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 1.102 2008/07/13 09:32:48 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, 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, 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);
 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                 if (error) {
 952                         kprintf("error %d\n", error);
 953                         Debugger("hammer_update_inode");
 954                 }
 955
 956                 if (error == 0) {
 957                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
 958                         if (hammer_debug_inode)
 959                                 kprintf(" error %d\n", error);
 960                         if (error && error != EDEADLK) {
 961                                 kprintf("error %d\n", error);
 962                                 Debugger("hammer_update_inode2");
 963                         }
 964                         if (error == 0) {
 965                                 ip->flags |= HAMMER_INODE_DELONDISK;
 966                         }
 967                         if (cursor->node)
 968                                 hammer_cache_node(&ip->cache[0], cursor->node);
 969                 }
 970                 if (error == EDEADLK) {
 971                         hammer_done_cursor(cursor);
 972                         error = hammer_init_cursor(trans, cursor,
 973                                                    &ip->cache[0], ip);
 974                         if (hammer_debug_inode)
 975                                 kprintf("IPDED %p %d\n", ip, error);
 976                         if (error == 0)
 977                                 goto retry;
 978                 }
 979         }
 980
 981         /*
 982          * Ok, write out the initial record or a new record (after deleting
 983          * the old one), unless the DELETED flag is set.  This routine will
 984          * clear DELONDISK if it writes out a record.
 985          *
 986          * Update our inode statistics if this is the first application of
 987          * the inode on-disk.
 988          */
 989         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 990                 /*
 991                  * Generate a record and write it to the media.  We clean-up
 992                  * the state before releasing so we do not have to set-up
 993                  * a flush_group.
 994                  */
 995                 record = hammer_alloc_mem_record(ip, 0);
 996                 record->type = HAMMER_MEM_RECORD_INODE;
 997                 record->flush_state = HAMMER_FST_FLUSH;
 998                 record->leaf = ip->sync_ino_leaf;
 999                 record->leaf.base.create_tid = trans->tid;
1000                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1001                 record->leaf.create_ts = trans->time32;
1002                 record->data = (void *)&ip->sync_ino_data;
1003                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1004
1005                 /*
1006                  * If this flag is set we cannot sync the new file size
1007                  * because we haven't finished related truncations.  The
1008                  * inode will be flushed in another flush group to finish
1009                  * the job.
1010                  */
1011                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1012                     ip->sync_ino_data.size != ip->ino_data.size) {
1013                         redirty = 1;
1014                         ip->sync_ino_data.size = ip->ino_data.size;
1015                 } else {
1016                         redirty = 0;
1017                 }
1018
1019                 for (;;) {
1020                         error = hammer_ip_sync_record_cursor(cursor, record);
1021                         if (hammer_debug_inode)
1022                                 kprintf("GENREC %p rec %08x %d\n",
1023                                         ip, record->flags, error);
1024                         if (error != EDEADLK)
1025                                 break;
1026                         hammer_done_cursor(cursor);
1027                         error = hammer_init_cursor(trans, cursor,
1028                                                    &ip->cache[0], ip);
1029                         if (hammer_debug_inode)
1030                                 kprintf("GENREC reinit %d\n", error);
1031                         if (error)
1032                                 break;
1033                 }
1034                 if (error) {
1035                         kprintf("error %d\n", error);
1036                         Debugger("hammer_update_inode3");
1037                 }
1038
1039                 /*
1040                  * The record isn't managed by the inode's record tree,
1041                  * destroy it whether we succeed or fail.
1042                  */
1043                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1044                 record->flags |= HAMMER_RECF_DELETED_FE;
1045                 record->flush_state = HAMMER_FST_IDLE;
1046                 hammer_rel_mem_record(record);
1047
1048                 /*
1049                  * Finish up.
1050                  */
1051                 if (error == 0) {
1052                         if (hammer_debug_inode)
1053                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1054                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1055                                             HAMMER_INODE_ATIME |
1056                                             HAMMER_INODE_MTIME);
1057                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1058                         if (redirty)
1059                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1060
1061                         /*
1062                          * Root volume count of inodes
1063                          */
1064                         hammer_sync_lock_sh(trans);
1065                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1066                                 hammer_modify_volume_field(trans,
1067                                                            trans->rootvol,
1068                                                            vol0_stat_inodes);
1069                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1070                                 hammer_modify_volume_done(trans->rootvol);
1071                                 ip->flags |= HAMMER_INODE_ONDISK;
1072                                 if (hammer_debug_inode)
1073                                         kprintf("NOWONDISK %p\n", ip);
1074                         }
1075                         hammer_sync_unlock(trans);
1076                 }
1077         }
1078
1079         /*
1080          * If the inode has been destroyed, clean out any left-over flags
1081          * that may have been set by the frontend.
1082          */
1083         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1084                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1085                                     HAMMER_INODE_ATIME |
1086                                     HAMMER_INODE_MTIME);
1087         }
1088         return(error);
1089 }
1090
1091 /*
1092  * Update only the itimes fields.
1093  *
1094  * ATIME can be updated without generating any UNDO.  MTIME is updated
1095  * with UNDO so it is guaranteed to be synchronized properly in case of
1096  * a crash.
1097  *
1098  * Neither field is included in the B-Tree leaf element's CRC, which is how
1099  * we can get away with updating ATIME the way we do.
1100  */
1101 static int
1102 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1103 {
1104         hammer_transaction_t trans = cursor->trans;
1105         int error;
1106
1107 retry:
1108         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1109             HAMMER_INODE_ONDISK) {
1110                 return(0);
1111         }
1112
1113         hammer_normalize_cursor(cursor);
1114         cursor->key_beg.localization = ip->obj_localization +
1115                                        HAMMER_LOCALIZE_INODE;
1116         cursor->key_beg.obj_id = ip->obj_id;
1117         cursor->key_beg.key = 0;
1118         cursor->key_beg.create_tid = 0;
1119         cursor->key_beg.delete_tid = 0;
1120         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1121         cursor->key_beg.obj_type = 0;
1122         cursor->asof = ip->obj_asof;
1123         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1124         cursor->flags |= HAMMER_CURSOR_ASOF;
1125         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1126         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1127         cursor->flags |= HAMMER_CURSOR_BACKEND;
1128
1129         error = hammer_btree_lookup(cursor);
1130         if (error) {
1131                 kprintf("error %d\n", error);
1132                 Debugger("hammer_update_itimes1");
1133         }
1134         if (error == 0) {
1135                 hammer_cache_node(&ip->cache[0], cursor->node);
1136                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1137                         /*
1138                          * Updating MTIME requires an UNDO.  Just cover
1139                          * both atime and mtime.
1140                          */
1141                         hammer_sync_lock_sh(trans);
1142                         hammer_modify_buffer(trans, cursor->data_buffer,
1143                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1144                                      HAMMER_ITIMES_BYTES);
1145                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1146                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1147                         hammer_modify_buffer_done(cursor->data_buffer);
1148                         hammer_sync_unlock(trans);
1149                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1150                         /*
1151                          * Updating atime only can be done in-place with
1152                          * no UNDO.
1153                          */
1154                         hammer_sync_lock_sh(trans);
1155                         hammer_modify_buffer(trans, cursor->data_buffer,
1156                                              NULL, 0);
1157                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1158                         hammer_modify_buffer_done(cursor->data_buffer);
1159                         hammer_sync_unlock(trans);
1160                 }
1161                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1162         }
1163         if (error == EDEADLK) {
1164                 hammer_done_cursor(cursor);
1165                 error = hammer_init_cursor(trans, cursor,
1166                                            &ip->cache[0], ip);
1167                 if (error == 0)
1168                         goto retry;
1169         }
1170         return(error);
1171 }
1172
1173 /*
1174  * Release a reference on an inode, flush as requested.
1175  *
1176  * On the last reference we queue the inode to the flusher for its final
1177  * disposition.
1178  */
1179 void
1180 hammer_rel_inode(struct hammer_inode *ip, int flush)
1181 {
1182         hammer_mount_t hmp = ip->hmp;
1183
1184         /*
1185          * Handle disposition when dropping the last ref.
1186          */
1187         for (;;) {
1188                 if (ip->lock.refs == 1) {
1189                         /*
1190                          * Determine whether on-disk action is needed for
1191                          * the inode's final disposition.
1192                          */
1193                         KKASSERT(ip->vp == NULL);
1194                         hammer_inode_unloadable_check(ip, 0);
1195                         if (ip->flags & HAMMER_INODE_MODMASK) {
1196                                 if (hmp->rsv_inodes > desiredvnodes) {
1197                                         hammer_flush_inode(ip,
1198                                                            HAMMER_FLUSH_SIGNAL);
1199                                 } else {
1200                                         hammer_flush_inode(ip, 0);
1201                                 }
1202                         } else if (ip->lock.refs == 1) {
1203                                 hammer_unload_inode(ip);
1204                                 break;
1205                         }
1206                 } else {
1207                         if (flush)
1208                                 hammer_flush_inode(ip, 0);
1209
1210                         /*
1211                          * The inode still has multiple refs, try to drop
1212                          * one ref.
1213                          */
1214                         KKASSERT(ip->lock.refs >= 1);
1215                         if (ip->lock.refs > 1) {
1216                                 hammer_unref(&ip->lock);
1217                                 break;
1218                         }
1219                 }
1220         }
1221 }
1222
1223 /*
1224  * Unload and destroy the specified inode.  Must be called with one remaining
1225  * reference.  The reference is disposed of.
1226  *
1227  * This can only be called in the context of the flusher.
1228  */
1229 static int
1230 hammer_unload_inode(struct hammer_inode *ip)
1231 {
1232         hammer_mount_t hmp = ip->hmp;
1233
1234         KASSERT(ip->lock.refs == 1,
1235                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1236         KKASSERT(ip->vp == NULL);
1237         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1238         KKASSERT(ip->cursor_ip_refs == 0);
1239         KKASSERT(ip->lock.lockcount == 0);
1240         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1241
1242         KKASSERT(RB_EMPTY(&ip->rec_tree));
1243         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1244
1245         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1246
1247         hammer_free_inode(ip);
1248         return(0);
1249 }
1250
1251 /*
1252  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1253  * the read-only flag for cached inodes.
1254  *
1255  * This routine is called from a RB_SCAN().
1256  */
1257 int
1258 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1259 {
1260         hammer_mount_t hmp = ip->hmp;
1261
1262         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1263                 ip->flags |= HAMMER_INODE_RO;
1264         else
1265                 ip->flags &= ~HAMMER_INODE_RO;
1266         return(0);
1267 }
1268
1269 /*
1270  * A transaction has modified an inode, requiring updates as specified by
1271  * the passed flags.
1272  *
1273  * HAMMER_INODE_DDIRTY: Inode data has been updated
1274  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1275  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1276  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1277  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1278  */
1279 void
1280 hammer_modify_inode(hammer_inode_t ip, int flags)
1281 {
1282         KKASSERT(ip->hmp->ronly == 0 ||
1283                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1284                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1285                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1286         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1287                 ip->flags |= HAMMER_INODE_RSV_INODES;
1288                 ++ip->hmp->rsv_inodes;
1289         }
1290
1291         ip->flags |= flags;
1292 }
1293
1294 /*
1295  * Request that an inode be flushed.  This whole mess cannot block and may
1296  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1297  * actively flush the inode until the flush can be done.
1298  *
1299  * The inode may already be flushing, or may be in a setup state.  We can
1300  * place the inode in a flushing state if it is currently idle and flag it
1301  * to reflush if it is currently flushing.
1302  *
1303  * If the HAMMER_FLUSH_SYNCHRONOUS flag is specified we will attempt to
1304  * flush the indoe synchronously using the caller's context.
1305  */
1306 void
1307 hammer_flush_inode(hammer_inode_t ip, int flags)
1308 {
1309         hammer_mount_t hmp;
1310         hammer_flush_group_t flg;
1311         int good;
1312
1313         /*
1314          * Setup a flush group.  It remains cached so it is ok if we
1315          * wind up not flushing the inode.
1316          */
1317         hmp = ip->hmp;
1318         flg = TAILQ_LAST(&ip->hmp->flush_group_list, hammer_flush_group_list);
1319
1320         if (flg) {
1321                 if (flg->running) {
1322                         flg = NULL;
1323                 } else if (flg->total_count + flg->refs >
1324                            ip->hmp->undo_rec_limit) {
1325                         hammer_flusher_async(ip->hmp, flg);
1326                         flg = NULL;
1327                 }
1328         }
1329         if (flg == NULL) {
1330                 flg = kmalloc(sizeof(*flg), M_HAMMER, M_WAITOK|M_ZERO);
1331                 TAILQ_INIT(&flg->flush_list);
1332                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1333         }
1334
1335         /*
1336          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1337          * state we have to put it back into an IDLE state so we can
1338          * drop the extra ref.
1339          *
1340          * If we have a parent dependancy we must still fall through
1341          * so we can run it.
1342          */
1343         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1344                 if (ip->flush_state == HAMMER_FST_SETUP &&
1345                     TAILQ_EMPTY(&ip->target_list)) {
1346                         ip->flush_state = HAMMER_FST_IDLE;
1347                         hammer_rel_inode(ip, 0);
1348                 }
1349                 if (ip->flush_state == HAMMER_FST_IDLE)
1350                         return;
1351         }
1352
1353         /*
1354          * Our flush action will depend on the current state.
1355          */
1356         switch(ip->flush_state) {
1357         case HAMMER_FST_IDLE:
1358                 /*
1359                  * We have no dependancies and can flush immediately.  Some
1360                  * our children may not be flushable so we have to re-test
1361                  * with that additional knowledge.
1362                  */
1363                 hammer_flush_inode_core(ip, flg, flags);
1364                 break;
1365         case HAMMER_FST_SETUP:
1366                 /*
1367                  * Recurse upwards through dependancies via target_list
1368                  * and start their flusher actions going if possible.
1369                  *
1370                  * 'good' is our connectivity.  -1 means we have none and
1371                  * can't flush, 0 means there weren't any dependancies, and
1372                  * 1 means we have good connectivity.
1373                  */
1374                 good = hammer_setup_parent_inodes(ip, flg);
1375
1376                 /*
1377                  * We can continue if good >= 0.  Determine how many records
1378                  * under our inode can be flushed (and mark them).
1379                  */
1380                 if (good >= 0) {
1381                         hammer_flush_inode_core(ip, flg, flags);
1382                 } else {
1383                         ip->flags |= HAMMER_INODE_REFLUSH;
1384                         if (flags & HAMMER_FLUSH_SIGNAL) {
1385                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1386                                 hammer_flusher_async(ip->hmp, flg);
1387                         }
1388                 }
1389                 break;
1390         default:
1391                 /*
1392                  * We are already flushing, flag the inode to reflush
1393                  * if needed after it completes its current flush.
1394                  */
1395                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1396                         ip->flags |= HAMMER_INODE_REFLUSH;
1397                 if (flags & HAMMER_FLUSH_SIGNAL) {
1398                         ip->flags |= HAMMER_INODE_RESIGNAL;
1399                         hammer_flusher_async(ip->hmp, flg);
1400                 }
1401                 break;
1402         }
1403 }
1404
1405 /*
1406  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1407  * ip which reference our ip.
1408  *
1409  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1410  *     so for now do not ref/deref the structures.  Note that if we use the
1411  *     ref/rel code later, the rel CAN block.
1412  */
1413 static int
1414 hammer_setup_parent_inodes(hammer_inode_t ip, hammer_flush_group_t flg)
1415 {
1416         hammer_record_t depend;
1417         int good;
1418         int r;
1419
1420         good = 0;
1421         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1422                 r = hammer_setup_parent_inodes_helper(depend, flg);
1423                 KKASSERT(depend->target_ip == ip);
1424                 if (r < 0 && good == 0)
1425                         good = -1;
1426                 if (r > 0)
1427                         good = 1;
1428         }
1429         return(good);
1430 }
1431
1432 /*
1433  * This helper function takes a record representing the dependancy between
1434  * the parent inode and child inode.
1435  *
1436  * record->ip           = parent inode
1437  * record->target_ip    = child inode
1438  *
1439  * We are asked to recurse upwards and convert the record from SETUP
1440  * to FLUSH if possible.
1441  *
1442  * Return 1 if the record gives us connectivity
1443  *
1444  * Return 0 if the record is not relevant
1445  *
1446  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1447  */
1448 static int
1449 hammer_setup_parent_inodes_helper(hammer_record_t record,
1450                                   hammer_flush_group_t flg)
1451 {
1452         hammer_mount_t hmp;
1453         hammer_inode_t pip;
1454         int good;
1455
1456         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1457         pip = record->ip;
1458         hmp = pip->hmp;
1459
1460         /*
1461          * If the record is already flushing, is it in our flush group?
1462          *
1463          * If it is in our flush group but it is a general record or a
1464          * delete-on-disk, it does not improve our connectivity (return 0),
1465          * and if the target inode is not trying to destroy itself we can't
1466          * allow the operation yet anyway (the second return -1).
1467          */
1468         if (record->flush_state == HAMMER_FST_FLUSH) {
1469                 if (record->flush_group != flg) {
1470                         pip->flags |= HAMMER_INODE_REFLUSH;
1471                         return(-1);
1472                 }
1473                 if (record->type == HAMMER_MEM_RECORD_ADD)
1474                         return(1);
1475                 /* GENERAL or DEL */
1476                 return(0);
1477         }
1478
1479         /*
1480          * It must be a setup record.  Try to resolve the setup dependancies
1481          * by recursing upwards so we can place ip on the flush list.
1482          */
1483         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1484
1485         good = hammer_setup_parent_inodes(pip, flg);
1486
1487         /*
1488          * We can't flush ip because it has no connectivity (XXX also check
1489          * nlinks for pre-existing connectivity!).  Flag it so any resolution
1490          * recurses back down.
1491          */
1492         if (good < 0) {
1493                 pip->flags |= HAMMER_INODE_REFLUSH;
1494                 return(good);
1495         }
1496
1497         /*
1498          * We are go, place the parent inode in a flushing state so we can
1499          * place its record in a flushing state.  Note that the parent
1500          * may already be flushing.  The record must be in the same flush
1501          * group as the parent.
1502          */
1503         if (pip->flush_state != HAMMER_FST_FLUSH)
1504                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1505         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1506         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1507
1508 #if 0
1509         if (record->type == HAMMER_MEM_RECORD_DEL &&
1510             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1511                 /*
1512                  * Regardless of flushing state we cannot sync this path if the
1513                  * record represents a delete-on-disk but the target inode
1514                  * is not ready to sync its own deletion.
1515                  *
1516                  * XXX need to count effective nlinks to determine whether
1517                  * the flush is ok, otherwise removing a hardlink will
1518                  * just leave the DEL record to rot.
1519                  */
1520                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1521                 return(-1);
1522         } else
1523 #endif
1524         if (pip->flush_group == flg) {
1525                 /*
1526                  * This is the record we wanted to synchronize.  If the
1527                  * record went into a flush state while we blocked it
1528                  * had better be in the correct flush group.
1529                  */
1530                 if (record->flush_state != HAMMER_FST_FLUSH) {
1531                         record->flush_state = HAMMER_FST_FLUSH;
1532                         record->flush_group = pip->flush_group;
1533                         ++record->flush_group->refs;
1534                         hammer_ref(&record->lock);
1535                 } else {
1536                         KKASSERT(record->flush_group == pip->flush_group);
1537                 }
1538                 if (record->type == HAMMER_MEM_RECORD_ADD)
1539                         return(1);
1540
1541                 /*
1542                  * A general or delete-on-disk record does not contribute
1543                  * to our visibility.  We can still flush it, however.
1544                  */
1545                 return(0);
1546         } else {
1547                 /*
1548                  * We couldn't resolve the dependancies, request that the
1549                  * inode be flushed when the dependancies can be resolved.
1550                  */
1551                 pip->flags |= HAMMER_INODE_REFLUSH;
1552                 return(-1);
1553         }
1554 }
1555
1556 /*
1557  * This is the core routine placing an inode into the FST_FLUSH state.
1558  */
1559 static void
1560 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1561 {
1562         int go_count;
1563
1564         /*
1565          * Set flush state and prevent the flusher from cycling into
1566          * the next flush group.  Do not place the ip on the list yet.
1567          * Inodes not in the idle state get an extra reference.
1568          */
1569         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1570         if (ip->flush_state == HAMMER_FST_IDLE)
1571                 hammer_ref(&ip->lock);
1572         ip->flush_state = HAMMER_FST_FLUSH;
1573         ip->flush_group = flg;
1574         ++ip->hmp->flusher.group_lock;
1575         ++ip->hmp->count_iqueued;
1576         ++hammer_count_iqueued;
1577         ++flg->total_count;
1578
1579         /*
1580          * We need to be able to vfsync/truncate from the backend.
1581          */
1582         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1583         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1584                 ip->flags |= HAMMER_INODE_VHELD;
1585                 vref(ip->vp);
1586         }
1587
1588         /*
1589          * Figure out how many in-memory records we can actually flush
1590          * (not including inode meta-data, buffers, etc).
1591          */
1592         if (flags & HAMMER_FLUSH_RECURSION) {
1593                 /*
1594                  * If this is a upwards recursion we do not want to
1595                  * recurse down again!
1596                  */
1597                 go_count = 1;
1598         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1599                 /*
1600                  * No new records are added if we must complete a flush
1601                  * from a previous cycle, but we do have to move the records
1602                  * from the previous cycle to the current one.
1603                  */
1604 #if 0
1605                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1606                                    hammer_syncgrp_child_callback, NULL);
1607 #endif
1608                 go_count = 1;
1609         } else {
1610                 /*
1611                  * Normal flush, scan records and bring them into the flush.
1612                  * Directory adds and deletes are usually skipped (they are
1613                  * grouped with the related inode rather then with the
1614                  * directory).
1615                  *
1616                  * go_count can be negative, which means the scan aborted
1617                  * due to the flush group being over-full and we should
1618                  * flush what we have.
1619                  */
1620                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1621                                    hammer_setup_child_callback, NULL);
1622         }
1623
1624         /*
1625          * This is a more involved test that includes go_count.  If we
1626          * can't flush, flag the inode and return.  If go_count is 0 we
1627          * were are unable to flush any records in our rec_tree and
1628          * must ignore the XDIRTY flag.
1629          */
1630         if (go_count == 0) {
1631                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1632                         ip->flags |= HAMMER_INODE_REFLUSH;
1633
1634                         --ip->hmp->count_iqueued;
1635                         --hammer_count_iqueued;
1636
1637                         ip->flush_state = HAMMER_FST_SETUP;
1638                         ip->flush_group = NULL;
1639                         if (ip->flags & HAMMER_INODE_VHELD) {
1640                                 ip->flags &= ~HAMMER_INODE_VHELD;
1641                                 vrele(ip->vp);
1642                         }
1643                         if (flags & HAMMER_FLUSH_SIGNAL) {
1644                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1645                                 hammer_flusher_async(ip->hmp, flg);
1646                         }
1647                         if (--ip->hmp->flusher.group_lock == 0)
1648                                 wakeup(&ip->hmp->flusher.group_lock);
1649                         return;
1650                 }
1651         }
1652
1653         /*
1654          * Snapshot the state of the inode for the backend flusher.
1655          *
1656          * We continue to retain save_trunc_off even when all truncations
1657          * have been resolved as an optimization to determine if we can
1658          * skip the B-Tree lookup for overwrite deletions.
1659          *
1660          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1661          * and stays in ip->flags.  Once set, it stays set until the
1662          * inode is destroyed.
1663          *
1664          * NOTE: If a truncation from a previous flush cycle had to be
1665          * continued into this one, the TRUNCATED flag will still be
1666          * set in sync_flags as will WOULDBLOCK.  When this occurs
1667          * we CANNOT safely integrate a new truncation from the front-end
1668          * because there may be data records in-memory assigned a flush
1669          * state from the previous cycle that are supposed to be flushed
1670          * before the next frontend truncation.
1671          */
1672         if ((ip->flags & (HAMMER_INODE_TRUNCATED | HAMMER_INODE_WOULDBLOCK)) ==
1673             HAMMER_INODE_TRUNCATED) {
1674                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1675                 ip->sync_trunc_off = ip->trunc_off;
1676                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1677                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1678                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1679
1680                 /*
1681                  * The save_trunc_off used to cache whether the B-Tree
1682                  * holds any records past that point is not used until
1683                  * after the truncation has succeeded, so we can safely
1684                  * set it now.
1685                  */
1686                 if (ip->save_trunc_off > ip->sync_trunc_off)
1687                         ip->save_trunc_off = ip->sync_trunc_off;
1688         }
1689         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1690                            ~HAMMER_INODE_TRUNCATED);
1691         ip->sync_ino_leaf = ip->ino_leaf;
1692         ip->sync_ino_data = ip->ino_data;
1693         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1694 #ifdef DEBUG_TRUNCATE
1695         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1696                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1697 #endif
1698
1699         /*
1700          * The flusher list inherits our inode and reference.
1701          */
1702         KKASSERT(flg->running == 0);
1703         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
1704         if (--ip->hmp->flusher.group_lock == 0)
1705                 wakeup(&ip->hmp->flusher.group_lock);
1706
1707         if (flags & HAMMER_FLUSH_SIGNAL) {
1708                 hammer_flusher_async(ip->hmp, flg);
1709         }
1710 }
1711
1712 /*
1713  * Callback for scan of ip->rec_tree.  Try to include each record in our
1714  * flush.  ip->flush_group has been set but the inode has not yet been
1715  * moved into a flushing state.
1716  *
1717  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1718  * both inodes.
1719  *
1720  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1721  * the caller from shortcutting the flush.
1722  */
1723 static int
1724 hammer_setup_child_callback(hammer_record_t rec, void *data)
1725 {
1726         hammer_flush_group_t flg;
1727         hammer_inode_t target_ip;
1728         hammer_inode_t ip;
1729         int r;
1730
1731         /*
1732          * Deleted records are ignored.  Note that the flush detects deleted
1733          * front-end records at multiple points to deal with races.  This is
1734          * just the first line of defense.  The only time DELETED_FE cannot
1735          * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1736          *
1737          * Don't get confused between record deletion and, say, directory
1738          * entry deletion.  The deletion of a directory entry that is on
1739          * the media has nothing to do with the record deletion flags.
1740          */
1741         if (rec->flags & (HAMMER_RECF_DELETED_FE|HAMMER_RECF_DELETED_BE)) {
1742                 if (rec->flush_state == HAMMER_FST_FLUSH) {
1743                         KKASSERT(rec->flush_group == rec->ip->flush_group);
1744                         r = 1;
1745                 } else {
1746                         r = 0;
1747                 }
1748                 return(r);
1749         }
1750
1751         /*
1752          * If the record is in an idle state it has no dependancies and
1753          * can be flushed.
1754          */
1755         ip = rec->ip;
1756         flg = ip->flush_group;
1757         r = 0;
1758
1759         switch(rec->flush_state) {
1760         case HAMMER_FST_IDLE:
1761                 /*
1762                  * The record has no setup dependancy, we can flush it.
1763                  */
1764                 KKASSERT(rec->target_ip == NULL);
1765                 rec->flush_state = HAMMER_FST_FLUSH;
1766                 rec->flush_group = flg;
1767                 ++flg->refs;
1768                 hammer_ref(&rec->lock);
1769                 r = 1;
1770                 break;
1771         case HAMMER_FST_SETUP:
1772                 /*
1773                  * The record has a setup dependancy.  These are typically
1774                  * directory entry adds and deletes.  Such entries will be
1775                  * flushed when their inodes are flushed so we do not have
1776                  * to add them to the flush here.
1777                  */
1778                 target_ip = rec->target_ip;
1779                 KKASSERT(target_ip != NULL);
1780                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1781
1782                 /*
1783                  * If the target IP is already flushing in our group
1784                  * we are golden, otherwise make sure the target
1785                  * reflushes.
1786                  */
1787                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1788                         if (target_ip->flush_group == flg) {
1789                                 rec->flush_state = HAMMER_FST_FLUSH;
1790                                 rec->flush_group = flg;
1791                                 ++flg->refs;
1792                                 hammer_ref(&rec->lock);
1793                                 r = 1;
1794                         } else {
1795                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1796                         }
1797                         break;
1798                 }
1799
1800                 /*
1801                  * Target IP is not yet flushing.  This can get complex
1802                  * because we have to be careful about the recursion.
1803                  */
1804                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
1805                     (target_ip->flags & HAMMER_INODE_REFLUSH) == 0) {
1806                         /*
1807                          * We aren't reclaiming or trying to flush target_ip.
1808                          * Let the record flush with the target.
1809                          */
1810                         /*r = 0;*/
1811                 } else if (flg->total_count + flg->refs >
1812                            ip->hmp->undo_rec_limit) {
1813                         /*
1814                          * Our flush group is over-full and we risk blowing
1815                          * out the UNDO FIFO.  Stop the scan, flush what we
1816                          * have, then reflush the directory.
1817                          *
1818                          * The directory may be forced through multiple
1819                          * flush groups before it can be completely
1820                          * flushed.
1821                          */
1822                         ip->flags |= HAMMER_INODE_REFLUSH;
1823                         ip->flags |= HAMMER_INODE_RESIGNAL;
1824                         r = -1;
1825                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1826                         /*
1827                          * If the target IP is not flushing we can force
1828                          * it to flush, even if it is unable to write out
1829                          * any of its own records we have at least one in
1830                          * hand that we CAN deal with.
1831                          */
1832                         rec->flush_state = HAMMER_FST_FLUSH;
1833                         rec->flush_group = flg;
1834                         ++flg->refs;
1835                         hammer_ref(&rec->lock);
1836                         hammer_flush_inode_core(target_ip, flg,
1837                                                 HAMMER_FLUSH_RECURSION);
1838                         r = 1;
1839                 } else {
1840                         /*
1841                          * General or delete-on-disk record.
1842                          *
1843                          * XXX this needs help.  If a delete-on-disk we could
1844                          * disconnect the target.  If the target has its own
1845                          * dependancies they really need to be flushed.
1846                          *
1847                          * XXX
1848                          */
1849                         rec->flush_state = HAMMER_FST_FLUSH;
1850                         rec->flush_group = flg;
1851                         ++flg->refs;
1852                         hammer_ref(&rec->lock);
1853                         hammer_flush_inode_core(target_ip, flg,
1854                                                 HAMMER_FLUSH_RECURSION);
1855                         r = 1;
1856                 }
1857                 break;
1858         case HAMMER_FST_FLUSH:
1859                 /*
1860                  * If the WOULDBLOCK flag is set records may have been left
1861                  * over from a previous flush attempt.  The flush group will
1862                  * have been left intact - we are probably reflushing it
1863                  * now.
1864                  */
1865                 KKASSERT(rec->flush_group == flg);
1866                 r = 1;
1867                 break;
1868         }
1869         return(r);
1870 }
1871
1872 #if 0
1873 /*
1874  * This version just moves records already in a flush state to the new
1875  * flush group and that is it.
1876  */
1877 static int
1878 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
1879 {
1880         hammer_inode_t ip = rec->ip;
1881
1882         switch(rec->flush_state) {
1883         case HAMMER_FST_FLUSH:
1884                 KKASSERT(rec->flush_group == ip->flush_group);
1885                 break;
1886         default:
1887                 break;
1888         }
1889         return(0);
1890 }
1891 #endif
1892
1893 /*
1894  * Wait for a previously queued flush to complete.
1895  */
1896 void
1897 hammer_wait_inode(hammer_inode_t ip)
1898 {
1899         hammer_flush_group_t flg;
1900
1901         flg = NULL;
1902         if (ip->flush_state == HAMMER_FST_SETUP) {
1903                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1904         }
1905         while (ip->flush_state != HAMMER_FST_IDLE) {
1906                 ip->flags |= HAMMER_INODE_FLUSHW;
1907                 tsleep(&ip->flags, 0, "hmrwin", 0);
1908         }
1909 }
1910
1911 /*
1912  * Called by the backend code when a flush has been completed.
1913  * The inode has already been removed from the flush list.
1914  *
1915  * A pipelined flush can occur, in which case we must re-enter the
1916  * inode on the list and re-copy its fields.
1917  */
1918 void
1919 hammer_flush_inode_done(hammer_inode_t ip)
1920 {
1921         hammer_mount_t hmp;
1922         int dorel;
1923
1924         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1925
1926         hmp = ip->hmp;
1927
1928         /*
1929          * Merge left-over flags back into the frontend and fix the state.
1930          * Incomplete truncations are retained by the backend.
1931          */
1932         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
1933         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
1934
1935         /*
1936          * The backend may have adjusted nlinks, so if the adjusted nlinks
1937          * does not match the fronttend set the frontend's RDIRTY flag again.
1938          */
1939         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
1940                 ip->flags |= HAMMER_INODE_DDIRTY;
1941
1942         /*
1943          * Fix up the dirty buffer status.
1944          */
1945         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
1946                 ip->flags |= HAMMER_INODE_BUFS;
1947         }
1948
1949         /*
1950          * Re-set the XDIRTY flag if some of the inode's in-memory records
1951          * could not be flushed.
1952          */
1953         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
1954                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
1955                  (!RB_EMPTY(&ip->rec_tree) &&
1956                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
1957
1958         /*
1959          * Do not lose track of inodes which no longer have vnode
1960          * assocations, otherwise they may never get flushed again.
1961          */
1962         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1963                 ip->flags |= HAMMER_INODE_REFLUSH;
1964
1965         /*
1966          * Adjust the flush state.
1967          */
1968         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1969                 /*
1970                  * We were unable to flush out all our records, leave the
1971                  * inode in a flush state and in the current flush group.
1972                  *
1973                  * This occurs if the UNDO block gets too full
1974                  * or there is too much dirty meta-data and allows the
1975                  * flusher to finalize the UNDO block and then re-flush.
1976                  */
1977                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
1978                 dorel = 0;
1979         } else {
1980                 /*
1981                  * Remove from the flush_group
1982                  */
1983                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1984                 ip->flush_group = NULL;
1985
1986                 /*
1987                  * Clean up the vnode ref and tracking counts.
1988                  */
1989                 if (ip->flags & HAMMER_INODE_VHELD) {
1990                         ip->flags &= ~HAMMER_INODE_VHELD;
1991                         vrele(ip->vp);
1992                 }
1993                 --hmp->count_iqueued;
1994                 --hammer_count_iqueued;
1995
1996                 /*
1997                  * And adjust the state.
1998                  */
1999                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2000                         ip->flush_state = HAMMER_FST_IDLE;
2001                         dorel = 1;
2002                 } else {
2003                         ip->flush_state = HAMMER_FST_SETUP;
2004                         dorel = 0;
2005                 }
2006
2007                 /*
2008                  * If the frontend is waiting for a flush to complete,
2009                  * wake it up.
2010                  */
2011                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2012                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2013                         wakeup(&ip->flags);
2014                 }
2015         }
2016
2017         /*
2018          * If the frontend made more changes and requested another flush,
2019          * then try to get it running.
2020          */
2021         if (ip->flags & HAMMER_INODE_REFLUSH) {
2022                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2023                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2024                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2025                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2026                 } else {
2027                         hammer_flush_inode(ip, 0);
2028                 }
2029         }
2030
2031         /*
2032          * If the inode is now clean drop the space reservation.
2033          */
2034         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2035             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2036                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2037                 --hmp->rsv_inodes;
2038         }
2039
2040         if (dorel)
2041                 hammer_rel_inode(ip, 0);
2042 }
2043
2044 /*
2045  * Called from hammer_sync_inode() to synchronize in-memory records
2046  * to the media.
2047  */
2048 static int
2049 hammer_sync_record_callback(hammer_record_t record, void *data)
2050 {
2051         hammer_cursor_t cursor = data;
2052         hammer_transaction_t trans = cursor->trans;
2053         hammer_mount_t hmp = trans->hmp;
2054         int error;
2055
2056         /*
2057          * Skip records that do not belong to the current flush.
2058          */
2059         ++hammer_stats_record_iterations;
2060         if (record->flush_state != HAMMER_FST_FLUSH)
2061                 return(0);
2062
2063 #if 1
2064         if (record->flush_group != record->ip->flush_group) {
2065                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2066                 Debugger("blah2");
2067                 return(0);
2068         }
2069 #endif
2070         KKASSERT(record->flush_group == record->ip->flush_group);
2071
2072         /*
2073          * Interlock the record using the BE flag.  Once BE is set the
2074          * frontend cannot change the state of FE.
2075          *
2076          * NOTE: If FE is set prior to us setting BE we still sync the
2077          * record out, but the flush completion code converts it to
2078          * a delete-on-disk record instead of destroying it.
2079          */
2080         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2081         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2082
2083         /*
2084          * The backend may have already disposed of the record.
2085          */
2086         if (record->flags & HAMMER_RECF_DELETED_BE) {
2087                 error = 0;
2088                 goto done;
2089         }
2090
2091         /*
2092          * If the whole inode is being deleting all on-disk records will
2093          * be deleted very soon, we can't sync any new records to disk
2094          * because they will be deleted in the same transaction they were
2095          * created in (delete_tid == create_tid), which will assert.
2096          *
2097          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2098          * that we currently panic on.
2099          */
2100         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2101                 switch(record->type) {
2102                 case HAMMER_MEM_RECORD_DATA:
2103                         /*
2104                          * We don't have to do anything, if the record was
2105                          * committed the space will have been accounted for
2106                          * in the blockmap.
2107                          */
2108                         /* fall through */
2109                 case HAMMER_MEM_RECORD_GENERAL:
2110                         record->flags |= HAMMER_RECF_DELETED_FE;
2111                         record->flags |= HAMMER_RECF_DELETED_BE;
2112                         error = 0;
2113                         goto done;
2114                 case HAMMER_MEM_RECORD_ADD:
2115                         panic("hammer_sync_record_callback: illegal add "
2116                               "during inode deletion record %p", record);
2117                         break; /* NOT REACHED */
2118                 case HAMMER_MEM_RECORD_INODE:
2119                         panic("hammer_sync_record_callback: attempt to "
2120                               "sync inode record %p?", record);
2121                         break; /* NOT REACHED */
2122                 case HAMMER_MEM_RECORD_DEL:
2123                         /*
2124                          * Follow through and issue the on-disk deletion
2125                          */
2126                         break;
2127                 }
2128         }
2129
2130         /*
2131          * If DELETED_FE is set special handling is needed for directory
2132          * entries.  Dependant pieces related to the directory entry may
2133          * have already been synced to disk.  If this occurs we have to
2134          * sync the directory entry and then change the in-memory record
2135          * from an ADD to a DELETE to cover the fact that it's been
2136          * deleted by the frontend.
2137          *
2138          * A directory delete covering record (MEM_RECORD_DEL) can never
2139          * be deleted by the frontend.
2140          *
2141          * Any other record type (aka DATA) can be deleted by the frontend.
2142          * XXX At the moment the flusher must skip it because there may
2143          * be another data record in the flush group for the same block,
2144          * meaning that some frontend data changes can leak into the backend's
2145          * synchronization point.
2146          */
2147         if (record->flags & HAMMER_RECF_DELETED_FE) {
2148                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2149                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2150                 } else {
2151                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2152                         record->flags |= HAMMER_RECF_DELETED_BE;
2153                         error = 0;
2154                         goto done;
2155                 }
2156         }
2157
2158         /*
2159          * Assign the create_tid for new records.  Deletions already
2160          * have the record's entire key properly set up.
2161          */
2162         if (record->type != HAMMER_MEM_RECORD_DEL)
2163                 record->leaf.base.create_tid = trans->tid;
2164                 record->leaf.create_ts = trans->time32;
2165         for (;;) {
2166                 error = hammer_ip_sync_record_cursor(cursor, record);
2167                 if (error != EDEADLK)
2168                         break;
2169                 hammer_done_cursor(cursor);
2170                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2171                                            record->ip);
2172                 if (error)
2173                         break;
2174         }
2175         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2176
2177         if (error) {
2178                 error = -error;
2179                 if (error != -ENOSPC) {
2180                         kprintf("hammer_sync_record_callback: sync failed rec "
2181                                 "%p, error %d\n", record, error);
2182                         Debugger("sync failed rec");
2183                 }
2184         }
2185 done:
2186         hammer_flush_record_done(record, error);
2187
2188         /*
2189          * Do partial finalization if we have built up too many dirty
2190          * buffers.  Otherwise a buffer cache deadlock can occur when
2191          * doing things like creating tens of thousands of tiny files.
2192          *
2193          * We must release our cursor lock to avoid a 3-way deadlock
2194          * due to the exclusive sync lock the finalizer must get.
2195          */
2196         if (hammer_flusher_meta_limit(hmp)) {
2197                 hammer_unlock_cursor(cursor, 0);
2198                 hammer_flusher_finalize(trans, 0);
2199                 hammer_lock_cursor(cursor, 0);
2200         }
2201
2202         return(error);
2203 }
2204
2205 /*
2206  * XXX error handling
2207  */
2208 int
2209 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2210 {
2211         struct hammer_cursor cursor;
2212         hammer_node_t tmp_node;
2213         hammer_record_t depend;
2214         hammer_record_t next;
2215         int error, tmp_error;
2216         u_int64_t nlinks;
2217
2218         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2219                 return(0);
2220
2221         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2222         if (error)
2223                 goto done;
2224
2225         /*
2226          * Any directory records referencing this inode which are not in
2227          * our current flush group must adjust our nlink count for the
2228          * purposes of synchronization to disk.
2229          *
2230          * Records which are in our flush group can be unlinked from our
2231          * inode now, potentially allowing the inode to be physically
2232          * deleted.
2233          *
2234          * This cannot block.
2235          */
2236         nlinks = ip->ino_data.nlinks;
2237         next = TAILQ_FIRST(&ip->target_list);
2238         while ((depend = next) != NULL) {
2239                 next = TAILQ_NEXT(depend, target_entry);
2240                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2241                     depend->flush_group == ip->flush_group) {
2242                         /*
2243                          * If this is an ADD that was deleted by the frontend
2244                          * the frontend nlinks count will have already been
2245                          * decremented, but the backend is going to sync its
2246                          * directory entry and must account for it.  The
2247                          * record will be converted to a delete-on-disk when
2248                          * it gets synced.
2249                          *
2250                          * If the ADD was not deleted by the frontend we
2251                          * can remove the dependancy from our target_list.
2252                          */
2253                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2254                                 ++nlinks;
2255                         } else {
2256                                 TAILQ_REMOVE(&ip->target_list, depend,
2257                                              target_entry);
2258                                 depend->target_ip = NULL;
2259                         }
2260                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2261                         /*
2262                          * Not part of our flush group
2263                          */
2264                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2265                         switch(depend->type) {
2266                         case HAMMER_MEM_RECORD_ADD:
2267                                 --nlinks;
2268                                 break;
2269                         case HAMMER_MEM_RECORD_DEL:
2270                                 ++nlinks;
2271                                 break;
2272                         default:
2273                                 break;
2274                         }
2275                 }
2276         }
2277
2278         /*
2279          * Set dirty if we had to modify the link count.
2280          */
2281         if (ip->sync_ino_data.nlinks != nlinks) {
2282                 KKASSERT((int64_t)nlinks >= 0);
2283                 ip->sync_ino_data.nlinks = nlinks;
2284                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2285         }
2286
2287         /*
2288          * If there is a trunction queued destroy any data past the (aligned)
2289          * truncation point.  Userland will have dealt with the buffer
2290          * containing the truncation point for us.
2291          *
2292          * We don't flush pending frontend data buffers until after we've
2293          * dealt with the truncation.
2294          */
2295         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2296                 /*
2297                  * Interlock trunc_off.  The VOP front-end may continue to
2298                  * make adjustments to it while we are blocked.
2299                  */
2300                 off_t trunc_off;
2301                 off_t aligned_trunc_off;
2302                 int blkmask;
2303
2304                 trunc_off = ip->sync_trunc_off;
2305                 blkmask = hammer_blocksize(trunc_off) - 1;
2306                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2307
2308                 /*
2309                  * Delete any whole blocks on-media.  The front-end has
2310                  * already cleaned out any partial block and made it
2311                  * pending.  The front-end may have updated trunc_off
2312                  * while we were blocked so we only use sync_trunc_off.
2313                  *
2314                  * This operation can blow out the buffer cache, EWOULDBLOCK
2315                  * means we were unable to complete the deletion.  The
2316                  * deletion will update sync_trunc_off in that case.
2317                  */
2318                 error = hammer_ip_delete_range(&cursor, ip,
2319                                                 aligned_trunc_off,
2320                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2321                 if (error == EWOULDBLOCK) {
2322                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2323                         error = 0;
2324                         goto defer_buffer_flush;
2325                 }
2326
2327                 if (error)
2328                         Debugger("hammer_ip_delete_range errored");
2329
2330                 /*
2331                  * Clear the truncation flag on the backend after we have
2332                  * complete the deletions.  Backend data is now good again
2333                  * (including new records we are about to sync, below).
2334                  *
2335                  * Leave sync_trunc_off intact.  As we write additional
2336                  * records the backend will update sync_trunc_off.  This
2337                  * tells the backend whether it can skip the overwrite
2338                  * test.  This should work properly even when the backend
2339                  * writes full blocks where the truncation point straddles
2340                  * the block because the comparison is against the base
2341                  * offset of the record.
2342                  */
2343                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2344                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2345         } else {
2346                 error = 0;
2347         }
2348
2349         /*
2350          * Now sync related records.  These will typically be directory
2351          * entries, records tracking direct-writes, or delete-on-disk records.
2352          */
2353         if (error == 0) {
2354                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2355                                     hammer_sync_record_callback, &cursor);
2356                 if (tmp_error < 0)
2357                         tmp_error = -error;
2358                 if (tmp_error)
2359                         error = tmp_error;
2360         }
2361         hammer_cache_node(&ip->cache[1], cursor.node);
2362
2363         /*
2364          * Re-seek for inode update, assuming our cache hasn't been ripped
2365          * out from under us.
2366          */
2367         if (error == 0) {
2368                 tmp_node = hammer_ref_node_safe(ip->hmp, &ip->cache[0], &error);
2369                 if (tmp_node) {
2370                         hammer_cursor_downgrade(&cursor);
2371                         hammer_lock_sh(&tmp_node->lock);
2372                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2373                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2374                         hammer_unlock(&tmp_node->lock);
2375                         hammer_rel_node(tmp_node);
2376                 }
2377                 error = 0;
2378         }
2379
2380         /*
2381          * If we are deleting the inode the frontend had better not have
2382          * any active references on elements making up the inode.
2383          *
2384          * The call to hammer_ip_delete_clean() cleans up auxillary records
2385          * but not DB or DATA records.  Those must have already been deleted
2386          * by the normal truncation mechanic.
2387          */
2388         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2389                 RB_EMPTY(&ip->rec_tree)  &&
2390             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2391             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2392                 int count1 = 0;
2393
2394                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2395                 if (error == 0) {
2396                         ip->flags |= HAMMER_INODE_DELETED;
2397                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2398                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2399                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2400
2401                         /*
2402                          * Set delete_tid in both the frontend and backend
2403                          * copy of the inode record.  The DELETED flag handles
2404                          * this, do not set RDIRTY.
2405                          */
2406                         ip->ino_leaf.base.delete_tid = trans->tid;
2407                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2408                         ip->ino_leaf.delete_ts = trans->time32;
2409                         ip->sync_ino_leaf.delete_ts = trans->time32;
2410
2411
2412                         /*
2413                          * Adjust the inode count in the volume header
2414                          */
2415                         hammer_sync_lock_sh(trans);
2416                         if (ip->flags & HAMMER_INODE_ONDISK) {
2417                                 hammer_modify_volume_field(trans,
2418                                                            trans->rootvol,
2419                                                            vol0_stat_inodes);
2420                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2421                                 hammer_modify_volume_done(trans->rootvol);
2422                         }
2423                         hammer_sync_unlock(trans);
2424                 } else {
2425                         Debugger("hammer_ip_delete_clean errored");
2426                 }
2427         }
2428
2429         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2430
2431         if (error)
2432                 Debugger("RB_SCAN errored");
2433
2434 defer_buffer_flush:
2435         /*
2436          * Now update the inode's on-disk inode-data and/or on-disk record.
2437          * DELETED and ONDISK are managed only in ip->flags.
2438          *
2439          * In the case of a defered buffer flush we still update the on-disk
2440          * inode to satisfy visibility requirements if there happen to be
2441          * directory dependancies.
2442          */
2443         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2444         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2445                 /*
2446                  * If deleted and on-disk, don't set any additional flags.
2447                  * the delete flag takes care of things.
2448                  *
2449                  * Clear flags which may have been set by the frontend.
2450                  */
2451                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2452                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2453                                     HAMMER_INODE_DELETING);
2454                 break;
2455         case HAMMER_INODE_DELETED:
2456                 /*
2457                  * Take care of the case where a deleted inode was never
2458                  * flushed to the disk in the first place.
2459                  *
2460                  * Clear flags which may have been set by the frontend.
2461                  */
2462                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2463                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2464                                     HAMMER_INODE_DELETING);
2465                 while (RB_ROOT(&ip->rec_tree)) {
2466                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2467                         hammer_ref(&record->lock);
2468                         KKASSERT(record->lock.refs == 1);
2469                         record->flags |= HAMMER_RECF_DELETED_FE;
2470                         record->flags |= HAMMER_RECF_DELETED_BE;
2471                         hammer_rel_mem_record(record);
2472                 }
2473                 break;
2474         case HAMMER_INODE_ONDISK:
2475                 /*
2476                  * If already on-disk, do not set any additional flags.
2477                  */
2478                 break;
2479         default:
2480                 /*
2481                  * If not on-disk and not deleted, set DDIRTY to force
2482                  * an initial record to be written.
2483                  *
2484                  * Also set the create_tid in both the frontend and backend
2485                  * copy of the inode record.
2486                  */
2487                 ip->ino_leaf.base.create_tid = trans->tid;
2488                 ip->ino_leaf.create_ts = trans->time32;
2489                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2490                 ip->sync_ino_leaf.create_ts = trans->time32;
2491                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2492                 break;
2493         }
2494
2495         /*
2496          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2497          * is already on-disk the old record is marked as deleted.
2498          *
2499          * If DELETED is set hammer_update_inode() will delete the existing
2500          * record without writing out a new one.
2501          *
2502          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2503          */
2504         if (ip->flags & HAMMER_INODE_DELETED) {
2505                 error = hammer_update_inode(&cursor, ip);
2506         } else
2507         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2508             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2509                 error = hammer_update_itimes(&cursor, ip);
2510         } else
2511         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2512                 error = hammer_update_inode(&cursor, ip);
2513         }
2514         if (error)
2515                 Debugger("hammer_update_itimes/inode errored");
2516 done:
2517         /*
2518          * Save the TID we used to sync the inode with to make sure we
2519          * do not improperly reuse it.
2520          */
2521         hammer_done_cursor(&cursor);
2522         return(error);
2523 }
2524
2525 /*
2526  * This routine is called when the OS is no longer actively referencing
2527  * the inode (but might still be keeping it cached), or when releasing
2528  * the last reference to an inode.
2529  *
2530  * At this point if the inode's nlinks count is zero we want to destroy
2531  * it, which may mean destroying it on-media too.
2532  */
2533 void
2534 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2535 {
2536         struct vnode *vp;
2537
2538         /*
2539          * Set the DELETING flag when the link count drops to 0 and the
2540          * OS no longer has any opens on the inode.
2541          *
2542          * The backend will clear DELETING (a mod flag) and set DELETED
2543          * (a state flag) when it is actually able to perform the
2544          * operation.
2545          */
2546         if (ip->ino_data.nlinks == 0 &&
2547             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2548                 ip->flags |= HAMMER_INODE_DELETING;
2549                 ip->flags |= HAMMER_INODE_TRUNCATED;
2550                 ip->trunc_off = 0;
2551                 vp = NULL;
2552                 if (getvp) {
2553                         if (hammer_get_vnode(ip, &vp) != 0)
2554                                 return;
2555                 }
2556
2557                 /*
2558                  * Final cleanup
2559                  */
2560                 if (ip->vp) {
2561                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2562                         vnode_pager_setsize(ip->vp, 0);
2563                 }
2564                 if (getvp) {
2565                         vput(vp);
2566                 }
2567         }
2568 }
2569
2570 /*
2571  * Re-test an inode when a dependancy had gone away to see if we
2572  * can chain flush it.
2573  */
2574 void
2575 hammer_test_inode(hammer_inode_t ip)
2576 {
2577         if (ip->flags & HAMMER_INODE_REFLUSH) {
2578                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2579                 hammer_ref(&ip->lock);
2580                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2581                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2582                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2583                 } else {
2584                         hammer_flush_inode(ip, 0);
2585                 }
2586                 hammer_rel_inode(ip, 0);
2587         }
2588 }
2589
2590 /*
2591  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2592  * reassociated with a vp or just before it gets freed.
2593  *
2594  * Wakeup one thread blocked waiting on reclaims to complete.  Note that
2595  * the inode the thread is waiting on behalf of is a different inode then
2596  * the inode we are called with.  This is to create a pipeline.
2597  */
2598 static void
2599 hammer_inode_wakereclaims(hammer_inode_t ip)
2600 {
2601         struct hammer_reclaim *reclaim;
2602         hammer_mount_t hmp = ip->hmp;
2603
2604         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2605                 return;
2606
2607         --hammer_count_reclaiming;
2608         --hmp->inode_reclaims;
2609         ip->flags &= ~HAMMER_INODE_RECLAIM;
2610
2611         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
2612                 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2613                 reclaim->okydoky = 1;
2614                 wakeup(reclaim);
2615         }
2616 }
2617
2618 /*
2619  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
2620  * inodes build up before we start blocking.
2621  *
2622  * When we block we don't care *which* inode has finished reclaiming,
2623  * as lone as one does.  This is somewhat heuristical... we also put a
2624  * cap on how long we are willing to wait.
2625  */
2626 void
2627 hammer_inode_waitreclaims(hammer_mount_t hmp)
2628 {
2629         struct hammer_reclaim reclaim;
2630         int delay;
2631
2632         if (hmp->inode_reclaims > HAMMER_RECLAIM_WAIT) {
2633                 reclaim.okydoky = 0;
2634                 TAILQ_INSERT_TAIL(&hmp->reclaim_list,
2635                                   &reclaim, entry);
2636         } else {
2637                 reclaim.okydoky = 1;
2638         }
2639
2640         if (reclaim.okydoky == 0) {
2641                 delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2642                         HAMMER_RECLAIM_WAIT;
2643                 if (delay >= 0)
2644                         tsleep(&reclaim, 0, "hmrrcm", delay + 1);
2645                 if (reclaim.okydoky == 0)
2646                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2647         }
2648 }
2649