
2614         }
2615 done:
2616         if (error) {
2617                 hammer_critical_error(ip->hmp, ip, error,
2618                                       "while syncing inode");
2619         }
2620         hammer_done_cursor(&cursor);
2621         return(error);
2622 }
2623
2624 /*
2625  * This routine is called when the OS is no longer actively referencing
2626  * the inode (but might still be keeping it cached), or when releasing
2627  * the last reference to an inode.
2628  *
2629  * At this point if the inode's nlinks count is zero we want to destroy
2630  * it, which may mean destroying it on-media too.
2631  */
2632 void
2633 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2634 {
2635         struct vnode *vp;
2636
2637         /*
2638          * Set the DELETING flag when the link count drops to 0 and the
2639          * OS no longer has any opens on the inode.
2640          *
2641          * The backend will clear DELETING (a mod flag) and set DELETED
2642          * (a state flag) when it is actually able to perform the
2643          * operation.
2644          */
2645         if (ip->ino_data.nlinks == 0 &&
2646             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2647                 ip->flags |= HAMMER_INODE_DELETING;
2648                 ip->flags |= HAMMER_INODE_TRUNCATED;
2649                 ip->trunc_off = 0;
2650                 vp = NULL;
2651                 if (getvp) {
2652                         if (hammer_get_vnode(ip, &vp) != 0)
2653                                 return;
2654                 }
2655
2656                 /*
2657                  * Final cleanup
2658                  */
2659                 if (ip->vp) {
2660                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2661                         vnode_pager_setsize(ip->vp, 0);
2662                 }
2663                 if (getvp) {
2664                         vput(vp);
2665                 }
2666         }
2667 }
2668
2669 /*
2670  * After potentially resolving a dependancy the inode is tested
2671  * to determine whether it needs to be reflushed.
2672  */
2673 void
2674 hammer_test_inode(hammer_inode_t ip)
2675 {
2676         if (ip->flags & HAMMER_INODE_REFLUSH) {
2677                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2678                 hammer_ref(&ip->lock);
2679                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2680                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2681                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2682                 } else {
2683                         hammer_flush_inode(ip, 0);
2684                 }
2685                 hammer_rel_inode(ip, 0);
2686         }
2687 }
2688
2689 /*
2690  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2691  * reassociated with a vp or just before it gets freed.
2692  *
2693  * Wakeup one thread blocked waiting on reclaims to complete.  Note that
2694  * the inode the thread is waiting on behalf of is a different inode then
2695  * the inode we are called with.  This is to create a pipeline.
2696  */
2697 static void
2698 hammer_inode_wakereclaims(hammer_inode_t ip)
2699 {
2700         struct hammer_reclaim *reclaim;
2701         hammer_mount_t hmp = ip->hmp;
2702
2703         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2704                 return;
2705
2706         --hammer_count_reclaiming;
2707         --hmp->inode_reclaims;
2708         ip->flags &= ~HAMMER_INODE_RECLAIM;
2709
2710         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
2711                 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2712                 reclaim->okydoky = 1;
2713                 wakeup(reclaim);
2714         }
2715 }
2716
2717 /*
2718  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
2719  * inodes build up before we start blocking.
2720  *
2721  * When we block we don't care *which* inode has finished reclaiming,
2722  * as lone as one does.  This is somewhat heuristical... we also put a
2723  * cap on how long we are willing to wait.
2724  */
2725 void
2726 hammer_inode_waitreclaims(hammer_mount_t hmp)
2727 {
2728         struct hammer_reclaim reclaim;
2729         int delay;
2730
2731         if (hmp->inode_reclaims > HAMMER_RECLAIM_WAIT) {
2732                 reclaim.okydoky = 0;
2733                 TAILQ_INSERT_TAIL(&hmp->reclaim_list,
2734                                   &reclaim, entry);
2735         } else {
2736                 reclaim.okydoky = 1;
2737         }
2738
2739         if (reclaim.okydoky == 0) {
2740                 delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2741                         HAMMER_RECLAIM_WAIT;
2742                 if (delay >= 0)
2743                         tsleep(&reclaim, 0, "hmrrcm", delay + 1);
2744                 if (reclaim.okydoky == 0)
2745                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2746         }
2747 }
2748