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 1.111 2008/09/17 21:44:20 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 can be accessed via
 297                          * non-root filesystem paths and setting VROOT may
 298                          * confuse the namecache.  Set VPFSROOT instead.
 299                          */
 300                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 301                             ip->obj_asof == hmp->asof) {
 302                                 if (ip->obj_localization == 0)
 303                                         vp->v_flag |= VROOT;
 304                                 else
 305                                         vp->v_flag |= VPFSROOT;
 306                         }
 307
 308                         vp->v_data = (void *)ip;
 309                         /* vnode locked by getnewvnode() */
 310                         /* make related vnode dirty if inode dirty? */
 311                         hammer_unlock(&ip->lock);
 312                         if (vp->v_type == VREG)
 313                                 vinitvmio(vp, ip->ino_data.size);
 314                         break;
 315                 }
 316
 317                 /*
 318                  * loop if the vget fails (aka races), or if the vp
 319                  * no longer matches ip->vp.
 320                  */
 321                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 322                         if (vp == ip->vp)
 323                                 break;
 324                         vput(vp);
 325                 }
 326         }
 327         *vpp = vp;
 328         return(error);
 329 }
 330
 331 /*
 332  * Locate all copies of the inode for obj_id compatible with the specified
 333  * asof, reference, and issue the related call-back.  This routine is used
 334  * for direct-io invalidation and does not create any new inodes.
 335  */
 336 void
 337 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
 338                             int (*callback)(hammer_inode_t ip, void *data),
 339                             void *data)
 340 {
 341         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
 342                                    hammer_inode_info_cmp_all_history,
 343                                    callback, iinfo);
 344 }
 345
 346 /*
 347  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 348  * do not attach or detach the related vnode (use hammer_get_vnode() for
 349  * that).
 350  *
 351  * The flags argument is only applied for newly created inodes, and only
 352  * certain flags are inherited.
 353  *
 354  * Called from the frontend.
 355  */
 356 struct hammer_inode *
 357 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
 358                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 359                  int flags, int *errorp)
 360 {
 361         hammer_mount_t hmp = trans->hmp;
 362         struct hammer_inode_info iinfo;
 363         struct hammer_cursor cursor;
 364         struct hammer_inode *ip;
 365
 366
 367         /*
 368          * Determine if we already have an inode cached.  If we do then
 369          * we are golden.
 370          */
 371         iinfo.obj_id = obj_id;
 372         iinfo.obj_asof = asof;
 373         iinfo.obj_localization = localization;
 374 loop:
 375         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 376         if (ip) {
 377                 hammer_ref(&ip->lock);
 378                 *errorp = 0;
 379                 return(ip);
 380         }
 381
 382         /*
 383          * Allocate a new inode structure and deal with races later.
 384          */
 385         ip = kmalloc(sizeof(*ip), M_HAMMER_INO, M_WAITOK|M_ZERO);
 386         ++hammer_count_inodes;
 387         ++hmp->count_inodes;
 388         ip->obj_id = obj_id;
 389         ip->obj_asof = iinfo.obj_asof;
 390         ip->obj_localization = localization;
 391         ip->hmp = hmp;
 392         ip->flags = flags & HAMMER_INODE_RO;
 393         ip->cache[0].ip = ip;
 394         ip->cache[1].ip = ip;
 395         if (hmp->ronly)
 396                 ip->flags |= HAMMER_INODE_RO;
 397         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 398                 0x7FFFFFFFFFFFFFFFLL;
 399         RB_INIT(&ip->rec_tree);
 400         TAILQ_INIT(&ip->target_list);
 401         hammer_ref(&ip->lock);
 402
 403         /*
 404          * Locate the on-disk inode.  If this is a PFS root we always
 405          * access the current version of the root inode and (if it is not
 406          * a master) always access information under it with a snapshot
 407          * TID.
 408          */
 409 retry:
 410         hammer_init_cursor(trans, &cursor, (dip ? &dip->cache[0] : NULL), NULL);
 411         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
 412         cursor.key_beg.obj_id = ip->obj_id;
 413         cursor.key_beg.key = 0;
 414         cursor.key_beg.create_tid = 0;
 415         cursor.key_beg.delete_tid = 0;
 416         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 417         cursor.key_beg.obj_type = 0;
 418
 419         cursor.asof = iinfo.obj_asof;
 420         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 421                        HAMMER_CURSOR_ASOF;
 422
 423         *errorp = hammer_btree_lookup(&cursor);
 424         if (*errorp == EDEADLK) {
 425                 hammer_done_cursor(&cursor);
 426                 goto retry;
 427         }
 428
 429         /*
 430          * On success the B-Tree lookup will hold the appropriate
 431          * buffer cache buffers and provide a pointer to the requested
 432          * information.  Copy the information to the in-memory inode
 433          * and cache the B-Tree node to improve future operations.
 434          */
 435         if (*errorp == 0) {
 436                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 437                 ip->ino_data = cursor.data->inode;
 438
 439                 /*
 440                  * cache[0] tries to cache the location of the object inode.
 441                  * The assumption is that it is near the directory inode.
 442                  *
 443                  * cache[1] tries to cache the location of the object data.
 444                  * The assumption is that it is near the directory data.
 445                  */
 446                 hammer_cache_node(&ip->cache[0], cursor.node);
 447                 if (dip && dip->cache[1].node)
 448                         hammer_cache_node(&ip->cache[1], dip->cache[1].node);
 449
 450                 /*
 451                  * The file should not contain any data past the file size
 452                  * stored in the inode.  Setting save_trunc_off to the
 453                  * file size instead of max reduces B-Tree lookup overheads
 454                  * on append by allowing the flusher to avoid checking for
 455                  * record overwrites.
 456                  */
 457                 ip->save_trunc_off = ip->ino_data.size;
 458
 459                 /*
 460                  * Locate and assign the pseudofs management structure to
 461                  * the inode.
 462                  */
 463                 if (dip && dip->obj_localization == ip->obj_localization) {
 464                         ip->pfsm = dip->pfsm;
 465                         hammer_ref(&ip->pfsm->lock);
 466                 } else {
 467                         ip->pfsm = hammer_load_pseudofs(trans,
 468                                                         ip->obj_localization,
 469                                                         errorp);
 470                         *errorp = 0;    /* ignore ENOENT */
 471                 }
 472         }
 473
 474         /*
 475          * The inode is placed on the red-black tree and will be synced to
 476          * the media when flushed or by the filesystem sync.  If this races
 477          * another instantiation/lookup the insertion will fail.
 478          */
 479         if (*errorp == 0) {
 480                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 481                         hammer_free_inode(ip);
 482                         hammer_done_cursor(&cursor);
 483                         goto loop;
 484                 }
 485                 ip->flags |= HAMMER_INODE_ONDISK;
 486         } else {
 487                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 488                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 489                         --hmp->rsv_inodes;
 490                 }
 491
 492                 hammer_free_inode(ip);
 493                 ip = NULL;
 494         }
 495         hammer_done_cursor(&cursor);
 496         return (ip);
 497 }
 498
 499 /*
 500  * Create a new filesystem object, returning the inode in *ipp.  The
 501  * returned inode will be referenced.  The inode is created in-memory.
 502  *
 503  * If pfsm is non-NULL the caller wishes to create the root inode for
 504  * a master PFS.
 505  */
 506 int
 507 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 508                     struct ucred *cred, hammer_inode_t dip,
 509                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
 510 {
 511         hammer_mount_t hmp;
 512         hammer_inode_t ip;
 513         uid_t xuid;
 514         int error;
 515
 516         hmp = trans->hmp;
 517
 518         ip = kmalloc(sizeof(*ip), M_HAMMER_INO, M_WAITOK|M_ZERO);
 519         ++hammer_count_inodes;
 520         ++hmp->count_inodes;
 521
 522         if (pfsm) {
 523                 KKASSERT(pfsm->localization != 0);
 524                 ip->obj_id = HAMMER_OBJID_ROOT;
 525                 ip->obj_localization = pfsm->localization;
 526         } else {
 527                 KKASSERT(dip != NULL);
 528                 ip->obj_id = hammer_alloc_objid(hmp, dip);
 529                 ip->obj_localization = dip->obj_localization;
 530         }
 531
 532         KKASSERT(ip->obj_id != 0);
 533         ip->obj_asof = hmp->asof;
 534         ip->hmp = hmp;
 535         ip->flush_state = HAMMER_FST_IDLE;
 536         ip->flags = HAMMER_INODE_DDIRTY |
 537                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
 538         ip->cache[0].ip = ip;
 539         ip->cache[1].ip = ip;
 540
 541         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 542         /* ip->save_trunc_off = 0; (already zero) */
 543         RB_INIT(&ip->rec_tree);
 544         TAILQ_INIT(&ip->target_list);
 545
 546         ip->ino_data.atime = trans->time;
 547         ip->ino_data.mtime = trans->time;
 548         ip->ino_data.size = 0;
 549         ip->ino_data.nlinks = 0;
 550
 551         /*
 552          * A nohistory designator on the parent directory is inherited by
 553          * the child.  We will do this even for pseudo-fs creation... the
 554          * sysad can turn it off.
 555          */
 556         if (dip) {
 557                 ip->ino_data.uflags = dip->ino_data.uflags &
 558                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 559         }
 560
 561         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 562         ip->ino_leaf.base.localization = ip->obj_localization +
 563                                          HAMMER_LOCALIZE_INODE;
 564         ip->ino_leaf.base.obj_id = ip->obj_id;
 565         ip->ino_leaf.base.key = 0;
 566         ip->ino_leaf.base.create_tid = 0;
 567         ip->ino_leaf.base.delete_tid = 0;
 568         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 569         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 570
 571         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 572         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 573         ip->ino_data.mode = vap->va_mode;
 574         ip->ino_data.ctime = trans->time;
 575
 576         /*
 577          * Setup the ".." pointer.  This only needs to be done for directories
 578          * but we do it for all objects as a recovery aid.
 579          */
 580         if (dip)
 581                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
 582 #if 0
 583         /*
 584          * The parent_obj_localization field only applies to pseudo-fs roots.
 585          * XXX this is no longer applicable, PFSs are no longer directly
 586          * tied into the parent's directory structure.
 587          */
 588         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
 589             ip->obj_id == HAMMER_OBJID_ROOT) {
 590                 ip->ino_data.ext.obj.parent_obj_localization =
 591                                                 dip->obj_localization;
 592         }
 593 #endif
 594
 595         switch(ip->ino_leaf.base.obj_type) {
 596         case HAMMER_OBJTYPE_CDEV:
 597         case HAMMER_OBJTYPE_BDEV:
 598                 ip->ino_data.rmajor = vap->va_rmajor;
 599                 ip->ino_data.rminor = vap->va_rminor;
 600                 break;
 601         default:
 602                 break;
 603         }
 604
 605         /*
 606          * Calculate default uid/gid and overwrite with information from
 607          * the vap.
 608          */
 609         if (dip) {
 610                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 611                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
 612                                              xuid, cred, &vap->va_mode);
 613         } else {
 614                 xuid = 0;
 615         }
 616         ip->ino_data.mode = vap->va_mode;
 617
 618         if (vap->va_vaflags & VA_UID_UUID_VALID)
 619                 ip->ino_data.uid = vap->va_uid_uuid;
 620         else if (vap->va_uid != (uid_t)VNOVAL)
 621                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 622         else
 623                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 624
 625         if (vap->va_vaflags & VA_GID_UUID_VALID)
 626                 ip->ino_data.gid = vap->va_gid_uuid;
 627         else if (vap->va_gid != (gid_t)VNOVAL)
 628                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 629         else if (dip)
 630                 ip->ino_data.gid = dip->ino_data.gid;
 631
 632         hammer_ref(&ip->lock);
 633
 634         if (pfsm) {
 635                 ip->pfsm = pfsm;
 636                 hammer_ref(&pfsm->lock);
 637                 error = 0;
 638         } else if (dip->obj_localization == ip->obj_localization) {
 639                 ip->pfsm = dip->pfsm;
 640                 hammer_ref(&ip->pfsm->lock);
 641                 error = 0;
 642         } else {
 643                 ip->pfsm = hammer_load_pseudofs(trans,
 644                                                 ip->obj_localization,
 645                                                 &error);
 646                 error = 0;      /* ignore ENOENT */
 647         }
 648
 649         if (error) {
 650                 hammer_free_inode(ip);
 651                 ip = NULL;
 652         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 653                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 654                 /* not reached */
 655                 hammer_free_inode(ip);
 656         }
 657         *ipp = ip;
 658         return(error);
 659 }
 660
 661 /*
 662  * Final cleanup / freeing of an inode structure
 663  */
 664 static void
 665 hammer_free_inode(hammer_inode_t ip)
 666 {
 667         KKASSERT(ip->lock.refs == 1);
 668         hammer_uncache_node(&ip->cache[0]);
 669         hammer_uncache_node(&ip->cache[1]);
 670         hammer_inode_wakereclaims(ip);
 671         if (ip->objid_cache)
 672                 hammer_clear_objid(ip);
 673         --hammer_count_inodes;
 674         --ip->hmp->count_inodes;
 675         if (ip->pfsm) {
 676                 hammer_rel_pseudofs(ip->hmp, ip->pfsm);
 677                 ip->pfsm = NULL;
 678         }
 679         kfree(ip, M_HAMMER_INO);
 680         ip = NULL;
 681 }
 682
 683 /*
 684  * Retrieve pseudo-fs data.  NULL will never be returned.
 685  *
 686  * If an error occurs *errorp will be set and a default template is returned,
 687  * otherwise *errorp is set to 0.  Typically when an error occurs it will
 688  * be ENOENT.
 689  */
 690 hammer_pseudofs_inmem_t
 691 hammer_load_pseudofs(hammer_transaction_t trans,
 692                      u_int32_t localization, int *errorp)
 693 {
 694         hammer_mount_t hmp = trans->hmp;
 695         hammer_inode_t ip;
 696         hammer_pseudofs_inmem_t pfsm;
 697         struct hammer_cursor cursor;
 698         int bytes;
 699
 700 retry:
 701         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
 702         if (pfsm) {
 703                 hammer_ref(&pfsm->lock);
 704                 *errorp = 0;
 705                 return(pfsm);
 706         }
 707
 708         /*
 709          * PFS records are stored in the root inode (not the PFS root inode,
 710          * but the real root).  Avoid an infinite recursion if loading
 711          * the PFS for the real root.
 712          */
 713         if (localization) {
 714                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
 715                                       HAMMER_MAX_TID,
 716                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
 717         } else {
 718                 ip = NULL;
 719         }
 720
 721         pfsm = kmalloc(sizeof(*pfsm), M_HAMMER, M_WAITOK | M_ZERO);
 722         pfsm->localization = localization;
 723         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
 724         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
 725
 726         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
 727         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
 728                                       HAMMER_LOCALIZE_MISC;
 729         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 730         cursor.key_beg.create_tid = 0;
 731         cursor.key_beg.delete_tid = 0;
 732         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 733         cursor.key_beg.obj_type = 0;
 734         cursor.key_beg.key = localization;
 735         cursor.asof = HAMMER_MAX_TID;
 736         cursor.flags |= HAMMER_CURSOR_ASOF;
 737
 738         if (ip)
 739                 *errorp = hammer_ip_lookup(&cursor);
 740         else
 741                 *errorp = hammer_btree_lookup(&cursor);
 742         if (*errorp == 0) {
 743                 *errorp = hammer_ip_resolve_data(&cursor);
 744                 if (*errorp == 0) {
 745                         if (cursor.data->pfsd.mirror_flags &
 746                             HAMMER_PFSD_DELETED) {
 747                                 *errorp = ENOENT;
 748                         } else {
 749                                 bytes = cursor.leaf->data_len;
 750                                 if (bytes > sizeof(pfsm->pfsd))
 751                                         bytes = sizeof(pfsm->pfsd);
 752                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
 753                         }
 754                 }
 755         }
 756         hammer_done_cursor(&cursor);
 757
 758         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 759         hammer_ref(&pfsm->lock);
 760         if (ip)
 761                 hammer_rel_inode(ip, 0);
 762         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
 763                 kfree(pfsm, M_HAMMER);
 764                 goto retry;
 765         }
 766         return(pfsm);
 767 }
 768
 769 /*
 770  * Store pseudo-fs data.  The backend will automatically delete any prior
 771  * on-disk pseudo-fs data but we have to delete in-memory versions.
 772  */
 773 int
 774 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
 775 {
 776         struct hammer_cursor cursor;
 777         hammer_record_t record;
 778         hammer_inode_t ip;
 779         int error;
 780
 781         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 782                               HAMMER_DEF_LOCALIZATION, 0, &error);
 783 retry:
 784         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 785         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
 786         cursor.key_beg.localization = ip->obj_localization +
 787                                       HAMMER_LOCALIZE_MISC;
 788         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 789         cursor.key_beg.create_tid = 0;
 790         cursor.key_beg.delete_tid = 0;
 791         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 792         cursor.key_beg.obj_type = 0;
 793         cursor.key_beg.key = pfsm->localization;
 794         cursor.asof = HAMMER_MAX_TID;
 795         cursor.flags |= HAMMER_CURSOR_ASOF;
 796
 797         error = hammer_ip_lookup(&cursor);
 798         if (error == 0 && hammer_cursor_inmem(&cursor)) {
 799                 record = cursor.iprec;
 800                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 801                         KKASSERT(cursor.deadlk_rec == NULL);
 802                         hammer_ref(&record->lock);
 803                         cursor.deadlk_rec = record;
 804                         error = EDEADLK;
 805                 } else {
 806                         record->flags |= HAMMER_RECF_DELETED_FE;
 807                         error = 0;
 808                 }
 809         }
 810         if (error == 0 || error == ENOENT) {
 811                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
 812                 record->type = HAMMER_MEM_RECORD_GENERAL;
 813
 814                 record->leaf.base.localization = ip->obj_localization +
 815                                                  HAMMER_LOCALIZE_MISC;
 816                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
 817                 record->leaf.base.key = pfsm->localization;
 818                 record->leaf.data_len = sizeof(pfsm->pfsd);
 819                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
 820                 error = hammer_ip_add_record(trans, record);
 821         }
 822         hammer_done_cursor(&cursor);
 823         if (error == EDEADLK)
 824                 goto retry;
 825         hammer_rel_inode(ip, 0);
 826         return(error);
 827 }
 828
 829 /*
 830  * Create a root directory for a PFS if one does not alredy exist.
 831  *
 832  * The PFS root stands alone so we must also bump the nlinks count
 833  * to prevent it from being destroyed on release.
 834  */
 835 int
 836 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
 837                        hammer_pseudofs_inmem_t pfsm)
 838 {
 839         hammer_inode_t ip;
 840         struct vattr vap;
 841         int error;
 842
 843         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 844                               pfsm->localization, 0, &error);
 845         if (ip == NULL) {
 846                 vattr_null(&vap);
 847                 vap.va_mode = 0755;
 848                 vap.va_type = VDIR;
 849                 error = hammer_create_inode(trans, &vap, cred, NULL, pfsm, &ip);
 850                 if (error == 0) {
 851                         ++ip->ino_data.nlinks;
 852                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 853                 }
 854         }
 855         if (ip)
 856                 hammer_rel_inode(ip, 0);
 857         return(error);
 858 }
 859
 860 /*
 861  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
 862  * if we are unable to disassociate all the inodes.
 863  */
 864 static
 865 int
 866 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
 867 {
 868         int res;
 869
 870         hammer_ref(&ip->lock);
 871         if (ip->lock.refs == 2 && ip->vp)
 872                 vclean_unlocked(ip->vp);
 873         if (ip->lock.refs == 1 && ip->vp == NULL)
 874                 res = 0;
 875         else
 876                 res = -1;       /* stop, someone is using the inode */
 877         hammer_rel_inode(ip, 0);
 878         return(res);
 879 }
 880
 881 int
 882 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
 883 {
 884         int res;
 885         int try;
 886
 887         for (try = res = 0; try < 4; ++try) {
 888                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
 889                                            hammer_inode_pfs_cmp,
 890                                            hammer_unload_pseudofs_callback,
 891                                            &localization);
 892                 if (res == 0 && try > 1)
 893                         break;
 894                 hammer_flusher_sync(trans->hmp);
 895         }
 896         if (res != 0)
 897                 res = ENOTEMPTY;
 898         return(res);
 899 }
 900
 901
 902 /*
 903  * Release a reference on a PFS
 904  */
 905 void
 906 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
 907 {
 908         hammer_unref(&pfsm->lock);
 909         if (pfsm->lock.refs == 0) {
 910                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
 911                 kfree(pfsm, M_HAMMER);
 912         }
 913 }
 914
 915 /*
 916  * Called by hammer_sync_inode().
 917  */
 918 static int
 919 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 920 {
 921         hammer_transaction_t trans = cursor->trans;
 922         hammer_record_t record;
 923         int error;
 924         int redirty;
 925
 926 retry:
 927         error = 0;
 928
 929         /*
 930          * If the inode has a presence on-disk then locate it and mark
 931          * it deleted, setting DELONDISK.
 932          *
 933          * The record may or may not be physically deleted, depending on
 934          * the retention policy.
 935          */
 936         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 937             HAMMER_INODE_ONDISK) {
 938                 hammer_normalize_cursor(cursor);
 939                 cursor->key_beg.localization = ip->obj_localization +
 940                                                HAMMER_LOCALIZE_INODE;
 941                 cursor->key_beg.obj_id = ip->obj_id;
 942                 cursor->key_beg.key = 0;
 943                 cursor->key_beg.create_tid = 0;
 944                 cursor->key_beg.delete_tid = 0;
 945                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 946                 cursor->key_beg.obj_type = 0;
 947                 cursor->asof = ip->obj_asof;
 948                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 949                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 950                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 951
 952                 error = hammer_btree_lookup(cursor);
 953                 if (hammer_debug_inode)
 954                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 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 == 0) {
 961                                 ip->flags |= HAMMER_INODE_DELONDISK;
 962                         }
 963                         if (cursor->node)
 964                                 hammer_cache_node(&ip->cache[0], cursor->node);
 965                 }
 966                 if (error == EDEADLK) {
 967                         hammer_done_cursor(cursor);
 968                         error = hammer_init_cursor(trans, cursor,
 969                                                    &ip->cache[0], ip);
 970                         if (hammer_debug_inode)
 971                                 kprintf("IPDED %p %d\n", ip, error);
 972                         if (error == 0)
 973                                 goto retry;
 974                 }
 975         }
 976
 977         /*
 978          * Ok, write out the initial record or a new record (after deleting
 979          * the old one), unless the DELETED flag is set.  This routine will
 980          * clear DELONDISK if it writes out a record.
 981          *
 982          * Update our inode statistics if this is the first application of
 983          * the inode on-disk.
 984          */
 985         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 986                 /*
 987                  * Generate a record and write it to the media.  We clean-up
 988                  * the state before releasing so we do not have to set-up
 989                  * a flush_group.
 990                  */
 991                 record = hammer_alloc_mem_record(ip, 0);
 992                 record->type = HAMMER_MEM_RECORD_INODE;
 993                 record->flush_state = HAMMER_FST_FLUSH;
 994                 record->leaf = ip->sync_ino_leaf;
 995                 record->leaf.base.create_tid = trans->tid;
 996                 record->leaf.data_len = sizeof(ip->sync_ino_data);
 997                 record->leaf.create_ts = trans->time32;
 998                 record->data = (void *)&ip->sync_ino_data;
 999                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1000
1001                 /*
1002                  * If this flag is set we cannot sync the new file size
1003                  * because we haven't finished related truncations.  The
1004                  * inode will be flushed in another flush group to finish
1005                  * the job.
1006                  */
1007                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1008                     ip->sync_ino_data.size != ip->ino_data.size) {
1009                         redirty = 1;
1010                         ip->sync_ino_data.size = ip->ino_data.size;
1011                 } else {
1012                         redirty = 0;
1013                 }
1014
1015                 for (;;) {
1016                         error = hammer_ip_sync_record_cursor(cursor, record);
1017                         if (hammer_debug_inode)
1018                                 kprintf("GENREC %p rec %08x %d\n",
1019                                         ip, record->flags, error);
1020                         if (error != EDEADLK)
1021                                 break;
1022                         hammer_done_cursor(cursor);
1023                         error = hammer_init_cursor(trans, cursor,
1024                                                    &ip->cache[0], ip);
1025                         if (hammer_debug_inode)
1026                                 kprintf("GENREC reinit %d\n", error);
1027                         if (error)
1028                                 break;
1029                 }
1030
1031                 /*
1032                  * The record isn't managed by the inode's record tree,
1033                  * destroy it whether we succeed or fail.
1034                  */
1035                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1036                 record->flags |= HAMMER_RECF_DELETED_FE | HAMMER_RECF_COMMITTED;
1037                 record->flush_state = HAMMER_FST_IDLE;
1038                 hammer_rel_mem_record(record);
1039
1040                 /*
1041                  * Finish up.
1042                  */
1043                 if (error == 0) {
1044                         if (hammer_debug_inode)
1045                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1046                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1047                                             HAMMER_INODE_ATIME |
1048                                             HAMMER_INODE_MTIME);
1049                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1050                         if (redirty)
1051                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1052
1053                         /*
1054                          * Root volume count of inodes
1055                          */
1056                         hammer_sync_lock_sh(trans);
1057                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1058                                 hammer_modify_volume_field(trans,
1059                                                            trans->rootvol,
1060                                                            vol0_stat_inodes);
1061                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1062                                 hammer_modify_volume_done(trans->rootvol);
1063                                 ip->flags |= HAMMER_INODE_ONDISK;
1064                                 if (hammer_debug_inode)
1065                                         kprintf("NOWONDISK %p\n", ip);
1066                         }
1067                         hammer_sync_unlock(trans);
1068                 }
1069         }
1070
1071         /*
1072          * If the inode has been destroyed, clean out any left-over flags
1073          * that may have been set by the frontend.
1074          */
1075         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1076                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1077                                     HAMMER_INODE_ATIME |
1078                                     HAMMER_INODE_MTIME);
1079         }
1080         return(error);
1081 }
1082
1083 /*
1084  * Update only the itimes fields.
1085  *
1086  * ATIME can be updated without generating any UNDO.  MTIME is updated
1087  * with UNDO so it is guaranteed to be synchronized properly in case of
1088  * a crash.
1089  *
1090  * Neither field is included in the B-Tree leaf element's CRC, which is how
1091  * we can get away with updating ATIME the way we do.
1092  */
1093 static int
1094 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1095 {
1096         hammer_transaction_t trans = cursor->trans;
1097         int error;
1098
1099 retry:
1100         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1101             HAMMER_INODE_ONDISK) {
1102                 return(0);
1103         }
1104
1105         hammer_normalize_cursor(cursor);
1106         cursor->key_beg.localization = ip->obj_localization +
1107                                        HAMMER_LOCALIZE_INODE;
1108         cursor->key_beg.obj_id = ip->obj_id;
1109         cursor->key_beg.key = 0;
1110         cursor->key_beg.create_tid = 0;
1111         cursor->key_beg.delete_tid = 0;
1112         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1113         cursor->key_beg.obj_type = 0;
1114         cursor->asof = ip->obj_asof;
1115         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1116         cursor->flags |= HAMMER_CURSOR_ASOF;
1117         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1118         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1119         cursor->flags |= HAMMER_CURSOR_BACKEND;
1120
1121         error = hammer_btree_lookup(cursor);
1122         if (error == 0) {
1123                 hammer_cache_node(&ip->cache[0], cursor->node);
1124                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1125                         /*
1126                          * Updating MTIME requires an UNDO.  Just cover
1127                          * both atime and mtime.
1128                          */
1129                         hammer_sync_lock_sh(trans);
1130                         hammer_modify_buffer(trans, cursor->data_buffer,
1131                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1132                                      HAMMER_ITIMES_BYTES);
1133                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1134                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1135                         hammer_modify_buffer_done(cursor->data_buffer);
1136                         hammer_sync_unlock(trans);
1137                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1138                         /*
1139                          * Updating atime only can be done in-place with
1140                          * no UNDO.
1141                          */
1142                         hammer_sync_lock_sh(trans);
1143                         hammer_modify_buffer(trans, cursor->data_buffer,
1144                                              NULL, 0);
1145                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1146                         hammer_modify_buffer_done(cursor->data_buffer);
1147                         hammer_sync_unlock(trans);
1148                 }
1149                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1150         }
1151         if (error == EDEADLK) {
1152                 hammer_done_cursor(cursor);
1153                 error = hammer_init_cursor(trans, cursor,
1154                                            &ip->cache[0], ip);
1155                 if (error == 0)
1156                         goto retry;
1157         }
1158         return(error);
1159 }
1160
1161 /*
1162  * Release a reference on an inode, flush as requested.
1163  *
1164  * On the last reference we queue the inode to the flusher for its final
1165  * disposition.
1166  */
1167 void
1168 hammer_rel_inode(struct hammer_inode *ip, int flush)
1169 {
1170         hammer_mount_t hmp = ip->hmp;
1171
1172         /*
1173          * Handle disposition when dropping the last ref.
1174          */
1175         for (;;) {
1176                 if (ip->lock.refs == 1) {
1177                         /*
1178                          * Determine whether on-disk action is needed for
1179                          * the inode's final disposition.
1180                          */
1181                         KKASSERT(ip->vp == NULL);
1182                         hammer_inode_unloadable_check(ip, 0);
1183                         if (ip->flags & HAMMER_INODE_MODMASK) {
1184                                 if (hmp->rsv_inodes > desiredvnodes) {
1185                                         hammer_flush_inode(ip,
1186                                                            HAMMER_FLUSH_SIGNAL);
1187                                 } else {
1188                                         hammer_flush_inode(ip, 0);
1189                                 }
1190                         } else if (ip->lock.refs == 1) {
1191                                 hammer_unload_inode(ip);
1192                                 break;
1193                         }
1194                 } else {
1195                         if (flush)
1196                                 hammer_flush_inode(ip, 0);
1197
1198                         /*
1199                          * The inode still has multiple refs, try to drop
1200                          * one ref.
1201                          */
1202                         KKASSERT(ip->lock.refs >= 1);
1203                         if (ip->lock.refs > 1) {
1204                                 hammer_unref(&ip->lock);
1205                                 break;
1206                         }
1207                 }
1208         }
1209 }
1210
1211 /*
1212  * Unload and destroy the specified inode.  Must be called with one remaining
1213  * reference.  The reference is disposed of.
1214  *
1215  * The inode must be completely clean.
1216  */
1217 static int
1218 hammer_unload_inode(struct hammer_inode *ip)
1219 {
1220         hammer_mount_t hmp = ip->hmp;
1221
1222         KASSERT(ip->lock.refs == 1,
1223                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1224         KKASSERT(ip->vp == NULL);
1225         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1226         KKASSERT(ip->cursor_ip_refs == 0);
1227         KKASSERT(ip->lock.lockcount == 0);
1228         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1229
1230         KKASSERT(RB_EMPTY(&ip->rec_tree));
1231         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1232
1233         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1234
1235         hammer_free_inode(ip);
1236         return(0);
1237 }
1238
1239 /*
1240  * Called during unmounting if a critical error occured.  The in-memory
1241  * inode and all related structures are destroyed.
1242  *
1243  * If a critical error did not occur the unmount code calls the standard
1244  * release and asserts that the inode is gone.
1245  */
1246 int
1247 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1248 {
1249         hammer_record_t rec;
1250
1251         /*
1252          * Get rid of the inodes in-memory records, regardless of their
1253          * state, and clear the mod-mask.
1254          */
1255         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1256                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1257                 rec->target_ip = NULL;
1258                 if (rec->flush_state == HAMMER_FST_SETUP)
1259                         rec->flush_state = HAMMER_FST_IDLE;
1260         }
1261         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1262                 if (rec->flush_state == HAMMER_FST_FLUSH)
1263                         --rec->flush_group->refs;
1264                 else
1265                         hammer_ref(&rec->lock);
1266                 KKASSERT(rec->lock.refs == 1);
1267                 rec->flush_state = HAMMER_FST_IDLE;
1268                 rec->flush_group = NULL;
1269                 rec->flags |= HAMMER_RECF_DELETED_FE;
1270                 rec->flags |= HAMMER_RECF_DELETED_BE;
1271                 hammer_rel_mem_record(rec);
1272         }
1273         ip->flags &= ~HAMMER_INODE_MODMASK;
1274         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1275         KKASSERT(ip->vp == NULL);
1276
1277         /*
1278          * Remove the inode from any flush group, force it idle.  FLUSH
1279          * and SETUP states have an inode ref.
1280          */
1281         switch(ip->flush_state) {
1282         case HAMMER_FST_FLUSH:
1283                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1284                 --ip->flush_group->refs;
1285                 ip->flush_group = NULL;
1286                 /* fall through */
1287         case HAMMER_FST_SETUP:
1288                 hammer_unref(&ip->lock);
1289                 ip->flush_state = HAMMER_FST_IDLE;
1290                 /* fall through */
1291         case HAMMER_FST_IDLE:
1292                 break;
1293         }
1294
1295         /*
1296          * There shouldn't be any associated vnode.  The unload needs at
1297          * least one ref, if we do have a vp steal its ip ref.
1298          */
1299         if (ip->vp) {
1300                 kprintf("hammer_destroy_inode_callback: Unexpected "
1301                         "vnode association ip %p vp %p\n", ip, ip->vp);
1302                 ip->vp->v_data = NULL;
1303                 ip->vp = NULL;
1304         } else {
1305                 hammer_ref(&ip->lock);
1306         }
1307         hammer_unload_inode(ip);
1308         return(0);
1309 }
1310
1311 /*
1312  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1313  * the read-only flag for cached inodes.
1314  *
1315  * This routine is called from a RB_SCAN().
1316  */
1317 int
1318 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1319 {
1320         hammer_mount_t hmp = ip->hmp;
1321
1322         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1323                 ip->flags |= HAMMER_INODE_RO;
1324         else
1325                 ip->flags &= ~HAMMER_INODE_RO;
1326         return(0);
1327 }
1328
1329 /*
1330  * A transaction has modified an inode, requiring updates as specified by
1331  * the passed flags.
1332  *
1333  * HAMMER_INODE_DDIRTY: Inode data has been updated
1334  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1335  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1336  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1337  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1338  */
1339 void
1340 hammer_modify_inode(hammer_inode_t ip, int flags)
1341 {
1342         /*
1343          * ronly of 0 or 2 does not trigger assertion.
1344          * 2 is a special error state
1345          */
1346         KKASSERT(ip->hmp->ronly != 1 ||
1347                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1348                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1349                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1350         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1351                 ip->flags |= HAMMER_INODE_RSV_INODES;
1352                 ++ip->hmp->rsv_inodes;
1353         }
1354
1355         ip->flags |= flags;
1356 }
1357
1358 /*
1359  * Request that an inode be flushed.  This whole mess cannot block and may
1360  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1361  * actively flush the inode until the flush can be done.
1362  *
1363  * The inode may already be flushing, or may be in a setup state.  We can
1364  * place the inode in a flushing state if it is currently idle and flag it
1365  * to reflush if it is currently flushing.
1366  *
1367  * Upon return if the inode could not be flushed due to a setup
1368  * dependancy, then it will be automatically flushed when the dependancy
1369  * is satisfied.
1370  */
1371 void
1372 hammer_flush_inode(hammer_inode_t ip, int flags)
1373 {
1374         hammer_mount_t hmp;
1375         hammer_flush_group_t flg;
1376         int good;
1377
1378         /*
1379          * next_flush_group is the first flush group we can place the inode
1380          * in.  It may be NULL.  If it becomes full we append a new flush
1381          * group and make that the next_flush_group.
1382          */
1383         hmp = ip->hmp;
1384         while ((flg = hmp->next_flush_group) != NULL) {
1385                 KKASSERT(flg->running == 0);
1386                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1387                         break;
1388                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1389                 hammer_flusher_async(ip->hmp, flg);
1390         }
1391         if (flg == NULL) {
1392                 flg = kmalloc(sizeof(*flg), M_HAMMER, M_WAITOK|M_ZERO);
1393                 hmp->next_flush_group = flg;
1394                 TAILQ_INIT(&flg->flush_list);
1395                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1396         }
1397
1398         /*
1399          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1400          * state we have to put it back into an IDLE state so we can
1401          * drop the extra ref.
1402          *
1403          * If we have a parent dependancy we must still fall through
1404          * so we can run it.
1405          */
1406         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1407                 if (ip->flush_state == HAMMER_FST_SETUP &&
1408                     TAILQ_EMPTY(&ip->target_list)) {
1409                         ip->flush_state = HAMMER_FST_IDLE;
1410                         hammer_rel_inode(ip, 0);
1411                 }
1412                 if (ip->flush_state == HAMMER_FST_IDLE)
1413                         return;
1414         }
1415
1416         /*
1417          * Our flush action will depend on the current state.
1418          */
1419         switch(ip->flush_state) {
1420         case HAMMER_FST_IDLE:
1421                 /*
1422                  * We have no dependancies and can flush immediately.  Some
1423                  * our children may not be flushable so we have to re-test
1424                  * with that additional knowledge.
1425                  */
1426                 hammer_flush_inode_core(ip, flg, flags);
1427                 break;
1428         case HAMMER_FST_SETUP:
1429                 /*
1430                  * Recurse upwards through dependancies via target_list
1431                  * and start their flusher actions going if possible.
1432                  *
1433                  * 'good' is our connectivity.  -1 means we have none and
1434                  * can't flush, 0 means there weren't any dependancies, and
1435                  * 1 means we have good connectivity.
1436                  */
1437                 good = hammer_setup_parent_inodes(ip, flg);
1438
1439                 if (good >= 0) {
1440                         /*
1441                          * We can continue if good >= 0.  Determine how
1442                          * many records under our inode can be flushed (and
1443                          * mark them).
1444                          */
1445                         hammer_flush_inode_core(ip, flg, flags);
1446                 } else {
1447                         /*
1448                          * Parent has no connectivity, tell it to flush
1449                          * us as soon as it does.
1450                          *
1451                          * The REFLUSH flag is also needed to trigger
1452                          * dependancy wakeups.
1453                          */
1454                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1455                                      HAMMER_INODE_REFLUSH;
1456                         if (flags & HAMMER_FLUSH_SIGNAL) {
1457                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1458                                 hammer_flusher_async(ip->hmp, flg);
1459                         }
1460                 }
1461                 break;
1462         case HAMMER_FST_FLUSH:
1463                 /*
1464                  * We are already flushing, flag the inode to reflush
1465                  * if needed after it completes its current flush.
1466                  *
1467                  * The REFLUSH flag is also needed to trigger
1468                  * dependancy wakeups.
1469                  */
1470                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1471                         ip->flags |= HAMMER_INODE_REFLUSH;
1472                 if (flags & HAMMER_FLUSH_SIGNAL) {
1473                         ip->flags |= HAMMER_INODE_RESIGNAL;
1474                         hammer_flusher_async(ip->hmp, flg);
1475                 }
1476                 break;
1477         }
1478 }
1479
1480 /*
1481  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1482  * ip which reference our ip.
1483  *
1484  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1485  *     so for now do not ref/deref the structures.  Note that if we use the
1486  *     ref/rel code later, the rel CAN block.
1487  */
1488 static int
1489 hammer_setup_parent_inodes(hammer_inode_t ip, hammer_flush_group_t flg)
1490 {
1491         hammer_record_t depend;
1492         int good;
1493         int r;
1494
1495         good = 0;
1496         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1497                 r = hammer_setup_parent_inodes_helper(depend, flg);
1498                 KKASSERT(depend->target_ip == ip);
1499                 if (r < 0 && good == 0)
1500                         good = -1;
1501                 if (r > 0)
1502                         good = 1;
1503         }
1504         return(good);
1505 }
1506
1507 /*
1508  * This helper function takes a record representing the dependancy between
1509  * the parent inode and child inode.
1510  *
1511  * record->ip           = parent inode
1512  * record->target_ip    = child inode
1513  *
1514  * We are asked to recurse upwards and convert the record from SETUP
1515  * to FLUSH if possible.
1516  *
1517  * Return 1 if the record gives us connectivity
1518  *
1519  * Return 0 if the record is not relevant
1520  *
1521  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1522  */
1523 static int
1524 hammer_setup_parent_inodes_helper(hammer_record_t record,
1525                                   hammer_flush_group_t flg)
1526 {
1527         hammer_mount_t hmp;
1528         hammer_inode_t pip;
1529         int good;
1530
1531         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1532         pip = record->ip;
1533         hmp = pip->hmp;
1534
1535         /*
1536          * If the record is already flushing, is it in our flush group?
1537          *
1538          * If it is in our flush group but it is a general record or a
1539          * delete-on-disk, it does not improve our connectivity (return 0),
1540          * and if the target inode is not trying to destroy itself we can't
1541          * allow the operation yet anyway (the second return -1).
1542          */
1543         if (record->flush_state == HAMMER_FST_FLUSH) {
1544                 /*
1545                  * If not in our flush group ask the parent to reflush
1546                  * us as soon as possible.
1547                  */
1548                 if (record->flush_group != flg) {
1549                         pip->flags |= HAMMER_INODE_REFLUSH;
1550                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1551                         return(-1);
1552                 }
1553
1554                 /*
1555                  * If in our flush group everything is already set up,
1556                  * just return whether the record will improve our
1557                  * visibility or not.
1558                  */
1559                 if (record->type == HAMMER_MEM_RECORD_ADD)
1560                         return(1);
1561                 return(0);
1562         }
1563
1564         /*
1565          * It must be a setup record.  Try to resolve the setup dependancies
1566          * by recursing upwards so we can place ip on the flush list.
1567          */
1568         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1569
1570         good = hammer_setup_parent_inodes(pip, flg);
1571
1572         /*
1573          * If good < 0 the parent has no connectivity and we cannot safely
1574          * flush the directory entry, which also means we can't flush our
1575          * ip.  Flag the parent and us for downward recursion once the
1576          * parent's connectivity is resolved.
1577          */
1578         if (good < 0) {
1579                 /* pip->flags |= HAMMER_INODE_CONN_DOWN; set by recursion */
1580                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1581                 return(good);
1582         }
1583
1584         /*
1585          * We are go, place the parent inode in a flushing state so we can
1586          * place its record in a flushing state.  Note that the parent
1587          * may already be flushing.  The record must be in the same flush
1588          * group as the parent.
1589          */
1590         if (pip->flush_state != HAMMER_FST_FLUSH)
1591                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1592         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1593         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1594
1595 #if 0
1596         if (record->type == HAMMER_MEM_RECORD_DEL &&
1597             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1598                 /*
1599                  * Regardless of flushing state we cannot sync this path if the
1600                  * record represents a delete-on-disk but the target inode
1601                  * is not ready to sync its own deletion.
1602                  *
1603                  * XXX need to count effective nlinks to determine whether
1604                  * the flush is ok, otherwise removing a hardlink will
1605                  * just leave the DEL record to rot.
1606                  */
1607                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1608                 return(-1);
1609         } else
1610 #endif
1611         if (pip->flush_group == flg) {
1612                 /*
1613                  * Because we have not calculated nlinks yet we can just
1614                  * set records to the flush state if the parent is in
1615                  * the same flush group as we are.
1616                  */
1617                 record->flush_state = HAMMER_FST_FLUSH;
1618                 record->flush_group = flg;
1619                 ++record->flush_group->refs;
1620                 hammer_ref(&record->lock);
1621
1622                 /*
1623                  * A general directory-add contributes to our visibility.
1624                  *
1625                  * Otherwise it is probably a directory-delete or
1626                  * delete-on-disk record and does not contribute to our
1627                  * visbility (but we can still flush it).
1628                  */
1629                 if (record->type == HAMMER_MEM_RECORD_ADD)
1630                         return(1);
1631                 return(0);
1632         } else {
1633                 /*
1634                  * If the parent is not in our flush group we cannot
1635                  * flush this record yet, there is no visibility.
1636                  * We tell the parent to reflush and mark ourselves
1637                  * so the parent knows it should flush us too.
1638                  */
1639                 pip->flags |= HAMMER_INODE_REFLUSH;
1640                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1641                 return(-1);
1642         }
1643 }
1644
1645 /*
1646  * This is the core routine placing an inode into the FST_FLUSH state.
1647  */
1648 static void
1649 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1650 {
1651         int go_count;
1652
1653         /*
1654          * Set flush state and prevent the flusher from cycling into
1655          * the next flush group.  Do not place the ip on the list yet.
1656          * Inodes not in the idle state get an extra reference.
1657          */
1658         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1659         if (ip->flush_state == HAMMER_FST_IDLE)
1660                 hammer_ref(&ip->lock);
1661         ip->flush_state = HAMMER_FST_FLUSH;
1662         ip->flush_group = flg;
1663         ++ip->hmp->flusher.group_lock;
1664         ++ip->hmp->count_iqueued;
1665         ++hammer_count_iqueued;
1666         ++flg->total_count;
1667
1668         /*
1669          * We need to be able to vfsync/truncate from the backend.
1670          */
1671         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1672         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1673                 ip->flags |= HAMMER_INODE_VHELD;
1674                 vref(ip->vp);
1675         }
1676
1677         /*
1678          * Figure out how many in-memory records we can actually flush
1679          * (not including inode meta-data, buffers, etc).
1680          */
1681         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1682         if (flags & HAMMER_FLUSH_RECURSION) {
1683                 /*
1684                  * If this is a upwards recursion we do not want to
1685                  * recurse down again!
1686                  */
1687                 go_count = 1;
1688 #if 0
1689         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1690                 /*
1691                  * No new records are added if we must complete a flush
1692                  * from a previous cycle, but we do have to move the records
1693                  * from the previous cycle to the current one.
1694                  */
1695 #if 0
1696                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1697                                    hammer_syncgrp_child_callback, NULL);
1698 #endif
1699                 go_count = 1;
1700 #endif
1701         } else {
1702                 /*
1703                  * Normal flush, scan records and bring them into the flush.
1704                  * Directory adds and deletes are usually skipped (they are
1705                  * grouped with the related inode rather then with the
1706                  * directory).
1707                  *
1708                  * go_count can be negative, which means the scan aborted
1709                  * due to the flush group being over-full and we should
1710                  * flush what we have.
1711                  */
1712                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1713                                    hammer_setup_child_callback, NULL);
1714         }
1715
1716         /*
1717          * This is a more involved test that includes go_count.  If we
1718          * can't flush, flag the inode and return.  If go_count is 0 we
1719          * were are unable to flush any records in our rec_tree and
1720          * must ignore the XDIRTY flag.
1721          */
1722         if (go_count == 0) {
1723                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1724                         --ip->hmp->count_iqueued;
1725                         --hammer_count_iqueued;
1726
1727                         --flg->total_count;
1728                         ip->flush_state = HAMMER_FST_SETUP;
1729                         ip->flush_group = NULL;
1730                         if (ip->flags & HAMMER_INODE_VHELD) {
1731                                 ip->flags &= ~HAMMER_INODE_VHELD;
1732                                 vrele(ip->vp);
1733                         }
1734
1735                         /*
1736                          * REFLUSH is needed to trigger dependancy wakeups
1737                          * when an inode is in SETUP.
1738                          */
1739                         ip->flags |= HAMMER_INODE_REFLUSH;
1740                         if (flags & HAMMER_FLUSH_SIGNAL) {
1741                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1742                                 hammer_flusher_async(ip->hmp, flg);
1743                         }
1744                         if (--ip->hmp->flusher.group_lock == 0)
1745                                 wakeup(&ip->hmp->flusher.group_lock);
1746                         return;
1747                 }
1748         }
1749
1750         /*
1751          * Snapshot the state of the inode for the backend flusher.
1752          *
1753          * We continue to retain save_trunc_off even when all truncations
1754          * have been resolved as an optimization to determine if we can
1755          * skip the B-Tree lookup for overwrite deletions.
1756          *
1757          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1758          * and stays in ip->flags.  Once set, it stays set until the
1759          * inode is destroyed.
1760          */
1761         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1762                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1763                 ip->sync_trunc_off = ip->trunc_off;
1764                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1765                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1766                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1767
1768                 /*
1769                  * The save_trunc_off used to cache whether the B-Tree
1770                  * holds any records past that point is not used until
1771                  * after the truncation has succeeded, so we can safely
1772                  * set it now.
1773                  */
1774                 if (ip->save_trunc_off > ip->sync_trunc_off)
1775                         ip->save_trunc_off = ip->sync_trunc_off;
1776         }
1777         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1778                            ~HAMMER_INODE_TRUNCATED);
1779         ip->sync_ino_leaf = ip->ino_leaf;
1780         ip->sync_ino_data = ip->ino_data;
1781         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1782 #ifdef DEBUG_TRUNCATE
1783         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1784                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1785 #endif
1786
1787         /*
1788          * The flusher list inherits our inode and reference.
1789          */
1790         KKASSERT(flg->running == 0);
1791         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
1792         if (--ip->hmp->flusher.group_lock == 0)
1793                 wakeup(&ip->hmp->flusher.group_lock);
1794
1795         if (flags & HAMMER_FLUSH_SIGNAL) {
1796                 hammer_flusher_async(ip->hmp, flg);
1797         }
1798 }
1799
1800 /*
1801  * Callback for scan of ip->rec_tree.  Try to include each record in our
1802  * flush.  ip->flush_group has been set but the inode has not yet been
1803  * moved into a flushing state.
1804  *
1805  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1806  * both inodes.
1807  *
1808  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1809  * the caller from shortcutting the flush.
1810  */
1811 static int
1812 hammer_setup_child_callback(hammer_record_t rec, void *data)
1813 {
1814         hammer_flush_group_t flg;
1815         hammer_inode_t target_ip;
1816         hammer_inode_t ip;
1817         int r;
1818
1819         /*
1820          * Deleted records are ignored.  Note that the flush detects deleted
1821          * front-end records at multiple points to deal with races.  This is
1822          * just the first line of defense.  The only time DELETED_FE cannot
1823          * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1824          *
1825          * Don't get confused between record deletion and, say, directory
1826          * entry deletion.  The deletion of a directory entry that is on
1827          * the media has nothing to do with the record deletion flags.
1828          */
1829         if (rec->flags & (HAMMER_RECF_DELETED_FE|HAMMER_RECF_DELETED_BE)) {
1830                 if (rec->flush_state == HAMMER_FST_FLUSH) {
1831                         KKASSERT(rec->flush_group == rec->ip->flush_group);
1832                         r = 1;
1833                 } else {
1834                         r = 0;
1835                 }
1836                 return(r);
1837         }
1838
1839         /*
1840          * If the record is in an idle state it has no dependancies and
1841          * can be flushed.
1842          */
1843         ip = rec->ip;
1844         flg = ip->flush_group;
1845         r = 0;
1846
1847         switch(rec->flush_state) {
1848         case HAMMER_FST_IDLE:
1849                 /*
1850                  * The record has no setup dependancy, we can flush it.
1851                  */
1852                 KKASSERT(rec->target_ip == NULL);
1853                 rec->flush_state = HAMMER_FST_FLUSH;
1854                 rec->flush_group = flg;
1855                 ++flg->refs;
1856                 hammer_ref(&rec->lock);
1857                 r = 1;
1858                 break;
1859         case HAMMER_FST_SETUP:
1860                 /*
1861                  * The record has a setup dependancy.  These are typically
1862                  * directory entry adds and deletes.  Such entries will be
1863                  * flushed when their inodes are flushed so we do not
1864                  * usually have to add them to the flush here.  However,
1865                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
1866                  * it is asking us to flush this record (and it).
1867                  */
1868                 target_ip = rec->target_ip;
1869                 KKASSERT(target_ip != NULL);
1870                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1871
1872                 /*
1873                  * If the target IP is already flushing in our group
1874                  * we could associate the record, but target_ip has
1875                  * already synced ino_data to sync_ino_data and we
1876                  * would also have to adjust nlinks.   Plus there are
1877                  * ordering issues for adds and deletes.
1878                  *
1879                  * Reflush downward if this is an ADD, and upward if
1880                  * this is a DEL.
1881                  */
1882                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1883                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
1884                                 ip->flags |= HAMMER_INODE_REFLUSH;
1885                         else
1886                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1887                         break;
1888                 }
1889
1890                 /*
1891                  * Target IP is not yet flushing.  This can get complex
1892                  * because we have to be careful about the recursion.
1893                  *
1894                  * Directories create an issue for us in that if a flush
1895                  * of a directory is requested the expectation is to flush
1896                  * any pending directory entries, but this will cause the
1897                  * related inodes to recursively flush as well.  We can't
1898                  * really defer the operation so just get as many as we
1899                  * can and
1900                  */
1901 #if 0
1902                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
1903                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
1904                         /*
1905                          * We aren't reclaiming and the target ip was not
1906                          * previously prevented from flushing due to this
1907                          * record dependancy.  Do not flush this record.
1908                          */
1909                         /*r = 0;*/
1910                 } else
1911 #endif
1912                 if (flg->total_count + flg->refs >
1913                            ip->hmp->undo_rec_limit) {
1914                         /*
1915                          * Our flush group is over-full and we risk blowing
1916                          * out the UNDO FIFO.  Stop the scan, flush what we
1917                          * have, then reflush the directory.
1918                          *
1919                          * The directory may be forced through multiple
1920                          * flush groups before it can be completely
1921                          * flushed.
1922                          */
1923                         ip->flags |= HAMMER_INODE_RESIGNAL |
1924                                      HAMMER_INODE_REFLUSH;
1925                         r = -1;
1926                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1927                         /*
1928                          * If the target IP is not flushing we can force
1929                          * it to flush, even if it is unable to write out
1930                          * any of its own records we have at least one in
1931                          * hand that we CAN deal with.
1932                          */
1933                         rec->flush_state = HAMMER_FST_FLUSH;
1934                         rec->flush_group = flg;
1935                         ++flg->refs;
1936                         hammer_ref(&rec->lock);
1937                         hammer_flush_inode_core(target_ip, flg,
1938                                                 HAMMER_FLUSH_RECURSION);
1939                         r = 1;
1940                 } else {
1941                         /*
1942                          * General or delete-on-disk record.
1943                          *
1944                          * XXX this needs help.  If a delete-on-disk we could
1945                          * disconnect the target.  If the target has its own
1946                          * dependancies they really need to be flushed.
1947                          *
1948                          * XXX
1949                          */
1950                         rec->flush_state = HAMMER_FST_FLUSH;
1951                         rec->flush_group = flg;
1952                         ++flg->refs;
1953                         hammer_ref(&rec->lock);
1954                         hammer_flush_inode_core(target_ip, flg,
1955                                                 HAMMER_FLUSH_RECURSION);
1956                         r = 1;
1957                 }
1958                 break;
1959         case HAMMER_FST_FLUSH:
1960                 /*
1961                  * The flush_group should already match.
1962                  */
1963                 KKASSERT(rec->flush_group == flg);
1964                 r = 1;
1965                 break;
1966         }
1967         return(r);
1968 }
1969
1970 #if 0
1971 /*
1972  * This version just moves records already in a flush state to the new
1973  * flush group and that is it.
1974  */
1975 static int
1976 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
1977 {
1978         hammer_inode_t ip = rec->ip;
1979
1980         switch(rec->flush_state) {
1981         case HAMMER_FST_FLUSH:
1982                 KKASSERT(rec->flush_group == ip->flush_group);
1983                 break;
1984         default:
1985                 break;
1986         }
1987         return(0);
1988 }
1989 #endif
1990
1991 /*
1992  * Wait for a previously queued flush to complete.
1993  *
1994  * If a critical error occured we don't try to wait.
1995  */
1996 void
1997 hammer_wait_inode(hammer_inode_t ip)
1998 {
1999         hammer_flush_group_t flg;
2000
2001         flg = NULL;
2002         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2003                 while (ip->flush_state != HAMMER_FST_IDLE &&
2004                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2005                         if (ip->flush_state == HAMMER_FST_SETUP)
2006                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2007                         if (ip->flush_state != HAMMER_FST_IDLE) {
2008                                 ip->flags |= HAMMER_INODE_FLUSHW;
2009                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2010                         }
2011                 }
2012         }
2013 }
2014
2015 /*
2016  * Called by the backend code when a flush has been completed.
2017  * The inode has already been removed from the flush list.
2018  *
2019  * A pipelined flush can occur, in which case we must re-enter the
2020  * inode on the list and re-copy its fields.
2021  */
2022 void
2023 hammer_flush_inode_done(hammer_inode_t ip, int error)
2024 {
2025         hammer_mount_t hmp;
2026         int dorel;
2027
2028         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2029
2030         hmp = ip->hmp;
2031
2032         /*
2033          * Merge left-over flags back into the frontend and fix the state.
2034          * Incomplete truncations are retained by the backend.
2035          */
2036         ip->error = error;
2037         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2038         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2039
2040         /*
2041          * The backend may have adjusted nlinks, so if the adjusted nlinks
2042          * does not match the fronttend set the frontend's RDIRTY flag again.
2043          */
2044         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2045                 ip->flags |= HAMMER_INODE_DDIRTY;
2046
2047         /*
2048          * Fix up the dirty buffer status.
2049          */
2050         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2051                 ip->flags |= HAMMER_INODE_BUFS;
2052         }
2053
2054         /*
2055          * Re-set the XDIRTY flag if some of the inode's in-memory records
2056          * could not be flushed.
2057          */
2058         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2059                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2060                  (!RB_EMPTY(&ip->rec_tree) &&
2061                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2062
2063         /*
2064          * Do not lose track of inodes which no longer have vnode
2065          * assocations, otherwise they may never get flushed again.
2066          */
2067         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
2068                 ip->flags |= HAMMER_INODE_REFLUSH;
2069
2070         /*
2071          * Adjust the flush state.
2072          */
2073         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2074                 /*
2075                  * We were unable to flush out all our records, leave the
2076                  * inode in a flush state and in the current flush group.
2077                  * The flush group will be re-run.
2078                  *
2079                  * This occurs if the UNDO block gets too full or there is
2080                  * too much dirty meta-data and allows the flusher to
2081                  * finalize the UNDO block and then re-flush.
2082                  */
2083                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2084                 dorel = 0;
2085         } else {
2086                 /*
2087                  * Remove from the flush_group
2088                  */
2089                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2090                 ip->flush_group = NULL;
2091
2092                 /*
2093                  * Clean up the vnode ref and tracking counts.
2094                  */
2095                 if (ip->flags & HAMMER_INODE_VHELD) {
2096                         ip->flags &= ~HAMMER_INODE_VHELD;
2097                         vrele(ip->vp);
2098                 }
2099                 --hmp->count_iqueued;
2100                 --hammer_count_iqueued;
2101
2102                 /*
2103                  * And adjust the state.
2104                  */
2105                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2106                         ip->flush_state = HAMMER_FST_IDLE;
2107                         dorel = 1;
2108                 } else {
2109                         ip->flush_state = HAMMER_FST_SETUP;
2110                         dorel = 0;
2111                 }
2112
2113                 /*
2114                  * If the frontend is waiting for a flush to complete,
2115                  * wake it up.
2116                  */
2117                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2118                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2119                         wakeup(&ip->flags);
2120                 }
2121
2122                 /*
2123                  * If the frontend made more changes and requested another
2124                  * flush, then try to get it running.
2125                  *
2126                  * Reflushes are aborted when the inode is errored out.
2127                  */
2128                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2129                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2130                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2131                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2132                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2133                         } else {
2134                                 hammer_flush_inode(ip, 0);
2135                         }
2136                 }
2137         }
2138
2139         /*
2140          * If we have no parent dependancies we can clear CONN_DOWN
2141          */
2142         if (TAILQ_EMPTY(&ip->target_list))
2143                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2144
2145         /*
2146          * If the inode is now clean drop the space reservation.
2147          */
2148         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2149             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2150                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2151                 --hmp->rsv_inodes;
2152         }
2153
2154         if (dorel)
2155                 hammer_rel_inode(ip, 0);
2156 }
2157
2158 /*
2159  * Called from hammer_sync_inode() to synchronize in-memory records
2160  * to the media.
2161  */
2162 static int
2163 hammer_sync_record_callback(hammer_record_t record, void *data)
2164 {
2165         hammer_cursor_t cursor = data;
2166         hammer_transaction_t trans = cursor->trans;
2167         hammer_mount_t hmp = trans->hmp;
2168         int error;
2169
2170         /*
2171          * Skip records that do not belong to the current flush.
2172          */
2173         ++hammer_stats_record_iterations;
2174         if (record->flush_state != HAMMER_FST_FLUSH)
2175                 return(0);
2176
2177 #if 1
2178         if (record->flush_group != record->ip->flush_group) {
2179                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2180                 Debugger("blah2");
2181                 return(0);
2182         }
2183 #endif
2184         KKASSERT(record->flush_group == record->ip->flush_group);
2185
2186         /*
2187          * Interlock the record using the BE flag.  Once BE is set the
2188          * frontend cannot change the state of FE.
2189          *
2190          * NOTE: If FE is set prior to us setting BE we still sync the
2191          * record out, but the flush completion code converts it to
2192          * a delete-on-disk record instead of destroying it.
2193          */
2194         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2195         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2196
2197         /*
2198          * The backend may have already disposed of the record.
2199          */
2200         if (record->flags & HAMMER_RECF_DELETED_BE) {
2201                 error = 0;
2202                 goto done;
2203         }
2204
2205         /*
2206          * If the whole inode is being deleting all on-disk records will
2207          * be deleted very soon, we can't sync any new records to disk
2208          * because they will be deleted in the same transaction they were
2209          * created in (delete_tid == create_tid), which will assert.
2210          *
2211          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2212          * that we currently panic on.
2213          */
2214         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2215                 switch(record->type) {
2216                 case HAMMER_MEM_RECORD_DATA:
2217                         /*
2218                          * We don't have to do anything, if the record was
2219                          * committed the space will have been accounted for
2220                          * in the blockmap.
2221                          */
2222                         /* fall through */
2223                 case HAMMER_MEM_RECORD_GENERAL:
2224                         record->flags |= HAMMER_RECF_DELETED_FE;
2225                         record->flags |= HAMMER_RECF_DELETED_BE;
2226                         error = 0;
2227                         goto done;
2228                 case HAMMER_MEM_RECORD_ADD:
2229                         panic("hammer_sync_record_callback: illegal add "
2230                               "during inode deletion record %p", record);
2231                         break; /* NOT REACHED */
2232                 case HAMMER_MEM_RECORD_INODE:
2233                         panic("hammer_sync_record_callback: attempt to "
2234                               "sync inode record %p?", record);
2235                         break; /* NOT REACHED */
2236                 case HAMMER_MEM_RECORD_DEL:
2237                         /*
2238                          * Follow through and issue the on-disk deletion
2239                          */
2240                         break;
2241                 }
2242         }
2243
2244         /*
2245          * If DELETED_FE is set special handling is needed for directory
2246          * entries.  Dependant pieces related to the directory entry may
2247          * have already been synced to disk.  If this occurs we have to
2248          * sync the directory entry and then change the in-memory record
2249          * from an ADD to a DELETE to cover the fact that it's been
2250          * deleted by the frontend.
2251          *
2252          * A directory delete covering record (MEM_RECORD_DEL) can never
2253          * be deleted by the frontend.
2254          *
2255          * Any other record type (aka DATA) can be deleted by the frontend.
2256          * XXX At the moment the flusher must skip it because there may
2257          * be another data record in the flush group for the same block,
2258          * meaning that some frontend data changes can leak into the backend's
2259          * synchronization point.
2260          */
2261         if (record->flags & HAMMER_RECF_DELETED_FE) {
2262                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2263                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2264                 } else {
2265                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2266                         record->flags |= HAMMER_RECF_DELETED_BE;
2267                         error = 0;
2268                         goto done;
2269                 }
2270         }
2271
2272         /*
2273          * Assign the create_tid for new records.  Deletions already
2274          * have the record's entire key properly set up.
2275          */
2276         if (record->type != HAMMER_MEM_RECORD_DEL)
2277                 record->leaf.base.create_tid = trans->tid;
2278                 record->leaf.create_ts = trans->time32;
2279         for (;;) {
2280                 error = hammer_ip_sync_record_cursor(cursor, record);
2281                 if (error != EDEADLK)
2282                         break;
2283                 hammer_done_cursor(cursor);
2284                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2285                                            record->ip);
2286                 if (error)
2287                         break;
2288         }
2289         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2290
2291         if (error)
2292                 error = -error;
2293 done:
2294         hammer_flush_record_done(record, error);
2295
2296         /*
2297          * Do partial finalization if we have built up too many dirty
2298          * buffers.  Otherwise a buffer cache deadlock can occur when
2299          * doing things like creating tens of thousands of tiny files.
2300          *
2301          * We must release our cursor lock to avoid a 3-way deadlock
2302          * due to the exclusive sync lock the finalizer must get.
2303          */
2304         if (hammer_flusher_meta_limit(hmp)) {
2305                 hammer_unlock_cursor(cursor, 0);
2306                 hammer_flusher_finalize(trans, 0);
2307                 hammer_lock_cursor(cursor, 0);
2308         }
2309
2310         return(error);
2311 }
2312
2313 /*
2314  * Backend function called by the flusher to sync an inode to media.
2315  */
2316 int
2317 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2318 {
2319         struct hammer_cursor cursor;
2320         hammer_node_t tmp_node;
2321         hammer_record_t depend;
2322         hammer_record_t next;
2323         int error, tmp_error;
2324         u_int64_t nlinks;
2325
2326         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2327                 return(0);
2328
2329         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2330         if (error)
2331                 goto done;
2332
2333         /*
2334          * Any directory records referencing this inode which are not in
2335          * our current flush group must adjust our nlink count for the
2336          * purposes of synchronization to disk.
2337          *
2338          * Records which are in our flush group can be unlinked from our
2339          * inode now, potentially allowing the inode to be physically
2340          * deleted.
2341          *
2342          * This cannot block.
2343          */
2344         nlinks = ip->ino_data.nlinks;
2345         next = TAILQ_FIRST(&ip->target_list);
2346         while ((depend = next) != NULL) {
2347                 next = TAILQ_NEXT(depend, target_entry);
2348                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2349                     depend->flush_group == ip->flush_group) {
2350                         /*
2351                          * If this is an ADD that was deleted by the frontend
2352                          * the frontend nlinks count will have already been
2353                          * decremented, but the backend is going to sync its
2354                          * directory entry and must account for it.  The
2355                          * record will be converted to a delete-on-disk when
2356                          * it gets synced.
2357                          *
2358                          * If the ADD was not deleted by the frontend we
2359                          * can remove the dependancy from our target_list.
2360                          */
2361                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2362                                 ++nlinks;
2363                         } else {
2364                                 TAILQ_REMOVE(&ip->target_list, depend,
2365                                              target_entry);
2366                                 depend->target_ip = NULL;
2367                         }
2368                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2369                         /*
2370                          * Not part of our flush group
2371                          */
2372                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2373                         switch(depend->type) {
2374                         case HAMMER_MEM_RECORD_ADD:
2375                                 --nlinks;
2376                                 break;
2377                         case HAMMER_MEM_RECORD_DEL:
2378                                 ++nlinks;
2379                                 break;
2380                         default:
2381                                 break;
2382                         }
2383                 }
2384         }
2385
2386         /*
2387          * Set dirty if we had to modify the link count.
2388          */
2389         if (ip->sync_ino_data.nlinks != nlinks) {
2390                 KKASSERT((int64_t)nlinks >= 0);
2391                 ip->sync_ino_data.nlinks = nlinks;
2392                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2393         }
2394
2395         /*
2396          * If there is a trunction queued destroy any data past the (aligned)
2397          * truncation point.  Userland will have dealt with the buffer
2398          * containing the truncation point for us.
2399          *
2400          * We don't flush pending frontend data buffers until after we've
2401          * dealt with the truncation.
2402          */
2403         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2404                 /*
2405                  * Interlock trunc_off.  The VOP front-end may continue to
2406                  * make adjustments to it while we are blocked.
2407                  */
2408                 off_t trunc_off;
2409                 off_t aligned_trunc_off;
2410                 int blkmask;
2411
2412                 trunc_off = ip->sync_trunc_off;
2413                 blkmask = hammer_blocksize(trunc_off) - 1;
2414                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2415
2416                 /*
2417                  * Delete any whole blocks on-media.  The front-end has
2418                  * already cleaned out any partial block and made it
2419                  * pending.  The front-end may have updated trunc_off
2420                  * while we were blocked so we only use sync_trunc_off.
2421                  *
2422                  * This operation can blow out the buffer cache, EWOULDBLOCK
2423                  * means we were unable to complete the deletion.  The
2424                  * deletion will update sync_trunc_off in that case.
2425                  */
2426                 error = hammer_ip_delete_range(&cursor, ip,
2427                                                 aligned_trunc_off,
2428                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2429                 if (error == EWOULDBLOCK) {
2430                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2431                         error = 0;
2432                         goto defer_buffer_flush;
2433                 }
2434
2435                 if (error)
2436                         goto done;
2437
2438                 /*
2439                  * Clear the truncation flag on the backend after we have
2440                  * complete the deletions.  Backend data is now good again
2441                  * (including new records we are about to sync, below).
2442                  *
2443                  * Leave sync_trunc_off intact.  As we write additional
2444                  * records the backend will update sync_trunc_off.  This
2445                  * tells the backend whether it can skip the overwrite
2446                  * test.  This should work properly even when the backend
2447                  * writes full blocks where the truncation point straddles
2448                  * the block because the comparison is against the base
2449                  * offset of the record.
2450                  */
2451                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2452                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2453         } else {
2454                 error = 0;
2455         }
2456
2457         /*
2458          * Now sync related records.  These will typically be directory
2459          * entries, records tracking direct-writes, or delete-on-disk records.
2460          */
2461         if (error == 0) {
2462                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2463                                     hammer_sync_record_callback, &cursor);
2464                 if (tmp_error < 0)
2465                         tmp_error = -error;
2466                 if (tmp_error)
2467                         error = tmp_error;
2468         }
2469         hammer_cache_node(&ip->cache[1], cursor.node);
2470
2471         /*
2472          * Re-seek for inode update, assuming our cache hasn't been ripped
2473          * out from under us.
2474          */
2475         if (error == 0) {
2476                 tmp_node = hammer_ref_node_safe(ip->hmp, &ip->cache[0], &error);
2477                 if (tmp_node) {
2478                         hammer_cursor_downgrade(&cursor);
2479                         hammer_lock_sh(&tmp_node->lock);
2480                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2481                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2482                         hammer_unlock(&tmp_node->lock);
2483                         hammer_rel_node(tmp_node);
2484                 }
2485                 error = 0;
2486         }
2487
2488         /*
2489          * If we are deleting the inode the frontend had better not have
2490          * any active references on elements making up the inode.
2491          *
2492          * The call to hammer_ip_delete_clean() cleans up auxillary records
2493          * but not DB or DATA records.  Those must have already been deleted
2494          * by the normal truncation mechanic.
2495          */
2496         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2497                 RB_EMPTY(&ip->rec_tree)  &&
2498             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2499             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2500                 int count1 = 0;
2501
2502                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2503                 if (error == 0) {
2504                         ip->flags |= HAMMER_INODE_DELETED;
2505                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2506                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2507                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2508
2509                         /*
2510                          * Set delete_tid in both the frontend and backend
2511                          * copy of the inode record.  The DELETED flag handles
2512                          * this, do not set RDIRTY.
2513                          */
2514                         ip->ino_leaf.base.delete_tid = trans->tid;
2515                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2516                         ip->ino_leaf.delete_ts = trans->time32;
2517                         ip->sync_ino_leaf.delete_ts = trans->time32;
2518
2519
2520                         /*
2521                          * Adjust the inode count in the volume header
2522                          */
2523                         hammer_sync_lock_sh(trans);
2524                         if (ip->flags & HAMMER_INODE_ONDISK) {
2525                                 hammer_modify_volume_field(trans,
2526                                                            trans->rootvol,
2527                                                            vol0_stat_inodes);
2528                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2529                                 hammer_modify_volume_done(trans->rootvol);
2530                         }
2531                         hammer_sync_unlock(trans);
2532                 }
2533         }
2534
2535         if (error)
2536                 goto done;
2537         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2538
2539 defer_buffer_flush:
2540         /*
2541          * Now update the inode's on-disk inode-data and/or on-disk record.
2542          * DELETED and ONDISK are managed only in ip->flags.
2543          *
2544          * In the case of a defered buffer flush we still update the on-disk
2545          * inode to satisfy visibility requirements if there happen to be
2546          * directory dependancies.
2547          */
2548         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2549         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2550                 /*
2551                  * If deleted and on-disk, don't set any additional flags.
2552                  * the delete flag takes care of things.
2553                  *
2554                  * Clear flags which may have been set by the frontend.
2555                  */
2556                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2557                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2558                                     HAMMER_INODE_DELETING);
2559                 break;
2560         case HAMMER_INODE_DELETED:
2561                 /*
2562                  * Take care of the case where a deleted inode was never
2563                  * flushed to the disk in the first place.
2564                  *
2565                  * Clear flags which may have been set by the frontend.
2566                  */
2567                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2568                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2569                                     HAMMER_INODE_DELETING);
2570                 while (RB_ROOT(&ip->rec_tree)) {
2571                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2572                         hammer_ref(&record->lock);
2573                         KKASSERT(record->lock.refs == 1);
2574                         record->flags |= HAMMER_RECF_DELETED_FE;
2575                         record->flags |= HAMMER_RECF_DELETED_BE;
2576                         hammer_rel_mem_record(record);
2577                 }
2578                 break;
2579         case HAMMER_INODE_ONDISK:
2580                 /*
2581                  * If already on-disk, do not set any additional flags.
2582                  */
2583                 break;
2584         default:
2585                 /*
2586                  * If not on-disk and not deleted, set DDIRTY to force
2587                  * an initial record to be written.
2588                  *
2589                  * Also set the create_tid in both the frontend and backend
2590                  * copy of the inode record.
2591                  */
2592                 ip->ino_leaf.base.create_tid = trans->tid;
2593                 ip->ino_leaf.create_ts = trans->time32;
2594                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2595                 ip->sync_ino_leaf.create_ts = trans->time32;
2596                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2597                 break;
2598         }
2599
2600         /*
2601          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2602          * is already on-disk the old record is marked as deleted.
2603          *
2604          * If DELETED is set hammer_update_inode() will delete the existing
2605          * record without writing out a new one.
2606          *
2607          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2608          */
2609         if (ip->flags & HAMMER_INODE_DELETED) {
2610                 error = hammer_update_inode(&cursor, ip);
2611         } else
2612         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2613             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2614                 error = hammer_update_itimes(&cursor, ip);
2615         } else
2616         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2617                 error = hammer_update_inode(&cursor, ip);
2618         }
2619 done:
2620         if (error) {
2621                 hammer_critical_error(ip->hmp, ip, error,
2622                                       "while syncing inode");
2623         }
2624         hammer_done_cursor(&cursor);
2625         return(error);
2626 }
2627
2628 /*
2629  * This routine is called when the OS is no longer actively referencing
2630  * the inode (but might still be keeping it cached), or when releasing
2631  * the last reference to an inode.
2632  *
2633  * At this point if the inode's nlinks count is zero we want to destroy
2634  * it, which may mean destroying it on-media too.
2635  */
2636 void
2637 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2638 {
2639         struct vnode *vp;
2640
2641         /*
2642          * Set the DELETING flag when the link count drops to 0 and the
2643          * OS no longer has any opens on the inode.
2644          *
2645          * The backend will clear DELETING (a mod flag) and set DELETED
2646          * (a state flag) when it is actually able to perform the
2647          * operation.
2648          */
2649         if (ip->ino_data.nlinks == 0 &&
2650             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2651                 ip->flags |= HAMMER_INODE_DELETING;
2652                 ip->flags |= HAMMER_INODE_TRUNCATED;
2653                 ip->trunc_off = 0;
2654                 vp = NULL;
2655                 if (getvp) {
2656                         if (hammer_get_vnode(ip, &vp) != 0)
2657                                 return;
2658                 }
2659
2660                 /*
2661                  * Final cleanup
2662                  */
2663                 if (ip->vp) {
2664                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2665                         vnode_pager_setsize(ip->vp, 0);
2666                 }
2667                 if (getvp) {
2668                         vput(vp);
2669                 }
2670         }
2671 }
2672
2673 /*
2674  * After potentially resolving a dependancy the inode is tested
2675  * to determine whether it needs to be reflushed.
2676  */
2677 void
2678 hammer_test_inode(hammer_inode_t ip)
2679 {
2680         if (ip->flags & HAMMER_INODE_REFLUSH) {
2681                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2682                 hammer_ref(&ip->lock);
2683                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2684                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2685                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2686                 } else {
2687                         hammer_flush_inode(ip, 0);
2688                 }
2689                 hammer_rel_inode(ip, 0);
2690         }
2691 }
2692
2693 /*
2694  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2695  * reassociated with a vp or just before it gets freed.
2696  *
2697  * Wakeup one thread blocked waiting on reclaims to complete.  Note that
2698  * the inode the thread is waiting on behalf of is a different inode then
2699  * the inode we are called with.  This is to create a pipeline.
2700  */
2701 static void
2702 hammer_inode_wakereclaims(hammer_inode_t ip)
2703 {
2704         struct hammer_reclaim *reclaim;
2705         hammer_mount_t hmp = ip->hmp;
2706
2707         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2708                 return;
2709
2710         --hammer_count_reclaiming;
2711         --hmp->inode_reclaims;
2712         ip->flags &= ~HAMMER_INODE_RECLAIM;
2713
2714         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
2715                 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2716                 reclaim->okydoky = 1;
2717                 wakeup(reclaim);
2718         }
2719 }
2720
2721 /*
2722  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
2723  * inodes build up before we start blocking.
2724  *
2725  * When we block we don't care *which* inode has finished reclaiming,
2726  * as lone as one does.  This is somewhat heuristical... we also put a
2727  * cap on how long we are willing to wait.
2728  */
2729 void
2730 hammer_inode_waitreclaims(hammer_mount_t hmp)
2731 {
2732         struct hammer_reclaim reclaim;
2733         int delay;
2734
2735         if (hmp->inode_reclaims > HAMMER_RECLAIM_WAIT) {
2736                 reclaim.okydoky = 0;
2737                 TAILQ_INSERT_TAIL(&hmp->reclaim_list,
2738                                   &reclaim, entry);
2739         } else {
2740                 reclaim.okydoky = 1;
2741         }
2742
2743         if (reclaim.okydoky == 0) {
2744                 delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2745                         HAMMER_RECLAIM_WAIT;
2746                 if (delay >= 0)
2747                         tsleep(&reclaim, 0, "hmrrcm", delay + 1);
2748                 if (reclaim.okydoky == 0)
2749                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2750         }
2751 }
2752