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, int depth,
  51                                         hammer_flush_group_t flg);
  52 static int      hammer_setup_parent_inodes_helper(hammer_record_t record,
  53                                         int depth, hammer_flush_group_t flg);
  54 static void     hammer_inode_wakereclaims(hammer_inode_t ip, int dowake);
  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 = *vpp;
 255                                 vp->v_type = VBAD;
 256                                 vx_put(vp);
 257                                 continue;
 258                         }
 259                         hammer_ref(&ip->lock);
 260                         vp = *vpp;
 261                         ip->vp = vp;
 262
 263                         obj_type = ip->ino_data.obj_type;
 264                         vp->v_type = hammer_get_vnode_type(obj_type);
 265
 266                         hammer_inode_wakereclaims(ip, 0);
 267
 268                         switch(ip->ino_data.obj_type) {
 269                         case HAMMER_OBJTYPE_CDEV:
 270                         case HAMMER_OBJTYPE_BDEV:
 271                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
 272                                 addaliasu(vp, ip->ino_data.rmajor,
 273                                           ip->ino_data.rminor);
 274                                 break;
 275                         case HAMMER_OBJTYPE_FIFO:
 276                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
 277                                 break;
 278                         case HAMMER_OBJTYPE_REGFILE:
 279                                 /*
 280                                  * MPSAFE read supported.
 281                                  */
 282                                 vp->v_flag |= VMP_READ;
 283                                 break;
 284                         default:
 285                                 break;
 286                         }
 287
 288                         /*
 289                          * Only mark as the root vnode if the ip is not
 290                          * historical, otherwise the VFS cache will get
 291                          * confused.  The other half of the special handling
 292                          * is in hammer_vop_nlookupdotdot().
 293                          *
 294                          * Pseudo-filesystem roots can be accessed via
 295                          * non-root filesystem paths and setting VROOT may
 296                          * confuse the namecache.  Set VPFSROOT instead.
 297                          */
 298                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 299                             ip->obj_asof == hmp->asof) {
 300                                 if (ip->obj_localization == 0)
 301                                         vp->v_flag |= VROOT;
 302                                 else
 303                                         vp->v_flag |= VPFSROOT;
 304                         }
 305
 306                         vp->v_data = (void *)ip;
 307                         /* vnode locked by getnewvnode() */
 308                         /* make related vnode dirty if inode dirty? */
 309                         hammer_unlock(&ip->lock);
 310                         if (vp->v_type == VREG)
 311                                 vinitvmio(vp, ip->ino_data.size);
 312                         break;
 313                 }
 314
 315                 /*
 316                  * loop if the vget fails (aka races), or if the vp
 317                  * no longer matches ip->vp.
 318                  */
 319                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 320                         if (vp == ip->vp)
 321                                 break;
 322                         vput(vp);
 323                 }
 324         }
 325         *vpp = vp;
 326         return(error);
 327 }
 328
 329 /*
 330  * Locate all copies of the inode for obj_id compatible with the specified
 331  * asof, reference, and issue the related call-back.  This routine is used
 332  * for direct-io invalidation and does not create any new inodes.
 333  */
 334 void
 335 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
 336                             int (*callback)(hammer_inode_t ip, void *data),
 337                             void *data)
 338 {
 339         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
 340                                    hammer_inode_info_cmp_all_history,
 341                                    callback, iinfo);
 342 }
 343
 344 /*
 345  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 346  * do not attach or detach the related vnode (use hammer_get_vnode() for
 347  * that).
 348  *
 349  * The flags argument is only applied for newly created inodes, and only
 350  * certain flags are inherited.
 351  *
 352  * Called from the frontend.
 353  */
 354 struct hammer_inode *
 355 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
 356                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 357                  int flags, int *errorp)
 358 {
 359         hammer_mount_t hmp = trans->hmp;
 360         struct hammer_node_cache *cachep;
 361         struct hammer_inode_info iinfo;
 362         struct hammer_cursor cursor;
 363         struct hammer_inode *ip;
 364
 365
 366         /*
 367          * Determine if we already have an inode cached.  If we do then
 368          * we are golden.
 369          *
 370          * If we find an inode with no vnode we have to mark the
 371          * transaction such that hammer_inode_waitreclaims() is
 372          * called later on to avoid building up an infinite number
 373          * of inodes.  Otherwise we can continue to * add new inodes
 374          * faster then they can be disposed of, even with the tsleep
 375          * delay.
 376          *
 377          * If we find a dummy inode we return a failure so dounlink
 378          * (which does another lookup) doesn't try to mess with the
 379          * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
 380          * to ref dummy inodes.
 381          */
 382         iinfo.obj_id = obj_id;
 383         iinfo.obj_asof = asof;
 384         iinfo.obj_localization = localization;
 385 loop:
 386         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 387         if (ip) {
 388                 if (ip->flags & HAMMER_INODE_DUMMY) {
 389                         *errorp = ENOENT;
 390                         return(NULL);
 391                 }
 392                 hammer_ref(&ip->lock);
 393                 *errorp = 0;
 394                 return(ip);
 395         }
 396
 397         /*
 398          * Allocate a new inode structure and deal with races later.
 399          */
 400         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 401         ++hammer_count_inodes;
 402         ++hmp->count_inodes;
 403         ip->obj_id = obj_id;
 404         ip->obj_asof = iinfo.obj_asof;
 405         ip->obj_localization = localization;
 406         ip->hmp = hmp;
 407         ip->flags = flags & HAMMER_INODE_RO;
 408         ip->cache[0].ip = ip;
 409         ip->cache[1].ip = ip;
 410         ip->cache[2].ip = ip;
 411         ip->cache[3].ip = ip;
 412         if (hmp->ronly)
 413                 ip->flags |= HAMMER_INODE_RO;
 414         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 415                 0x7FFFFFFFFFFFFFFFLL;
 416         RB_INIT(&ip->rec_tree);
 417         TAILQ_INIT(&ip->target_list);
 418         hammer_ref(&ip->lock);
 419
 420         /*
 421          * Locate the on-disk inode.  If this is a PFS root we always
 422          * access the current version of the root inode and (if it is not
 423          * a master) always access information under it with a snapshot
 424          * TID.
 425          *
 426          * We cache recent inode lookups in this directory in dip->cache[2].
 427          * If we can't find it we assume the inode we are looking for is
 428          * close to the directory inode.
 429          */
 430 retry:
 431         cachep = NULL;
 432         if (dip) {
 433                 if (dip->cache[2].node)
 434                         cachep = &dip->cache[2];
 435                 else
 436                         cachep = &dip->cache[0];
 437         }
 438         hammer_init_cursor(trans, &cursor, cachep, NULL);
 439         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
 440         cursor.key_beg.obj_id = ip->obj_id;
 441         cursor.key_beg.key = 0;
 442         cursor.key_beg.create_tid = 0;
 443         cursor.key_beg.delete_tid = 0;
 444         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 445         cursor.key_beg.obj_type = 0;
 446
 447         cursor.asof = iinfo.obj_asof;
 448         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 449                        HAMMER_CURSOR_ASOF;
 450
 451         *errorp = hammer_btree_lookup(&cursor);
 452         if (*errorp == EDEADLK) {
 453                 hammer_done_cursor(&cursor);
 454                 goto retry;
 455         }
 456
 457         /*
 458          * On success the B-Tree lookup will hold the appropriate
 459          * buffer cache buffers and provide a pointer to the requested
 460          * information.  Copy the information to the in-memory inode
 461          * and cache the B-Tree node to improve future operations.
 462          */
 463         if (*errorp == 0) {
 464                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 465                 ip->ino_data = cursor.data->inode;
 466
 467                 /*
 468                  * cache[0] tries to cache the location of the object inode.
 469                  * The assumption is that it is near the directory inode.
 470                  *
 471                  * cache[1] tries to cache the location of the object data.
 472                  * We might have something in the governing directory from
 473                  * scan optimizations (see the strategy code in
 474                  * hammer_vnops.c).
 475                  *
 476                  * We update dip->cache[2], if possible, with the location
 477                  * of the object inode for future directory shortcuts.
 478                  */
 479                 hammer_cache_node(&ip->cache[0], cursor.node);
 480                 if (dip) {
 481                         if (dip->cache[3].node) {
 482                                 hammer_cache_node(&ip->cache[1],
 483                                                   dip->cache[3].node);
 484                         }
 485                         hammer_cache_node(&dip->cache[2], cursor.node);
 486                 }
 487
 488                 /*
 489                  * The file should not contain any data past the file size
 490                  * stored in the inode.  Setting save_trunc_off to the
 491                  * file size instead of max reduces B-Tree lookup overheads
 492                  * on append by allowing the flusher to avoid checking for
 493                  * record overwrites.
 494                  */
 495                 ip->save_trunc_off = ip->ino_data.size;
 496
 497                 /*
 498                  * Locate and assign the pseudofs management structure to
 499                  * the inode.
 500                  */
 501                 if (dip && dip->obj_localization == ip->obj_localization) {
 502                         ip->pfsm = dip->pfsm;
 503                         hammer_ref(&ip->pfsm->lock);
 504                 } else {
 505                         ip->pfsm = hammer_load_pseudofs(trans,
 506                                                         ip->obj_localization,
 507                                                         errorp);
 508                         *errorp = 0;    /* ignore ENOENT */
 509                 }
 510         }
 511
 512         /*
 513          * The inode is placed on the red-black tree and will be synced to
 514          * the media when flushed or by the filesystem sync.  If this races
 515          * another instantiation/lookup the insertion will fail.
 516          */
 517         if (*errorp == 0) {
 518                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 519                         hammer_free_inode(ip);
 520                         hammer_done_cursor(&cursor);
 521                         goto loop;
 522                 }
 523                 ip->flags |= HAMMER_INODE_ONDISK;
 524         } else {
 525                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 526                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 527                         --hmp->rsv_inodes;
 528                 }
 529
 530                 hammer_free_inode(ip);
 531                 ip = NULL;
 532         }
 533         hammer_done_cursor(&cursor);
 534         trans->flags |= HAMMER_TRANSF_NEWINODE;
 535         return (ip);
 536 }
 537
 538 /*
 539  * Get a dummy inode to placemark a broken directory entry.
 540  */
 541 struct hammer_inode *
 542 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
 543                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 544                  int flags, int *errorp)
 545 {
 546         hammer_mount_t hmp = trans->hmp;
 547         struct hammer_inode_info iinfo;
 548         struct hammer_inode *ip;
 549
 550         /*
 551          * Determine if we already have an inode cached.  If we do then
 552          * we are golden.
 553          *
 554          * If we find an inode with no vnode we have to mark the
 555          * transaction such that hammer_inode_waitreclaims() is
 556          * called later on to avoid building up an infinite number
 557          * of inodes.  Otherwise we can continue to * add new inodes
 558          * faster then they can be disposed of, even with the tsleep
 559          * delay.
 560          *
 561          * If we find a non-fake inode we return an error.  Only fake
 562          * inodes can be returned by this routine.
 563          */
 564         iinfo.obj_id = obj_id;
 565         iinfo.obj_asof = asof;
 566         iinfo.obj_localization = localization;
 567 loop:
 568         *errorp = 0;
 569         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 570         if (ip) {
 571                 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
 572                         *errorp = ENOENT;
 573                         return(NULL);
 574                 }
 575                 hammer_ref(&ip->lock);
 576                 return(ip);
 577         }
 578
 579         /*
 580          * Allocate a new inode structure and deal with races later.
 581          */
 582         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 583         ++hammer_count_inodes;
 584         ++hmp->count_inodes;
 585         ip->obj_id = obj_id;
 586         ip->obj_asof = iinfo.obj_asof;
 587         ip->obj_localization = localization;
 588         ip->hmp = hmp;
 589         ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
 590         ip->cache[0].ip = ip;
 591         ip->cache[1].ip = ip;
 592         ip->cache[2].ip = ip;
 593         ip->cache[3].ip = ip;
 594         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 595                 0x7FFFFFFFFFFFFFFFLL;
 596         RB_INIT(&ip->rec_tree);
 597         TAILQ_INIT(&ip->target_list);
 598         hammer_ref(&ip->lock);
 599
 600         /*
 601          * Populate the dummy inode.  Leave everything zero'd out.
 602          *
 603          * (ip->ino_leaf and ip->ino_data)
 604          *
 605          * Make the dummy inode a FIFO object which most copy programs
 606          * will properly ignore.
 607          */
 608         ip->save_trunc_off = ip->ino_data.size;
 609         ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
 610
 611         /*
 612          * Locate and assign the pseudofs management structure to
 613          * the inode.
 614          */
 615         if (dip && dip->obj_localization == ip->obj_localization) {
 616                 ip->pfsm = dip->pfsm;
 617                 hammer_ref(&ip->pfsm->lock);
 618         } else {
 619                 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
 620                                                 errorp);
 621                 *errorp = 0;    /* ignore ENOENT */
 622         }
 623
 624         /*
 625          * The inode is placed on the red-black tree and will be synced to
 626          * the media when flushed or by the filesystem sync.  If this races
 627          * another instantiation/lookup the insertion will fail.
 628          *
 629          * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
 630          */
 631         if (*errorp == 0) {
 632                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 633                         hammer_free_inode(ip);
 634                         goto loop;
 635                 }
 636         } else {
 637                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 638                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 639                         --hmp->rsv_inodes;
 640                 }
 641                 hammer_free_inode(ip);
 642                 ip = NULL;
 643         }
 644         trans->flags |= HAMMER_TRANSF_NEWINODE;
 645         return (ip);
 646 }
 647
 648 /*
 649  * Return a referenced inode only if it is in our inode cache.
 650  *
 651  * Dummy inodes do not count.
 652  */
 653 struct hammer_inode *
 654 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
 655                   hammer_tid_t asof, u_int32_t localization)
 656 {
 657         hammer_mount_t hmp = trans->hmp;
 658         struct hammer_inode_info iinfo;
 659         struct hammer_inode *ip;
 660
 661         iinfo.obj_id = obj_id;
 662         iinfo.obj_asof = asof;
 663         iinfo.obj_localization = localization;
 664
 665         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 666         if (ip) {
 667                 if (ip->flags & HAMMER_INODE_DUMMY)
 668                         ip = NULL;
 669                 else
 670                         hammer_ref(&ip->lock);
 671         }
 672         return(ip);
 673 }
 674
 675 /*
 676  * Create a new filesystem object, returning the inode in *ipp.  The
 677  * returned inode will be referenced.  The inode is created in-memory.
 678  *
 679  * If pfsm is non-NULL the caller wishes to create the root inode for
 680  * a master PFS.
 681  */
 682 int
 683 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 684                     struct ucred *cred,
 685                     hammer_inode_t dip, const char *name, int namelen,
 686                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
 687 {
 688         hammer_mount_t hmp;
 689         hammer_inode_t ip;
 690         uid_t xuid;
 691         int error;
 692         int64_t namekey;
 693         u_int32_t dummy;
 694
 695         hmp = trans->hmp;
 696
 697         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 698         ++hammer_count_inodes;
 699         ++hmp->count_inodes;
 700         trans->flags |= HAMMER_TRANSF_NEWINODE;
 701
 702         if (pfsm) {
 703                 KKASSERT(pfsm->localization != 0);
 704                 ip->obj_id = HAMMER_OBJID_ROOT;
 705                 ip->obj_localization = pfsm->localization;
 706         } else {
 707                 KKASSERT(dip != NULL);
 708                 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
 709                 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
 710                 ip->obj_localization = dip->obj_localization;
 711         }
 712
 713         KKASSERT(ip->obj_id != 0);
 714         ip->obj_asof = hmp->asof;
 715         ip->hmp = hmp;
 716         ip->flush_state = HAMMER_FST_IDLE;
 717         ip->flags = HAMMER_INODE_DDIRTY |
 718                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
 719         ip->cache[0].ip = ip;
 720         ip->cache[1].ip = ip;
 721         ip->cache[2].ip = ip;
 722         ip->cache[3].ip = ip;
 723
 724         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 725         /* ip->save_trunc_off = 0; (already zero) */
 726         RB_INIT(&ip->rec_tree);
 727         TAILQ_INIT(&ip->target_list);
 728
 729         ip->ino_data.atime = trans->time;
 730         ip->ino_data.mtime = trans->time;
 731         ip->ino_data.size = 0;
 732         ip->ino_data.nlinks = 0;
 733
 734         /*
 735          * A nohistory designator on the parent directory is inherited by
 736          * the child.  We will do this even for pseudo-fs creation... the
 737          * sysad can turn it off.
 738          */
 739         if (dip) {
 740                 ip->ino_data.uflags = dip->ino_data.uflags &
 741                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 742         }
 743
 744         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 745         ip->ino_leaf.base.localization = ip->obj_localization +
 746                                          HAMMER_LOCALIZE_INODE;
 747         ip->ino_leaf.base.obj_id = ip->obj_id;
 748         ip->ino_leaf.base.key = 0;
 749         ip->ino_leaf.base.create_tid = 0;
 750         ip->ino_leaf.base.delete_tid = 0;
 751         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 752         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 753
 754         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 755         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 756         ip->ino_data.mode = vap->va_mode;
 757         ip->ino_data.ctime = trans->time;
 758
 759         /*
 760          * If we are running version 2 or greater directory entries are
 761          * inode-localized instead of data-localized.
 762          */
 763         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
 764                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
 765                         ip->ino_data.cap_flags |=
 766                                 HAMMER_INODE_CAP_DIR_LOCAL_INO;
 767                 }
 768         }
 769
 770         /*
 771          * Setup the ".." pointer.  This only needs to be done for directories
 772          * but we do it for all objects as a recovery aid.
 773          */
 774         if (dip)
 775                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
 776 #if 0
 777         /*
 778          * The parent_obj_localization field only applies to pseudo-fs roots.
 779          * XXX this is no longer applicable, PFSs are no longer directly
 780          * tied into the parent's directory structure.
 781          */
 782         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
 783             ip->obj_id == HAMMER_OBJID_ROOT) {
 784                 ip->ino_data.ext.obj.parent_obj_localization =
 785                                                 dip->obj_localization;
 786         }
 787 #endif
 788
 789         switch(ip->ino_leaf.base.obj_type) {
 790         case HAMMER_OBJTYPE_CDEV:
 791         case HAMMER_OBJTYPE_BDEV:
 792                 ip->ino_data.rmajor = vap->va_rmajor;
 793                 ip->ino_data.rminor = vap->va_rminor;
 794                 break;
 795         default:
 796                 break;
 797         }
 798
 799         /*
 800          * Calculate default uid/gid and overwrite with information from
 801          * the vap.
 802          */
 803         if (dip) {
 804                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 805                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
 806                                              xuid, cred, &vap->va_mode);
 807         } else {
 808                 xuid = 0;
 809         }
 810         ip->ino_data.mode = vap->va_mode;
 811
 812         if (vap->va_vaflags & VA_UID_UUID_VALID)
 813                 ip->ino_data.uid = vap->va_uid_uuid;
 814         else if (vap->va_uid != (uid_t)VNOVAL)
 815                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 816         else
 817                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 818
 819         if (vap->va_vaflags & VA_GID_UUID_VALID)
 820                 ip->ino_data.gid = vap->va_gid_uuid;
 821         else if (vap->va_gid != (gid_t)VNOVAL)
 822                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 823         else if (dip)
 824                 ip->ino_data.gid = dip->ino_data.gid;
 825
 826         hammer_ref(&ip->lock);
 827
 828         if (pfsm) {
 829                 ip->pfsm = pfsm;
 830                 hammer_ref(&pfsm->lock);
 831                 error = 0;
 832         } else if (dip->obj_localization == ip->obj_localization) {
 833                 ip->pfsm = dip->pfsm;
 834                 hammer_ref(&ip->pfsm->lock);
 835                 error = 0;
 836         } else {
 837                 ip->pfsm = hammer_load_pseudofs(trans,
 838                                                 ip->obj_localization,
 839                                                 &error);
 840                 error = 0;      /* ignore ENOENT */
 841         }
 842
 843         if (error) {
 844                 hammer_free_inode(ip);
 845                 ip = NULL;
 846         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 847                 panic("hammer_create_inode: duplicate obj_id %llx",
 848                       (long long)ip->obj_id);
 849                 /* not reached */
 850                 hammer_free_inode(ip);
 851         }
 852         *ipp = ip;
 853         return(error);
 854 }
 855
 856 /*
 857  * Final cleanup / freeing of an inode structure
 858  */
 859 static void
 860 hammer_free_inode(hammer_inode_t ip)
 861 {
 862         struct hammer_mount *hmp;
 863
 864         hmp = ip->hmp;
 865         KKASSERT(ip->lock.refs == 1);
 866         hammer_uncache_node(&ip->cache[0]);
 867         hammer_uncache_node(&ip->cache[1]);
 868         hammer_uncache_node(&ip->cache[2]);
 869         hammer_uncache_node(&ip->cache[3]);
 870         hammer_inode_wakereclaims(ip, 1);
 871         if (ip->objid_cache)
 872                 hammer_clear_objid(ip);
 873         --hammer_count_inodes;
 874         --hmp->count_inodes;
 875         if (ip->pfsm) {
 876                 hammer_rel_pseudofs(hmp, ip->pfsm);
 877                 ip->pfsm = NULL;
 878         }
 879         kfree(ip, hmp->m_inodes);
 880         ip = NULL;
 881 }
 882
 883 /*
 884  * Retrieve pseudo-fs data.  NULL will never be returned.
 885  *
 886  * If an error occurs *errorp will be set and a default template is returned,
 887  * otherwise *errorp is set to 0.  Typically when an error occurs it will
 888  * be ENOENT.
 889  */
 890 hammer_pseudofs_inmem_t
 891 hammer_load_pseudofs(hammer_transaction_t trans,
 892                      u_int32_t localization, int *errorp)
 893 {
 894         hammer_mount_t hmp = trans->hmp;
 895         hammer_inode_t ip;
 896         hammer_pseudofs_inmem_t pfsm;
 897         struct hammer_cursor cursor;
 898         int bytes;
 899
 900 retry:
 901         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
 902         if (pfsm) {
 903                 hammer_ref(&pfsm->lock);
 904                 *errorp = 0;
 905                 return(pfsm);
 906         }
 907
 908         /*
 909          * PFS records are stored in the root inode (not the PFS root inode,
 910          * but the real root).  Avoid an infinite recursion if loading
 911          * the PFS for the real root.
 912          */
 913         if (localization) {
 914                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
 915                                       HAMMER_MAX_TID,
 916                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
 917         } else {
 918                 ip = NULL;
 919         }
 920
 921         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
 922         pfsm->localization = localization;
 923         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
 924         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
 925
 926         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
 927         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
 928                                       HAMMER_LOCALIZE_MISC;
 929         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 930         cursor.key_beg.create_tid = 0;
 931         cursor.key_beg.delete_tid = 0;
 932         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 933         cursor.key_beg.obj_type = 0;
 934         cursor.key_beg.key = localization;
 935         cursor.asof = HAMMER_MAX_TID;
 936         cursor.flags |= HAMMER_CURSOR_ASOF;
 937
 938         if (ip)
 939                 *errorp = hammer_ip_lookup(&cursor);
 940         else
 941                 *errorp = hammer_btree_lookup(&cursor);
 942         if (*errorp == 0) {
 943                 *errorp = hammer_ip_resolve_data(&cursor);
 944                 if (*errorp == 0) {
 945                         if (cursor.data->pfsd.mirror_flags &
 946                             HAMMER_PFSD_DELETED) {
 947                                 *errorp = ENOENT;
 948                         } else {
 949                                 bytes = cursor.leaf->data_len;
 950                                 if (bytes > sizeof(pfsm->pfsd))
 951                                         bytes = sizeof(pfsm->pfsd);
 952                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
 953                         }
 954                 }
 955         }
 956         hammer_done_cursor(&cursor);
 957
 958         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 959         hammer_ref(&pfsm->lock);
 960         if (ip)
 961                 hammer_rel_inode(ip, 0);
 962         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
 963                 kfree(pfsm, hmp->m_misc);
 964                 goto retry;
 965         }
 966         return(pfsm);
 967 }
 968
 969 /*
 970  * Store pseudo-fs data.  The backend will automatically delete any prior
 971  * on-disk pseudo-fs data but we have to delete in-memory versions.
 972  */
 973 int
 974 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
 975 {
 976         struct hammer_cursor cursor;
 977         hammer_record_t record;
 978         hammer_inode_t ip;
 979         int error;
 980
 981         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 982                               HAMMER_DEF_LOCALIZATION, 0, &error);
 983 retry:
 984         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 985         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
 986         cursor.key_beg.localization = ip->obj_localization +
 987                                       HAMMER_LOCALIZE_MISC;
 988         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 989         cursor.key_beg.create_tid = 0;
 990         cursor.key_beg.delete_tid = 0;
 991         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 992         cursor.key_beg.obj_type = 0;
 993         cursor.key_beg.key = pfsm->localization;
 994         cursor.asof = HAMMER_MAX_TID;
 995         cursor.flags |= HAMMER_CURSOR_ASOF;
 996
 997         /*
 998          * Replace any in-memory version of the record.
 999          */
1000         error = hammer_ip_lookup(&cursor);
1001         if (error == 0 && hammer_cursor_inmem(&cursor)) {
1002                 record = cursor.iprec;
1003                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1004                         KKASSERT(cursor.deadlk_rec == NULL);
1005                         hammer_ref(&record->lock);
1006                         cursor.deadlk_rec = record;
1007                         error = EDEADLK;
1008                 } else {
1009                         record->flags |= HAMMER_RECF_DELETED_FE;
1010                         error = 0;
1011                 }
1012         }
1013
1014         /*
1015          * Allocate replacement general record.  The backend flush will
1016          * delete any on-disk version of the record.
1017          */
1018         if (error == 0 || error == ENOENT) {
1019                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1020                 record->type = HAMMER_MEM_RECORD_GENERAL;
1021
1022                 record->leaf.base.localization = ip->obj_localization +
1023                                                  HAMMER_LOCALIZE_MISC;
1024                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1025                 record->leaf.base.key = pfsm->localization;
1026                 record->leaf.data_len = sizeof(pfsm->pfsd);
1027                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1028                 error = hammer_ip_add_record(trans, record);
1029         }
1030         hammer_done_cursor(&cursor);
1031         if (error == EDEADLK)
1032                 goto retry;
1033         hammer_rel_inode(ip, 0);
1034         return(error);
1035 }
1036
1037 /*
1038  * Create a root directory for a PFS if one does not alredy exist.
1039  *
1040  * The PFS root stands alone so we must also bump the nlinks count
1041  * to prevent it from being destroyed on release.
1042  */
1043 int
1044 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1045                        hammer_pseudofs_inmem_t pfsm)
1046 {
1047         hammer_inode_t ip;
1048         struct vattr vap;
1049         int error;
1050
1051         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1052                               pfsm->localization, 0, &error);
1053         if (ip == NULL) {
1054                 vattr_null(&vap);
1055                 vap.va_mode = 0755;
1056                 vap.va_type = VDIR;
1057                 error = hammer_create_inode(trans, &vap, cred,
1058                                             NULL, NULL, 0,
1059                                             pfsm, &ip);
1060                 if (error == 0) {
1061                         ++ip->ino_data.nlinks;
1062                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1063                 }
1064         }
1065         if (ip)
1066                 hammer_rel_inode(ip, 0);
1067         return(error);
1068 }
1069
1070 /*
1071  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1072  * if we are unable to disassociate all the inodes.
1073  */
1074 static
1075 int
1076 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1077 {
1078         int res;
1079
1080         hammer_ref(&ip->lock);
1081         if (ip->lock.refs == 2 && ip->vp)
1082                 vclean_unlocked(ip->vp);
1083         if (ip->lock.refs == 1 && ip->vp == NULL)
1084                 res = 0;
1085         else
1086                 res = -1;       /* stop, someone is using the inode */
1087         hammer_rel_inode(ip, 0);
1088         return(res);
1089 }
1090
1091 int
1092 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1093 {
1094         int res;
1095         int try;
1096
1097         for (try = res = 0; try < 4; ++try) {
1098                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1099                                            hammer_inode_pfs_cmp,
1100                                            hammer_unload_pseudofs_callback,
1101                                            &localization);
1102                 if (res == 0 && try > 1)
1103                         break;
1104                 hammer_flusher_sync(trans->hmp);
1105         }
1106         if (res != 0)
1107                 res = ENOTEMPTY;
1108         return(res);
1109 }
1110
1111
1112 /*
1113  * Release a reference on a PFS
1114  */
1115 void
1116 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1117 {
1118         hammer_unref(&pfsm->lock);
1119         if (pfsm->lock.refs == 0) {
1120                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1121                 kfree(pfsm, hmp->m_misc);
1122         }
1123 }
1124
1125 /*
1126  * Called by hammer_sync_inode().
1127  */
1128 static int
1129 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1130 {
1131         hammer_transaction_t trans = cursor->trans;
1132         hammer_record_t record;
1133         int error;
1134         int redirty;
1135
1136 retry:
1137         error = 0;
1138
1139         /*
1140          * If the inode has a presence on-disk then locate it and mark
1141          * it deleted, setting DELONDISK.
1142          *
1143          * The record may or may not be physically deleted, depending on
1144          * the retention policy.
1145          */
1146         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1147             HAMMER_INODE_ONDISK) {
1148                 hammer_normalize_cursor(cursor);
1149                 cursor->key_beg.localization = ip->obj_localization +
1150                                                HAMMER_LOCALIZE_INODE;
1151                 cursor->key_beg.obj_id = ip->obj_id;
1152                 cursor->key_beg.key = 0;
1153                 cursor->key_beg.create_tid = 0;
1154                 cursor->key_beg.delete_tid = 0;
1155                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1156                 cursor->key_beg.obj_type = 0;
1157                 cursor->asof = ip->obj_asof;
1158                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1159                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1160                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1161
1162                 error = hammer_btree_lookup(cursor);
1163                 if (hammer_debug_inode)
1164                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1165
1166                 if (error == 0) {
1167                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1168                         if (hammer_debug_inode)
1169                                 kprintf(" error %d\n", error);
1170                         if (error == 0) {
1171                                 ip->flags |= HAMMER_INODE_DELONDISK;
1172                         }
1173                         if (cursor->node)
1174                                 hammer_cache_node(&ip->cache[0], cursor->node);
1175                 }
1176                 if (error == EDEADLK) {
1177                         hammer_done_cursor(cursor);
1178                         error = hammer_init_cursor(trans, cursor,
1179                                                    &ip->cache[0], ip);
1180                         if (hammer_debug_inode)
1181                                 kprintf("IPDED %p %d\n", ip, error);
1182                         if (error == 0)
1183                                 goto retry;
1184                 }
1185         }
1186
1187         /*
1188          * Ok, write out the initial record or a new record (after deleting
1189          * the old one), unless the DELETED flag is set.  This routine will
1190          * clear DELONDISK if it writes out a record.
1191          *
1192          * Update our inode statistics if this is the first application of
1193          * the inode on-disk.
1194          */
1195         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1196                 /*
1197                  * Generate a record and write it to the media.  We clean-up
1198                  * the state before releasing so we do not have to set-up
1199                  * a flush_group.
1200                  */
1201                 record = hammer_alloc_mem_record(ip, 0);
1202                 record->type = HAMMER_MEM_RECORD_INODE;
1203                 record->flush_state = HAMMER_FST_FLUSH;
1204                 record->leaf = ip->sync_ino_leaf;
1205                 record->leaf.base.create_tid = trans->tid;
1206                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1207                 record->leaf.create_ts = trans->time32;
1208                 record->data = (void *)&ip->sync_ino_data;
1209                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1210
1211                 /*
1212                  * If this flag is set we cannot sync the new file size
1213                  * because we haven't finished related truncations.  The
1214                  * inode will be flushed in another flush group to finish
1215                  * the job.
1216                  */
1217                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1218                     ip->sync_ino_data.size != ip->ino_data.size) {
1219                         redirty = 1;
1220                         ip->sync_ino_data.size = ip->ino_data.size;
1221                 } else {
1222                         redirty = 0;
1223                 }
1224
1225                 for (;;) {
1226                         error = hammer_ip_sync_record_cursor(cursor, record);
1227                         if (hammer_debug_inode)
1228                                 kprintf("GENREC %p rec %08x %d\n",
1229                                         ip, record->flags, error);
1230                         if (error != EDEADLK)
1231                                 break;
1232                         hammer_done_cursor(cursor);
1233                         error = hammer_init_cursor(trans, cursor,
1234                                                    &ip->cache[0], ip);
1235                         if (hammer_debug_inode)
1236                                 kprintf("GENREC reinit %d\n", error);
1237                         if (error)
1238                                 break;
1239                 }
1240
1241                 /*
1242                  * Note:  The record was never on the inode's record tree
1243                  * so just wave our hands importantly and destroy it.
1244                  */
1245                 record->flags |= HAMMER_RECF_COMMITTED;
1246                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1247                 record->flush_state = HAMMER_FST_IDLE;
1248                 ++ip->rec_generation;
1249                 hammer_rel_mem_record(record);
1250
1251                 /*
1252                  * Finish up.
1253                  */
1254                 if (error == 0) {
1255                         if (hammer_debug_inode)
1256                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1257                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1258                                             HAMMER_INODE_ATIME |
1259                                             HAMMER_INODE_MTIME);
1260                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1261                         if (redirty)
1262                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1263
1264                         /*
1265                          * Root volume count of inodes
1266                          */
1267                         hammer_sync_lock_sh(trans);
1268                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1269                                 hammer_modify_volume_field(trans,
1270                                                            trans->rootvol,
1271                                                            vol0_stat_inodes);
1272                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1273                                 hammer_modify_volume_done(trans->rootvol);
1274                                 ip->flags |= HAMMER_INODE_ONDISK;
1275                                 if (hammer_debug_inode)
1276                                         kprintf("NOWONDISK %p\n", ip);
1277                         }
1278                         hammer_sync_unlock(trans);
1279                 }
1280         }
1281
1282         /*
1283          * If the inode has been destroyed, clean out any left-over flags
1284          * that may have been set by the frontend.
1285          */
1286         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1287                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1288                                     HAMMER_INODE_ATIME |
1289                                     HAMMER_INODE_MTIME);
1290         }
1291         return(error);
1292 }
1293
1294 /*
1295  * Update only the itimes fields.
1296  *
1297  * ATIME can be updated without generating any UNDO.  MTIME is updated
1298  * with UNDO so it is guaranteed to be synchronized properly in case of
1299  * a crash.
1300  *
1301  * Neither field is included in the B-Tree leaf element's CRC, which is how
1302  * we can get away with updating ATIME the way we do.
1303  */
1304 static int
1305 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1306 {
1307         hammer_transaction_t trans = cursor->trans;
1308         int error;
1309
1310 retry:
1311         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1312             HAMMER_INODE_ONDISK) {
1313                 return(0);
1314         }
1315
1316         hammer_normalize_cursor(cursor);
1317         cursor->key_beg.localization = ip->obj_localization +
1318                                        HAMMER_LOCALIZE_INODE;
1319         cursor->key_beg.obj_id = ip->obj_id;
1320         cursor->key_beg.key = 0;
1321         cursor->key_beg.create_tid = 0;
1322         cursor->key_beg.delete_tid = 0;
1323         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1324         cursor->key_beg.obj_type = 0;
1325         cursor->asof = ip->obj_asof;
1326         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1327         cursor->flags |= HAMMER_CURSOR_ASOF;
1328         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1329         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1330         cursor->flags |= HAMMER_CURSOR_BACKEND;
1331
1332         error = hammer_btree_lookup(cursor);
1333         if (error == 0) {
1334                 hammer_cache_node(&ip->cache[0], cursor->node);
1335                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1336                         /*
1337                          * Updating MTIME requires an UNDO.  Just cover
1338                          * both atime and mtime.
1339                          */
1340                         hammer_sync_lock_sh(trans);
1341                         hammer_modify_buffer(trans, cursor->data_buffer,
1342                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1343                                      HAMMER_ITIMES_BYTES);
1344                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1345                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1346                         hammer_modify_buffer_done(cursor->data_buffer);
1347                         hammer_sync_unlock(trans);
1348                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1349                         /*
1350                          * Updating atime only can be done in-place with
1351                          * no UNDO.
1352                          */
1353                         hammer_sync_lock_sh(trans);
1354                         hammer_modify_buffer(trans, cursor->data_buffer,
1355                                              NULL, 0);
1356                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1357                         hammer_modify_buffer_done(cursor->data_buffer);
1358                         hammer_sync_unlock(trans);
1359                 }
1360                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1361         }
1362         if (error == EDEADLK) {
1363                 hammer_done_cursor(cursor);
1364                 error = hammer_init_cursor(trans, cursor,
1365                                            &ip->cache[0], ip);
1366                 if (error == 0)
1367                         goto retry;
1368         }
1369         return(error);
1370 }
1371
1372 /*
1373  * Release a reference on an inode, flush as requested.
1374  *
1375  * On the last reference we queue the inode to the flusher for its final
1376  * disposition.
1377  */
1378 void
1379 hammer_rel_inode(struct hammer_inode *ip, int flush)
1380 {
1381         /*hammer_mount_t hmp = ip->hmp;*/
1382
1383         /*
1384          * Handle disposition when dropping the last ref.
1385          */
1386         for (;;) {
1387                 if (ip->lock.refs == 1) {
1388                         /*
1389                          * Determine whether on-disk action is needed for
1390                          * the inode's final disposition.
1391                          */
1392                         KKASSERT(ip->vp == NULL);
1393                         hammer_inode_unloadable_check(ip, 0);
1394                         if (ip->flags & HAMMER_INODE_MODMASK) {
1395                                 hammer_flush_inode(ip, 0);
1396                         } else if (ip->lock.refs == 1) {
1397                                 hammer_unload_inode(ip);
1398                                 break;
1399                         }
1400                 } else {
1401                         if (flush)
1402                                 hammer_flush_inode(ip, 0);
1403
1404                         /*
1405                          * The inode still has multiple refs, try to drop
1406                          * one ref.
1407                          */
1408                         KKASSERT(ip->lock.refs >= 1);
1409                         if (ip->lock.refs > 1) {
1410                                 hammer_unref(&ip->lock);
1411                                 break;
1412                         }
1413                 }
1414         }
1415 }
1416
1417 /*
1418  * Unload and destroy the specified inode.  Must be called with one remaining
1419  * reference.  The reference is disposed of.
1420  *
1421  * The inode must be completely clean.
1422  */
1423 static int
1424 hammer_unload_inode(struct hammer_inode *ip)
1425 {
1426         hammer_mount_t hmp = ip->hmp;
1427
1428         KASSERT(ip->lock.refs == 1,
1429                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1430         KKASSERT(ip->vp == NULL);
1431         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1432         KKASSERT(ip->cursor_ip_refs == 0);
1433         KKASSERT(ip->lock.lockcount == 0);
1434         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1435
1436         KKASSERT(RB_EMPTY(&ip->rec_tree));
1437         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1438
1439         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1440
1441         hammer_free_inode(ip);
1442         return(0);
1443 }
1444
1445 /*
1446  * Called during unmounting if a critical error occured.  The in-memory
1447  * inode and all related structures are destroyed.
1448  *
1449  * If a critical error did not occur the unmount code calls the standard
1450  * release and asserts that the inode is gone.
1451  */
1452 int
1453 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1454 {
1455         hammer_record_t rec;
1456
1457         /*
1458          * Get rid of the inodes in-memory records, regardless of their
1459          * state, and clear the mod-mask.
1460          */
1461         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1462                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1463                 rec->target_ip = NULL;
1464                 if (rec->flush_state == HAMMER_FST_SETUP)
1465                         rec->flush_state = HAMMER_FST_IDLE;
1466         }
1467         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1468                 if (rec->flush_state == HAMMER_FST_FLUSH)
1469                         --rec->flush_group->refs;
1470                 else
1471                         hammer_ref(&rec->lock);
1472                 KKASSERT(rec->lock.refs == 1);
1473                 rec->flush_state = HAMMER_FST_IDLE;
1474                 rec->flush_group = NULL;
1475                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1476                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1477                 ++ip->rec_generation;
1478                 hammer_rel_mem_record(rec);
1479         }
1480         ip->flags &= ~HAMMER_INODE_MODMASK;
1481         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1482         KKASSERT(ip->vp == NULL);
1483
1484         /*
1485          * Remove the inode from any flush group, force it idle.  FLUSH
1486          * and SETUP states have an inode ref.
1487          */
1488         switch(ip->flush_state) {
1489         case HAMMER_FST_FLUSH:
1490                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1491                 --ip->flush_group->refs;
1492                 ip->flush_group = NULL;
1493                 /* fall through */
1494         case HAMMER_FST_SETUP:
1495                 hammer_unref(&ip->lock);
1496                 ip->flush_state = HAMMER_FST_IDLE;
1497                 /* fall through */
1498         case HAMMER_FST_IDLE:
1499                 break;
1500         }
1501
1502         /*
1503          * There shouldn't be any associated vnode.  The unload needs at
1504          * least one ref, if we do have a vp steal its ip ref.
1505          */
1506         if (ip->vp) {
1507                 kprintf("hammer_destroy_inode_callback: Unexpected "
1508                         "vnode association ip %p vp %p\n", ip, ip->vp);
1509                 ip->vp->v_data = NULL;
1510                 ip->vp = NULL;
1511         } else {
1512                 hammer_ref(&ip->lock);
1513         }
1514         hammer_unload_inode(ip);
1515         return(0);
1516 }
1517
1518 /*
1519  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1520  * the read-only flag for cached inodes.
1521  *
1522  * This routine is called from a RB_SCAN().
1523  */
1524 int
1525 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1526 {
1527         hammer_mount_t hmp = ip->hmp;
1528
1529         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1530                 ip->flags |= HAMMER_INODE_RO;
1531         else
1532                 ip->flags &= ~HAMMER_INODE_RO;
1533         return(0);
1534 }
1535
1536 /*
1537  * A transaction has modified an inode, requiring updates as specified by
1538  * the passed flags.
1539  *
1540  * HAMMER_INODE_DDIRTY: Inode data has been updated
1541  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1542  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1543  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1544  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1545  */
1546 void
1547 hammer_modify_inode(hammer_inode_t ip, int flags)
1548 {
1549         /*
1550          * ronly of 0 or 2 does not trigger assertion.
1551          * 2 is a special error state
1552          */
1553         KKASSERT(ip->hmp->ronly != 1 ||
1554                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1555                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1556                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1557         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1558                 ip->flags |= HAMMER_INODE_RSV_INODES;
1559                 ++ip->hmp->rsv_inodes;
1560         }
1561
1562         ip->flags |= flags;
1563 }
1564
1565 /*
1566  * Request that an inode be flushed.  This whole mess cannot block and may
1567  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1568  * actively flush the inode until the flush can be done.
1569  *
1570  * The inode may already be flushing, or may be in a setup state.  We can
1571  * place the inode in a flushing state if it is currently idle and flag it
1572  * to reflush if it is currently flushing.
1573  *
1574  * Upon return if the inode could not be flushed due to a setup
1575  * dependancy, then it will be automatically flushed when the dependancy
1576  * is satisfied.
1577  */
1578 void
1579 hammer_flush_inode(hammer_inode_t ip, int flags)
1580 {
1581         hammer_mount_t hmp;
1582         hammer_flush_group_t flg;
1583         int good;
1584
1585         /*
1586          * next_flush_group is the first flush group we can place the inode
1587          * in.  It may be NULL.  If it becomes full we append a new flush
1588          * group and make that the next_flush_group.
1589          */
1590         hmp = ip->hmp;
1591         while ((flg = hmp->next_flush_group) != NULL) {
1592                 KKASSERT(flg->running == 0);
1593                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1594                         break;
1595                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1596                 hammer_flusher_async(ip->hmp, flg);
1597         }
1598         if (flg == NULL) {
1599                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1600                 hmp->next_flush_group = flg;
1601                 TAILQ_INIT(&flg->flush_list);
1602                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1603         }
1604
1605         /*
1606          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1607          * state we have to put it back into an IDLE state so we can
1608          * drop the extra ref.
1609          *
1610          * If we have a parent dependancy we must still fall through
1611          * so we can run it.
1612          */
1613         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1614                 if (ip->flush_state == HAMMER_FST_SETUP &&
1615                     TAILQ_EMPTY(&ip->target_list)) {
1616                         ip->flush_state = HAMMER_FST_IDLE;
1617                         hammer_rel_inode(ip, 0);
1618                 }
1619                 if (ip->flush_state == HAMMER_FST_IDLE)
1620                         return;
1621         }
1622
1623         /*
1624          * Our flush action will depend on the current state.
1625          */
1626         switch(ip->flush_state) {
1627         case HAMMER_FST_IDLE:
1628                 /*
1629                  * We have no dependancies and can flush immediately.  Some
1630                  * our children may not be flushable so we have to re-test
1631                  * with that additional knowledge.
1632                  */
1633                 hammer_flush_inode_core(ip, flg, flags);
1634                 break;
1635         case HAMMER_FST_SETUP:
1636                 /*
1637                  * Recurse upwards through dependancies via target_list
1638                  * and start their flusher actions going if possible.
1639                  *
1640                  * 'good' is our connectivity.  -1 means we have none and
1641                  * can't flush, 0 means there weren't any dependancies, and
1642                  * 1 means we have good connectivity.
1643                  */
1644                 good = hammer_setup_parent_inodes(ip, 0, flg);
1645
1646                 if (good >= 0) {
1647                         /*
1648                          * We can continue if good >= 0.  Determine how
1649                          * many records under our inode can be flushed (and
1650                          * mark them).
1651                          */
1652                         hammer_flush_inode_core(ip, flg, flags);
1653                 } else {
1654                         /*
1655                          * Parent has no connectivity, tell it to flush
1656                          * us as soon as it does.
1657                          *
1658                          * The REFLUSH flag is also needed to trigger
1659                          * dependancy wakeups.
1660                          */
1661                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1662                                      HAMMER_INODE_REFLUSH;
1663                         if (flags & HAMMER_FLUSH_SIGNAL) {
1664                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1665                                 hammer_flusher_async(ip->hmp, flg);
1666                         }
1667                 }
1668                 break;
1669         case HAMMER_FST_FLUSH:
1670                 /*
1671                  * We are already flushing, flag the inode to reflush
1672                  * if needed after it completes its current flush.
1673                  *
1674                  * The REFLUSH flag is also needed to trigger
1675                  * dependancy wakeups.
1676                  */
1677                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1678                         ip->flags |= HAMMER_INODE_REFLUSH;
1679                 if (flags & HAMMER_FLUSH_SIGNAL) {
1680                         ip->flags |= HAMMER_INODE_RESIGNAL;
1681                         hammer_flusher_async(ip->hmp, flg);
1682                 }
1683                 break;
1684         }
1685 }
1686
1687 /*
1688  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1689  * ip which reference our ip.
1690  *
1691  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1692  *     so for now do not ref/deref the structures.  Note that if we use the
1693  *     ref/rel code later, the rel CAN block.
1694  */
1695 static int
1696 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1697                            hammer_flush_group_t flg)
1698 {
1699         hammer_record_t depend;
1700         int good;
1701         int r;
1702
1703         /*
1704          * If we hit our recursion limit and we have parent dependencies
1705          * We cannot continue.  Returning < 0 will cause us to be flagged
1706          * for reflush.  Returning -2 cuts off additional dependency checks
1707          * because they are likely to also hit the depth limit.
1708          *
1709          * We cannot return < 0 if there are no dependencies or there might
1710          * not be anything to wakeup (ip).
1711          */
1712         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1713                 kprintf("HAMMER Warning: depth limit reached on "
1714                         "setup recursion, inode %p %016llx\n",
1715                         ip, (long long)ip->obj_id);
1716                 return(-2);
1717         }
1718
1719         /*
1720          * Scan dependencies
1721          */
1722         good = 0;
1723         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1724                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1725                 KKASSERT(depend->target_ip == ip);
1726                 if (r < 0 && good == 0)
1727                         good = -1;
1728                 if (r > 0)
1729                         good = 1;
1730
1731                 /*
1732                  * If we failed due to the recursion depth limit then stop
1733                  * now.
1734                  */
1735                 if (r == -2)
1736                         break;
1737         }
1738         return(good);
1739 }
1740
1741 /*
1742  * This helper function takes a record representing the dependancy between
1743  * the parent inode and child inode.
1744  *
1745  * record->ip           = parent inode
1746  * record->target_ip    = child inode
1747  *
1748  * We are asked to recurse upwards and convert the record from SETUP
1749  * to FLUSH if possible.
1750  *
1751  * Return 1 if the record gives us connectivity
1752  *
1753  * Return 0 if the record is not relevant
1754  *
1755  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1756  */
1757 static int
1758 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1759                                   hammer_flush_group_t flg)
1760 {
1761         hammer_mount_t hmp;
1762         hammer_inode_t pip;
1763         int good;
1764
1765         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1766         pip = record->ip;
1767         hmp = pip->hmp;
1768
1769         /*
1770          * If the record is already flushing, is it in our flush group?
1771          *
1772          * If it is in our flush group but it is a general record or a
1773          * delete-on-disk, it does not improve our connectivity (return 0),
1774          * and if the target inode is not trying to destroy itself we can't
1775          * allow the operation yet anyway (the second return -1).
1776          */
1777         if (record->flush_state == HAMMER_FST_FLUSH) {
1778                 /*
1779                  * If not in our flush group ask the parent to reflush
1780                  * us as soon as possible.
1781                  */
1782                 if (record->flush_group != flg) {
1783                         pip->flags |= HAMMER_INODE_REFLUSH;
1784                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1785                         return(-1);
1786                 }
1787
1788                 /*
1789                  * If in our flush group everything is already set up,
1790                  * just return whether the record will improve our
1791                  * visibility or not.
1792                  */
1793                 if (record->type == HAMMER_MEM_RECORD_ADD)
1794                         return(1);
1795                 return(0);
1796         }
1797
1798         /*
1799          * It must be a setup record.  Try to resolve the setup dependancies
1800          * by recursing upwards so we can place ip on the flush list.
1801          *
1802          * Limit ourselves to 20 levels of recursion to avoid blowing out
1803          * the kernel stack.  If we hit the recursion limit we can't flush
1804          * until the parent flushes.  The parent will flush independantly
1805          * on its own and ultimately a deep recursion will be resolved.
1806          */
1807         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1808
1809         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1810
1811         /*
1812          * If good < 0 the parent has no connectivity and we cannot safely
1813          * flush the directory entry, which also means we can't flush our
1814          * ip.  Flag us for downward recursion once the parent's
1815          * connectivity is resolved.  Flag the parent for [re]flush or it
1816          * may not check for downward recursions.
1817          */
1818         if (good < 0) {
1819                 pip->flags |= HAMMER_INODE_REFLUSH;
1820                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1821                 return(good);
1822         }
1823
1824         /*
1825          * We are go, place the parent inode in a flushing state so we can
1826          * place its record in a flushing state.  Note that the parent
1827          * may already be flushing.  The record must be in the same flush
1828          * group as the parent.
1829          */
1830         if (pip->flush_state != HAMMER_FST_FLUSH)
1831                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1832         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1833         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1834
1835 #if 0
1836         if (record->type == HAMMER_MEM_RECORD_DEL &&
1837             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1838                 /*
1839                  * Regardless of flushing state we cannot sync this path if the
1840                  * record represents a delete-on-disk but the target inode
1841                  * is not ready to sync its own deletion.
1842                  *
1843                  * XXX need to count effective nlinks to determine whether
1844                  * the flush is ok, otherwise removing a hardlink will
1845                  * just leave the DEL record to rot.
1846                  */
1847                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1848                 return(-1);
1849         } else
1850 #endif
1851         if (pip->flush_group == flg) {
1852                 /*
1853                  * Because we have not calculated nlinks yet we can just
1854                  * set records to the flush state if the parent is in
1855                  * the same flush group as we are.
1856                  */
1857                 record->flush_state = HAMMER_FST_FLUSH;
1858                 record->flush_group = flg;
1859                 ++record->flush_group->refs;
1860                 hammer_ref(&record->lock);
1861
1862                 /*
1863                  * A general directory-add contributes to our visibility.
1864                  *
1865                  * Otherwise it is probably a directory-delete or
1866                  * delete-on-disk record and does not contribute to our
1867                  * visbility (but we can still flush it).
1868                  */
1869                 if (record->type == HAMMER_MEM_RECORD_ADD)
1870                         return(1);
1871                 return(0);
1872         } else {
1873                 /*
1874                  * If the parent is not in our flush group we cannot
1875                  * flush this record yet, there is no visibility.
1876                  * We tell the parent to reflush and mark ourselves
1877                  * so the parent knows it should flush us too.
1878                  */
1879                 pip->flags |= HAMMER_INODE_REFLUSH;
1880                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1881                 return(-1);
1882         }
1883 }
1884
1885 /*
1886  * This is the core routine placing an inode into the FST_FLUSH state.
1887  */
1888 static void
1889 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1890 {
1891         int go_count;
1892
1893         /*
1894          * Set flush state and prevent the flusher from cycling into
1895          * the next flush group.  Do not place the ip on the list yet.
1896          * Inodes not in the idle state get an extra reference.
1897          */
1898         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1899         if (ip->flush_state == HAMMER_FST_IDLE)
1900                 hammer_ref(&ip->lock);
1901         ip->flush_state = HAMMER_FST_FLUSH;
1902         ip->flush_group = flg;
1903         ++ip->hmp->flusher.group_lock;
1904         ++ip->hmp->count_iqueued;
1905         ++hammer_count_iqueued;
1906         ++flg->total_count;
1907
1908         /*
1909          * If the flush group reaches the autoflush limit we want to signal
1910          * the flusher.  This is particularly important for remove()s.
1911          */
1912         if (flg->total_count == hammer_autoflush)
1913                 flags |= HAMMER_FLUSH_SIGNAL;
1914
1915         /*
1916          * We need to be able to vfsync/truncate from the backend.
1917          */
1918         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1919         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1920                 ip->flags |= HAMMER_INODE_VHELD;
1921                 vref(ip->vp);
1922         }
1923
1924         /*
1925          * Figure out how many in-memory records we can actually flush
1926          * (not including inode meta-data, buffers, etc).
1927          */
1928         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1929         if (flags & HAMMER_FLUSH_RECURSION) {
1930                 /*
1931                  * If this is a upwards recursion we do not want to
1932                  * recurse down again!
1933                  */
1934                 go_count = 1;
1935 #if 0
1936         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1937                 /*
1938                  * No new records are added if we must complete a flush
1939                  * from a previous cycle, but we do have to move the records
1940                  * from the previous cycle to the current one.
1941                  */
1942 #if 0
1943                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1944                                    hammer_syncgrp_child_callback, NULL);
1945 #endif
1946                 go_count = 1;
1947 #endif
1948         } else {
1949                 /*
1950                  * Normal flush, scan records and bring them into the flush.
1951                  * Directory adds and deletes are usually skipped (they are
1952                  * grouped with the related inode rather then with the
1953                  * directory).
1954                  *
1955                  * go_count can be negative, which means the scan aborted
1956                  * due to the flush group being over-full and we should
1957                  * flush what we have.
1958                  */
1959                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1960                                    hammer_setup_child_callback, NULL);
1961         }
1962
1963         /*
1964          * This is a more involved test that includes go_count.  If we
1965          * can't flush, flag the inode and return.  If go_count is 0 we
1966          * were are unable to flush any records in our rec_tree and
1967          * must ignore the XDIRTY flag.
1968          */
1969         if (go_count == 0) {
1970                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1971                         --ip->hmp->count_iqueued;
1972                         --hammer_count_iqueued;
1973
1974                         --flg->total_count;
1975                         ip->flush_state = HAMMER_FST_SETUP;
1976                         ip->flush_group = NULL;
1977                         if (ip->flags & HAMMER_INODE_VHELD) {
1978                                 ip->flags &= ~HAMMER_INODE_VHELD;
1979                                 vrele(ip->vp);
1980                         }
1981
1982                         /*
1983                          * REFLUSH is needed to trigger dependancy wakeups
1984                          * when an inode is in SETUP.
1985                          */
1986                         ip->flags |= HAMMER_INODE_REFLUSH;
1987                         if (flags & HAMMER_FLUSH_SIGNAL) {
1988                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1989                                 hammer_flusher_async(ip->hmp, flg);
1990                         }
1991                         if (--ip->hmp->flusher.group_lock == 0)
1992                                 wakeup(&ip->hmp->flusher.group_lock);
1993                         return;
1994                 }
1995         }
1996
1997         /*
1998          * Snapshot the state of the inode for the backend flusher.
1999          *
2000          * We continue to retain save_trunc_off even when all truncations
2001          * have been resolved as an optimization to determine if we can
2002          * skip the B-Tree lookup for overwrite deletions.
2003          *
2004          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2005          * and stays in ip->flags.  Once set, it stays set until the
2006          * inode is destroyed.
2007          */
2008         if (ip->flags & HAMMER_INODE_TRUNCATED) {
2009                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2010                 ip->sync_trunc_off = ip->trunc_off;
2011                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2012                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2013                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2014
2015                 /*
2016                  * The save_trunc_off used to cache whether the B-Tree
2017                  * holds any records past that point is not used until
2018                  * after the truncation has succeeded, so we can safely
2019                  * set it now.
2020                  */
2021                 if (ip->save_trunc_off > ip->sync_trunc_off)
2022                         ip->save_trunc_off = ip->sync_trunc_off;
2023         }
2024         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2025                            ~HAMMER_INODE_TRUNCATED);
2026         ip->sync_ino_leaf = ip->ino_leaf;
2027         ip->sync_ino_data = ip->ino_data;
2028         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2029 #ifdef DEBUG_TRUNCATE
2030         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2031                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2032 #endif
2033
2034         /*
2035          * The flusher list inherits our inode and reference.
2036          */
2037         KKASSERT(flg->running == 0);
2038         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
2039         if (--ip->hmp->flusher.group_lock == 0)
2040                 wakeup(&ip->hmp->flusher.group_lock);
2041
2042         if (flags & HAMMER_FLUSH_SIGNAL) {
2043                 hammer_flusher_async(ip->hmp, flg);
2044         }
2045 }
2046
2047 /*
2048  * Callback for scan of ip->rec_tree.  Try to include each record in our
2049  * flush.  ip->flush_group has been set but the inode has not yet been
2050  * moved into a flushing state.
2051  *
2052  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2053  * both inodes.
2054  *
2055  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2056  * the caller from shortcutting the flush.
2057  */
2058 static int
2059 hammer_setup_child_callback(hammer_record_t rec, void *data)
2060 {
2061         hammer_flush_group_t flg;
2062         hammer_inode_t target_ip;
2063         hammer_inode_t ip;
2064         int r;
2065
2066         /*
2067          * Records deleted or committed by the backend are ignored.
2068          * Note that the flush detects deleted frontend records at
2069          * multiple points to deal with races.  This is just the first
2070          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2071          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2072          * messes up link-count calculations.
2073          *
2074          * NOTE: Don't get confused between record deletion and, say,
2075          * directory entry deletion.  The deletion of a directory entry
2076          * which is on-media has nothing to do with the record deletion
2077          * flags.
2078          */
2079         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2080                           HAMMER_RECF_COMMITTED)) {
2081                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2082                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2083                         r = 1;
2084                 } else {
2085                         r = 0;
2086                 }
2087                 return(r);
2088         }
2089
2090         /*
2091          * If the record is in an idle state it has no dependancies and
2092          * can be flushed.
2093          */
2094         ip = rec->ip;
2095         flg = ip->flush_group;
2096         r = 0;
2097
2098         switch(rec->flush_state) {
2099         case HAMMER_FST_IDLE:
2100                 /*
2101                  * The record has no setup dependancy, we can flush it.
2102                  */
2103                 KKASSERT(rec->target_ip == NULL);
2104                 rec->flush_state = HAMMER_FST_FLUSH;
2105                 rec->flush_group = flg;
2106                 ++flg->refs;
2107                 hammer_ref(&rec->lock);
2108                 r = 1;
2109                 break;
2110         case HAMMER_FST_SETUP:
2111                 /*
2112                  * The record has a setup dependancy.  These are typically
2113                  * directory entry adds and deletes.  Such entries will be
2114                  * flushed when their inodes are flushed so we do not
2115                  * usually have to add them to the flush here.  However,
2116                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2117                  * it is asking us to flush this record (and it).
2118                  */
2119                 target_ip = rec->target_ip;
2120                 KKASSERT(target_ip != NULL);
2121                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2122
2123                 /*
2124                  * If the target IP is already flushing in our group
2125                  * we could associate the record, but target_ip has
2126                  * already synced ino_data to sync_ino_data and we
2127                  * would also have to adjust nlinks.   Plus there are
2128                  * ordering issues for adds and deletes.
2129                  *
2130                  * Reflush downward if this is an ADD, and upward if
2131                  * this is a DEL.
2132                  */
2133                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2134                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2135                                 ip->flags |= HAMMER_INODE_REFLUSH;
2136                         else
2137                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2138                         break;
2139                 }
2140
2141                 /*
2142                  * Target IP is not yet flushing.  This can get complex
2143                  * because we have to be careful about the recursion.
2144                  *
2145                  * Directories create an issue for us in that if a flush
2146                  * of a directory is requested the expectation is to flush
2147                  * any pending directory entries, but this will cause the
2148                  * related inodes to recursively flush as well.  We can't
2149                  * really defer the operation so just get as many as we
2150                  * can and
2151                  */
2152 #if 0
2153                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2154                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2155                         /*
2156                          * We aren't reclaiming and the target ip was not
2157                          * previously prevented from flushing due to this
2158                          * record dependancy.  Do not flush this record.
2159                          */
2160                         /*r = 0;*/
2161                 } else
2162 #endif
2163                 if (flg->total_count + flg->refs >
2164                            ip->hmp->undo_rec_limit) {
2165                         /*
2166                          * Our flush group is over-full and we risk blowing
2167                          * out the UNDO FIFO.  Stop the scan, flush what we
2168                          * have, then reflush the directory.
2169                          *
2170                          * The directory may be forced through multiple
2171                          * flush groups before it can be completely
2172                          * flushed.
2173                          */
2174                         ip->flags |= HAMMER_INODE_RESIGNAL |
2175                                      HAMMER_INODE_REFLUSH;
2176                         r = -1;
2177                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2178                         /*
2179                          * If the target IP is not flushing we can force
2180                          * it to flush, even if it is unable to write out
2181                          * any of its own records we have at least one in
2182                          * hand that we CAN deal with.
2183                          */
2184                         rec->flush_state = HAMMER_FST_FLUSH;
2185                         rec->flush_group = flg;
2186                         ++flg->refs;
2187                         hammer_ref(&rec->lock);
2188                         hammer_flush_inode_core(target_ip, flg,
2189                                                 HAMMER_FLUSH_RECURSION);
2190                         r = 1;
2191                 } else {
2192                         /*
2193                          * General or delete-on-disk record.
2194                          *
2195                          * XXX this needs help.  If a delete-on-disk we could
2196                          * disconnect the target.  If the target has its own
2197                          * dependancies they really need to be flushed.
2198                          *
2199                          * XXX
2200                          */
2201                         rec->flush_state = HAMMER_FST_FLUSH;
2202                         rec->flush_group = flg;
2203                         ++flg->refs;
2204                         hammer_ref(&rec->lock);
2205                         hammer_flush_inode_core(target_ip, flg,
2206                                                 HAMMER_FLUSH_RECURSION);
2207                         r = 1;
2208                 }
2209                 break;
2210         case HAMMER_FST_FLUSH:
2211                 /*
2212                  * The flush_group should already match.
2213                  */
2214                 KKASSERT(rec->flush_group == flg);
2215                 r = 1;
2216                 break;
2217         }
2218         return(r);
2219 }
2220
2221 #if 0
2222 /*
2223  * This version just moves records already in a flush state to the new
2224  * flush group and that is it.
2225  */
2226 static int
2227 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2228 {
2229         hammer_inode_t ip = rec->ip;
2230
2231         switch(rec->flush_state) {
2232         case HAMMER_FST_FLUSH:
2233                 KKASSERT(rec->flush_group == ip->flush_group);
2234                 break;
2235         default:
2236                 break;
2237         }
2238         return(0);
2239 }
2240 #endif
2241
2242 /*
2243  * Wait for a previously queued flush to complete.
2244  *
2245  * If a critical error occured we don't try to wait.
2246  */
2247 void
2248 hammer_wait_inode(hammer_inode_t ip)
2249 {
2250         hammer_flush_group_t flg;
2251
2252         flg = NULL;
2253         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2254                 while (ip->flush_state != HAMMER_FST_IDLE &&
2255                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2256                         if (ip->flush_state == HAMMER_FST_SETUP)
2257                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2258                         if (ip->flush_state != HAMMER_FST_IDLE) {
2259                                 ip->flags |= HAMMER_INODE_FLUSHW;
2260                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2261                         }
2262                 }
2263         }
2264 }
2265
2266 /*
2267  * Called by the backend code when a flush has been completed.
2268  * The inode has already been removed from the flush list.
2269  *
2270  * A pipelined flush can occur, in which case we must re-enter the
2271  * inode on the list and re-copy its fields.
2272  */
2273 void
2274 hammer_flush_inode_done(hammer_inode_t ip, int error)
2275 {
2276         hammer_mount_t hmp;
2277         int dorel;
2278
2279         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2280
2281         hmp = ip->hmp;
2282
2283         /*
2284          * Auto-reflush if the backend could not completely flush
2285          * the inode.  This fixes a case where a deferred buffer flush
2286          * could cause fsync to return early.
2287          */
2288         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2289                 ip->flags |= HAMMER_INODE_REFLUSH;
2290
2291         /*
2292          * Merge left-over flags back into the frontend and fix the state.
2293          * Incomplete truncations are retained by the backend.
2294          */
2295         ip->error = error;
2296         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2297         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2298
2299         /*
2300          * The backend may have adjusted nlinks, so if the adjusted nlinks
2301          * does not match the fronttend set the frontend's RDIRTY flag again.
2302          */
2303         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2304                 ip->flags |= HAMMER_INODE_DDIRTY;
2305
2306         /*
2307          * Fix up the dirty buffer status.
2308          */
2309         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2310                 ip->flags |= HAMMER_INODE_BUFS;
2311         }
2312
2313         /*
2314          * Re-set the XDIRTY flag if some of the inode's in-memory records
2315          * could not be flushed.
2316          */
2317         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2318                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2319                  (!RB_EMPTY(&ip->rec_tree) &&
2320                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2321
2322         /*
2323          * Do not lose track of inodes which no longer have vnode
2324          * assocations, otherwise they may never get flushed again.
2325          *
2326          * The reflush flag can be set superfluously, causing extra pain
2327          * for no reason.  If the inode is no longer modified it no longer
2328          * needs to be flushed.
2329          */
2330         if (ip->flags & HAMMER_INODE_MODMASK) {
2331                 if (ip->vp == NULL)
2332                         ip->flags |= HAMMER_INODE_REFLUSH;
2333         } else {
2334                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2335         }
2336
2337         /*
2338          * Adjust the flush state.
2339          */
2340         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2341                 /*
2342                  * We were unable to flush out all our records, leave the
2343                  * inode in a flush state and in the current flush group.
2344                  * The flush group will be re-run.
2345                  *
2346                  * This occurs if the UNDO block gets too full or there is
2347                  * too much dirty meta-data and allows the flusher to
2348                  * finalize the UNDO block and then re-flush.
2349                  */
2350                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2351                 dorel = 0;
2352         } else {
2353                 /*
2354                  * Remove from the flush_group
2355                  */
2356                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2357                 ip->flush_group = NULL;
2358
2359                 /*
2360                  * Clean up the vnode ref and tracking counts.
2361                  */
2362                 if (ip->flags & HAMMER_INODE_VHELD) {
2363                         ip->flags &= ~HAMMER_INODE_VHELD;
2364                         vrele(ip->vp);
2365                 }
2366                 --hmp->count_iqueued;
2367                 --hammer_count_iqueued;
2368
2369                 /*
2370                  * And adjust the state.
2371                  */
2372                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2373                         ip->flush_state = HAMMER_FST_IDLE;
2374                         dorel = 1;
2375                 } else {
2376                         ip->flush_state = HAMMER_FST_SETUP;
2377                         dorel = 0;
2378                 }
2379
2380                 /*
2381                  * If the frontend is waiting for a flush to complete,
2382                  * wake it up.
2383                  */
2384                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2385                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2386                         wakeup(&ip->flags);
2387                 }
2388
2389                 /*
2390                  * If the frontend made more changes and requested another
2391                  * flush, then try to get it running.
2392                  *
2393                  * Reflushes are aborted when the inode is errored out.
2394                  */
2395                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2396                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2397                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2398                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2399                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2400                         } else {
2401                                 hammer_flush_inode(ip, 0);
2402                         }
2403                 }
2404         }
2405
2406         /*
2407          * If we have no parent dependancies we can clear CONN_DOWN
2408          */
2409         if (TAILQ_EMPTY(&ip->target_list))
2410                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2411
2412         /*
2413          * If the inode is now clean drop the space reservation.
2414          */
2415         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2416             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2417                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2418                 --hmp->rsv_inodes;
2419         }
2420
2421         if (dorel)
2422                 hammer_rel_inode(ip, 0);
2423 }
2424
2425 /*
2426  * Called from hammer_sync_inode() to synchronize in-memory records
2427  * to the media.
2428  */
2429 static int
2430 hammer_sync_record_callback(hammer_record_t record, void *data)
2431 {
2432         hammer_cursor_t cursor = data;
2433         hammer_transaction_t trans = cursor->trans;
2434         hammer_mount_t hmp = trans->hmp;
2435         int error;
2436
2437         /*
2438          * Skip records that do not belong to the current flush.
2439          */
2440         ++hammer_stats_record_iterations;
2441         if (record->flush_state != HAMMER_FST_FLUSH)
2442                 return(0);
2443
2444 #if 1
2445         if (record->flush_group != record->ip->flush_group) {
2446                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2447                 Debugger("blah2");
2448                 return(0);
2449         }
2450 #endif
2451         KKASSERT(record->flush_group == record->ip->flush_group);
2452
2453         /*
2454          * Interlock the record using the BE flag.  Once BE is set the
2455          * frontend cannot change the state of FE.
2456          *
2457          * NOTE: If FE is set prior to us setting BE we still sync the
2458          * record out, but the flush completion code converts it to
2459          * a delete-on-disk record instead of destroying it.
2460          */
2461         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2462         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2463
2464         /*
2465          * The backend has already disposed of the record.
2466          */
2467         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2468                 error = 0;
2469                 goto done;
2470         }
2471
2472         /*
2473          * If the whole inode is being deleting all on-disk records will
2474          * be deleted very soon, we can't sync any new records to disk
2475          * because they will be deleted in the same transaction they were
2476          * created in (delete_tid == create_tid), which will assert.
2477          *
2478          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2479          * that we currently panic on.
2480          */
2481         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2482                 switch(record->type) {
2483                 case HAMMER_MEM_RECORD_DATA:
2484                         /*
2485                          * We don't have to do anything, if the record was
2486                          * committed the space will have been accounted for
2487                          * in the blockmap.
2488                          */
2489                         /* fall through */
2490                 case HAMMER_MEM_RECORD_GENERAL:
2491                         /*
2492                          * Set deleted-by-backend flag.  Do not set the
2493                          * backend committed flag, because we are throwing
2494                          * the record away.
2495                          */
2496                         record->flags |= HAMMER_RECF_DELETED_BE;
2497                         ++record->ip->rec_generation;
2498                         error = 0;
2499                         goto done;
2500                 case HAMMER_MEM_RECORD_ADD:
2501                         panic("hammer_sync_record_callback: illegal add "
2502                               "during inode deletion record %p", record);
2503                         break; /* NOT REACHED */
2504                 case HAMMER_MEM_RECORD_INODE:
2505                         panic("hammer_sync_record_callback: attempt to "
2506                               "sync inode record %p?", record);
2507                         break; /* NOT REACHED */
2508                 case HAMMER_MEM_RECORD_DEL:
2509                         /*
2510                          * Follow through and issue the on-disk deletion
2511                          */
2512                         break;
2513                 }
2514         }
2515
2516         /*
2517          * If DELETED_FE is set special handling is needed for directory
2518          * entries.  Dependant pieces related to the directory entry may
2519          * have already been synced to disk.  If this occurs we have to
2520          * sync the directory entry and then change the in-memory record
2521          * from an ADD to a DELETE to cover the fact that it's been
2522          * deleted by the frontend.
2523          *
2524          * A directory delete covering record (MEM_RECORD_DEL) can never
2525          * be deleted by the frontend.
2526          *
2527          * Any other record type (aka DATA) can be deleted by the frontend.
2528          * XXX At the moment the flusher must skip it because there may
2529          * be another data record in the flush group for the same block,
2530          * meaning that some frontend data changes can leak into the backend's
2531          * synchronization point.
2532          */
2533         if (record->flags & HAMMER_RECF_DELETED_FE) {
2534                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2535                         /*
2536                          * Convert a front-end deleted directory-add to
2537                          * a directory-delete entry later.
2538                          */
2539                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2540                 } else {
2541                         /*
2542                          * Dispose of the record (race case).  Mark as
2543                          * deleted by backend (and not committed).
2544                          */
2545                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2546                         record->flags |= HAMMER_RECF_DELETED_BE;
2547                         ++record->ip->rec_generation;
2548                         error = 0;
2549                         goto done;
2550                 }
2551         }
2552
2553         /*
2554          * Assign the create_tid for new records.  Deletions already
2555          * have the record's entire key properly set up.
2556          */
2557         if (record->type != HAMMER_MEM_RECORD_DEL) {
2558                 record->leaf.base.create_tid = trans->tid;
2559                 record->leaf.create_ts = trans->time32;
2560         }
2561         for (;;) {
2562                 error = hammer_ip_sync_record_cursor(cursor, record);
2563                 if (error != EDEADLK)
2564                         break;
2565                 hammer_done_cursor(cursor);
2566                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2567                                            record->ip);
2568                 if (error)
2569                         break;
2570         }
2571         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2572
2573         if (error)
2574                 error = -error;
2575 done:
2576         hammer_flush_record_done(record, error);
2577
2578         /*
2579          * Do partial finalization if we have built up too many dirty
2580          * buffers.  Otherwise a buffer cache deadlock can occur when
2581          * doing things like creating tens of thousands of tiny files.
2582          *
2583          * We must release our cursor lock to avoid a 3-way deadlock
2584          * due to the exclusive sync lock the finalizer must get.
2585          */
2586         if (hammer_flusher_meta_limit(hmp)) {
2587                 hammer_unlock_cursor(cursor);
2588                 hammer_flusher_finalize(trans, 0);
2589                 hammer_lock_cursor(cursor);
2590         }
2591
2592         return(error);
2593 }
2594
2595 /*
2596  * Backend function called by the flusher to sync an inode to media.
2597  */
2598 int
2599 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2600 {
2601         struct hammer_cursor cursor;
2602         hammer_node_t tmp_node;
2603         hammer_record_t depend;
2604         hammer_record_t next;
2605         int error, tmp_error;
2606         u_int64_t nlinks;
2607
2608         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2609                 return(0);
2610
2611         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2612         if (error)
2613                 goto done;
2614
2615         /*
2616          * Any directory records referencing this inode which are not in
2617          * our current flush group must adjust our nlink count for the
2618          * purposes of synchronization to disk.
2619          *
2620          * Records which are in our flush group can be unlinked from our
2621          * inode now, potentially allowing the inode to be physically
2622          * deleted.
2623          *
2624          * This cannot block.
2625          */
2626         nlinks = ip->ino_data.nlinks;
2627         next = TAILQ_FIRST(&ip->target_list);
2628         while ((depend = next) != NULL) {
2629                 next = TAILQ_NEXT(depend, target_entry);
2630                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2631                     depend->flush_group == ip->flush_group) {
2632                         /*
2633                          * If this is an ADD that was deleted by the frontend
2634                          * the frontend nlinks count will have already been
2635                          * decremented, but the backend is going to sync its
2636                          * directory entry and must account for it.  The
2637                          * record will be converted to a delete-on-disk when
2638                          * it gets synced.
2639                          *
2640                          * If the ADD was not deleted by the frontend we
2641                          * can remove the dependancy from our target_list.
2642                          */
2643                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2644                                 ++nlinks;
2645                         } else {
2646                                 TAILQ_REMOVE(&ip->target_list, depend,
2647                                              target_entry);
2648                                 depend->target_ip = NULL;
2649                         }
2650                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2651                         /*
2652                          * Not part of our flush group and not deleted by
2653                          * the front-end, adjust the link count synced to
2654                          * the media (undo what the frontend did when it
2655                          * queued the record).
2656                          */
2657                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2658                         switch(depend->type) {
2659                         case HAMMER_MEM_RECORD_ADD:
2660                                 --nlinks;
2661                                 break;
2662                         case HAMMER_MEM_RECORD_DEL:
2663                                 ++nlinks;
2664                                 break;
2665                         default:
2666                                 break;
2667                         }
2668                 }
2669         }
2670
2671         /*
2672          * Set dirty if we had to modify the link count.
2673          */
2674         if (ip->sync_ino_data.nlinks != nlinks) {
2675                 KKASSERT((int64_t)nlinks >= 0);
2676                 ip->sync_ino_data.nlinks = nlinks;
2677                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2678         }
2679
2680         /*
2681          * If there is a trunction queued destroy any data past the (aligned)
2682          * truncation point.  Userland will have dealt with the buffer
2683          * containing the truncation point for us.
2684          *
2685          * We don't flush pending frontend data buffers until after we've
2686          * dealt with the truncation.
2687          */
2688         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2689                 /*
2690                  * Interlock trunc_off.  The VOP front-end may continue to
2691                  * make adjustments to it while we are blocked.
2692                  */
2693                 off_t trunc_off;
2694                 off_t aligned_trunc_off;
2695                 int blkmask;
2696
2697                 trunc_off = ip->sync_trunc_off;
2698                 blkmask = hammer_blocksize(trunc_off) - 1;
2699                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2700
2701                 /*
2702                  * Delete any whole blocks on-media.  The front-end has
2703                  * already cleaned out any partial block and made it
2704                  * pending.  The front-end may have updated trunc_off
2705                  * while we were blocked so we only use sync_trunc_off.
2706                  *
2707                  * This operation can blow out the buffer cache, EWOULDBLOCK
2708                  * means we were unable to complete the deletion.  The
2709                  * deletion will update sync_trunc_off in that case.
2710                  */
2711                 error = hammer_ip_delete_range(&cursor, ip,
2712                                                 aligned_trunc_off,
2713                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2714                 if (error == EWOULDBLOCK) {
2715                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2716                         error = 0;
2717                         goto defer_buffer_flush;
2718                 }
2719
2720                 if (error)
2721                         goto done;
2722
2723                 /*
2724                  * Clear the truncation flag on the backend after we have
2725                  * complete the deletions.  Backend data is now good again
2726                  * (including new records we are about to sync, below).
2727                  *
2728                  * Leave sync_trunc_off intact.  As we write additional
2729                  * records the backend will update sync_trunc_off.  This
2730                  * tells the backend whether it can skip the overwrite
2731                  * test.  This should work properly even when the backend
2732                  * writes full blocks where the truncation point straddles
2733                  * the block because the comparison is against the base
2734                  * offset of the record.
2735                  */
2736                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2737                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2738         } else {
2739                 error = 0;
2740         }
2741
2742         /*
2743          * Now sync related records.  These will typically be directory
2744          * entries, records tracking direct-writes, or delete-on-disk records.
2745          */
2746         if (error == 0) {
2747                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2748                                     hammer_sync_record_callback, &cursor);
2749                 if (tmp_error < 0)
2750                         tmp_error = -error;
2751                 if (tmp_error)
2752                         error = tmp_error;
2753         }
2754         hammer_cache_node(&ip->cache[1], cursor.node);
2755
2756         /*
2757          * Re-seek for inode update, assuming our cache hasn't been ripped
2758          * out from under us.
2759          */
2760         if (error == 0) {
2761                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2762                 if (tmp_node) {
2763                         hammer_cursor_downgrade(&cursor);
2764                         hammer_lock_sh(&tmp_node->lock);
2765                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2766                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2767                         hammer_unlock(&tmp_node->lock);
2768                         hammer_rel_node(tmp_node);
2769                 }
2770                 error = 0;
2771         }
2772
2773         /*
2774          * If we are deleting the inode the frontend had better not have
2775          * any active references on elements making up the inode.
2776          *
2777          * The call to hammer_ip_delete_clean() cleans up auxillary records
2778          * but not DB or DATA records.  Those must have already been deleted
2779          * by the normal truncation mechanic.
2780          */
2781         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2782                 RB_EMPTY(&ip->rec_tree)  &&
2783             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2784             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2785                 int count1 = 0;
2786
2787                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2788                 if (error == 0) {
2789                         ip->flags |= HAMMER_INODE_DELETED;
2790                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2791                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2792                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2793
2794                         /*
2795                          * Set delete_tid in both the frontend and backend
2796                          * copy of the inode record.  The DELETED flag handles
2797                          * this, do not set RDIRTY.
2798                          */
2799                         ip->ino_leaf.base.delete_tid = trans->tid;
2800                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2801                         ip->ino_leaf.delete_ts = trans->time32;
2802                         ip->sync_ino_leaf.delete_ts = trans->time32;
2803
2804
2805                         /*
2806                          * Adjust the inode count in the volume header
2807                          */
2808                         hammer_sync_lock_sh(trans);
2809                         if (ip->flags & HAMMER_INODE_ONDISK) {
2810                                 hammer_modify_volume_field(trans,
2811                                                            trans->rootvol,
2812                                                            vol0_stat_inodes);
2813                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2814                                 hammer_modify_volume_done(trans->rootvol);
2815                         }
2816                         hammer_sync_unlock(trans);
2817                 }
2818         }
2819
2820         if (error)
2821                 goto done;
2822         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2823
2824 defer_buffer_flush:
2825         /*
2826          * Now update the inode's on-disk inode-data and/or on-disk record.
2827          * DELETED and ONDISK are managed only in ip->flags.
2828          *
2829          * In the case of a defered buffer flush we still update the on-disk
2830          * inode to satisfy visibility requirements if there happen to be
2831          * directory dependancies.
2832          */
2833         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2834         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2835                 /*
2836                  * If deleted and on-disk, don't set any additional flags.
2837                  * the delete flag takes care of things.
2838                  *
2839                  * Clear flags which may have been set by the frontend.
2840                  */
2841                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2842                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2843                                     HAMMER_INODE_DELETING);
2844                 break;
2845         case HAMMER_INODE_DELETED:
2846                 /*
2847                  * Take care of the case where a deleted inode was never
2848                  * flushed to the disk in the first place.
2849                  *
2850                  * Clear flags which may have been set by the frontend.
2851                  */
2852                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2853                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2854                                     HAMMER_INODE_DELETING);
2855                 while (RB_ROOT(&ip->rec_tree)) {
2856                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2857                         hammer_ref(&record->lock);
2858                         KKASSERT(record->lock.refs == 1);
2859                         record->flags |= HAMMER_RECF_DELETED_BE;
2860                         ++record->ip->rec_generation;
2861                         hammer_rel_mem_record(record);
2862                 }
2863                 break;
2864         case HAMMER_INODE_ONDISK:
2865                 /*
2866                  * If already on-disk, do not set any additional flags.
2867                  */
2868                 break;
2869         default:
2870                 /*
2871                  * If not on-disk and not deleted, set DDIRTY to force
2872                  * an initial record to be written.
2873                  *
2874                  * Also set the create_tid in both the frontend and backend
2875                  * copy of the inode record.
2876                  */
2877                 ip->ino_leaf.base.create_tid = trans->tid;
2878                 ip->ino_leaf.create_ts = trans->time32;
2879                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2880                 ip->sync_ino_leaf.create_ts = trans->time32;
2881                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2882                 break;
2883         }
2884
2885         /*
2886          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2887          * is already on-disk the old record is marked as deleted.
2888          *
2889          * If DELETED is set hammer_update_inode() will delete the existing
2890          * record without writing out a new one.
2891          *
2892          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2893          */
2894         if (ip->flags & HAMMER_INODE_DELETED) {
2895                 error = hammer_update_inode(&cursor, ip);
2896         } else
2897         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2898             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2899                 error = hammer_update_itimes(&cursor, ip);
2900         } else
2901         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2902                 error = hammer_update_inode(&cursor, ip);
2903         }
2904 done:
2905         if (error) {
2906                 hammer_critical_error(ip->hmp, ip, error,
2907                                       "while syncing inode");
2908         }
2909         hammer_done_cursor(&cursor);
2910         return(error);
2911 }
2912
2913 /*
2914  * This routine is called when the OS is no longer actively referencing
2915  * the inode (but might still be keeping it cached), or when releasing
2916  * the last reference to an inode.
2917  *
2918  * At this point if the inode's nlinks count is zero we want to destroy
2919  * it, which may mean destroying it on-media too.
2920  */
2921 void
2922 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2923 {
2924         struct vnode *vp;
2925
2926         /*
2927          * Set the DELETING flag when the link count drops to 0 and the
2928          * OS no longer has any opens on the inode.
2929          *
2930          * The backend will clear DELETING (a mod flag) and set DELETED
2931          * (a state flag) when it is actually able to perform the
2932          * operation.
2933          *
2934          * Don't reflag the deletion if the flusher is currently syncing
2935          * one that was already flagged.  A previously set DELETING flag
2936          * may bounce around flags and sync_flags until the operation is
2937          * completely done.
2938          */
2939         if (ip->ino_data.nlinks == 0 &&
2940             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2941                 ip->flags |= HAMMER_INODE_DELETING;
2942                 ip->flags |= HAMMER_INODE_TRUNCATED;
2943                 ip->trunc_off = 0;
2944                 vp = NULL;
2945                 if (getvp) {
2946                         if (hammer_get_vnode(ip, &vp) != 0)
2947                                 return;
2948                 }
2949
2950                 /*
2951                  * Final cleanup
2952                  */
2953                 if (ip->vp) {
2954                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2955                         vnode_pager_setsize(ip->vp, 0);
2956                 }
2957                 if (getvp) {
2958                         vput(vp);
2959                 }
2960         }
2961 }
2962
2963 /*
2964  * After potentially resolving a dependancy the inode is tested
2965  * to determine whether it needs to be reflushed.
2966  */
2967 void
2968 hammer_test_inode(hammer_inode_t ip)
2969 {
2970         if (ip->flags & HAMMER_INODE_REFLUSH) {
2971                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2972                 hammer_ref(&ip->lock);
2973                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2974                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2975                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2976                 } else {
2977                         hammer_flush_inode(ip, 0);
2978                 }
2979                 hammer_rel_inode(ip, 0);
2980         }
2981 }
2982
2983 /*
2984  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2985  * reassociated with a vp or just before it gets freed.
2986  *
2987  * Pipeline wakeups to threads blocked due to an excessive number of
2988  * detached inodes.  The reclaim count generates a bit of negative
2989  * feedback.
2990  */
2991 static void
2992 hammer_inode_wakereclaims(hammer_inode_t ip, int dowake)
2993 {
2994         struct hammer_reclaim *reclaim;
2995         hammer_mount_t hmp = ip->hmp;
2996
2997         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2998                 return;
2999
3000         --hammer_count_reclaiming;
3001         --hmp->inode_reclaims;
3002         ip->flags &= ~HAMMER_INODE_RECLAIM;
3003
3004         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT || dowake) {
3005                 reclaim = TAILQ_FIRST(&hmp->reclaim_list);
3006                 if (reclaim && reclaim->count > 0 && --reclaim->count == 0) {
3007                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3008                         wakeup(reclaim);
3009                 }
3010         }
3011 }
3012
3013 /*
3014  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3015  * inodes build up before we start blocking.
3016  *
3017  * When we block we don't care *which* inode has finished reclaiming,
3018  * as lone as one does.  This is somewhat heuristical... we also put a
3019  * cap on how long we are willing to wait.
3020  */
3021 void
3022 hammer_inode_waitreclaims(hammer_mount_t hmp)
3023 {
3024         struct hammer_reclaim reclaim;
3025         int delay;
3026
3027         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT)
3028                 return;
3029         delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
3030                 (HAMMER_RECLAIM_WAIT * 3) + 1;
3031         if (delay > 0) {
3032                 reclaim.count = 2;
3033                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3034                 tsleep(&reclaim, 0, "hmrrcm", delay);
3035                 if (reclaim.count > 0)
3036                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3037         }
3038 }
3039
3040 /*
3041  * A larger then normal backlog of inodes is sitting in the flusher,
3042  * enforce a general slowdown to let it catch up.  This routine is only
3043  * called on completion of a non-flusher-related transaction which
3044  * performed B-Tree node I/O.
3045  *
3046  * It is possible for the flusher to stall in a continuous load.
3047  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3048  * If the flusher is unable to catch up the inode count can bloat until
3049  * we run out of kvm.
3050  *
3051  * This is a bit of a hack.
3052  */
3053 void
3054 hammer_inode_waithard(hammer_mount_t hmp)
3055 {
3056         /*
3057          * Hysteresis.
3058          */
3059         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3060                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT / 2 &&
3061                     hmp->count_iqueued < hmp->count_inodes / 20) {
3062                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3063                         return;
3064                 }
3065         } else {
3066                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT ||
3067                     hmp->count_iqueued < hmp->count_inodes / 10) {
3068                         return;
3069                 }
3070                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3071         }
3072
3073         /*
3074          * Block for one flush cycle.
3075          */
3076         hammer_flusher_wait_next(hmp);
3077 }
3078