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 1.114 2008/09/24 00:53:51 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                 hammer_rel_inode(ip, 1);
 225         }
 226         return(0);
 227 }
 228
 229 /*
 230  * Return a locked vnode for the specified inode.  The inode must be
 231  * referenced but NOT LOCKED on entry and will remain referenced on
 232  * return.
 233  *
 234  * Called from the frontend.
 235  */
 236 int
 237 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 238 {
 239         hammer_mount_t hmp;
 240         struct vnode *vp;
 241         int error = 0;
 242         u_int8_t obj_type;
 243
 244         hmp = ip->hmp;
 245
 246         for (;;) {
 247                 if ((vp = ip->vp) == NULL) {
 248                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
 249                         if (error)
 250                                 break;
 251                         hammer_lock_ex(&ip->lock);
 252                         if (ip->vp != NULL) {
 253                                 hammer_unlock(&ip->lock);
 254                                 vp->v_type = VBAD;
 255                                 vx_put(vp);
 256                                 continue;
 257                         }
 258                         hammer_ref(&ip->lock);
 259                         vp = *vpp;
 260                         ip->vp = vp;
 261
 262                         obj_type = ip->ino_data.obj_type;
 263                         vp->v_type = hammer_get_vnode_type(obj_type);
 264
 265                         hammer_inode_wakereclaims(ip);
 266
 267                         switch(ip->ino_data.obj_type) {
 268                         case HAMMER_OBJTYPE_CDEV:
 269                         case HAMMER_OBJTYPE_BDEV:
 270                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
 271                                 addaliasu(vp, ip->ino_data.rmajor,
 272                                           ip->ino_data.rminor);
 273                                 break;
 274                         case HAMMER_OBJTYPE_FIFO:
 275                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
 276                                 break;
 277                         default:
 278                                 break;
 279                         }
 280
 281                         /*
 282                          * Only mark as the root vnode if the ip is not
 283                          * historical, otherwise the VFS cache will get
 284                          * confused.  The other half of the special handling
 285                          * is in hammer_vop_nlookupdotdot().
 286                          *
 287                          * Pseudo-filesystem roots can be accessed via
 288                          * non-root filesystem paths and setting VROOT may
 289                          * confuse the namecache.  Set VPFSROOT instead.
 290                          */
 291                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 292                             ip->obj_asof == hmp->asof) {
 293                                 if (ip->obj_localization == 0)
 294                                         vp->v_flag |= VROOT;
 295                                 else
 296                                         vp->v_flag |= VPFSROOT;
 297                         }
 298
 299                         vp->v_data = (void *)ip;
 300                         /* vnode locked by getnewvnode() */
 301                         /* make related vnode dirty if inode dirty? */
 302                         hammer_unlock(&ip->lock);
 303                         if (vp->v_type == VREG)
 304                                 vinitvmio(vp, ip->ino_data.size);
 305                         break;
 306                 }
 307
 308                 /*
 309                  * loop if the vget fails (aka races), or if the vp
 310                  * no longer matches ip->vp.
 311                  */
 312                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 313                         if (vp == ip->vp)
 314                                 break;
 315                         vput(vp);
 316                 }
 317         }
 318         *vpp = vp;
 319         return(error);
 320 }
 321
 322 /*
 323  * Locate all copies of the inode for obj_id compatible with the specified
 324  * asof, reference, and issue the related call-back.  This routine is used
 325  * for direct-io invalidation and does not create any new inodes.
 326  */
 327 void
 328 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
 329                             int (*callback)(hammer_inode_t ip, void *data),
 330                             void *data)
 331 {
 332         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
 333                                    hammer_inode_info_cmp_all_history,
 334                                    callback, iinfo);
 335 }
 336
 337 /*
 338  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 339  * do not attach or detach the related vnode (use hammer_get_vnode() for
 340  * that).
 341  *
 342  * The flags argument is only applied for newly created inodes, and only
 343  * certain flags are inherited.
 344  *
 345  * Called from the frontend.
 346  */
 347 struct hammer_inode *
 348 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
 349                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 350                  int flags, int *errorp)
 351 {
 352         hammer_mount_t hmp = trans->hmp;
 353         struct hammer_inode_info iinfo;
 354         struct hammer_cursor cursor;
 355         struct hammer_inode *ip;
 356
 357
 358         /*
 359          * Determine if we already have an inode cached.  If we do then
 360          * we are golden.
 361          */
 362         iinfo.obj_id = obj_id;
 363         iinfo.obj_asof = asof;
 364         iinfo.obj_localization = localization;
 365 loop:
 366         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 367         if (ip) {
 368                 hammer_ref(&ip->lock);
 369                 *errorp = 0;
 370                 return(ip);
 371         }
 372
 373         /*
 374          * Allocate a new inode structure and deal with races later.
 375          */
 376         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 377         ++hammer_count_inodes;
 378         ++hmp->count_inodes;
 379         ip->obj_id = obj_id;
 380         ip->obj_asof = iinfo.obj_asof;
 381         ip->obj_localization = localization;
 382         ip->hmp = hmp;
 383         ip->flags = flags & HAMMER_INODE_RO;
 384         ip->cache[0].ip = ip;
 385         ip->cache[1].ip = ip;
 386         if (hmp->ronly)
 387                 ip->flags |= HAMMER_INODE_RO;
 388         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 389                 0x7FFFFFFFFFFFFFFFLL;
 390         RB_INIT(&ip->rec_tree);
 391         TAILQ_INIT(&ip->target_list);
 392         hammer_ref(&ip->lock);
 393
 394         /*
 395          * Locate the on-disk inode.  If this is a PFS root we always
 396          * access the current version of the root inode and (if it is not
 397          * a master) always access information under it with a snapshot
 398          * TID.
 399          */
 400 retry:
 401         hammer_init_cursor(trans, &cursor, (dip ? &dip->cache[0] : NULL), NULL);
 402         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
 403         cursor.key_beg.obj_id = ip->obj_id;
 404         cursor.key_beg.key = 0;
 405         cursor.key_beg.create_tid = 0;
 406         cursor.key_beg.delete_tid = 0;
 407         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 408         cursor.key_beg.obj_type = 0;
 409
 410         cursor.asof = iinfo.obj_asof;
 411         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 412                        HAMMER_CURSOR_ASOF;
 413
 414         *errorp = hammer_btree_lookup(&cursor);
 415         if (*errorp == EDEADLK) {
 416                 hammer_done_cursor(&cursor);
 417                 goto retry;
 418         }
 419
 420         /*
 421          * On success the B-Tree lookup will hold the appropriate
 422          * buffer cache buffers and provide a pointer to the requested
 423          * information.  Copy the information to the in-memory inode
 424          * and cache the B-Tree node to improve future operations.
 425          */
 426         if (*errorp == 0) {
 427                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 428                 ip->ino_data = cursor.data->inode;
 429
 430                 /*
 431                  * cache[0] tries to cache the location of the object inode.
 432                  * The assumption is that it is near the directory inode.
 433                  *
 434                  * cache[1] tries to cache the location of the object data.
 435                  * The assumption is that it is near the directory data.
 436                  */
 437                 hammer_cache_node(&ip->cache[0], cursor.node);
 438                 if (dip && dip->cache[1].node)
 439                         hammer_cache_node(&ip->cache[1], dip->cache[1].node);
 440
 441                 /*
 442                  * The file should not contain any data past the file size
 443                  * stored in the inode.  Setting save_trunc_off to the
 444                  * file size instead of max reduces B-Tree lookup overheads
 445                  * on append by allowing the flusher to avoid checking for
 446                  * record overwrites.
 447                  */
 448                 ip->save_trunc_off = ip->ino_data.size;
 449
 450                 /*
 451                  * Locate and assign the pseudofs management structure to
 452                  * the inode.
 453                  */
 454                 if (dip && dip->obj_localization == ip->obj_localization) {
 455                         ip->pfsm = dip->pfsm;
 456                         hammer_ref(&ip->pfsm->lock);
 457                 } else {
 458                         ip->pfsm = hammer_load_pseudofs(trans,
 459                                                         ip->obj_localization,
 460                                                         errorp);
 461                         *errorp = 0;    /* ignore ENOENT */
 462                 }
 463         }
 464
 465         /*
 466          * The inode is placed on the red-black tree and will be synced to
 467          * the media when flushed or by the filesystem sync.  If this races
 468          * another instantiation/lookup the insertion will fail.
 469          */
 470         if (*errorp == 0) {
 471                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 472                         hammer_free_inode(ip);
 473                         hammer_done_cursor(&cursor);
 474                         goto loop;
 475                 }
 476                 ip->flags |= HAMMER_INODE_ONDISK;
 477         } else {
 478                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 479                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 480                         --hmp->rsv_inodes;
 481                 }
 482
 483                 hammer_free_inode(ip);
 484                 ip = NULL;
 485         }
 486         hammer_done_cursor(&cursor);
 487         trans->flags |= HAMMER_TRANSF_NEWINODE;
 488         return (ip);
 489 }
 490
 491 /*
 492  * Create a new filesystem object, returning the inode in *ipp.  The
 493  * returned inode will be referenced.  The inode is created in-memory.
 494  *
 495  * If pfsm is non-NULL the caller wishes to create the root inode for
 496  * a master PFS.
 497  */
 498 int
 499 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 500                     struct ucred *cred, hammer_inode_t dip,
 501                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
 502 {
 503         hammer_mount_t hmp;
 504         hammer_inode_t ip;
 505         uid_t xuid;
 506         int error;
 507
 508         hmp = trans->hmp;
 509
 510         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 511         ++hammer_count_inodes;
 512         ++hmp->count_inodes;
 513
 514         if (pfsm) {
 515                 KKASSERT(pfsm->localization != 0);
 516                 ip->obj_id = HAMMER_OBJID_ROOT;
 517                 ip->obj_localization = pfsm->localization;
 518         } else {
 519                 KKASSERT(dip != NULL);
 520                 ip->obj_id = hammer_alloc_objid(hmp, dip);
 521                 ip->obj_localization = dip->obj_localization;
 522         }
 523
 524         KKASSERT(ip->obj_id != 0);
 525         ip->obj_asof = hmp->asof;
 526         ip->hmp = hmp;
 527         ip->flush_state = HAMMER_FST_IDLE;
 528         ip->flags = HAMMER_INODE_DDIRTY |
 529                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
 530         ip->cache[0].ip = ip;
 531         ip->cache[1].ip = ip;
 532
 533         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 534         /* ip->save_trunc_off = 0; (already zero) */
 535         RB_INIT(&ip->rec_tree);
 536         TAILQ_INIT(&ip->target_list);
 537
 538         ip->ino_data.atime = trans->time;
 539         ip->ino_data.mtime = trans->time;
 540         ip->ino_data.size = 0;
 541         ip->ino_data.nlinks = 0;
 542
 543         /*
 544          * A nohistory designator on the parent directory is inherited by
 545          * the child.  We will do this even for pseudo-fs creation... the
 546          * sysad can turn it off.
 547          */
 548         if (dip) {
 549                 ip->ino_data.uflags = dip->ino_data.uflags &
 550                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 551         }
 552
 553         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 554         ip->ino_leaf.base.localization = ip->obj_localization +
 555                                          HAMMER_LOCALIZE_INODE;
 556         ip->ino_leaf.base.obj_id = ip->obj_id;
 557         ip->ino_leaf.base.key = 0;
 558         ip->ino_leaf.base.create_tid = 0;
 559         ip->ino_leaf.base.delete_tid = 0;
 560         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 561         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 562
 563         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 564         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 565         ip->ino_data.mode = vap->va_mode;
 566         ip->ino_data.ctime = trans->time;
 567
 568         /*
 569          * If we are running version 2 or greater we use dirhash algorithm #1
 570          * which is semi-sorted.  Algorithm #0 was just a pure crc.
 571          */
 572         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
 573                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
 574                         ip->ino_data.cap_flags |= HAMMER_INODE_CAP_DIRHASH_ALG1;
 575                 }
 576         }
 577
 578         /*
 579          * Setup the ".." pointer.  This only needs to be done for directories
 580          * but we do it for all objects as a recovery aid.
 581          */
 582         if (dip)
 583                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
 584 #if 0
 585         /*
 586          * The parent_obj_localization field only applies to pseudo-fs roots.
 587          * XXX this is no longer applicable, PFSs are no longer directly
 588          * tied into the parent's directory structure.
 589          */
 590         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
 591             ip->obj_id == HAMMER_OBJID_ROOT) {
 592                 ip->ino_data.ext.obj.parent_obj_localization =
 593                                                 dip->obj_localization;
 594         }
 595 #endif
 596
 597         switch(ip->ino_leaf.base.obj_type) {
 598         case HAMMER_OBJTYPE_CDEV:
 599         case HAMMER_OBJTYPE_BDEV:
 600                 ip->ino_data.rmajor = vap->va_rmajor;
 601                 ip->ino_data.rminor = vap->va_rminor;
 602                 break;
 603         default:
 604                 break;
 605         }
 606
 607         /*
 608          * Calculate default uid/gid and overwrite with information from
 609          * the vap.
 610          */
 611         if (dip) {
 612                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 613                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
 614                                              xuid, cred, &vap->va_mode);
 615         } else {
 616                 xuid = 0;
 617         }
 618         ip->ino_data.mode = vap->va_mode;
 619
 620         if (vap->va_vaflags & VA_UID_UUID_VALID)
 621                 ip->ino_data.uid = vap->va_uid_uuid;
 622         else if (vap->va_uid != (uid_t)VNOVAL)
 623                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 624         else
 625                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 626
 627         if (vap->va_vaflags & VA_GID_UUID_VALID)
 628                 ip->ino_data.gid = vap->va_gid_uuid;
 629         else if (vap->va_gid != (gid_t)VNOVAL)
 630                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 631         else if (dip)
 632                 ip->ino_data.gid = dip->ino_data.gid;
 633
 634         hammer_ref(&ip->lock);
 635
 636         if (pfsm) {
 637                 ip->pfsm = pfsm;
 638                 hammer_ref(&pfsm->lock);
 639                 error = 0;
 640         } else if (dip->obj_localization == ip->obj_localization) {
 641                 ip->pfsm = dip->pfsm;
 642                 hammer_ref(&ip->pfsm->lock);
 643                 error = 0;
 644         } else {
 645                 ip->pfsm = hammer_load_pseudofs(trans,
 646                                                 ip->obj_localization,
 647                                                 &error);
 648                 error = 0;      /* ignore ENOENT */
 649         }
 650
 651         if (error) {
 652                 hammer_free_inode(ip);
 653                 ip = NULL;
 654         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 655                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 656                 /* not reached */
 657                 hammer_free_inode(ip);
 658         }
 659         *ipp = ip;
 660         return(error);
 661 }
 662
 663 /*
 664  * Final cleanup / freeing of an inode structure
 665  */
 666 static void
 667 hammer_free_inode(hammer_inode_t ip)
 668 {
 669         struct hammer_mount *hmp;
 670
 671         hmp = ip->hmp;
 672         KKASSERT(ip->lock.refs == 1);
 673         hammer_uncache_node(&ip->cache[0]);
 674         hammer_uncache_node(&ip->cache[1]);
 675         hammer_inode_wakereclaims(ip);
 676         if (ip->objid_cache)
 677                 hammer_clear_objid(ip);
 678         --hammer_count_inodes;
 679         --hmp->count_inodes;
 680         if (ip->pfsm) {
 681                 hammer_rel_pseudofs(hmp, ip->pfsm);
 682                 ip->pfsm = NULL;
 683         }
 684         kfree(ip, hmp->m_inodes);
 685         ip = NULL;
 686 }
 687
 688 /*
 689  * Retrieve pseudo-fs data.  NULL will never be returned.
 690  *
 691  * If an error occurs *errorp will be set and a default template is returned,
 692  * otherwise *errorp is set to 0.  Typically when an error occurs it will
 693  * be ENOENT.
 694  */
 695 hammer_pseudofs_inmem_t
 696 hammer_load_pseudofs(hammer_transaction_t trans,
 697                      u_int32_t localization, int *errorp)
 698 {
 699         hammer_mount_t hmp = trans->hmp;
 700         hammer_inode_t ip;
 701         hammer_pseudofs_inmem_t pfsm;
 702         struct hammer_cursor cursor;
 703         int bytes;
 704
 705 retry:
 706         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
 707         if (pfsm) {
 708                 hammer_ref(&pfsm->lock);
 709                 *errorp = 0;
 710                 return(pfsm);
 711         }
 712
 713         /*
 714          * PFS records are stored in the root inode (not the PFS root inode,
 715          * but the real root).  Avoid an infinite recursion if loading
 716          * the PFS for the real root.
 717          */
 718         if (localization) {
 719                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
 720                                       HAMMER_MAX_TID,
 721                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
 722         } else {
 723                 ip = NULL;
 724         }
 725
 726         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
 727         pfsm->localization = localization;
 728         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
 729         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
 730
 731         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
 732         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
 733                                       HAMMER_LOCALIZE_MISC;
 734         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 735         cursor.key_beg.create_tid = 0;
 736         cursor.key_beg.delete_tid = 0;
 737         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 738         cursor.key_beg.obj_type = 0;
 739         cursor.key_beg.key = localization;
 740         cursor.asof = HAMMER_MAX_TID;
 741         cursor.flags |= HAMMER_CURSOR_ASOF;
 742
 743         if (ip)
 744                 *errorp = hammer_ip_lookup(&cursor);
 745         else
 746                 *errorp = hammer_btree_lookup(&cursor);
 747         if (*errorp == 0) {
 748                 *errorp = hammer_ip_resolve_data(&cursor);
 749                 if (*errorp == 0) {
 750                         if (cursor.data->pfsd.mirror_flags &
 751                             HAMMER_PFSD_DELETED) {
 752                                 *errorp = ENOENT;
 753                         } else {
 754                                 bytes = cursor.leaf->data_len;
 755                                 if (bytes > sizeof(pfsm->pfsd))
 756                                         bytes = sizeof(pfsm->pfsd);
 757                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
 758                         }
 759                 }
 760         }
 761         hammer_done_cursor(&cursor);
 762
 763         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 764         hammer_ref(&pfsm->lock);
 765         if (ip)
 766                 hammer_rel_inode(ip, 0);
 767         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
 768                 kfree(pfsm, hmp->m_misc);
 769                 goto retry;
 770         }
 771         return(pfsm);
 772 }
 773
 774 /*
 775  * Store pseudo-fs data.  The backend will automatically delete any prior
 776  * on-disk pseudo-fs data but we have to delete in-memory versions.
 777  */
 778 int
 779 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
 780 {
 781         struct hammer_cursor cursor;
 782         hammer_record_t record;
 783         hammer_inode_t ip;
 784         int error;
 785
 786         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 787                               HAMMER_DEF_LOCALIZATION, 0, &error);
 788 retry:
 789         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 790         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
 791         cursor.key_beg.localization = ip->obj_localization +
 792                                       HAMMER_LOCALIZE_MISC;
 793         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 794         cursor.key_beg.create_tid = 0;
 795         cursor.key_beg.delete_tid = 0;
 796         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 797         cursor.key_beg.obj_type = 0;
 798         cursor.key_beg.key = pfsm->localization;
 799         cursor.asof = HAMMER_MAX_TID;
 800         cursor.flags |= HAMMER_CURSOR_ASOF;
 801
 802         error = hammer_ip_lookup(&cursor);
 803         if (error == 0 && hammer_cursor_inmem(&cursor)) {
 804                 record = cursor.iprec;
 805                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 806                         KKASSERT(cursor.deadlk_rec == NULL);
 807                         hammer_ref(&record->lock);
 808                         cursor.deadlk_rec = record;
 809                         error = EDEADLK;
 810                 } else {
 811                         record->flags |= HAMMER_RECF_DELETED_FE;
 812                         error = 0;
 813                 }
 814         }
 815         if (error == 0 || error == ENOENT) {
 816                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
 817                 record->type = HAMMER_MEM_RECORD_GENERAL;
 818
 819                 record->leaf.base.localization = ip->obj_localization +
 820                                                  HAMMER_LOCALIZE_MISC;
 821                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
 822                 record->leaf.base.key = pfsm->localization;
 823                 record->leaf.data_len = sizeof(pfsm->pfsd);
 824                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
 825                 error = hammer_ip_add_record(trans, record);
 826         }
 827         hammer_done_cursor(&cursor);
 828         if (error == EDEADLK)
 829                 goto retry;
 830         hammer_rel_inode(ip, 0);
 831         return(error);
 832 }
 833
 834 /*
 835  * Create a root directory for a PFS if one does not alredy exist.
 836  *
 837  * The PFS root stands alone so we must also bump the nlinks count
 838  * to prevent it from being destroyed on release.
 839  */
 840 int
 841 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
 842                        hammer_pseudofs_inmem_t pfsm)
 843 {
 844         hammer_inode_t ip;
 845         struct vattr vap;
 846         int error;
 847
 848         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 849                               pfsm->localization, 0, &error);
 850         if (ip == NULL) {
 851                 vattr_null(&vap);
 852                 vap.va_mode = 0755;
 853                 vap.va_type = VDIR;
 854                 error = hammer_create_inode(trans, &vap, cred, NULL, pfsm, &ip);
 855                 if (error == 0) {
 856                         ++ip->ino_data.nlinks;
 857                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 858                 }
 859         }
 860         if (ip)
 861                 hammer_rel_inode(ip, 0);
 862         return(error);
 863 }
 864
 865 /*
 866  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
 867  * if we are unable to disassociate all the inodes.
 868  */
 869 static
 870 int
 871 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
 872 {
 873         int res;
 874
 875         hammer_ref(&ip->lock);
 876         if (ip->lock.refs == 2 && ip->vp)
 877                 vclean_unlocked(ip->vp);
 878         if (ip->lock.refs == 1 && ip->vp == NULL)
 879                 res = 0;
 880         else
 881                 res = -1;       /* stop, someone is using the inode */
 882         hammer_rel_inode(ip, 0);
 883         return(res);
 884 }
 885
 886 int
 887 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
 888 {
 889         int res;
 890         int try;
 891
 892         for (try = res = 0; try < 4; ++try) {
 893                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
 894                                            hammer_inode_pfs_cmp,
 895                                            hammer_unload_pseudofs_callback,
 896                                            &localization);
 897                 if (res == 0 && try > 1)
 898                         break;
 899                 hammer_flusher_sync(trans->hmp);
 900         }
 901         if (res != 0)
 902                 res = ENOTEMPTY;
 903         return(res);
 904 }
 905
 906
 907 /*
 908  * Release a reference on a PFS
 909  */
 910 void
 911 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
 912 {
 913         hammer_unref(&pfsm->lock);
 914         if (pfsm->lock.refs == 0) {
 915                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
 916                 kfree(pfsm, hmp->m_misc);
 917         }
 918 }
 919
 920 /*
 921  * Called by hammer_sync_inode().
 922  */
 923 static int
 924 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 925 {
 926         hammer_transaction_t trans = cursor->trans;
 927         hammer_record_t record;
 928         int error;
 929         int redirty;
 930
 931 retry:
 932         error = 0;
 933
 934         /*
 935          * If the inode has a presence on-disk then locate it and mark
 936          * it deleted, setting DELONDISK.
 937          *
 938          * The record may or may not be physically deleted, depending on
 939          * the retention policy.
 940          */
 941         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 942             HAMMER_INODE_ONDISK) {
 943                 hammer_normalize_cursor(cursor);
 944                 cursor->key_beg.localization = ip->obj_localization +
 945                                                HAMMER_LOCALIZE_INODE;
 946                 cursor->key_beg.obj_id = ip->obj_id;
 947                 cursor->key_beg.key = 0;
 948                 cursor->key_beg.create_tid = 0;
 949                 cursor->key_beg.delete_tid = 0;
 950                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 951                 cursor->key_beg.obj_type = 0;
 952                 cursor->asof = ip->obj_asof;
 953                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 954                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 955                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 956
 957                 error = hammer_btree_lookup(cursor);
 958                 if (hammer_debug_inode)
 959                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 960
 961                 if (error == 0) {
 962                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
 963                         if (hammer_debug_inode)
 964                                 kprintf(" error %d\n", error);
 965                         if (error == 0) {
 966                                 ip->flags |= HAMMER_INODE_DELONDISK;
 967                         }
 968                         if (cursor->node)
 969                                 hammer_cache_node(&ip->cache[0], cursor->node);
 970                 }
 971                 if (error == EDEADLK) {
 972                         hammer_done_cursor(cursor);
 973                         error = hammer_init_cursor(trans, cursor,
 974                                                    &ip->cache[0], ip);
 975                         if (hammer_debug_inode)
 976                                 kprintf("IPDED %p %d\n", ip, error);
 977                         if (error == 0)
 978                                 goto retry;
 979                 }
 980         }
 981
 982         /*
 983          * Ok, write out the initial record or a new record (after deleting
 984          * the old one), unless the DELETED flag is set.  This routine will
 985          * clear DELONDISK if it writes out a record.
 986          *
 987          * Update our inode statistics if this is the first application of
 988          * the inode on-disk.
 989          */
 990         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 991                 /*
 992                  * Generate a record and write it to the media.  We clean-up
 993                  * the state before releasing so we do not have to set-up
 994                  * a flush_group.
 995                  */
 996                 record = hammer_alloc_mem_record(ip, 0);
 997                 record->type = HAMMER_MEM_RECORD_INODE;
 998                 record->flush_state = HAMMER_FST_FLUSH;
 999                 record->leaf = ip->sync_ino_leaf;
1000                 record->leaf.base.create_tid = trans->tid;
1001                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1002                 record->leaf.create_ts = trans->time32;
1003                 record->data = (void *)&ip->sync_ino_data;
1004                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1005
1006                 /*
1007                  * If this flag is set we cannot sync the new file size
1008                  * because we haven't finished related truncations.  The
1009                  * inode will be flushed in another flush group to finish
1010                  * the job.
1011                  */
1012                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1013                     ip->sync_ino_data.size != ip->ino_data.size) {
1014                         redirty = 1;
1015                         ip->sync_ino_data.size = ip->ino_data.size;
1016                 } else {
1017                         redirty = 0;
1018                 }
1019
1020                 for (;;) {
1021                         error = hammer_ip_sync_record_cursor(cursor, record);
1022                         if (hammer_debug_inode)
1023                                 kprintf("GENREC %p rec %08x %d\n",
1024                                         ip, record->flags, error);
1025                         if (error != EDEADLK)
1026                                 break;
1027                         hammer_done_cursor(cursor);
1028                         error = hammer_init_cursor(trans, cursor,
1029                                                    &ip->cache[0], ip);
1030                         if (hammer_debug_inode)
1031                                 kprintf("GENREC reinit %d\n", error);
1032                         if (error)
1033                                 break;
1034                 }
1035
1036                 /*
1037                  * The record isn't managed by the inode's record tree,
1038                  * destroy it whether we succeed or fail.
1039                  */
1040                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1041                 record->flags |= HAMMER_RECF_DELETED_FE | HAMMER_RECF_COMMITTED;
1042                 record->flush_state = HAMMER_FST_IDLE;
1043                 hammer_rel_mem_record(record);
1044
1045                 /*
1046                  * Finish up.
1047                  */
1048                 if (error == 0) {
1049                         if (hammer_debug_inode)
1050                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1051                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1052                                             HAMMER_INODE_ATIME |
1053                                             HAMMER_INODE_MTIME);
1054                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1055                         if (redirty)
1056                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1057
1058                         /*
1059                          * Root volume count of inodes
1060                          */
1061                         hammer_sync_lock_sh(trans);
1062                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1063                                 hammer_modify_volume_field(trans,
1064                                                            trans->rootvol,
1065                                                            vol0_stat_inodes);
1066                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1067                                 hammer_modify_volume_done(trans->rootvol);
1068                                 ip->flags |= HAMMER_INODE_ONDISK;
1069                                 if (hammer_debug_inode)
1070                                         kprintf("NOWONDISK %p\n", ip);
1071                         }
1072                         hammer_sync_unlock(trans);
1073                 }
1074         }
1075
1076         /*
1077          * If the inode has been destroyed, clean out any left-over flags
1078          * that may have been set by the frontend.
1079          */
1080         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1081                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1082                                     HAMMER_INODE_ATIME |
1083                                     HAMMER_INODE_MTIME);
1084         }
1085         return(error);
1086 }
1087
1088 /*
1089  * Update only the itimes fields.
1090  *
1091  * ATIME can be updated without generating any UNDO.  MTIME is updated
1092  * with UNDO so it is guaranteed to be synchronized properly in case of
1093  * a crash.
1094  *
1095  * Neither field is included in the B-Tree leaf element's CRC, which is how
1096  * we can get away with updating ATIME the way we do.
1097  */
1098 static int
1099 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1100 {
1101         hammer_transaction_t trans = cursor->trans;
1102         int error;
1103
1104 retry:
1105         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1106             HAMMER_INODE_ONDISK) {
1107                 return(0);
1108         }
1109
1110         hammer_normalize_cursor(cursor);
1111         cursor->key_beg.localization = ip->obj_localization +
1112                                        HAMMER_LOCALIZE_INODE;
1113         cursor->key_beg.obj_id = ip->obj_id;
1114         cursor->key_beg.key = 0;
1115         cursor->key_beg.create_tid = 0;
1116         cursor->key_beg.delete_tid = 0;
1117         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1118         cursor->key_beg.obj_type = 0;
1119         cursor->asof = ip->obj_asof;
1120         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1121         cursor->flags |= HAMMER_CURSOR_ASOF;
1122         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1123         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1124         cursor->flags |= HAMMER_CURSOR_BACKEND;
1125
1126         error = hammer_btree_lookup(cursor);
1127         if (error == 0) {
1128                 hammer_cache_node(&ip->cache[0], cursor->node);
1129                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1130                         /*
1131                          * Updating MTIME requires an UNDO.  Just cover
1132                          * both atime and mtime.
1133                          */
1134                         hammer_sync_lock_sh(trans);
1135                         hammer_modify_buffer(trans, cursor->data_buffer,
1136                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1137                                      HAMMER_ITIMES_BYTES);
1138                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1139                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1140                         hammer_modify_buffer_done(cursor->data_buffer);
1141                         hammer_sync_unlock(trans);
1142                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1143                         /*
1144                          * Updating atime only can be done in-place with
1145                          * no UNDO.
1146                          */
1147                         hammer_sync_lock_sh(trans);
1148                         hammer_modify_buffer(trans, cursor->data_buffer,
1149                                              NULL, 0);
1150                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1151                         hammer_modify_buffer_done(cursor->data_buffer);
1152                         hammer_sync_unlock(trans);
1153                 }
1154                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1155         }
1156         if (error == EDEADLK) {
1157                 hammer_done_cursor(cursor);
1158                 error = hammer_init_cursor(trans, cursor,
1159                                            &ip->cache[0], ip);
1160                 if (error == 0)
1161                         goto retry;
1162         }
1163         return(error);
1164 }
1165
1166 /*
1167  * Release a reference on an inode, flush as requested.
1168  *
1169  * On the last reference we queue the inode to the flusher for its final
1170  * disposition.
1171  */
1172 void
1173 hammer_rel_inode(struct hammer_inode *ip, int flush)
1174 {
1175         /*hammer_mount_t hmp = ip->hmp;*/
1176
1177         /*
1178          * Handle disposition when dropping the last ref.
1179          */
1180         for (;;) {
1181                 if (ip->lock.refs == 1) {
1182                         /*
1183                          * Determine whether on-disk action is needed for
1184                          * the inode's final disposition.
1185                          */
1186                         KKASSERT(ip->vp == NULL);
1187                         hammer_inode_unloadable_check(ip, 0);
1188                         if (ip->flags & HAMMER_INODE_MODMASK) {
1189                                 hammer_flush_inode(ip, 0);
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), hmp->m_misc, 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          * If the flush group reaches the autoflush limit we want to signal
1670          * the flusher.  This is particularly important for remove()s.
1671          */
1672         if (flg->total_count == hammer_autoflush)
1673                 flags |= HAMMER_FLUSH_SIGNAL;
1674
1675         /*
1676          * We need to be able to vfsync/truncate from the backend.
1677          */
1678         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1679         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1680                 ip->flags |= HAMMER_INODE_VHELD;
1681                 vref(ip->vp);
1682         }
1683
1684         /*
1685          * Figure out how many in-memory records we can actually flush
1686          * (not including inode meta-data, buffers, etc).
1687          */
1688         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1689         if (flags & HAMMER_FLUSH_RECURSION) {
1690                 /*
1691                  * If this is a upwards recursion we do not want to
1692                  * recurse down again!
1693                  */
1694                 go_count = 1;
1695 #if 0
1696         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1697                 /*
1698                  * No new records are added if we must complete a flush
1699                  * from a previous cycle, but we do have to move the records
1700                  * from the previous cycle to the current one.
1701                  */
1702 #if 0
1703                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1704                                    hammer_syncgrp_child_callback, NULL);
1705 #endif
1706                 go_count = 1;
1707 #endif
1708         } else {
1709                 /*
1710                  * Normal flush, scan records and bring them into the flush.
1711                  * Directory adds and deletes are usually skipped (they are
1712                  * grouped with the related inode rather then with the
1713                  * directory).
1714                  *
1715                  * go_count can be negative, which means the scan aborted
1716                  * due to the flush group being over-full and we should
1717                  * flush what we have.
1718                  */
1719                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1720                                    hammer_setup_child_callback, NULL);
1721         }
1722
1723         /*
1724          * This is a more involved test that includes go_count.  If we
1725          * can't flush, flag the inode and return.  If go_count is 0 we
1726          * were are unable to flush any records in our rec_tree and
1727          * must ignore the XDIRTY flag.
1728          */
1729         if (go_count == 0) {
1730                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1731                         --ip->hmp->count_iqueued;
1732                         --hammer_count_iqueued;
1733
1734                         --flg->total_count;
1735                         ip->flush_state = HAMMER_FST_SETUP;
1736                         ip->flush_group = NULL;
1737                         if (ip->flags & HAMMER_INODE_VHELD) {
1738                                 ip->flags &= ~HAMMER_INODE_VHELD;
1739                                 vrele(ip->vp);
1740                         }
1741
1742                         /*
1743                          * REFLUSH is needed to trigger dependancy wakeups
1744                          * when an inode is in SETUP.
1745                          */
1746                         ip->flags |= HAMMER_INODE_REFLUSH;
1747                         if (flags & HAMMER_FLUSH_SIGNAL) {
1748                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1749                                 hammer_flusher_async(ip->hmp, flg);
1750                         }
1751                         if (--ip->hmp->flusher.group_lock == 0)
1752                                 wakeup(&ip->hmp->flusher.group_lock);
1753                         return;
1754                 }
1755         }
1756
1757         /*
1758          * Snapshot the state of the inode for the backend flusher.
1759          *
1760          * We continue to retain save_trunc_off even when all truncations
1761          * have been resolved as an optimization to determine if we can
1762          * skip the B-Tree lookup for overwrite deletions.
1763          *
1764          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1765          * and stays in ip->flags.  Once set, it stays set until the
1766          * inode is destroyed.
1767          */
1768         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1769                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1770                 ip->sync_trunc_off = ip->trunc_off;
1771                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1772                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1773                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1774
1775                 /*
1776                  * The save_trunc_off used to cache whether the B-Tree
1777                  * holds any records past that point is not used until
1778                  * after the truncation has succeeded, so we can safely
1779                  * set it now.
1780                  */
1781                 if (ip->save_trunc_off > ip->sync_trunc_off)
1782                         ip->save_trunc_off = ip->sync_trunc_off;
1783         }
1784         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1785                            ~HAMMER_INODE_TRUNCATED);
1786         ip->sync_ino_leaf = ip->ino_leaf;
1787         ip->sync_ino_data = ip->ino_data;
1788         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1789 #ifdef DEBUG_TRUNCATE
1790         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1791                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1792 #endif
1793
1794         /*
1795          * The flusher list inherits our inode and reference.
1796          */
1797         KKASSERT(flg->running == 0);
1798         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
1799         if (--ip->hmp->flusher.group_lock == 0)
1800                 wakeup(&ip->hmp->flusher.group_lock);
1801
1802         if (flags & HAMMER_FLUSH_SIGNAL) {
1803                 hammer_flusher_async(ip->hmp, flg);
1804         }
1805 }
1806
1807 /*
1808  * Callback for scan of ip->rec_tree.  Try to include each record in our
1809  * flush.  ip->flush_group has been set but the inode has not yet been
1810  * moved into a flushing state.
1811  *
1812  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1813  * both inodes.
1814  *
1815  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1816  * the caller from shortcutting the flush.
1817  */
1818 static int
1819 hammer_setup_child_callback(hammer_record_t rec, void *data)
1820 {
1821         hammer_flush_group_t flg;
1822         hammer_inode_t target_ip;
1823         hammer_inode_t ip;
1824         int r;
1825
1826         /*
1827          * Deleted records are ignored.  Note that the flush detects deleted
1828          * front-end records at multiple points to deal with races.  This is
1829          * just the first line of defense.  The only time DELETED_FE cannot
1830          * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1831          *
1832          * Don't get confused between record deletion and, say, directory
1833          * entry deletion.  The deletion of a directory entry that is on
1834          * the media has nothing to do with the record deletion flags.
1835          */
1836         if (rec->flags & (HAMMER_RECF_DELETED_FE|HAMMER_RECF_DELETED_BE)) {
1837                 if (rec->flush_state == HAMMER_FST_FLUSH) {
1838                         KKASSERT(rec->flush_group == rec->ip->flush_group);
1839                         r = 1;
1840                 } else {
1841                         r = 0;
1842                 }
1843                 return(r);
1844         }
1845
1846         /*
1847          * If the record is in an idle state it has no dependancies and
1848          * can be flushed.
1849          */
1850         ip = rec->ip;
1851         flg = ip->flush_group;
1852         r = 0;
1853
1854         switch(rec->flush_state) {
1855         case HAMMER_FST_IDLE:
1856                 /*
1857                  * The record has no setup dependancy, we can flush it.
1858                  */
1859                 KKASSERT(rec->target_ip == NULL);
1860                 rec->flush_state = HAMMER_FST_FLUSH;
1861                 rec->flush_group = flg;
1862                 ++flg->refs;
1863                 hammer_ref(&rec->lock);
1864                 r = 1;
1865                 break;
1866         case HAMMER_FST_SETUP:
1867                 /*
1868                  * The record has a setup dependancy.  These are typically
1869                  * directory entry adds and deletes.  Such entries will be
1870                  * flushed when their inodes are flushed so we do not
1871                  * usually have to add them to the flush here.  However,
1872                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
1873                  * it is asking us to flush this record (and it).
1874                  */
1875                 target_ip = rec->target_ip;
1876                 KKASSERT(target_ip != NULL);
1877                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1878
1879                 /*
1880                  * If the target IP is already flushing in our group
1881                  * we could associate the record, but target_ip has
1882                  * already synced ino_data to sync_ino_data and we
1883                  * would also have to adjust nlinks.   Plus there are
1884                  * ordering issues for adds and deletes.
1885                  *
1886                  * Reflush downward if this is an ADD, and upward if
1887                  * this is a DEL.
1888                  */
1889                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1890                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
1891                                 ip->flags |= HAMMER_INODE_REFLUSH;
1892                         else
1893                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1894                         break;
1895                 }
1896
1897                 /*
1898                  * Target IP is not yet flushing.  This can get complex
1899                  * because we have to be careful about the recursion.
1900                  *
1901                  * Directories create an issue for us in that if a flush
1902                  * of a directory is requested the expectation is to flush
1903                  * any pending directory entries, but this will cause the
1904                  * related inodes to recursively flush as well.  We can't
1905                  * really defer the operation so just get as many as we
1906                  * can and
1907                  */
1908 #if 0
1909                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
1910                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
1911                         /*
1912                          * We aren't reclaiming and the target ip was not
1913                          * previously prevented from flushing due to this
1914                          * record dependancy.  Do not flush this record.
1915                          */
1916                         /*r = 0;*/
1917                 } else
1918 #endif
1919                 if (flg->total_count + flg->refs >
1920                            ip->hmp->undo_rec_limit) {
1921                         /*
1922                          * Our flush group is over-full and we risk blowing
1923                          * out the UNDO FIFO.  Stop the scan, flush what we
1924                          * have, then reflush the directory.
1925                          *
1926                          * The directory may be forced through multiple
1927                          * flush groups before it can be completely
1928                          * flushed.
1929                          */
1930                         ip->flags |= HAMMER_INODE_RESIGNAL |
1931                                      HAMMER_INODE_REFLUSH;
1932                         r = -1;
1933                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1934                         /*
1935                          * If the target IP is not flushing we can force
1936                          * it to flush, even if it is unable to write out
1937                          * any of its own records we have at least one in
1938                          * hand that we CAN deal with.
1939                          */
1940                         rec->flush_state = HAMMER_FST_FLUSH;
1941                         rec->flush_group = flg;
1942                         ++flg->refs;
1943                         hammer_ref(&rec->lock);
1944                         hammer_flush_inode_core(target_ip, flg,
1945                                                 HAMMER_FLUSH_RECURSION);
1946                         r = 1;
1947                 } else {
1948                         /*
1949                          * General or delete-on-disk record.
1950                          *
1951                          * XXX this needs help.  If a delete-on-disk we could
1952                          * disconnect the target.  If the target has its own
1953                          * dependancies they really need to be flushed.
1954                          *
1955                          * XXX
1956                          */
1957                         rec->flush_state = HAMMER_FST_FLUSH;
1958                         rec->flush_group = flg;
1959                         ++flg->refs;
1960                         hammer_ref(&rec->lock);
1961                         hammer_flush_inode_core(target_ip, flg,
1962                                                 HAMMER_FLUSH_RECURSION);
1963                         r = 1;
1964                 }
1965                 break;
1966         case HAMMER_FST_FLUSH:
1967                 /*
1968                  * The flush_group should already match.
1969                  */
1970                 KKASSERT(rec->flush_group == flg);
1971                 r = 1;
1972                 break;
1973         }
1974         return(r);
1975 }
1976
1977 #if 0
1978 /*
1979  * This version just moves records already in a flush state to the new
1980  * flush group and that is it.
1981  */
1982 static int
1983 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
1984 {
1985         hammer_inode_t ip = rec->ip;
1986
1987         switch(rec->flush_state) {
1988         case HAMMER_FST_FLUSH:
1989                 KKASSERT(rec->flush_group == ip->flush_group);
1990                 break;
1991         default:
1992                 break;
1993         }
1994         return(0);
1995 }
1996 #endif
1997
1998 /*
1999  * Wait for a previously queued flush to complete.
2000  *
2001  * If a critical error occured we don't try to wait.
2002  */
2003 void
2004 hammer_wait_inode(hammer_inode_t ip)
2005 {
2006         hammer_flush_group_t flg;
2007
2008         flg = NULL;
2009         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2010                 while (ip->flush_state != HAMMER_FST_IDLE &&
2011                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2012                         if (ip->flush_state == HAMMER_FST_SETUP)
2013                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2014                         if (ip->flush_state != HAMMER_FST_IDLE) {
2015                                 ip->flags |= HAMMER_INODE_FLUSHW;
2016                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2017                         }
2018                 }
2019         }
2020 }
2021
2022 /*
2023  * Called by the backend code when a flush has been completed.
2024  * The inode has already been removed from the flush list.
2025  *
2026  * A pipelined flush can occur, in which case we must re-enter the
2027  * inode on the list and re-copy its fields.
2028  */
2029 void
2030 hammer_flush_inode_done(hammer_inode_t ip, int error)
2031 {
2032         hammer_mount_t hmp;
2033         int dorel;
2034
2035         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2036
2037         hmp = ip->hmp;
2038
2039         /*
2040          * Merge left-over flags back into the frontend and fix the state.
2041          * Incomplete truncations are retained by the backend.
2042          */
2043         ip->error = error;
2044         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2045         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2046
2047         /*
2048          * The backend may have adjusted nlinks, so if the adjusted nlinks
2049          * does not match the fronttend set the frontend's RDIRTY flag again.
2050          */
2051         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2052                 ip->flags |= HAMMER_INODE_DDIRTY;
2053
2054         /*
2055          * Fix up the dirty buffer status.
2056          */
2057         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2058                 ip->flags |= HAMMER_INODE_BUFS;
2059         }
2060
2061         /*
2062          * Re-set the XDIRTY flag if some of the inode's in-memory records
2063          * could not be flushed.
2064          */
2065         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2066                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2067                  (!RB_EMPTY(&ip->rec_tree) &&
2068                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2069
2070         /*
2071          * Do not lose track of inodes which no longer have vnode
2072          * assocations, otherwise they may never get flushed again.
2073          *
2074          * The reflush flag can be set superfluously, causing extra pain
2075          * for no reason.  If the inode is no longer modified it no longer
2076          * needs to be flushed.
2077          */
2078         if (ip->flags & HAMMER_INODE_MODMASK) {
2079                 if (ip->vp == NULL)
2080                         ip->flags |= HAMMER_INODE_REFLUSH;
2081         } else {
2082                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2083         }
2084
2085         /*
2086          * Adjust the flush state.
2087          */
2088         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2089                 /*
2090                  * We were unable to flush out all our records, leave the
2091                  * inode in a flush state and in the current flush group.
2092                  * The flush group will be re-run.
2093                  *
2094                  * This occurs if the UNDO block gets too full or there is
2095                  * too much dirty meta-data and allows the flusher to
2096                  * finalize the UNDO block and then re-flush.
2097                  */
2098                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2099                 dorel = 0;
2100         } else {
2101                 /*
2102                  * Remove from the flush_group
2103                  */
2104                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2105                 ip->flush_group = NULL;
2106
2107                 /*
2108                  * Clean up the vnode ref and tracking counts.
2109                  */
2110                 if (ip->flags & HAMMER_INODE_VHELD) {
2111                         ip->flags &= ~HAMMER_INODE_VHELD;
2112                         vrele(ip->vp);
2113                 }
2114                 --hmp->count_iqueued;
2115                 --hammer_count_iqueued;
2116
2117                 /*
2118                  * And adjust the state.
2119                  */
2120                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2121                         ip->flush_state = HAMMER_FST_IDLE;
2122                         dorel = 1;
2123                 } else {
2124                         ip->flush_state = HAMMER_FST_SETUP;
2125                         dorel = 0;
2126                 }
2127
2128                 /*
2129                  * If the frontend is waiting for a flush to complete,
2130                  * wake it up.
2131                  */
2132                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2133                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2134                         wakeup(&ip->flags);
2135                 }
2136
2137                 /*
2138                  * If the frontend made more changes and requested another
2139                  * flush, then try to get it running.
2140                  *
2141                  * Reflushes are aborted when the inode is errored out.
2142                  */
2143                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2144                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2145                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2146                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2147                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2148                         } else {
2149                                 hammer_flush_inode(ip, 0);
2150                         }
2151                 }
2152         }
2153
2154         /*
2155          * If we have no parent dependancies we can clear CONN_DOWN
2156          */
2157         if (TAILQ_EMPTY(&ip->target_list))
2158                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2159
2160         /*
2161          * If the inode is now clean drop the space reservation.
2162          */
2163         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2164             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2165                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2166                 --hmp->rsv_inodes;
2167         }
2168
2169         if (dorel)
2170                 hammer_rel_inode(ip, 0);
2171 }
2172
2173 /*
2174  * Called from hammer_sync_inode() to synchronize in-memory records
2175  * to the media.
2176  */
2177 static int
2178 hammer_sync_record_callback(hammer_record_t record, void *data)
2179 {
2180         hammer_cursor_t cursor = data;
2181         hammer_transaction_t trans = cursor->trans;
2182         hammer_mount_t hmp = trans->hmp;
2183         int error;
2184
2185         /*
2186          * Skip records that do not belong to the current flush.
2187          */
2188         ++hammer_stats_record_iterations;
2189         if (record->flush_state != HAMMER_FST_FLUSH)
2190                 return(0);
2191
2192 #if 1
2193         if (record->flush_group != record->ip->flush_group) {
2194                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2195                 Debugger("blah2");
2196                 return(0);
2197         }
2198 #endif
2199         KKASSERT(record->flush_group == record->ip->flush_group);
2200
2201         /*
2202          * Interlock the record using the BE flag.  Once BE is set the
2203          * frontend cannot change the state of FE.
2204          *
2205          * NOTE: If FE is set prior to us setting BE we still sync the
2206          * record out, but the flush completion code converts it to
2207          * a delete-on-disk record instead of destroying it.
2208          */
2209         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2210         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2211
2212         /*
2213          * The backend may have already disposed of the record.
2214          */
2215         if (record->flags & HAMMER_RECF_DELETED_BE) {
2216                 error = 0;
2217                 goto done;
2218         }
2219
2220         /*
2221          * If the whole inode is being deleting all on-disk records will
2222          * be deleted very soon, we can't sync any new records to disk
2223          * because they will be deleted in the same transaction they were
2224          * created in (delete_tid == create_tid), which will assert.
2225          *
2226          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2227          * that we currently panic on.
2228          */
2229         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2230                 switch(record->type) {
2231                 case HAMMER_MEM_RECORD_DATA:
2232                         /*
2233                          * We don't have to do anything, if the record was
2234                          * committed the space will have been accounted for
2235                          * in the blockmap.
2236                          */
2237                         /* fall through */
2238                 case HAMMER_MEM_RECORD_GENERAL:
2239                         record->flags |= HAMMER_RECF_DELETED_FE;
2240                         record->flags |= HAMMER_RECF_DELETED_BE;
2241                         error = 0;
2242                         goto done;
2243                 case HAMMER_MEM_RECORD_ADD:
2244                         panic("hammer_sync_record_callback: illegal add "
2245                               "during inode deletion record %p", record);
2246                         break; /* NOT REACHED */
2247                 case HAMMER_MEM_RECORD_INODE:
2248                         panic("hammer_sync_record_callback: attempt to "
2249                               "sync inode record %p?", record);
2250                         break; /* NOT REACHED */
2251                 case HAMMER_MEM_RECORD_DEL:
2252                         /*
2253                          * Follow through and issue the on-disk deletion
2254                          */
2255                         break;
2256                 }
2257         }
2258
2259         /*
2260          * If DELETED_FE is set special handling is needed for directory
2261          * entries.  Dependant pieces related to the directory entry may
2262          * have already been synced to disk.  If this occurs we have to
2263          * sync the directory entry and then change the in-memory record
2264          * from an ADD to a DELETE to cover the fact that it's been
2265          * deleted by the frontend.
2266          *
2267          * A directory delete covering record (MEM_RECORD_DEL) can never
2268          * be deleted by the frontend.
2269          *
2270          * Any other record type (aka DATA) can be deleted by the frontend.
2271          * XXX At the moment the flusher must skip it because there may
2272          * be another data record in the flush group for the same block,
2273          * meaning that some frontend data changes can leak into the backend's
2274          * synchronization point.
2275          */
2276         if (record->flags & HAMMER_RECF_DELETED_FE) {
2277                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2278                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2279                 } else {
2280                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2281                         record->flags |= HAMMER_RECF_DELETED_BE;
2282                         error = 0;
2283                         goto done;
2284                 }
2285         }
2286
2287         /*
2288          * Assign the create_tid for new records.  Deletions already
2289          * have the record's entire key properly set up.
2290          */
2291         if (record->type != HAMMER_MEM_RECORD_DEL)
2292                 record->leaf.base.create_tid = trans->tid;
2293                 record->leaf.create_ts = trans->time32;
2294         for (;;) {
2295                 error = hammer_ip_sync_record_cursor(cursor, record);
2296                 if (error != EDEADLK)
2297                         break;
2298                 hammer_done_cursor(cursor);
2299                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2300                                            record->ip);
2301                 if (error)
2302                         break;
2303         }
2304         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2305
2306         if (error)
2307                 error = -error;
2308 done:
2309         hammer_flush_record_done(record, error);
2310
2311         /*
2312          * Do partial finalization if we have built up too many dirty
2313          * buffers.  Otherwise a buffer cache deadlock can occur when
2314          * doing things like creating tens of thousands of tiny files.
2315          *
2316          * We must release our cursor lock to avoid a 3-way deadlock
2317          * due to the exclusive sync lock the finalizer must get.
2318          */
2319         if (hammer_flusher_meta_limit(hmp)) {
2320                 hammer_unlock_cursor(cursor, 0);
2321                 hammer_flusher_finalize(trans, 0);
2322                 hammer_lock_cursor(cursor, 0);
2323         }
2324
2325         return(error);
2326 }
2327
2328 /*
2329  * Backend function called by the flusher to sync an inode to media.
2330  */
2331 int
2332 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2333 {
2334         struct hammer_cursor cursor;
2335         hammer_node_t tmp_node;
2336         hammer_record_t depend;
2337         hammer_record_t next;
2338         int error, tmp_error;
2339         u_int64_t nlinks;
2340
2341         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2342                 return(0);
2343
2344         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2345         if (error)
2346                 goto done;
2347
2348         /*
2349          * Any directory records referencing this inode which are not in
2350          * our current flush group must adjust our nlink count for the
2351          * purposes of synchronization to disk.
2352          *
2353          * Records which are in our flush group can be unlinked from our
2354          * inode now, potentially allowing the inode to be physically
2355          * deleted.
2356          *
2357          * This cannot block.
2358          */
2359         nlinks = ip->ino_data.nlinks;
2360         next = TAILQ_FIRST(&ip->target_list);
2361         while ((depend = next) != NULL) {
2362                 next = TAILQ_NEXT(depend, target_entry);
2363                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2364                     depend->flush_group == ip->flush_group) {
2365                         /*
2366                          * If this is an ADD that was deleted by the frontend
2367                          * the frontend nlinks count will have already been
2368                          * decremented, but the backend is going to sync its
2369                          * directory entry and must account for it.  The
2370                          * record will be converted to a delete-on-disk when
2371                          * it gets synced.
2372                          *
2373                          * If the ADD was not deleted by the frontend we
2374                          * can remove the dependancy from our target_list.
2375                          */
2376                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2377                                 ++nlinks;
2378                         } else {
2379                                 TAILQ_REMOVE(&ip->target_list, depend,
2380                                              target_entry);
2381                                 depend->target_ip = NULL;
2382                         }
2383                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2384                         /*
2385                          * Not part of our flush group
2386                          */
2387                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2388                         switch(depend->type) {
2389                         case HAMMER_MEM_RECORD_ADD:
2390                                 --nlinks;
2391                                 break;
2392                         case HAMMER_MEM_RECORD_DEL:
2393                                 ++nlinks;
2394                                 break;
2395                         default:
2396                                 break;
2397                         }
2398                 }
2399         }
2400
2401         /*
2402          * Set dirty if we had to modify the link count.
2403          */
2404         if (ip->sync_ino_data.nlinks != nlinks) {
2405                 KKASSERT((int64_t)nlinks >= 0);
2406                 ip->sync_ino_data.nlinks = nlinks;
2407                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2408         }
2409
2410         /*
2411          * If there is a trunction queued destroy any data past the (aligned)
2412          * truncation point.  Userland will have dealt with the buffer
2413          * containing the truncation point for us.
2414          *
2415          * We don't flush pending frontend data buffers until after we've
2416          * dealt with the truncation.
2417          */
2418         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2419                 /*
2420                  * Interlock trunc_off.  The VOP front-end may continue to
2421                  * make adjustments to it while we are blocked.
2422                  */
2423                 off_t trunc_off;
2424                 off_t aligned_trunc_off;
2425                 int blkmask;
2426
2427                 trunc_off = ip->sync_trunc_off;
2428                 blkmask = hammer_blocksize(trunc_off) - 1;
2429                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2430
2431                 /*
2432                  * Delete any whole blocks on-media.  The front-end has
2433                  * already cleaned out any partial block and made it
2434                  * pending.  The front-end may have updated trunc_off
2435                  * while we were blocked so we only use sync_trunc_off.
2436                  *
2437                  * This operation can blow out the buffer cache, EWOULDBLOCK
2438                  * means we were unable to complete the deletion.  The
2439                  * deletion will update sync_trunc_off in that case.
2440                  */
2441                 error = hammer_ip_delete_range(&cursor, ip,
2442                                                 aligned_trunc_off,
2443                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2444                 if (error == EWOULDBLOCK) {
2445                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2446                         error = 0;
2447                         goto defer_buffer_flush;
2448                 }
2449
2450                 if (error)
2451                         goto done;
2452
2453                 /*
2454                  * Clear the truncation flag on the backend after we have
2455                  * complete the deletions.  Backend data is now good again
2456                  * (including new records we are about to sync, below).
2457                  *
2458                  * Leave sync_trunc_off intact.  As we write additional
2459                  * records the backend will update sync_trunc_off.  This
2460                  * tells the backend whether it can skip the overwrite
2461                  * test.  This should work properly even when the backend
2462                  * writes full blocks where the truncation point straddles
2463                  * the block because the comparison is against the base
2464                  * offset of the record.
2465                  */
2466                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2467                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2468         } else {
2469                 error = 0;
2470         }
2471
2472         /*
2473          * Now sync related records.  These will typically be directory
2474          * entries, records tracking direct-writes, or delete-on-disk records.
2475          */
2476         if (error == 0) {
2477                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2478                                     hammer_sync_record_callback, &cursor);
2479                 if (tmp_error < 0)
2480                         tmp_error = -error;
2481                 if (tmp_error)
2482                         error = tmp_error;
2483         }
2484         hammer_cache_node(&ip->cache[1], cursor.node);
2485
2486         /*
2487          * Re-seek for inode update, assuming our cache hasn't been ripped
2488          * out from under us.
2489          */
2490         if (error == 0) {
2491                 tmp_node = hammer_ref_node_safe(ip->hmp, &ip->cache[0], &error);
2492                 if (tmp_node) {
2493                         hammer_cursor_downgrade(&cursor);
2494                         hammer_lock_sh(&tmp_node->lock);
2495                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2496                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2497                         hammer_unlock(&tmp_node->lock);
2498                         hammer_rel_node(tmp_node);
2499                 }
2500                 error = 0;
2501         }
2502
2503         /*
2504          * If we are deleting the inode the frontend had better not have
2505          * any active references on elements making up the inode.
2506          *
2507          * The call to hammer_ip_delete_clean() cleans up auxillary records
2508          * but not DB or DATA records.  Those must have already been deleted
2509          * by the normal truncation mechanic.
2510          */
2511         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2512                 RB_EMPTY(&ip->rec_tree)  &&
2513             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2514             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2515                 int count1 = 0;
2516
2517                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2518                 if (error == 0) {
2519                         ip->flags |= HAMMER_INODE_DELETED;
2520                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2521                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2522                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2523
2524                         /*
2525                          * Set delete_tid in both the frontend and backend
2526                          * copy of the inode record.  The DELETED flag handles
2527                          * this, do not set RDIRTY.
2528                          */
2529                         ip->ino_leaf.base.delete_tid = trans->tid;
2530                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2531                         ip->ino_leaf.delete_ts = trans->time32;
2532                         ip->sync_ino_leaf.delete_ts = trans->time32;
2533
2534
2535                         /*
2536                          * Adjust the inode count in the volume header
2537                          */
2538                         hammer_sync_lock_sh(trans);
2539                         if (ip->flags & HAMMER_INODE_ONDISK) {
2540                                 hammer_modify_volume_field(trans,
2541                                                            trans->rootvol,
2542                                                            vol0_stat_inodes);
2543                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2544                                 hammer_modify_volume_done(trans->rootvol);
2545                         }
2546                         hammer_sync_unlock(trans);
2547                 }
2548         }
2549
2550         if (error)
2551                 goto done;
2552         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2553
2554 defer_buffer_flush:
2555         /*
2556          * Now update the inode's on-disk inode-data and/or on-disk record.
2557          * DELETED and ONDISK are managed only in ip->flags.
2558          *
2559          * In the case of a defered buffer flush we still update the on-disk
2560          * inode to satisfy visibility requirements if there happen to be
2561          * directory dependancies.
2562          */
2563         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2564         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2565                 /*
2566                  * If deleted and on-disk, don't set any additional flags.
2567                  * the delete flag takes care of things.
2568                  *
2569                  * Clear flags which may have been set by the frontend.
2570                  */
2571                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2572                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2573                                     HAMMER_INODE_DELETING);
2574                 break;
2575         case HAMMER_INODE_DELETED:
2576                 /*
2577                  * Take care of the case where a deleted inode was never
2578                  * flushed to the disk in the first place.
2579                  *
2580                  * Clear flags which may have been set by the frontend.
2581                  */
2582                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2583                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2584                                     HAMMER_INODE_DELETING);
2585                 while (RB_ROOT(&ip->rec_tree)) {
2586                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2587                         hammer_ref(&record->lock);
2588                         KKASSERT(record->lock.refs == 1);
2589                         record->flags |= HAMMER_RECF_DELETED_FE;
2590                         record->flags |= HAMMER_RECF_DELETED_BE;
2591                         hammer_rel_mem_record(record);
2592                 }
2593                 break;
2594         case HAMMER_INODE_ONDISK:
2595                 /*
2596                  * If already on-disk, do not set any additional flags.
2597                  */
2598                 break;
2599         default:
2600                 /*
2601                  * If not on-disk and not deleted, set DDIRTY to force
2602                  * an initial record to be written.
2603                  *
2604                  * Also set the create_tid in both the frontend and backend
2605                  * copy of the inode record.
2606                  */
2607                 ip->ino_leaf.base.create_tid = trans->tid;
2608                 ip->ino_leaf.create_ts = trans->time32;
2609                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2610                 ip->sync_ino_leaf.create_ts = trans->time32;
2611                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2612                 break;
2613         }
2614
2615         /*
2616          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2617          * is already on-disk the old record is marked as deleted.
2618          *
2619          * If DELETED is set hammer_update_inode() will delete the existing
2620          * record without writing out a new one.
2621          *
2622          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2623          */
2624         if (ip->flags & HAMMER_INODE_DELETED) {
2625                 error = hammer_update_inode(&cursor, ip);
2626         } else
2627         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2628             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2629                 error = hammer_update_itimes(&cursor, ip);
2630         } else
2631         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2632                 error = hammer_update_inode(&cursor, ip);
2633         }
2634 done:
2635         if (error) {
2636                 hammer_critical_error(ip->hmp, ip, error,
2637                                       "while syncing inode");
2638         }
2639         hammer_done_cursor(&cursor);
2640         return(error);
2641 }
2642
2643 /*
2644  * This routine is called when the OS is no longer actively referencing
2645  * the inode (but might still be keeping it cached), or when releasing
2646  * the last reference to an inode.
2647  *
2648  * At this point if the inode's nlinks count is zero we want to destroy
2649  * it, which may mean destroying it on-media too.
2650  */
2651 void
2652 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2653 {
2654         struct vnode *vp;
2655
2656         /*
2657          * Set the DELETING flag when the link count drops to 0 and the
2658          * OS no longer has any opens on the inode.
2659          *
2660          * The backend will clear DELETING (a mod flag) and set DELETED
2661          * (a state flag) when it is actually able to perform the
2662          * operation.
2663          *
2664          * Don't reflag the deletion if the flusher is currently syncing
2665          * one that was already flagged.  A previously set DELETING flag
2666          * may bounce around flags and sync_flags until the operation is
2667          * completely done.
2668          */
2669         if (ip->ino_data.nlinks == 0 &&
2670             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2671                 ip->flags |= HAMMER_INODE_DELETING;
2672                 ip->flags |= HAMMER_INODE_TRUNCATED;
2673                 ip->trunc_off = 0;
2674                 vp = NULL;
2675                 if (getvp) {
2676                         if (hammer_get_vnode(ip, &vp) != 0)
2677                                 return;
2678                 }
2679
2680                 /*
2681                  * Final cleanup
2682                  */
2683                 if (ip->vp) {
2684                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2685                         vnode_pager_setsize(ip->vp, 0);
2686                 }
2687                 if (getvp) {
2688                         vput(vp);
2689                 }
2690         }
2691 }
2692
2693 /*
2694  * After potentially resolving a dependancy the inode is tested
2695  * to determine whether it needs to be reflushed.
2696  */
2697 void
2698 hammer_test_inode(hammer_inode_t ip)
2699 {
2700         if (ip->flags & HAMMER_INODE_REFLUSH) {
2701                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2702                 hammer_ref(&ip->lock);
2703                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2704                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2705                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2706                 } else {
2707                         hammer_flush_inode(ip, 0);
2708                 }
2709                 hammer_rel_inode(ip, 0);
2710         }
2711 }
2712
2713 /*
2714  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2715  * reassociated with a vp or just before it gets freed.
2716  *
2717  * Wakeup one thread blocked waiting on reclaims to complete.  Note that
2718  * the inode the thread is waiting on behalf of is a different inode then
2719  * the inode we are called with.  This is to create a pipeline.
2720  */
2721 static void
2722 hammer_inode_wakereclaims(hammer_inode_t ip)
2723 {
2724         struct hammer_reclaim *reclaim;
2725         hammer_mount_t hmp = ip->hmp;
2726
2727         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2728                 return;
2729
2730         --hammer_count_reclaiming;
2731         --hmp->inode_reclaims;
2732         ip->flags &= ~HAMMER_INODE_RECLAIM;
2733
2734         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
2735                 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2736                 reclaim->okydoky = 1;
2737                 wakeup(reclaim);
2738         }
2739 }
2740
2741 /*
2742  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
2743  * inodes build up before we start blocking.
2744  *
2745  * When we block we don't care *which* inode has finished reclaiming,
2746  * as lone as one does.  This is somewhat heuristical... we also put a
2747  * cap on how long we are willing to wait.
2748  */
2749 void
2750 hammer_inode_waitreclaims(hammer_mount_t hmp)
2751 {
2752         struct hammer_reclaim reclaim;
2753         int delay;
2754
2755         if (hmp->inode_reclaims > HAMMER_RECLAIM_WAIT) {
2756                 reclaim.okydoky = 0;
2757                 TAILQ_INSERT_TAIL(&hmp->reclaim_list,
2758                                   &reclaim, entry);
2759         } else {
2760                 reclaim.okydoky = 1;
2761         }
2762
2763         if (reclaim.okydoky == 0) {
2764                 delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2765                         (HAMMER_RECLAIM_WAIT * 5);
2766                 if (delay >= 0)
2767                         tsleep(&reclaim, 0, "hmrrcm", delay + 1);
2768                 if (reclaim.okydoky == 0)
2769                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2770         }
2771 }
2772