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
  40 static int      hammer_unload_inode(struct hammer_inode *ip);
  41 static void     hammer_free_inode(hammer_inode_t ip);
  42 static void     hammer_flush_inode_core(hammer_inode_t ip,
  43                                         hammer_flush_group_t flg, int flags);
  44 static int      hammer_setup_child_callback(hammer_record_t rec, void *data);
  45 #if 0
  46 static int      hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
  47 #endif
  48 static int      hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
  49                                         hammer_flush_group_t flg);
  50 static int      hammer_setup_parent_inodes_helper(hammer_record_t record,
  51                                         int depth, hammer_flush_group_t flg);
  52 static void     hammer_inode_wakereclaims(hammer_inode_t ip);
  53 static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
  54                                         pid_t pid);
  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 int
  82 hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  83 {
  84         if (ip1->redo_fifo_start < ip2->redo_fifo_start)
  85                 return(-1);
  86         if (ip1->redo_fifo_start > ip2->redo_fifo_start)
  87                 return(1);
  88         return(0);
  89 }
  90
  91 /*
  92  * RB-Tree support for inode structures / special LOOKUP_INFO
  93  */
  94 static int
  95 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  96 {
  97         if (info->obj_localization < ip->obj_localization)
  98                 return(-1);
  99         if (info->obj_localization > ip->obj_localization)
 100                 return(1);
 101         if (info->obj_id < ip->obj_id)
 102                 return(-1);
 103         if (info->obj_id > ip->obj_id)
 104                 return(1);
 105         if (info->obj_asof < ip->obj_asof)
 106                 return(-1);
 107         if (info->obj_asof > ip->obj_asof)
 108                 return(1);
 109         return(0);
 110 }
 111
 112 /*
 113  * Used by hammer_scan_inode_snapshots() to locate all of an object's
 114  * snapshots.  Note that the asof field is not tested, which we can get
 115  * away with because it is the lowest-priority field.
 116  */
 117 static int
 118 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
 119 {
 120         hammer_inode_info_t info = data;
 121
 122         if (ip->obj_localization > info->obj_localization)
 123                 return(1);
 124         if (ip->obj_localization < info->obj_localization)
 125                 return(-1);
 126         if (ip->obj_id > info->obj_id)
 127                 return(1);
 128         if (ip->obj_id < info->obj_id)
 129                 return(-1);
 130         return(0);
 131 }
 132
 133 /*
 134  * Used by hammer_unload_pseudofs() to locate all inodes associated with
 135  * a particular PFS.
 136  */
 137 static int
 138 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
 139 {
 140         u_int32_t localization = *(u_int32_t *)data;
 141         if (ip->obj_localization > localization)
 142                 return(1);
 143         if (ip->obj_localization < localization)
 144                 return(-1);
 145         return(0);
 146 }
 147
 148 /*
 149  * RB-Tree support for pseudofs structures
 150  */
 151 static int
 152 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
 153 {
 154         if (p1->localization < p2->localization)
 155                 return(-1);
 156         if (p1->localization > p2->localization)
 157                 return(1);
 158         return(0);
 159 }
 160
 161
 162 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 163 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 164                 hammer_inode_info_cmp, hammer_inode_info_t);
 165 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 166              hammer_pfs_rb_compare, u_int32_t, localization);
 167
 168 /*
 169  * The kernel is not actively referencing this vnode but is still holding
 170  * it cached.
 171  *
 172  * This is called from the frontend.
 173  *
 174  * MPALMOSTSAFE
 175  */
 176 int
 177 hammer_vop_inactive(struct vop_inactive_args *ap)
 178 {
 179         struct hammer_inode *ip = VTOI(ap->a_vp);
 180         hammer_mount_t hmp;
 181
 182         /*
 183          * Degenerate case
 184          */
 185         if (ip == NULL) {
 186                 vrecycle(ap->a_vp);
 187                 return(0);
 188         }
 189
 190         /*
 191          * If the inode no longer has visibility in the filesystem try to
 192          * recycle it immediately, even if the inode is dirty.  Recycling
 193          * it quickly allows the system to reclaim buffer cache and VM
 194          * resources which can matter a lot in a heavily loaded system.
 195          *
 196          * This can deadlock in vfsync() if we aren't careful.
 197          *
 198          * Do not queue the inode to the flusher if we still have visibility,
 199          * otherwise namespace calls such as chmod will unnecessarily generate
 200          * multiple inode updates.
 201          */
 202         if (ip->ino_data.nlinks == 0) {
 203                 hmp = ip->hmp;
 204                 lwkt_gettoken(&hmp->fs_token);
 205                 hammer_inode_unloadable_check(ip, 0);
 206                 if (ip->flags & HAMMER_INODE_MODMASK)
 207                         hammer_flush_inode(ip, 0);
 208                 lwkt_reltoken(&hmp->fs_token);
 209                 vrecycle(ap->a_vp);
 210         }
 211         return(0);
 212 }
 213
 214 /*
 215  * Release the vnode association.  This is typically (but not always)
 216  * the last reference on the inode.
 217  *
 218  * Once the association is lost we are on our own with regards to
 219  * flushing the inode.
 220  *
 221  * We must interlock ip->vp so hammer_get_vnode() can avoid races.
 222  */
 223 int
 224 hammer_vop_reclaim(struct vop_reclaim_args *ap)
 225 {
 226         struct hammer_inode *ip;
 227         hammer_mount_t hmp;
 228         struct vnode *vp;
 229
 230         vp = ap->a_vp;
 231
 232         if ((ip = vp->v_data) != NULL) {
 233                 hmp = ip->hmp;
 234                 lwkt_gettoken(&hmp->fs_token);
 235                 hammer_lock_ex(&ip->lock);
 236                 vp->v_data = NULL;
 237                 ip->vp = NULL;
 238
 239                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
 240                         ++hammer_count_reclaims;
 241                         ++hmp->count_reclaims;
 242                         ip->flags |= HAMMER_INODE_RECLAIM;
 243                 }
 244                 hammer_unlock(&ip->lock);
 245                 hammer_rel_inode(ip, 1);
 246                 lwkt_reltoken(&hmp->fs_token);
 247         }
 248         return(0);
 249 }
 250
 251 /*
 252  * Return a locked vnode for the specified inode.  The inode must be
 253  * referenced but NOT LOCKED on entry and will remain referenced on
 254  * return.
 255  *
 256  * Called from the frontend.
 257  */
 258 int
 259 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 260 {
 261         hammer_mount_t hmp;
 262         struct vnode *vp;
 263         int error = 0;
 264         u_int8_t obj_type;
 265
 266         hmp = ip->hmp;
 267
 268         for (;;) {
 269                 if ((vp = ip->vp) == NULL) {
 270                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
 271                         if (error)
 272                                 break;
 273                         hammer_lock_ex(&ip->lock);
 274                         if (ip->vp != NULL) {
 275                                 hammer_unlock(&ip->lock);
 276                                 vp = *vpp;
 277                                 vp->v_type = VBAD;
 278                                 vx_put(vp);
 279                                 continue;
 280                         }
 281                         hammer_ref(&ip->lock);
 282                         vp = *vpp;
 283                         ip->vp = vp;
 284
 285                         obj_type = ip->ino_data.obj_type;
 286                         vp->v_type = hammer_get_vnode_type(obj_type);
 287
 288                         hammer_inode_wakereclaims(ip);
 289
 290                         switch(ip->ino_data.obj_type) {
 291                         case HAMMER_OBJTYPE_CDEV:
 292                         case HAMMER_OBJTYPE_BDEV:
 293                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
 294                                 addaliasu(vp, ip->ino_data.rmajor,
 295                                           ip->ino_data.rminor);
 296                                 break;
 297                         case HAMMER_OBJTYPE_FIFO:
 298                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
 299                                 break;
 300                         case HAMMER_OBJTYPE_REGFILE:
 301                                 break;
 302                         default:
 303                                 break;
 304                         }
 305
 306                         /*
 307                          * Only mark as the root vnode if the ip is not
 308                          * historical, otherwise the VFS cache will get
 309                          * confused.  The other half of the special handling
 310                          * is in hammer_vop_nlookupdotdot().
 311                          *
 312                          * Pseudo-filesystem roots can be accessed via
 313                          * non-root filesystem paths and setting VROOT may
 314                          * confuse the namecache.  Set VPFSROOT instead.
 315                          */
 316                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 317                             ip->obj_asof == hmp->asof) {
 318                                 if (ip->obj_localization == 0)
 319                                         vsetflags(vp, VROOT);
 320                                 else
 321                                         vsetflags(vp, VPFSROOT);
 322                         }
 323
 324                         vp->v_data = (void *)ip;
 325                         /* vnode locked by getnewvnode() */
 326                         /* make related vnode dirty if inode dirty? */
 327                         hammer_unlock(&ip->lock);
 328                         if (vp->v_type == VREG) {
 329                                 vinitvmio(vp, ip->ino_data.size,
 330                                           hammer_blocksize(ip->ino_data.size),
 331                                           hammer_blockoff(ip->ino_data.size));
 332                         }
 333                         break;
 334                 }
 335
 336                 /*
 337                  * Interlock vnode clearing.  This does not prevent the
 338                  * vnode from going into a reclaimed state but it does
 339                  * prevent it from being destroyed or reused so the vget()
 340                  * will properly fail.
 341                  */
 342                 hammer_lock_ex(&ip->lock);
 343                 if ((vp = ip->vp) == NULL) {
 344                         hammer_unlock(&ip->lock);
 345                         continue;
 346                 }
 347                 vhold_interlocked(vp);
 348                 hammer_unlock(&ip->lock);
 349
 350                 /*
 351                  * loop if the vget fails (aka races), or if the vp
 352                  * no longer matches ip->vp.
 353                  */
 354                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 355                         if (vp == ip->vp) {
 356                                 vdrop(vp);
 357                                 break;
 358                         }
 359                         vput(vp);
 360                 }
 361                 vdrop(vp);
 362         }
 363         *vpp = vp;
 364         return(error);
 365 }
 366
 367 /*
 368  * Locate all copies of the inode for obj_id compatible with the specified
 369  * asof, reference, and issue the related call-back.  This routine is used
 370  * for direct-io invalidation and does not create any new inodes.
 371  */
 372 void
 373 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
 374                             int (*callback)(hammer_inode_t ip, void *data),
 375                             void *data)
 376 {
 377         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
 378                                    hammer_inode_info_cmp_all_history,
 379                                    callback, iinfo);
 380 }
 381
 382 /*
 383  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 384  * do not attach or detach the related vnode (use hammer_get_vnode() for
 385  * that).
 386  *
 387  * The flags argument is only applied for newly created inodes, and only
 388  * certain flags are inherited.
 389  *
 390  * Called from the frontend.
 391  */
 392 struct hammer_inode *
 393 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
 394                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 395                  int flags, int *errorp)
 396 {
 397         hammer_mount_t hmp = trans->hmp;
 398         struct hammer_node_cache *cachep;
 399         struct hammer_inode_info iinfo;
 400         struct hammer_cursor cursor;
 401         struct hammer_inode *ip;
 402
 403
 404         /*
 405          * Determine if we already have an inode cached.  If we do then
 406          * we are golden.
 407          *
 408          * If we find an inode with no vnode we have to mark the
 409          * transaction such that hammer_inode_waitreclaims() is
 410          * called later on to avoid building up an infinite number
 411          * of inodes.  Otherwise we can continue to * add new inodes
 412          * faster then they can be disposed of, even with the tsleep
 413          * delay.
 414          *
 415          * If we find a dummy inode we return a failure so dounlink
 416          * (which does another lookup) doesn't try to mess with the
 417          * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
 418          * to ref dummy inodes.
 419          */
 420         iinfo.obj_id = obj_id;
 421         iinfo.obj_asof = asof;
 422         iinfo.obj_localization = localization;
 423 loop:
 424         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 425         if (ip) {
 426                 if (ip->flags & HAMMER_INODE_DUMMY) {
 427                         *errorp = ENOENT;
 428                         return(NULL);
 429                 }
 430                 hammer_ref(&ip->lock);
 431                 *errorp = 0;
 432                 return(ip);
 433         }
 434
 435         /*
 436          * Allocate a new inode structure and deal with races later.
 437          */
 438         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 439         ++hammer_count_inodes;
 440         ++hmp->count_inodes;
 441         ip->obj_id = obj_id;
 442         ip->obj_asof = iinfo.obj_asof;
 443         ip->obj_localization = localization;
 444         ip->hmp = hmp;
 445         ip->flags = flags & HAMMER_INODE_RO;
 446         ip->cache[0].ip = ip;
 447         ip->cache[1].ip = ip;
 448         ip->cache[2].ip = ip;
 449         ip->cache[3].ip = ip;
 450         if (hmp->ronly)
 451                 ip->flags |= HAMMER_INODE_RO;
 452         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 453                 0x7FFFFFFFFFFFFFFFLL;
 454         RB_INIT(&ip->rec_tree);
 455         TAILQ_INIT(&ip->target_list);
 456         hammer_ref(&ip->lock);
 457
 458         /*
 459          * Locate the on-disk inode.  If this is a PFS root we always
 460          * access the current version of the root inode and (if it is not
 461          * a master) always access information under it with a snapshot
 462          * TID.
 463          *
 464          * We cache recent inode lookups in this directory in dip->cache[2].
 465          * If we can't find it we assume the inode we are looking for is
 466          * close to the directory inode.
 467          */
 468 retry:
 469         cachep = NULL;
 470         if (dip) {
 471                 if (dip->cache[2].node)
 472                         cachep = &dip->cache[2];
 473                 else
 474                         cachep = &dip->cache[0];
 475         }
 476         hammer_init_cursor(trans, &cursor, cachep, NULL);
 477         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
 478         cursor.key_beg.obj_id = ip->obj_id;
 479         cursor.key_beg.key = 0;
 480         cursor.key_beg.create_tid = 0;
 481         cursor.key_beg.delete_tid = 0;
 482         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 483         cursor.key_beg.obj_type = 0;
 484
 485         cursor.asof = iinfo.obj_asof;
 486         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 487                        HAMMER_CURSOR_ASOF;
 488
 489         *errorp = hammer_btree_lookup(&cursor);
 490         if (*errorp == EDEADLK) {
 491                 hammer_done_cursor(&cursor);
 492                 goto retry;
 493         }
 494
 495         /*
 496          * On success the B-Tree lookup will hold the appropriate
 497          * buffer cache buffers and provide a pointer to the requested
 498          * information.  Copy the information to the in-memory inode
 499          * and cache the B-Tree node to improve future operations.
 500          */
 501         if (*errorp == 0) {
 502                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 503                 ip->ino_data = cursor.data->inode;
 504
 505                 /*
 506                  * cache[0] tries to cache the location of the object inode.
 507                  * The assumption is that it is near the directory inode.
 508                  *
 509                  * cache[1] tries to cache the location of the object data.
 510                  * We might have something in the governing directory from
 511                  * scan optimizations (see the strategy code in
 512                  * hammer_vnops.c).
 513                  *
 514                  * We update dip->cache[2], if possible, with the location
 515                  * of the object inode for future directory shortcuts.
 516                  */
 517                 hammer_cache_node(&ip->cache[0], cursor.node);
 518                 if (dip) {
 519                         if (dip->cache[3].node) {
 520                                 hammer_cache_node(&ip->cache[1],
 521                                                   dip->cache[3].node);
 522                         }
 523                         hammer_cache_node(&dip->cache[2], cursor.node);
 524                 }
 525
 526                 /*
 527                  * The file should not contain any data past the file size
 528                  * stored in the inode.  Setting save_trunc_off to the
 529                  * file size instead of max reduces B-Tree lookup overheads
 530                  * on append by allowing the flusher to avoid checking for
 531                  * record overwrites.
 532                  */
 533                 ip->save_trunc_off = ip->ino_data.size;
 534
 535                 /*
 536                  * Locate and assign the pseudofs management structure to
 537                  * the inode.
 538                  */
 539                 if (dip && dip->obj_localization == ip->obj_localization) {
 540                         ip->pfsm = dip->pfsm;
 541                         hammer_ref(&ip->pfsm->lock);
 542                 } else {
 543                         ip->pfsm = hammer_load_pseudofs(trans,
 544                                                         ip->obj_localization,
 545                                                         errorp);
 546                         *errorp = 0;    /* ignore ENOENT */
 547                 }
 548         }
 549
 550         /*
 551          * The inode is placed on the red-black tree and will be synced to
 552          * the media when flushed or by the filesystem sync.  If this races
 553          * another instantiation/lookup the insertion will fail.
 554          */
 555         if (*errorp == 0) {
 556                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 557                         hammer_free_inode(ip);
 558                         hammer_done_cursor(&cursor);
 559                         goto loop;
 560                 }
 561                 ip->flags |= HAMMER_INODE_ONDISK;
 562         } else {
 563                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 564                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 565                         --hmp->rsv_inodes;
 566                 }
 567
 568                 hammer_free_inode(ip);
 569                 ip = NULL;
 570         }
 571         hammer_done_cursor(&cursor);
 572
 573         /*
 574          * NEWINODE is only set if the inode becomes dirty later,
 575          * setting it here just leads to unnecessary stalls.
 576          *
 577          * trans->flags |= HAMMER_TRANSF_NEWINODE;
 578          */
 579         return (ip);
 580 }
 581
 582 /*
 583  * Get a dummy inode to placemark a broken directory entry.
 584  */
 585 struct hammer_inode *
 586 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
 587                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 588                  int flags, int *errorp)
 589 {
 590         hammer_mount_t hmp = trans->hmp;
 591         struct hammer_inode_info iinfo;
 592         struct hammer_inode *ip;
 593
 594         /*
 595          * Determine if we already have an inode cached.  If we do then
 596          * we are golden.
 597          *
 598          * If we find an inode with no vnode we have to mark the
 599          * transaction such that hammer_inode_waitreclaims() is
 600          * called later on to avoid building up an infinite number
 601          * of inodes.  Otherwise we can continue to * add new inodes
 602          * faster then they can be disposed of, even with the tsleep
 603          * delay.
 604          *
 605          * If we find a non-fake inode we return an error.  Only fake
 606          * inodes can be returned by this routine.
 607          */
 608         iinfo.obj_id = obj_id;
 609         iinfo.obj_asof = asof;
 610         iinfo.obj_localization = localization;
 611 loop:
 612         *errorp = 0;
 613         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 614         if (ip) {
 615                 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
 616                         *errorp = ENOENT;
 617                         return(NULL);
 618                 }
 619                 hammer_ref(&ip->lock);
 620                 return(ip);
 621         }
 622
 623         /*
 624          * Allocate a new inode structure and deal with races later.
 625          */
 626         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 627         ++hammer_count_inodes;
 628         ++hmp->count_inodes;
 629         ip->obj_id = obj_id;
 630         ip->obj_asof = iinfo.obj_asof;
 631         ip->obj_localization = localization;
 632         ip->hmp = hmp;
 633         ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
 634         ip->cache[0].ip = ip;
 635         ip->cache[1].ip = ip;
 636         ip->cache[2].ip = ip;
 637         ip->cache[3].ip = ip;
 638         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 639                 0x7FFFFFFFFFFFFFFFLL;
 640         RB_INIT(&ip->rec_tree);
 641         TAILQ_INIT(&ip->target_list);
 642         hammer_ref(&ip->lock);
 643
 644         /*
 645          * Populate the dummy inode.  Leave everything zero'd out.
 646          *
 647          * (ip->ino_leaf and ip->ino_data)
 648          *
 649          * Make the dummy inode a FIFO object which most copy programs
 650          * will properly ignore.
 651          */
 652         ip->save_trunc_off = ip->ino_data.size;
 653         ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
 654
 655         /*
 656          * Locate and assign the pseudofs management structure to
 657          * the inode.
 658          */
 659         if (dip && dip->obj_localization == ip->obj_localization) {
 660                 ip->pfsm = dip->pfsm;
 661                 hammer_ref(&ip->pfsm->lock);
 662         } else {
 663                 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
 664                                                 errorp);
 665                 *errorp = 0;    /* ignore ENOENT */
 666         }
 667
 668         /*
 669          * The inode is placed on the red-black tree and will be synced to
 670          * the media when flushed or by the filesystem sync.  If this races
 671          * another instantiation/lookup the insertion will fail.
 672          *
 673          * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
 674          */
 675         if (*errorp == 0) {
 676                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 677                         hammer_free_inode(ip);
 678                         goto loop;
 679                 }
 680         } else {
 681                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 682                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 683                         --hmp->rsv_inodes;
 684                 }
 685                 hammer_free_inode(ip);
 686                 ip = NULL;
 687         }
 688         trans->flags |= HAMMER_TRANSF_NEWINODE;
 689         return (ip);
 690 }
 691
 692 /*
 693  * Return a referenced inode only if it is in our inode cache.
 694  *
 695  * Dummy inodes do not count.
 696  */
 697 struct hammer_inode *
 698 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
 699                   hammer_tid_t asof, u_int32_t localization)
 700 {
 701         hammer_mount_t hmp = trans->hmp;
 702         struct hammer_inode_info iinfo;
 703         struct hammer_inode *ip;
 704
 705         iinfo.obj_id = obj_id;
 706         iinfo.obj_asof = asof;
 707         iinfo.obj_localization = localization;
 708
 709         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 710         if (ip) {
 711                 if (ip->flags & HAMMER_INODE_DUMMY)
 712                         ip = NULL;
 713                 else
 714                         hammer_ref(&ip->lock);
 715         }
 716         return(ip);
 717 }
 718
 719 /*
 720  * Create a new filesystem object, returning the inode in *ipp.  The
 721  * returned inode will be referenced.  The inode is created in-memory.
 722  *
 723  * If pfsm is non-NULL the caller wishes to create the root inode for
 724  * a master PFS.
 725  */
 726 int
 727 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 728                     struct ucred *cred,
 729                     hammer_inode_t dip, const char *name, int namelen,
 730                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
 731 {
 732         hammer_mount_t hmp;
 733         hammer_inode_t ip;
 734         uid_t xuid;
 735         int error;
 736         int64_t namekey;
 737         u_int32_t dummy;
 738
 739         hmp = trans->hmp;
 740
 741         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 742         ++hammer_count_inodes;
 743         ++hmp->count_inodes;
 744         trans->flags |= HAMMER_TRANSF_NEWINODE;
 745
 746         if (pfsm) {
 747                 KKASSERT(pfsm->localization != 0);
 748                 ip->obj_id = HAMMER_OBJID_ROOT;
 749                 ip->obj_localization = pfsm->localization;
 750         } else {
 751                 KKASSERT(dip != NULL);
 752                 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
 753                 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
 754                 ip->obj_localization = dip->obj_localization;
 755         }
 756
 757         KKASSERT(ip->obj_id != 0);
 758         ip->obj_asof = hmp->asof;
 759         ip->hmp = hmp;
 760         ip->flush_state = HAMMER_FST_IDLE;
 761         ip->flags = HAMMER_INODE_DDIRTY |
 762                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
 763         ip->cache[0].ip = ip;
 764         ip->cache[1].ip = ip;
 765         ip->cache[2].ip = ip;
 766         ip->cache[3].ip = ip;
 767
 768         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 769         /* ip->save_trunc_off = 0; (already zero) */
 770         RB_INIT(&ip->rec_tree);
 771         TAILQ_INIT(&ip->target_list);
 772
 773         ip->ino_data.atime = trans->time;
 774         ip->ino_data.mtime = trans->time;
 775         ip->ino_data.size = 0;
 776         ip->ino_data.nlinks = 0;
 777
 778         /*
 779          * A nohistory designator on the parent directory is inherited by
 780          * the child.  We will do this even for pseudo-fs creation... the
 781          * sysad can turn it off.
 782          */
 783         if (dip) {
 784                 ip->ino_data.uflags = dip->ino_data.uflags &
 785                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 786         }
 787
 788         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 789         ip->ino_leaf.base.localization = ip->obj_localization +
 790                                          HAMMER_LOCALIZE_INODE;
 791         ip->ino_leaf.base.obj_id = ip->obj_id;
 792         ip->ino_leaf.base.key = 0;
 793         ip->ino_leaf.base.create_tid = 0;
 794         ip->ino_leaf.base.delete_tid = 0;
 795         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 796         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 797
 798         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 799         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 800         ip->ino_data.mode = vap->va_mode;
 801         ip->ino_data.ctime = trans->time;
 802
 803         /*
 804          * If we are running version 2 or greater directory entries are
 805          * inode-localized instead of data-localized.
 806          */
 807         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
 808                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
 809                         ip->ino_data.cap_flags |=
 810                                 HAMMER_INODE_CAP_DIR_LOCAL_INO;
 811                 }
 812         }
 813         if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
 814                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
 815                         ip->ino_data.cap_flags |=
 816                                 HAMMER_INODE_CAP_DIRHASH_ALG1;
 817                 }
 818         }
 819
 820         /*
 821          * Setup the ".." pointer.  This only needs to be done for directories
 822          * but we do it for all objects as a recovery aid.
 823          */
 824         if (dip)
 825                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
 826 #if 0
 827         /*
 828          * The parent_obj_localization field only applies to pseudo-fs roots.
 829          * XXX this is no longer applicable, PFSs are no longer directly
 830          * tied into the parent's directory structure.
 831          */
 832         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
 833             ip->obj_id == HAMMER_OBJID_ROOT) {
 834                 ip->ino_data.ext.obj.parent_obj_localization =
 835                                                 dip->obj_localization;
 836         }
 837 #endif
 838
 839         switch(ip->ino_leaf.base.obj_type) {
 840         case HAMMER_OBJTYPE_CDEV:
 841         case HAMMER_OBJTYPE_BDEV:
 842                 ip->ino_data.rmajor = vap->va_rmajor;
 843                 ip->ino_data.rminor = vap->va_rminor;
 844                 break;
 845         default:
 846                 break;
 847         }
 848
 849         /*
 850          * Calculate default uid/gid and overwrite with information from
 851          * the vap.
 852          */
 853         if (dip) {
 854                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 855                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
 856                                              xuid, cred, &vap->va_mode);
 857         } else {
 858                 xuid = 0;
 859         }
 860         ip->ino_data.mode = vap->va_mode;
 861
 862         if (vap->va_vaflags & VA_UID_UUID_VALID)
 863                 ip->ino_data.uid = vap->va_uid_uuid;
 864         else if (vap->va_uid != (uid_t)VNOVAL)
 865                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 866         else
 867                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 868
 869         if (vap->va_vaflags & VA_GID_UUID_VALID)
 870                 ip->ino_data.gid = vap->va_gid_uuid;
 871         else if (vap->va_gid != (gid_t)VNOVAL)
 872                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 873         else if (dip)
 874                 ip->ino_data.gid = dip->ino_data.gid;
 875
 876         hammer_ref(&ip->lock);
 877
 878         if (pfsm) {
 879                 ip->pfsm = pfsm;
 880                 hammer_ref(&pfsm->lock);
 881                 error = 0;
 882         } else if (dip->obj_localization == ip->obj_localization) {
 883                 ip->pfsm = dip->pfsm;
 884                 hammer_ref(&ip->pfsm->lock);
 885                 error = 0;
 886         } else {
 887                 ip->pfsm = hammer_load_pseudofs(trans,
 888                                                 ip->obj_localization,
 889                                                 &error);
 890                 error = 0;      /* ignore ENOENT */
 891         }
 892
 893         if (error) {
 894                 hammer_free_inode(ip);
 895                 ip = NULL;
 896         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 897                 panic("hammer_create_inode: duplicate obj_id %llx",
 898                       (long long)ip->obj_id);
 899                 /* not reached */
 900                 hammer_free_inode(ip);
 901         }
 902         *ipp = ip;
 903         return(error);
 904 }
 905
 906 /*
 907  * Final cleanup / freeing of an inode structure
 908  */
 909 static void
 910 hammer_free_inode(hammer_inode_t ip)
 911 {
 912         struct hammer_mount *hmp;
 913
 914         hmp = ip->hmp;
 915         KKASSERT(hammer_oneref(&ip->lock));
 916         hammer_uncache_node(&ip->cache[0]);
 917         hammer_uncache_node(&ip->cache[1]);
 918         hammer_uncache_node(&ip->cache[2]);
 919         hammer_uncache_node(&ip->cache[3]);
 920         hammer_inode_wakereclaims(ip);
 921         if (ip->objid_cache)
 922                 hammer_clear_objid(ip);
 923         --hammer_count_inodes;
 924         --hmp->count_inodes;
 925         if (ip->pfsm) {
 926                 hammer_rel_pseudofs(hmp, ip->pfsm);
 927                 ip->pfsm = NULL;
 928         }
 929         kfree(ip, hmp->m_inodes);
 930         ip = NULL;
 931 }
 932
 933 /*
 934  * Retrieve pseudo-fs data.  NULL will never be returned.
 935  *
 936  * If an error occurs *errorp will be set and a default template is returned,
 937  * otherwise *errorp is set to 0.  Typically when an error occurs it will
 938  * be ENOENT.
 939  */
 940 hammer_pseudofs_inmem_t
 941 hammer_load_pseudofs(hammer_transaction_t trans,
 942                      u_int32_t localization, int *errorp)
 943 {
 944         hammer_mount_t hmp = trans->hmp;
 945         hammer_inode_t ip;
 946         hammer_pseudofs_inmem_t pfsm;
 947         struct hammer_cursor cursor;
 948         int bytes;
 949
 950 retry:
 951         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
 952         if (pfsm) {
 953                 hammer_ref(&pfsm->lock);
 954                 *errorp = 0;
 955                 return(pfsm);
 956         }
 957
 958         /*
 959          * PFS records are stored in the root inode (not the PFS root inode,
 960          * but the real root).  Avoid an infinite recursion if loading
 961          * the PFS for the real root.
 962          */
 963         if (localization) {
 964                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
 965                                       HAMMER_MAX_TID,
 966                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
 967         } else {
 968                 ip = NULL;
 969         }
 970
 971         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
 972         pfsm->localization = localization;
 973         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
 974         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
 975
 976         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
 977         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
 978                                       HAMMER_LOCALIZE_MISC;
 979         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 980         cursor.key_beg.create_tid = 0;
 981         cursor.key_beg.delete_tid = 0;
 982         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 983         cursor.key_beg.obj_type = 0;
 984         cursor.key_beg.key = localization;
 985         cursor.asof = HAMMER_MAX_TID;
 986         cursor.flags |= HAMMER_CURSOR_ASOF;
 987
 988         if (ip)
 989                 *errorp = hammer_ip_lookup(&cursor);
 990         else
 991                 *errorp = hammer_btree_lookup(&cursor);
 992         if (*errorp == 0) {
 993                 *errorp = hammer_ip_resolve_data(&cursor);
 994                 if (*errorp == 0) {
 995                         if (cursor.data->pfsd.mirror_flags &
 996                             HAMMER_PFSD_DELETED) {
 997                                 *errorp = ENOENT;
 998                         } else {
 999                                 bytes = cursor.leaf->data_len;
1000                                 if (bytes > sizeof(pfsm->pfsd))
1001                                         bytes = sizeof(pfsm->pfsd);
1002                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
1003                         }
1004                 }
1005         }
1006         hammer_done_cursor(&cursor);
1007
1008         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1009         hammer_ref(&pfsm->lock);
1010         if (ip)
1011                 hammer_rel_inode(ip, 0);
1012         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
1013                 kfree(pfsm, hmp->m_misc);
1014                 goto retry;
1015         }
1016         return(pfsm);
1017 }
1018
1019 /*
1020  * Store pseudo-fs data.  The backend will automatically delete any prior
1021  * on-disk pseudo-fs data but we have to delete in-memory versions.
1022  */
1023 int
1024 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
1025 {
1026         struct hammer_cursor cursor;
1027         hammer_record_t record;
1028         hammer_inode_t ip;
1029         int error;
1030
1031         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1032                               HAMMER_DEF_LOCALIZATION, 0, &error);
1033 retry:
1034         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1035         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1036         cursor.key_beg.localization = ip->obj_localization +
1037                                       HAMMER_LOCALIZE_MISC;
1038         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1039         cursor.key_beg.create_tid = 0;
1040         cursor.key_beg.delete_tid = 0;
1041         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1042         cursor.key_beg.obj_type = 0;
1043         cursor.key_beg.key = pfsm->localization;
1044         cursor.asof = HAMMER_MAX_TID;
1045         cursor.flags |= HAMMER_CURSOR_ASOF;
1046
1047         /*
1048          * Replace any in-memory version of the record.
1049          */
1050         error = hammer_ip_lookup(&cursor);
1051         if (error == 0 && hammer_cursor_inmem(&cursor)) {
1052                 record = cursor.iprec;
1053                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1054                         KKASSERT(cursor.deadlk_rec == NULL);
1055                         hammer_ref(&record->lock);
1056                         cursor.deadlk_rec = record;
1057                         error = EDEADLK;
1058                 } else {
1059                         record->flags |= HAMMER_RECF_DELETED_FE;
1060                         error = 0;
1061                 }
1062         }
1063
1064         /*
1065          * Allocate replacement general record.  The backend flush will
1066          * delete any on-disk version of the record.
1067          */
1068         if (error == 0 || error == ENOENT) {
1069                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1070                 record->type = HAMMER_MEM_RECORD_GENERAL;
1071
1072                 record->leaf.base.localization = ip->obj_localization +
1073                                                  HAMMER_LOCALIZE_MISC;
1074                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1075                 record->leaf.base.key = pfsm->localization;
1076                 record->leaf.data_len = sizeof(pfsm->pfsd);
1077                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1078                 error = hammer_ip_add_record(trans, record);
1079         }
1080         hammer_done_cursor(&cursor);
1081         if (error == EDEADLK)
1082                 goto retry;
1083         hammer_rel_inode(ip, 0);
1084         return(error);
1085 }
1086
1087 /*
1088  * Create a root directory for a PFS if one does not alredy exist.
1089  *
1090  * The PFS root stands alone so we must also bump the nlinks count
1091  * to prevent it from being destroyed on release.
1092  */
1093 int
1094 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1095                        hammer_pseudofs_inmem_t pfsm)
1096 {
1097         hammer_inode_t ip;
1098         struct vattr vap;
1099         int error;
1100
1101         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1102                               pfsm->localization, 0, &error);
1103         if (ip == NULL) {
1104                 vattr_null(&vap);
1105                 vap.va_mode = 0755;
1106                 vap.va_type = VDIR;
1107                 error = hammer_create_inode(trans, &vap, cred,
1108                                             NULL, NULL, 0,
1109                                             pfsm, &ip);
1110                 if (error == 0) {
1111                         ++ip->ino_data.nlinks;
1112                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1113                 }
1114         }
1115         if (ip)
1116                 hammer_rel_inode(ip, 0);
1117         return(error);
1118 }
1119
1120 /*
1121  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1122  * if we are unable to disassociate all the inodes.
1123  */
1124 static
1125 int
1126 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1127 {
1128         int res;
1129
1130         hammer_ref(&ip->lock);
1131         if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1132                 vclean_unlocked(ip->vp);
1133         if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1134                 res = 0;
1135         else
1136                 res = -1;       /* stop, someone is using the inode */
1137         hammer_rel_inode(ip, 0);
1138         return(res);
1139 }
1140
1141 int
1142 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1143 {
1144         int res;
1145         int try;
1146
1147         for (try = res = 0; try < 4; ++try) {
1148                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1149                                            hammer_inode_pfs_cmp,
1150                                            hammer_unload_pseudofs_callback,
1151                                            &localization);
1152                 if (res == 0 && try > 1)
1153                         break;
1154                 hammer_flusher_sync(trans->hmp);
1155         }
1156         if (res != 0)
1157                 res = ENOTEMPTY;
1158         return(res);
1159 }
1160
1161
1162 /*
1163  * Release a reference on a PFS
1164  */
1165 void
1166 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1167 {
1168         hammer_rel(&pfsm->lock);
1169         if (hammer_norefs(&pfsm->lock)) {
1170                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1171                 kfree(pfsm, hmp->m_misc);
1172         }
1173 }
1174
1175 /*
1176  * Called by hammer_sync_inode().
1177  */
1178 static int
1179 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1180 {
1181         hammer_transaction_t trans = cursor->trans;
1182         hammer_record_t record;
1183         int error;
1184         int redirty;
1185
1186 retry:
1187         error = 0;
1188
1189         /*
1190          * If the inode has a presence on-disk then locate it and mark
1191          * it deleted, setting DELONDISK.
1192          *
1193          * The record may or may not be physically deleted, depending on
1194          * the retention policy.
1195          */
1196         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1197             HAMMER_INODE_ONDISK) {
1198                 hammer_normalize_cursor(cursor);
1199                 cursor->key_beg.localization = ip->obj_localization +
1200                                                HAMMER_LOCALIZE_INODE;
1201                 cursor->key_beg.obj_id = ip->obj_id;
1202                 cursor->key_beg.key = 0;
1203                 cursor->key_beg.create_tid = 0;
1204                 cursor->key_beg.delete_tid = 0;
1205                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1206                 cursor->key_beg.obj_type = 0;
1207                 cursor->asof = ip->obj_asof;
1208                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1209                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1210                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1211
1212                 error = hammer_btree_lookup(cursor);
1213                 if (hammer_debug_inode)
1214                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1215
1216                 if (error == 0) {
1217                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1218                         if (hammer_debug_inode)
1219                                 kprintf(" error %d\n", error);
1220                         if (error == 0) {
1221                                 ip->flags |= HAMMER_INODE_DELONDISK;
1222                         }
1223                         if (cursor->node)
1224                                 hammer_cache_node(&ip->cache[0], cursor->node);
1225                 }
1226                 if (error == EDEADLK) {
1227                         hammer_done_cursor(cursor);
1228                         error = hammer_init_cursor(trans, cursor,
1229                                                    &ip->cache[0], ip);
1230                         if (hammer_debug_inode)
1231                                 kprintf("IPDED %p %d\n", ip, error);
1232                         if (error == 0)
1233                                 goto retry;
1234                 }
1235         }
1236
1237         /*
1238          * Ok, write out the initial record or a new record (after deleting
1239          * the old one), unless the DELETED flag is set.  This routine will
1240          * clear DELONDISK if it writes out a record.
1241          *
1242          * Update our inode statistics if this is the first application of
1243          * the inode on-disk.
1244          */
1245         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1246                 /*
1247                  * Generate a record and write it to the media.  We clean-up
1248                  * the state before releasing so we do not have to set-up
1249                  * a flush_group.
1250                  */
1251                 record = hammer_alloc_mem_record(ip, 0);
1252                 record->type = HAMMER_MEM_RECORD_INODE;
1253                 record->flush_state = HAMMER_FST_FLUSH;
1254                 record->leaf = ip->sync_ino_leaf;
1255                 record->leaf.base.create_tid = trans->tid;
1256                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1257                 record->leaf.create_ts = trans->time32;
1258                 record->data = (void *)&ip->sync_ino_data;
1259                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1260
1261                 /*
1262                  * If this flag is set we cannot sync the new file size
1263                  * because we haven't finished related truncations.  The
1264                  * inode will be flushed in another flush group to finish
1265                  * the job.
1266                  */
1267                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1268                     ip->sync_ino_data.size != ip->ino_data.size) {
1269                         redirty = 1;
1270                         ip->sync_ino_data.size = ip->ino_data.size;
1271                 } else {
1272                         redirty = 0;
1273                 }
1274
1275                 for (;;) {
1276                         error = hammer_ip_sync_record_cursor(cursor, record);
1277                         if (hammer_debug_inode)
1278                                 kprintf("GENREC %p rec %08x %d\n",
1279                                         ip, record->flags, error);
1280                         if (error != EDEADLK)
1281                                 break;
1282                         hammer_done_cursor(cursor);
1283                         error = hammer_init_cursor(trans, cursor,
1284                                                    &ip->cache[0], ip);
1285                         if (hammer_debug_inode)
1286                                 kprintf("GENREC reinit %d\n", error);
1287                         if (error)
1288                                 break;
1289                 }
1290
1291                 /*
1292                  * Note:  The record was never on the inode's record tree
1293                  * so just wave our hands importantly and destroy it.
1294                  */
1295                 record->flags |= HAMMER_RECF_COMMITTED;
1296                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1297                 record->flush_state = HAMMER_FST_IDLE;
1298                 ++ip->rec_generation;
1299                 hammer_rel_mem_record(record);
1300
1301                 /*
1302                  * Finish up.
1303                  */
1304                 if (error == 0) {
1305                         if (hammer_debug_inode)
1306                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1307                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1308                                             HAMMER_INODE_SDIRTY |
1309                                             HAMMER_INODE_ATIME |
1310                                             HAMMER_INODE_MTIME);
1311                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1312                         if (redirty)
1313                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1314
1315                         /*
1316                          * Root volume count of inodes
1317                          */
1318                         hammer_sync_lock_sh(trans);
1319                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1320                                 hammer_modify_volume_field(trans,
1321                                                            trans->rootvol,
1322                                                            vol0_stat_inodes);
1323                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1324                                 hammer_modify_volume_done(trans->rootvol);
1325                                 ip->flags |= HAMMER_INODE_ONDISK;
1326                                 if (hammer_debug_inode)
1327                                         kprintf("NOWONDISK %p\n", ip);
1328                         }
1329                         hammer_sync_unlock(trans);
1330                 }
1331         }
1332
1333         /*
1334          * If the inode has been destroyed, clean out any left-over flags
1335          * that may have been set by the frontend.
1336          */
1337         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1338                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1339                                     HAMMER_INODE_SDIRTY |
1340                                     HAMMER_INODE_ATIME |
1341                                     HAMMER_INODE_MTIME);
1342         }
1343         return(error);
1344 }
1345
1346 /*
1347  * Update only the itimes fields.
1348  *
1349  * ATIME can be updated without generating any UNDO.  MTIME is updated
1350  * with UNDO so it is guaranteed to be synchronized properly in case of
1351  * a crash.
1352  *
1353  * Neither field is included in the B-Tree leaf element's CRC, which is how
1354  * we can get away with updating ATIME the way we do.
1355  */
1356 static int
1357 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1358 {
1359         hammer_transaction_t trans = cursor->trans;
1360         int error;
1361
1362 retry:
1363         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1364             HAMMER_INODE_ONDISK) {
1365                 return(0);
1366         }
1367
1368         hammer_normalize_cursor(cursor);
1369         cursor->key_beg.localization = ip->obj_localization +
1370                                        HAMMER_LOCALIZE_INODE;
1371         cursor->key_beg.obj_id = ip->obj_id;
1372         cursor->key_beg.key = 0;
1373         cursor->key_beg.create_tid = 0;
1374         cursor->key_beg.delete_tid = 0;
1375         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1376         cursor->key_beg.obj_type = 0;
1377         cursor->asof = ip->obj_asof;
1378         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1379         cursor->flags |= HAMMER_CURSOR_ASOF;
1380         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1381         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1382         cursor->flags |= HAMMER_CURSOR_BACKEND;
1383
1384         error = hammer_btree_lookup(cursor);
1385         if (error == 0) {
1386                 hammer_cache_node(&ip->cache[0], cursor->node);
1387                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1388                         /*
1389                          * Updating MTIME requires an UNDO.  Just cover
1390                          * both atime and mtime.
1391                          */
1392                         hammer_sync_lock_sh(trans);
1393                         hammer_modify_buffer(trans, cursor->data_buffer,
1394                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1395                                      HAMMER_ITIMES_BYTES);
1396                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1397                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1398                         hammer_modify_buffer_done(cursor->data_buffer);
1399                         hammer_sync_unlock(trans);
1400                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1401                         /*
1402                          * Updating atime only can be done in-place with
1403                          * no UNDO.
1404                          */
1405                         hammer_sync_lock_sh(trans);
1406                         hammer_modify_buffer(trans, cursor->data_buffer,
1407                                              NULL, 0);
1408                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1409                         hammer_modify_buffer_done(cursor->data_buffer);
1410                         hammer_sync_unlock(trans);
1411                 }
1412                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1413         }
1414         if (error == EDEADLK) {
1415                 hammer_done_cursor(cursor);
1416                 error = hammer_init_cursor(trans, cursor,
1417                                            &ip->cache[0], ip);
1418                 if (error == 0)
1419                         goto retry;
1420         }
1421         return(error);
1422 }
1423
1424 /*
1425  * Release a reference on an inode, flush as requested.
1426  *
1427  * On the last reference we queue the inode to the flusher for its final
1428  * disposition.
1429  */
1430 void
1431 hammer_rel_inode(struct hammer_inode *ip, int flush)
1432 {
1433         /*hammer_mount_t hmp = ip->hmp;*/
1434
1435         /*
1436          * Handle disposition when dropping the last ref.
1437          */
1438         for (;;) {
1439                 if (hammer_oneref(&ip->lock)) {
1440                         /*
1441                          * Determine whether on-disk action is needed for
1442                          * the inode's final disposition.
1443                          */
1444                         KKASSERT(ip->vp == NULL);
1445                         hammer_inode_unloadable_check(ip, 0);
1446                         if (ip->flags & HAMMER_INODE_MODMASK) {
1447                                 hammer_flush_inode(ip, 0);
1448                         } else if (hammer_oneref(&ip->lock)) {
1449                                 hammer_unload_inode(ip);
1450                                 break;
1451                         }
1452                 } else {
1453                         if (flush)
1454                                 hammer_flush_inode(ip, 0);
1455
1456                         /*
1457                          * The inode still has multiple refs, try to drop
1458                          * one ref.
1459                          */
1460                         KKASSERT(hammer_isactive(&ip->lock) >= 1);
1461                         if (hammer_isactive(&ip->lock) > 1) {
1462                                 hammer_rel(&ip->lock);
1463                                 break;
1464                         }
1465                 }
1466         }
1467 }
1468
1469 /*
1470  * Unload and destroy the specified inode.  Must be called with one remaining
1471  * reference.  The reference is disposed of.
1472  *
1473  * The inode must be completely clean.
1474  */
1475 static int
1476 hammer_unload_inode(struct hammer_inode *ip)
1477 {
1478         hammer_mount_t hmp = ip->hmp;
1479
1480         KASSERT(hammer_oneref(&ip->lock),
1481                 ("hammer_unload_inode: %d refs\n", hammer_isactive(&ip->lock)));
1482         KKASSERT(ip->vp == NULL);
1483         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1484         KKASSERT(ip->cursor_ip_refs == 0);
1485         KKASSERT(hammer_notlocked(&ip->lock));
1486         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1487
1488         KKASSERT(RB_EMPTY(&ip->rec_tree));
1489         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1490
1491         if (ip->flags & HAMMER_INODE_RDIRTY) {
1492                 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1493                 ip->flags &= ~HAMMER_INODE_RDIRTY;
1494         }
1495         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1496
1497         hammer_free_inode(ip);
1498         return(0);
1499 }
1500
1501 /*
1502  * Called during unmounting if a critical error occured.  The in-memory
1503  * inode and all related structures are destroyed.
1504  *
1505  * If a critical error did not occur the unmount code calls the standard
1506  * release and asserts that the inode is gone.
1507  */
1508 int
1509 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1510 {
1511         hammer_record_t rec;
1512
1513         /*
1514          * Get rid of the inodes in-memory records, regardless of their
1515          * state, and clear the mod-mask.
1516          */
1517         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1518                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1519                 rec->target_ip = NULL;
1520                 if (rec->flush_state == HAMMER_FST_SETUP)
1521                         rec->flush_state = HAMMER_FST_IDLE;
1522         }
1523         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1524                 if (rec->flush_state == HAMMER_FST_FLUSH)
1525                         --rec->flush_group->refs;
1526                 else
1527                         hammer_ref(&rec->lock);
1528                 KKASSERT(hammer_oneref(&rec->lock));
1529                 rec->flush_state = HAMMER_FST_IDLE;
1530                 rec->flush_group = NULL;
1531                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1532                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1533                 ++ip->rec_generation;
1534                 hammer_rel_mem_record(rec);
1535         }
1536         ip->flags &= ~HAMMER_INODE_MODMASK;
1537         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1538         KKASSERT(ip->vp == NULL);
1539
1540         /*
1541          * Remove the inode from any flush group, force it idle.  FLUSH
1542          * and SETUP states have an inode ref.
1543          */
1544         switch(ip->flush_state) {
1545         case HAMMER_FST_FLUSH:
1546                 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1547                 --ip->flush_group->refs;
1548                 ip->flush_group = NULL;
1549                 /* fall through */
1550         case HAMMER_FST_SETUP:
1551                 hammer_rel(&ip->lock);
1552                 ip->flush_state = HAMMER_FST_IDLE;
1553                 /* fall through */
1554         case HAMMER_FST_IDLE:
1555                 break;
1556         }
1557
1558         /*
1559          * There shouldn't be any associated vnode.  The unload needs at
1560          * least one ref, if we do have a vp steal its ip ref.
1561          */
1562         if (ip->vp) {
1563                 kprintf("hammer_destroy_inode_callback: Unexpected "
1564                         "vnode association ip %p vp %p\n", ip, ip->vp);
1565                 ip->vp->v_data = NULL;
1566                 ip->vp = NULL;
1567         } else {
1568                 hammer_ref(&ip->lock);
1569         }
1570         hammer_unload_inode(ip);
1571         return(0);
1572 }
1573
1574 /*
1575  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1576  * the read-only flag for cached inodes.
1577  *
1578  * This routine is called from a RB_SCAN().
1579  */
1580 int
1581 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1582 {
1583         hammer_mount_t hmp = ip->hmp;
1584
1585         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1586                 ip->flags |= HAMMER_INODE_RO;
1587         else
1588                 ip->flags &= ~HAMMER_INODE_RO;
1589         return(0);
1590 }
1591
1592 /*
1593  * A transaction has modified an inode, requiring updates as specified by
1594  * the passed flags.
1595  *
1596  * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1597  *                      and not including size changes due to write-append
1598  *                      (but other size changes are included).
1599  * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1600  *                      write-append.
1601  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1602  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1603  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1604  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1605  */
1606 void
1607 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1608 {
1609         /*
1610          * ronly of 0 or 2 does not trigger assertion.
1611          * 2 is a special error state
1612          */
1613         KKASSERT(ip->hmp->ronly != 1 ||
1614                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1615                             HAMMER_INODE_SDIRTY |
1616                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1617                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1618         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1619                 ip->flags |= HAMMER_INODE_RSV_INODES;
1620                 ++ip->hmp->rsv_inodes;
1621         }
1622
1623         /*
1624          * Set the NEWINODE flag in the transaction if the inode
1625          * transitions to a dirty state.  This is used to track
1626          * the load on the inode cache.
1627          */
1628         if (trans &&
1629             (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1630             (flags & HAMMER_INODE_MODMASK)) {
1631                 trans->flags |= HAMMER_TRANSF_NEWINODE;
1632         }
1633
1634         ip->flags |= flags;
1635 }
1636
1637 /*
1638  * Attempt to quickly update the atime for a hammer inode.  Return 0 on
1639  * success, -1 on failure.
1640  *
1641  * We attempt to update the atime with only the ip lock and not the
1642  * whole filesystem lock in order to improve concurrency.  We can only
1643  * do this safely if the ATIME flag is already pending on the inode.
1644  *
1645  * This function is called via a vnops path (ip pointer is stable) without
1646  * fs_token held.
1647  */
1648 int
1649 hammer_update_atime_quick(hammer_inode_t ip)
1650 {
1651         struct timeval tv;
1652         int res = -1;
1653
1654         if ((ip->flags & HAMMER_INODE_RO) ||
1655             (ip->hmp->mp->mnt_flag & MNT_NOATIME)) {
1656                 /*
1657                  * Silently indicate success on read-only mount/snap
1658                  */
1659                 res = 0;
1660         } else if (ip->flags & HAMMER_INODE_ATIME) {
1661                 /*
1662                  * Double check with inode lock held against backend.  This
1663                  * is only safe if all we need to do is update
1664                  * ino_data.atime.
1665                  */
1666                 getmicrotime(&tv);
1667                 hammer_lock_ex(&ip->lock);
1668                 if (ip->flags & HAMMER_INODE_ATIME) {
1669                         ip->ino_data.atime =
1670                             (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec;
1671                         res = 0;
1672                 }
1673                 hammer_unlock(&ip->lock);
1674         }
1675         return res;
1676 }
1677
1678 /*
1679  * Request that an inode be flushed.  This whole mess cannot block and may
1680  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1681  * actively flush the inode until the flush can be done.
1682  *
1683  * The inode may already be flushing, or may be in a setup state.  We can
1684  * place the inode in a flushing state if it is currently idle and flag it
1685  * to reflush if it is currently flushing.
1686  *
1687  * Upon return if the inode could not be flushed due to a setup
1688  * dependancy, then it will be automatically flushed when the dependancy
1689  * is satisfied.
1690  */
1691 void
1692 hammer_flush_inode(hammer_inode_t ip, int flags)
1693 {
1694         hammer_mount_t hmp;
1695         hammer_flush_group_t flg;
1696         int good;
1697
1698         /*
1699          * fill_flush_group is the first flush group we may be able to
1700          * continue filling, it may be open or closed but it will always
1701          * be past the currently flushing (running) flg.
1702          *
1703          * next_flush_group is the next open flush group.
1704          */
1705         hmp = ip->hmp;
1706         while ((flg = hmp->fill_flush_group) != NULL) {
1707                 KKASSERT(flg->running == 0);
1708                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1709                     flg->total_count <= hammer_autoflush) {
1710                         break;
1711                 }
1712                 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1713                 hammer_flusher_async(ip->hmp, flg);
1714         }
1715         if (flg == NULL) {
1716                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1717                 flg->seq = hmp->flusher.next++;
1718                 if (hmp->next_flush_group == NULL)
1719                         hmp->next_flush_group = flg;
1720                 if (hmp->fill_flush_group == NULL)
1721                         hmp->fill_flush_group = flg;
1722                 RB_INIT(&flg->flush_tree);
1723                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1724         }
1725
1726         /*
1727          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1728          * state we have to put it back into an IDLE state so we can
1729          * drop the extra ref.
1730          *
1731          * If we have a parent dependancy we must still fall through
1732          * so we can run it.
1733          */
1734         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1735                 if (ip->flush_state == HAMMER_FST_SETUP &&
1736                     TAILQ_EMPTY(&ip->target_list)) {
1737                         ip->flush_state = HAMMER_FST_IDLE;
1738                         hammer_rel_inode(ip, 0);
1739                 }
1740                 if (ip->flush_state == HAMMER_FST_IDLE)
1741                         return;
1742         }
1743
1744         /*
1745          * Our flush action will depend on the current state.
1746          */
1747         switch(ip->flush_state) {
1748         case HAMMER_FST_IDLE:
1749                 /*
1750                  * We have no dependancies and can flush immediately.  Some
1751                  * our children may not be flushable so we have to re-test
1752                  * with that additional knowledge.
1753                  */
1754                 hammer_flush_inode_core(ip, flg, flags);
1755                 break;
1756         case HAMMER_FST_SETUP:
1757                 /*
1758                  * Recurse upwards through dependancies via target_list
1759                  * and start their flusher actions going if possible.
1760                  *
1761                  * 'good' is our connectivity.  -1 means we have none and
1762                  * can't flush, 0 means there weren't any dependancies, and
1763                  * 1 means we have good connectivity.
1764                  */
1765                 good = hammer_setup_parent_inodes(ip, 0, flg);
1766
1767                 if (good >= 0) {
1768                         /*
1769                          * We can continue if good >= 0.  Determine how
1770                          * many records under our inode can be flushed (and
1771                          * mark them).
1772                          */
1773                         hammer_flush_inode_core(ip, flg, flags);
1774                 } else {
1775                         /*
1776                          * Parent has no connectivity, tell it to flush
1777                          * us as soon as it does.
1778                          *
1779                          * The REFLUSH flag is also needed to trigger
1780                          * dependancy wakeups.
1781                          */
1782                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1783                                      HAMMER_INODE_REFLUSH;
1784                         if (flags & HAMMER_FLUSH_SIGNAL) {
1785                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1786                                 hammer_flusher_async(ip->hmp, flg);
1787                         }
1788                 }
1789                 break;
1790         case HAMMER_FST_FLUSH:
1791                 /*
1792                  * We are already flushing, flag the inode to reflush
1793                  * if needed after it completes its current flush.
1794                  *
1795                  * The REFLUSH flag is also needed to trigger
1796                  * dependancy wakeups.
1797                  */
1798                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1799                         ip->flags |= HAMMER_INODE_REFLUSH;
1800                 if (flags & HAMMER_FLUSH_SIGNAL) {
1801                         ip->flags |= HAMMER_INODE_RESIGNAL;
1802                         hammer_flusher_async(ip->hmp, flg);
1803                 }
1804                 break;
1805         }
1806 }
1807
1808 /*
1809  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1810  * ip which reference our ip.
1811  *
1812  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1813  *     so for now do not ref/deref the structures.  Note that if we use the
1814  *     ref/rel code later, the rel CAN block.
1815  */
1816 static int
1817 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1818                            hammer_flush_group_t flg)
1819 {
1820         hammer_record_t depend;
1821         int good;
1822         int r;
1823
1824         /*
1825          * If we hit our recursion limit and we have parent dependencies
1826          * We cannot continue.  Returning < 0 will cause us to be flagged
1827          * for reflush.  Returning -2 cuts off additional dependency checks
1828          * because they are likely to also hit the depth limit.
1829          *
1830          * We cannot return < 0 if there are no dependencies or there might
1831          * not be anything to wakeup (ip).
1832          */
1833         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1834                 kprintf("HAMMER Warning: depth limit reached on "
1835                         "setup recursion, inode %p %016llx\n",
1836                         ip, (long long)ip->obj_id);
1837                 return(-2);
1838         }
1839
1840         /*
1841          * Scan dependencies
1842          */
1843         good = 0;
1844         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1845                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1846                 KKASSERT(depend->target_ip == ip);
1847                 if (r < 0 && good == 0)
1848                         good = -1;
1849                 if (r > 0)
1850                         good = 1;
1851
1852                 /*
1853                  * If we failed due to the recursion depth limit then stop
1854                  * now.
1855                  */
1856                 if (r == -2)
1857                         break;
1858         }
1859         return(good);
1860 }
1861
1862 /*
1863  * This helper function takes a record representing the dependancy between
1864  * the parent inode and child inode.
1865  *
1866  * record->ip           = parent inode
1867  * record->target_ip    = child inode
1868  *
1869  * We are asked to recurse upwards and convert the record from SETUP
1870  * to FLUSH if possible.
1871  *
1872  * Return 1 if the record gives us connectivity
1873  *
1874  * Return 0 if the record is not relevant
1875  *
1876  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1877  */
1878 static int
1879 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1880                                   hammer_flush_group_t flg)
1881 {
1882         hammer_mount_t hmp;
1883         hammer_inode_t pip;
1884         int good;
1885
1886         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1887         pip = record->ip;
1888         hmp = pip->hmp;
1889
1890         /*
1891          * If the record is already flushing, is it in our flush group?
1892          *
1893          * If it is in our flush group but it is a general record or a
1894          * delete-on-disk, it does not improve our connectivity (return 0),
1895          * and if the target inode is not trying to destroy itself we can't
1896          * allow the operation yet anyway (the second return -1).
1897          */
1898         if (record->flush_state == HAMMER_FST_FLUSH) {
1899                 /*
1900                  * If not in our flush group ask the parent to reflush
1901                  * us as soon as possible.
1902                  */
1903                 if (record->flush_group != flg) {
1904                         pip->flags |= HAMMER_INODE_REFLUSH;
1905                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1906                         return(-1);
1907                 }
1908
1909                 /*
1910                  * If in our flush group everything is already set up,
1911                  * just return whether the record will improve our
1912                  * visibility or not.
1913                  */
1914                 if (record->type == HAMMER_MEM_RECORD_ADD)
1915                         return(1);
1916                 return(0);
1917         }
1918
1919         /*
1920          * It must be a setup record.  Try to resolve the setup dependancies
1921          * by recursing upwards so we can place ip on the flush list.
1922          *
1923          * Limit ourselves to 20 levels of recursion to avoid blowing out
1924          * the kernel stack.  If we hit the recursion limit we can't flush
1925          * until the parent flushes.  The parent will flush independantly
1926          * on its own and ultimately a deep recursion will be resolved.
1927          */
1928         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1929
1930         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1931
1932         /*
1933          * If good < 0 the parent has no connectivity and we cannot safely
1934          * flush the directory entry, which also means we can't flush our
1935          * ip.  Flag us for downward recursion once the parent's
1936          * connectivity is resolved.  Flag the parent for [re]flush or it
1937          * may not check for downward recursions.
1938          */
1939         if (good < 0) {
1940                 pip->flags |= HAMMER_INODE_REFLUSH;
1941                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1942                 return(good);
1943         }
1944
1945         /*
1946          * We are go, place the parent inode in a flushing state so we can
1947          * place its record in a flushing state.  Note that the parent
1948          * may already be flushing.  The record must be in the same flush
1949          * group as the parent.
1950          */
1951         if (pip->flush_state != HAMMER_FST_FLUSH)
1952                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1953         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1954
1955         /*
1956          * It is possible for a rename to create a loop in the recursion
1957          * and revisit a record.  This will result in the record being
1958          * placed in a flush state unexpectedly.  This check deals with
1959          * the case.
1960          */
1961         if (record->flush_state == HAMMER_FST_FLUSH) {
1962                 if (record->type == HAMMER_MEM_RECORD_ADD)
1963                         return(1);
1964                 return(0);
1965         }
1966
1967         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1968
1969 #if 0
1970         if (record->type == HAMMER_MEM_RECORD_DEL &&
1971             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1972                 /*
1973                  * Regardless of flushing state we cannot sync this path if the
1974                  * record represents a delete-on-disk but the target inode
1975                  * is not ready to sync its own deletion.
1976                  *
1977                  * XXX need to count effective nlinks to determine whether
1978                  * the flush is ok, otherwise removing a hardlink will
1979                  * just leave the DEL record to rot.
1980                  */
1981                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1982                 return(-1);
1983         } else
1984 #endif
1985         if (pip->flush_group == flg) {
1986                 /*
1987                  * Because we have not calculated nlinks yet we can just
1988                  * set records to the flush state if the parent is in
1989                  * the same flush group as we are.
1990                  */
1991                 record->flush_state = HAMMER_FST_FLUSH;
1992                 record->flush_group = flg;
1993                 ++record->flush_group->refs;
1994                 hammer_ref(&record->lock);
1995
1996                 /*
1997                  * A general directory-add contributes to our visibility.
1998                  *
1999                  * Otherwise it is probably a directory-delete or
2000                  * delete-on-disk record and does not contribute to our
2001                  * visbility (but we can still flush it).
2002                  */
2003                 if (record->type == HAMMER_MEM_RECORD_ADD)
2004                         return(1);
2005                 return(0);
2006         } else {
2007                 /*
2008                  * If the parent is not in our flush group we cannot
2009                  * flush this record yet, there is no visibility.
2010                  * We tell the parent to reflush and mark ourselves
2011                  * so the parent knows it should flush us too.
2012                  */
2013                 pip->flags |= HAMMER_INODE_REFLUSH;
2014                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2015                 return(-1);
2016         }
2017 }
2018
2019 /*
2020  * This is the core routine placing an inode into the FST_FLUSH state.
2021  */
2022 static void
2023 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
2024 {
2025         hammer_mount_t hmp = ip->hmp;
2026         int go_count;
2027
2028         /*
2029          * Set flush state and prevent the flusher from cycling into
2030          * the next flush group.  Do not place the ip on the list yet.
2031          * Inodes not in the idle state get an extra reference.
2032          */
2033         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
2034         if (ip->flush_state == HAMMER_FST_IDLE)
2035                 hammer_ref(&ip->lock);
2036         ip->flush_state = HAMMER_FST_FLUSH;
2037         ip->flush_group = flg;
2038         ++hmp->flusher.group_lock;
2039         ++hmp->count_iqueued;
2040         ++hammer_count_iqueued;
2041         ++flg->total_count;
2042         hammer_redo_fifo_start_flush(ip);
2043
2044 #if 0
2045         /*
2046          * We need to be able to vfsync/truncate from the backend.
2047          *
2048          * XXX Any truncation from the backend will acquire the vnode
2049          *     independently.
2050          */
2051         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2052         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2053                 ip->flags |= HAMMER_INODE_VHELD;
2054                 vref(ip->vp);
2055         }
2056 #endif
2057
2058         /*
2059          * Figure out how many in-memory records we can actually flush
2060          * (not including inode meta-data, buffers, etc).
2061          */
2062         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2063         if (flags & HAMMER_FLUSH_RECURSION) {
2064                 /*
2065                  * If this is a upwards recursion we do not want to
2066                  * recurse down again!
2067                  */
2068                 go_count = 1;
2069 #if 0
2070         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2071                 /*
2072                  * No new records are added if we must complete a flush
2073                  * from a previous cycle, but we do have to move the records
2074                  * from the previous cycle to the current one.
2075                  */
2076 #if 0
2077                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2078                                    hammer_syncgrp_child_callback, NULL);
2079 #endif
2080                 go_count = 1;
2081 #endif
2082         } else {
2083                 /*
2084                  * Normal flush, scan records and bring them into the flush.
2085                  * Directory adds and deletes are usually skipped (they are
2086                  * grouped with the related inode rather then with the
2087                  * directory).
2088                  *
2089                  * go_count can be negative, which means the scan aborted
2090                  * due to the flush group being over-full and we should
2091                  * flush what we have.
2092                  */
2093                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2094                                    hammer_setup_child_callback, NULL);
2095         }
2096
2097         /*
2098          * This is a more involved test that includes go_count.  If we
2099          * can't flush, flag the inode and return.  If go_count is 0 we
2100          * were are unable to flush any records in our rec_tree and
2101          * must ignore the XDIRTY flag.
2102          */
2103         if (go_count == 0) {
2104                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2105                         --hmp->count_iqueued;
2106                         --hammer_count_iqueued;
2107
2108                         --flg->total_count;
2109                         ip->flush_state = HAMMER_FST_SETUP;
2110                         ip->flush_group = NULL;
2111                         if (flags & HAMMER_FLUSH_SIGNAL) {
2112                                 ip->flags |= HAMMER_INODE_REFLUSH |
2113                                              HAMMER_INODE_RESIGNAL;
2114                         } else {
2115                                 ip->flags |= HAMMER_INODE_REFLUSH;
2116                         }
2117 #if 0
2118                         if (ip->flags & HAMMER_INODE_VHELD) {
2119                                 ip->flags &= ~HAMMER_INODE_VHELD;
2120                                 vrele(ip->vp);
2121                         }
2122 #endif
2123
2124                         /*
2125                          * REFLUSH is needed to trigger dependancy wakeups
2126                          * when an inode is in SETUP.
2127                          */
2128                         ip->flags |= HAMMER_INODE_REFLUSH;
2129                         if (--hmp->flusher.group_lock == 0)
2130                                 wakeup(&hmp->flusher.group_lock);
2131                         return;
2132                 }
2133         }
2134
2135         /*
2136          * Snapshot the state of the inode for the backend flusher.
2137          *
2138          * We continue to retain save_trunc_off even when all truncations
2139          * have been resolved as an optimization to determine if we can
2140          * skip the B-Tree lookup for overwrite deletions.
2141          *
2142          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2143          * and stays in ip->flags.  Once set, it stays set until the
2144          * inode is destroyed.
2145          */
2146         if (ip->flags & HAMMER_INODE_TRUNCATED) {
2147                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2148                 ip->sync_trunc_off = ip->trunc_off;
2149                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2150                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2151                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2152
2153                 /*
2154                  * The save_trunc_off used to cache whether the B-Tree
2155                  * holds any records past that point is not used until
2156                  * after the truncation has succeeded, so we can safely
2157                  * set it now.
2158                  */
2159                 if (ip->save_trunc_off > ip->sync_trunc_off)
2160                         ip->save_trunc_off = ip->sync_trunc_off;
2161         }
2162         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2163                            ~HAMMER_INODE_TRUNCATED);
2164         ip->sync_ino_leaf = ip->ino_leaf;
2165         ip->sync_ino_data = ip->ino_data;
2166         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2167 #ifdef DEBUG_TRUNCATE
2168         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2169                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2170 #endif
2171
2172         /*
2173          * The flusher list inherits our inode and reference.
2174          */
2175         KKASSERT(flg->running == 0);
2176         RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2177         if (--hmp->flusher.group_lock == 0)
2178                 wakeup(&hmp->flusher.group_lock);
2179
2180         /*
2181          * Auto-flush the group if it grows too large.  Make sure the
2182          * inode reclaim wait pipeline continues to work.
2183          */
2184         if (flg->total_count >= hammer_autoflush ||
2185             flg->total_count >= hammer_limit_reclaims / 4) {
2186                 if (hmp->fill_flush_group == flg)
2187                         hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2188                 hammer_flusher_async(hmp, flg);
2189         }
2190 }
2191
2192 /*
2193  * Callback for scan of ip->rec_tree.  Try to include each record in our
2194  * flush.  ip->flush_group has been set but the inode has not yet been
2195  * moved into a flushing state.
2196  *
2197  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2198  * both inodes.
2199  *
2200  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2201  * the caller from shortcutting the flush.
2202  */
2203 static int
2204 hammer_setup_child_callback(hammer_record_t rec, void *data)
2205 {
2206         hammer_flush_group_t flg;
2207         hammer_inode_t target_ip;
2208         hammer_inode_t ip;
2209         int r;
2210
2211         /*
2212          * Records deleted or committed by the backend are ignored.
2213          * Note that the flush detects deleted frontend records at
2214          * multiple points to deal with races.  This is just the first
2215          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2216          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2217          * messes up link-count calculations.
2218          *
2219          * NOTE: Don't get confused between record deletion and, say,
2220          * directory entry deletion.  The deletion of a directory entry
2221          * which is on-media has nothing to do with the record deletion
2222          * flags.
2223          */
2224         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2225                           HAMMER_RECF_COMMITTED)) {
2226                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2227                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2228                         r = 1;
2229                 } else {
2230                         r = 0;
2231                 }
2232                 return(r);
2233         }
2234
2235         /*
2236          * If the record is in an idle state it has no dependancies and
2237          * can be flushed.
2238          */
2239         ip = rec->ip;
2240         flg = ip->flush_group;
2241         r = 0;
2242
2243         switch(rec->flush_state) {
2244         case HAMMER_FST_IDLE:
2245                 /*
2246                  * The record has no setup dependancy, we can flush it.
2247                  */
2248                 KKASSERT(rec->target_ip == NULL);
2249                 rec->flush_state = HAMMER_FST_FLUSH;
2250                 rec->flush_group = flg;
2251                 ++flg->refs;
2252                 hammer_ref(&rec->lock);
2253                 r = 1;
2254                 break;
2255         case HAMMER_FST_SETUP:
2256                 /*
2257                  * The record has a setup dependancy.  These are typically
2258                  * directory entry adds and deletes.  Such entries will be
2259                  * flushed when their inodes are flushed so we do not
2260                  * usually have to add them to the flush here.  However,
2261                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2262                  * it is asking us to flush this record (and it).
2263                  */
2264                 target_ip = rec->target_ip;
2265                 KKASSERT(target_ip != NULL);
2266                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2267
2268                 /*
2269                  * If the target IP is already flushing in our group
2270                  * we could associate the record, but target_ip has
2271                  * already synced ino_data to sync_ino_data and we
2272                  * would also have to adjust nlinks.   Plus there are
2273                  * ordering issues for adds and deletes.
2274                  *
2275                  * Reflush downward if this is an ADD, and upward if
2276                  * this is a DEL.
2277                  */
2278                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2279                         if (rec->type == HAMMER_MEM_RECORD_ADD)
2280                                 ip->flags |= HAMMER_INODE_REFLUSH;
2281                         else
2282                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2283                         break;
2284                 }
2285
2286                 /*
2287                  * Target IP is not yet flushing.  This can get complex
2288                  * because we have to be careful about the recursion.
2289                  *
2290                  * Directories create an issue for us in that if a flush
2291                  * of a directory is requested the expectation is to flush
2292                  * any pending directory entries, but this will cause the
2293                  * related inodes to recursively flush as well.  We can't
2294                  * really defer the operation so just get as many as we
2295                  * can and
2296                  */
2297 #if 0
2298                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2299                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2300                         /*
2301                          * We aren't reclaiming and the target ip was not
2302                          * previously prevented from flushing due to this
2303                          * record dependancy.  Do not flush this record.
2304                          */
2305                         /*r = 0;*/
2306                 } else
2307 #endif
2308                 if (flg->total_count + flg->refs >
2309                            ip->hmp->undo_rec_limit) {
2310                         /*
2311                          * Our flush group is over-full and we risk blowing
2312                          * out the UNDO FIFO.  Stop the scan, flush what we
2313                          * have, then reflush the directory.
2314                          *
2315                          * The directory may be forced through multiple
2316                          * flush groups before it can be completely
2317                          * flushed.
2318                          */
2319                         ip->flags |= HAMMER_INODE_RESIGNAL |
2320                                      HAMMER_INODE_REFLUSH;
2321                         r = -1;
2322                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2323                         /*
2324                          * If the target IP is not flushing we can force
2325                          * it to flush, even if it is unable to write out
2326                          * any of its own records we have at least one in
2327                          * hand that we CAN deal with.
2328                          */
2329                         rec->flush_state = HAMMER_FST_FLUSH;
2330                         rec->flush_group = flg;
2331                         ++flg->refs;
2332                         hammer_ref(&rec->lock);
2333                         hammer_flush_inode_core(target_ip, flg,
2334                                                 HAMMER_FLUSH_RECURSION);
2335                         r = 1;
2336                 } else {
2337                         /*
2338                          * General or delete-on-disk record.
2339                          *
2340                          * XXX this needs help.  If a delete-on-disk we could
2341                          * disconnect the target.  If the target has its own
2342                          * dependancies they really need to be flushed.
2343                          *
2344                          * XXX
2345                          */
2346                         rec->flush_state = HAMMER_FST_FLUSH;
2347                         rec->flush_group = flg;
2348                         ++flg->refs;
2349                         hammer_ref(&rec->lock);
2350                         hammer_flush_inode_core(target_ip, flg,
2351                                                 HAMMER_FLUSH_RECURSION);
2352                         r = 1;
2353                 }
2354                 break;
2355         case HAMMER_FST_FLUSH:
2356                 /*
2357                  * The record could be part of a previous flush group if the
2358                  * inode is a directory (the record being a directory entry).
2359                  * Once the flush group was closed a hammer_test_inode()
2360                  * function can cause a new flush group to be setup, placing
2361                  * the directory inode itself in a new flush group.
2362                  *
2363                  * When associated with a previous flush group we count it
2364                  * as if it were in our current flush group, since it will
2365                  * effectively be flushed by the time we flush our current
2366                  * flush group.
2367                  */
2368                 KKASSERT(
2369                     rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2370                     rec->flush_group == flg);
2371                 r = 1;
2372                 break;
2373         }
2374         return(r);
2375 }
2376
2377 #if 0
2378 /*
2379  * This version just moves records already in a flush state to the new
2380  * flush group and that is it.
2381  */
2382 static int
2383 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2384 {
2385         hammer_inode_t ip = rec->ip;
2386
2387         switch(rec->flush_state) {
2388         case HAMMER_FST_FLUSH:
2389                 KKASSERT(rec->flush_group == ip->flush_group);
2390                 break;
2391         default:
2392                 break;
2393         }
2394         return(0);
2395 }
2396 #endif
2397
2398 /*
2399  * Wait for a previously queued flush to complete.
2400  *
2401  * If a critical error occured we don't try to wait.
2402  */
2403 void
2404 hammer_wait_inode(hammer_inode_t ip)
2405 {
2406         hammer_flush_group_t flg;
2407
2408         flg = NULL;
2409
2410         /*
2411          * The inode can be in a SETUP state in which case RESIGNAL
2412          * should be set.  If RESIGNAL is not set then the previous
2413          * flush completed and a later operation placed the inode
2414          * in a passive setup state again, so we're done.
2415          *
2416          * The inode can be in a FLUSH state in which case we
2417          * can just wait for completion.
2418          */
2419         while (ip->flush_state == HAMMER_FST_FLUSH ||
2420             (ip->flush_state == HAMMER_FST_SETUP &&
2421              (ip->flags & HAMMER_INODE_RESIGNAL))) {
2422                 /*
2423                  * Don't try to flush on a critical error
2424                  */
2425                 if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
2426                         break;
2427
2428                 /*
2429                  * If the inode was already being flushed its flg
2430                  * may not have been queued to the backend.  We
2431                  * have to make sure it gets queued or we can wind
2432                  * up blocked or deadlocked (particularly if we are
2433                  * the vnlru thread).
2434                  */
2435                 if (ip->flush_state == HAMMER_FST_FLUSH) {
2436                         KKASSERT(ip->flush_group);
2437                         if (ip->flush_group->closed == 0) {
2438                                 kprintf("hammer: debug: forcing async "
2439                                         "flush ip %016jx\n",
2440                                         (intmax_t)ip->obj_id);
2441                                 hammer_flusher_async(ip->hmp,
2442                                                      ip->flush_group);
2443                                 continue; /* retest */
2444                         }
2445                 }
2446
2447                 /*
2448                  * In a flush state with the flg queued to the backend
2449                  * or in a setup state with RESIGNAL set, we can safely
2450                  * wait.
2451                  */
2452                 ip->flags |= HAMMER_INODE_FLUSHW;
2453                 tsleep(&ip->flags, 0, "hmrwin", 0);
2454         }
2455
2456 #if 0
2457         /*
2458          * The inode may have been in a passive setup state,
2459          * call flush to make sure we get signaled.
2460          */
2461         if (ip->flush_state == HAMMER_FST_SETUP)
2462                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2463 #endif
2464
2465 }
2466
2467 /*
2468  * Called by the backend code when a flush has been completed.
2469  * The inode has already been removed from the flush list.
2470  *
2471  * A pipelined flush can occur, in which case we must re-enter the
2472  * inode on the list and re-copy its fields.
2473  */
2474 void
2475 hammer_flush_inode_done(hammer_inode_t ip, int error)
2476 {
2477         hammer_mount_t hmp;
2478         int dorel;
2479
2480         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2481
2482         hmp = ip->hmp;
2483
2484         /*
2485          * Auto-reflush if the backend could not completely flush
2486          * the inode.  This fixes a case where a deferred buffer flush
2487          * could cause fsync to return early.
2488          */
2489         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2490                 ip->flags |= HAMMER_INODE_REFLUSH;
2491
2492         /*
2493          * Merge left-over flags back into the frontend and fix the state.
2494          * Incomplete truncations are retained by the backend.
2495          */
2496         ip->error = error;
2497         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2498         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2499
2500         /*
2501          * The backend may have adjusted nlinks, so if the adjusted nlinks
2502          * does not match the fronttend set the frontend's DDIRTY flag again.
2503          */
2504         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2505                 ip->flags |= HAMMER_INODE_DDIRTY;
2506
2507         /*
2508          * Fix up the dirty buffer status.
2509          */
2510         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2511                 ip->flags |= HAMMER_INODE_BUFS;
2512         }
2513         hammer_redo_fifo_end_flush(ip);
2514
2515         /*
2516          * Re-set the XDIRTY flag if some of the inode's in-memory records
2517          * could not be flushed.
2518          */
2519         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2520                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2521                  (!RB_EMPTY(&ip->rec_tree) &&
2522                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2523
2524         /*
2525          * Do not lose track of inodes which no longer have vnode
2526          * assocations, otherwise they may never get flushed again.
2527          *
2528          * The reflush flag can be set superfluously, causing extra pain
2529          * for no reason.  If the inode is no longer modified it no longer
2530          * needs to be flushed.
2531          */
2532         if (ip->flags & HAMMER_INODE_MODMASK) {
2533                 if (ip->vp == NULL)
2534                         ip->flags |= HAMMER_INODE_REFLUSH;
2535         } else {
2536                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2537         }
2538
2539         /*
2540          * Adjust the flush state.
2541          */
2542         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2543                 /*
2544                  * We were unable to flush out all our records, leave the
2545                  * inode in a flush state and in the current flush group.
2546                  * The flush group will be re-run.
2547                  *
2548                  * This occurs if the UNDO block gets too full or there is
2549                  * too much dirty meta-data and allows the flusher to
2550                  * finalize the UNDO block and then re-flush.
2551                  */
2552                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2553                 dorel = 0;
2554         } else {
2555                 /*
2556                  * Remove from the flush_group
2557                  */
2558                 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2559                 ip->flush_group = NULL;
2560
2561 #if 0
2562                 /*
2563                  * Clean up the vnode ref and tracking counts.
2564                  */
2565                 if (ip->flags & HAMMER_INODE_VHELD) {
2566                         ip->flags &= ~HAMMER_INODE_VHELD;
2567                         vrele(ip->vp);
2568                 }
2569 #endif
2570                 --hmp->count_iqueued;
2571                 --hammer_count_iqueued;
2572
2573                 /*
2574                  * And adjust the state.
2575                  */
2576                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2577                         ip->flush_state = HAMMER_FST_IDLE;
2578                         dorel = 1;
2579                 } else {
2580                         ip->flush_state = HAMMER_FST_SETUP;
2581                         dorel = 0;
2582                 }
2583
2584                 /*
2585                  * If the frontend is waiting for a flush to complete,
2586                  * wake it up.
2587                  */
2588                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2589                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2590                         wakeup(&ip->flags);
2591                 }
2592
2593                 /*
2594                  * If the frontend made more changes and requested another
2595                  * flush, then try to get it running.
2596                  *
2597                  * Reflushes are aborted when the inode is errored out.
2598                  */
2599                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2600                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2601                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2602                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2603                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2604                         } else {
2605                                 hammer_flush_inode(ip, 0);
2606                         }
2607                 }
2608         }
2609
2610         /*
2611          * If we have no parent dependancies we can clear CONN_DOWN
2612          */
2613         if (TAILQ_EMPTY(&ip->target_list))
2614                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2615
2616         /*
2617          * If the inode is now clean drop the space reservation.
2618          */
2619         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2620             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2621                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2622                 --hmp->rsv_inodes;
2623         }
2624
2625         ip->flags &= ~HAMMER_INODE_SLAVEFLUSH;
2626
2627         if (dorel)
2628                 hammer_rel_inode(ip, 0);
2629 }
2630
2631 /*
2632  * Called from hammer_sync_inode() to synchronize in-memory records
2633  * to the media.
2634  */
2635 static int
2636 hammer_sync_record_callback(hammer_record_t record, void *data)
2637 {
2638         hammer_cursor_t cursor = data;
2639         hammer_transaction_t trans = cursor->trans;
2640         hammer_mount_t hmp = trans->hmp;
2641         int error;
2642
2643         /*
2644          * Skip records that do not belong to the current flush.
2645          */
2646         ++hammer_stats_record_iterations;
2647         if (record->flush_state != HAMMER_FST_FLUSH)
2648                 return(0);
2649
2650 #if 1
2651         if (record->flush_group != record->ip->flush_group) {
2652                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2653                 if (hammer_debug_critical)
2654                         Debugger("blah2");
2655                 return(0);
2656         }
2657 #endif
2658         KKASSERT(record->flush_group == record->ip->flush_group);
2659
2660         /*
2661          * Interlock the record using the BE flag.  Once BE is set the
2662          * frontend cannot change the state of FE.
2663          *
2664          * NOTE: If FE is set prior to us setting BE we still sync the
2665          * record out, but the flush completion code converts it to
2666          * a delete-on-disk record instead of destroying it.
2667          */
2668         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2669         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2670
2671         /*
2672          * The backend has already disposed of the record.
2673          */
2674         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2675                 error = 0;
2676                 goto done;
2677         }
2678
2679         /*
2680          * If the whole inode is being deleted and all on-disk records will
2681          * be deleted very soon, we can't sync any new records to disk
2682          * because they will be deleted in the same transaction they were
2683          * created in (delete_tid == create_tid), which will assert.
2684          *
2685          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2686          * that we currently panic on.
2687          */
2688         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2689                 switch(record->type) {
2690                 case HAMMER_MEM_RECORD_DATA:
2691                         /*
2692                          * We don't have to do anything, if the record was
2693                          * committed the space will have been accounted for
2694                          * in the blockmap.
2695                          */
2696                         /* fall through */
2697                 case HAMMER_MEM_RECORD_GENERAL:
2698                         /*
2699                          * Set deleted-by-backend flag.  Do not set the
2700                          * backend committed flag, because we are throwing
2701                          * the record away.
2702                          */
2703                         record->flags |= HAMMER_RECF_DELETED_BE;
2704                         ++record->ip->rec_generation;
2705                         error = 0;
2706                         goto done;
2707                 case HAMMER_MEM_RECORD_ADD:
2708                         panic("hammer_sync_record_callback: illegal add "
2709                               "during inode deletion record %p", record);
2710                         break; /* NOT REACHED */
2711                 case HAMMER_MEM_RECORD_INODE:
2712                         panic("hammer_sync_record_callback: attempt to "
2713                               "sync inode record %p?", record);
2714                         break; /* NOT REACHED */
2715                 case HAMMER_MEM_RECORD_DEL:
2716                         /*
2717                          * Follow through and issue the on-disk deletion
2718                          */
2719                         break;
2720                 }
2721         }
2722
2723         /*
2724          * If DELETED_FE is set special handling is needed for directory
2725          * entries.  Dependant pieces related to the directory entry may
2726          * have already been synced to disk.  If this occurs we have to
2727          * sync the directory entry and then change the in-memory record
2728          * from an ADD to a DELETE to cover the fact that it's been
2729          * deleted by the frontend.
2730          *
2731          * A directory delete covering record (MEM_RECORD_DEL) can never
2732          * be deleted by the frontend.
2733          *
2734          * Any other record type (aka DATA) can be deleted by the frontend.
2735          * XXX At the moment the flusher must skip it because there may
2736          * be another data record in the flush group for the same block,
2737          * meaning that some frontend data changes can leak into the backend's
2738          * synchronization point.
2739          */
2740         if (record->flags & HAMMER_RECF_DELETED_FE) {
2741                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2742                         /*
2743                          * Convert a front-end deleted directory-add to
2744                          * a directory-delete entry later.
2745                          */
2746                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2747                 } else {
2748                         /*
2749                          * Dispose of the record (race case).  Mark as
2750                          * deleted by backend (and not committed).
2751                          */
2752                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2753                         record->flags |= HAMMER_RECF_DELETED_BE;
2754                         ++record->ip->rec_generation;
2755                         error = 0;
2756                         goto done;
2757                 }
2758         }
2759
2760         /*
2761          * Assign the create_tid for new records.  Deletions already
2762          * have the record's entire key properly set up.
2763          */
2764         if (record->type != HAMMER_MEM_RECORD_DEL) {
2765                 record->leaf.base.create_tid = trans->tid;
2766                 record->leaf.create_ts = trans->time32;
2767         }
2768
2769         /*
2770          * This actually moves the record to the on-media B-Tree.  We
2771          * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2772          * indicating that the related REDO_WRITE(s) have been committed.
2773          *
2774          * During recovery any REDO_TERM's within the nominal recovery span
2775          * are ignored since the related meta-data is being undone, causing
2776          * any matching REDO_WRITEs to execute.  The REDO_TERMs outside
2777          * the nominal recovery span will match against REDO_WRITEs and
2778          * prevent them from being executed (because the meta-data has
2779          * already been synchronized).
2780          */
2781         if (record->flags & HAMMER_RECF_REDO) {
2782                 KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2783                 hammer_generate_redo(trans, record->ip,
2784                                      record->leaf.base.key -
2785                                          record->leaf.data_len,
2786                                      HAMMER_REDO_TERM_WRITE,
2787                                      NULL,
2788                                      record->leaf.data_len);
2789         }
2790
2791         for (;;) {
2792                 error = hammer_ip_sync_record_cursor(cursor, record);
2793                 if (error != EDEADLK)
2794                         break;
2795                 hammer_done_cursor(cursor);
2796                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2797                                            record->ip);
2798                 if (error)
2799                         break;
2800         }
2801         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2802
2803         if (error)
2804                 error = -error;
2805 done:
2806         hammer_flush_record_done(record, error);
2807
2808         /*
2809          * Do partial finalization if we have built up too many dirty
2810          * buffers.  Otherwise a buffer cache deadlock can occur when
2811          * doing things like creating tens of thousands of tiny files.
2812          *
2813          * We must release our cursor lock to avoid a 3-way deadlock
2814          * due to the exclusive sync lock the finalizer must get.
2815          *
2816          * WARNING: See warnings in hammer_unlock_cursor() function.
2817          */
2818         if (hammer_flusher_meta_limit(hmp) ||
2819             vm_page_count_severe()) {
2820                 hammer_unlock_cursor(cursor);
2821                 hammer_flusher_finalize(trans, 0);
2822                 hammer_lock_cursor(cursor);
2823         }
2824         return(error);
2825 }
2826
2827 /*
2828  * Backend function called by the flusher to sync an inode to media.
2829  */
2830 int
2831 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2832 {
2833         struct hammer_cursor cursor;
2834         hammer_node_t tmp_node;
2835         hammer_record_t depend;
2836         hammer_record_t next;
2837         int error, tmp_error;
2838         u_int64_t nlinks;
2839
2840         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2841                 return(0);
2842
2843         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2844         if (error)
2845                 goto done;
2846
2847         /*
2848          * Any directory records referencing this inode which are not in
2849          * our current flush group must adjust our nlink count for the
2850          * purposes of synchronizating to disk.
2851          *
2852          * Records which are in our flush group can be unlinked from our
2853          * inode now, potentially allowing the inode to be physically
2854          * deleted.
2855          *
2856          * This cannot block.
2857          */
2858         nlinks = ip->ino_data.nlinks;
2859         next = TAILQ_FIRST(&ip->target_list);
2860         while ((depend = next) != NULL) {
2861                 next = TAILQ_NEXT(depend, target_entry);
2862                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2863                     depend->flush_group == ip->flush_group) {
2864                         /*
2865                          * If this is an ADD that was deleted by the frontend
2866                          * the frontend nlinks count will have already been
2867                          * decremented, but the backend is going to sync its
2868                          * directory entry and must account for it.  The
2869                          * record will be converted to a delete-on-disk when
2870                          * it gets synced.
2871                          *
2872                          * If the ADD was not deleted by the frontend we
2873                          * can remove the dependancy from our target_list.
2874                          */
2875                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2876                                 ++nlinks;
2877                         } else {
2878                                 TAILQ_REMOVE(&ip->target_list, depend,
2879                                              target_entry);
2880                                 depend->target_ip = NULL;
2881                         }
2882                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2883                         /*
2884                          * Not part of our flush group and not deleted by
2885                          * the front-end, adjust the link count synced to
2886                          * the media (undo what the frontend did when it
2887                          * queued the record).
2888                          */
2889                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2890                         switch(depend->type) {
2891                         case HAMMER_MEM_RECORD_ADD:
2892                                 --nlinks;
2893                                 break;
2894                         case HAMMER_MEM_RECORD_DEL:
2895                                 ++nlinks;
2896                                 break;
2897                         default:
2898                                 break;
2899                         }
2900                 }
2901         }
2902
2903         /*
2904          * Set dirty if we had to modify the link count.
2905          */
2906         if (ip->sync_ino_data.nlinks != nlinks) {
2907                 KKASSERT((int64_t)nlinks >= 0);
2908                 ip->sync_ino_data.nlinks = nlinks;
2909                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2910         }
2911
2912         /*
2913          * If there is a trunction queued destroy any data past the (aligned)
2914          * truncation point.  Userland will have dealt with the buffer
2915          * containing the truncation point for us.
2916          *
2917          * We don't flush pending frontend data buffers until after we've
2918          * dealt with the truncation.
2919          */
2920         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2921                 /*
2922                  * Interlock trunc_off.  The VOP front-end may continue to
2923                  * make adjustments to it while we are blocked.
2924                  */
2925                 off_t trunc_off;
2926                 off_t aligned_trunc_off;
2927                 int blkmask;
2928
2929                 trunc_off = ip->sync_trunc_off;
2930                 blkmask = hammer_blocksize(trunc_off) - 1;
2931                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2932
2933                 /*
2934                  * Delete any whole blocks on-media.  The front-end has
2935                  * already cleaned out any partial block and made it
2936                  * pending.  The front-end may have updated trunc_off
2937                  * while we were blocked so we only use sync_trunc_off.
2938                  *
2939                  * This operation can blow out the buffer cache, EWOULDBLOCK
2940                  * means we were unable to complete the deletion.  The
2941                  * deletion will update sync_trunc_off in that case.
2942                  */
2943                 error = hammer_ip_delete_range(&cursor, ip,
2944                                                 aligned_trunc_off,
2945                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2946                 if (error == EWOULDBLOCK) {
2947                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2948                         error = 0;
2949                         goto defer_buffer_flush;
2950                 }
2951
2952                 if (error)
2953                         goto done;
2954
2955                 /*
2956                  * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
2957                  *
2958                  * XXX we do this even if we did not previously generate
2959                  * a REDO_TRUNC record.  This operation may enclosed the
2960                  * range for multiple prior truncation entries in the REDO
2961                  * log.
2962                  */
2963                 if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
2964                     (ip->flags & HAMMER_INODE_RDIRTY)) {
2965                         hammer_generate_redo(trans, ip, aligned_trunc_off,
2966                                              HAMMER_REDO_TERM_TRUNC,
2967                                              NULL, 0);
2968                 }
2969
2970                 /*
2971                  * Clear the truncation flag on the backend after we have
2972                  * completed the deletions.  Backend data is now good again
2973                  * (including new records we are about to sync, below).
2974                  *
2975                  * Leave sync_trunc_off intact.  As we write additional
2976                  * records the backend will update sync_trunc_off.  This
2977                  * tells the backend whether it can skip the overwrite
2978                  * test.  This should work properly even when the backend
2979                  * writes full blocks where the truncation point straddles
2980                  * the block because the comparison is against the base
2981                  * offset of the record.
2982                  */
2983                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2984                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2985         } else {
2986                 error = 0;
2987         }
2988
2989         /*
2990          * Now sync related records.  These will typically be directory
2991          * entries, records tracking direct-writes, or delete-on-disk records.
2992          */
2993         if (error == 0) {
2994                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2995                                     hammer_sync_record_callback, &cursor);
2996                 if (tmp_error < 0)
2997                         tmp_error = -error;
2998                 if (tmp_error)
2999                         error = tmp_error;
3000         }
3001         hammer_cache_node(&ip->cache[1], cursor.node);
3002
3003         /*
3004          * Re-seek for inode update, assuming our cache hasn't been ripped
3005          * out from under us.
3006          */
3007         if (error == 0) {
3008                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
3009                 if (tmp_node) {
3010                         hammer_cursor_downgrade(&cursor);
3011                         hammer_lock_sh(&tmp_node->lock);
3012                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
3013                                 hammer_cursor_seek(&cursor, tmp_node, 0);
3014                         hammer_unlock(&tmp_node->lock);
3015                         hammer_rel_node(tmp_node);
3016                 }
3017                 error = 0;
3018         }
3019
3020         /*
3021          * If we are deleting the inode the frontend had better not have
3022          * any active references on elements making up the inode.
3023          *
3024          * The call to hammer_ip_delete_clean() cleans up auxillary records
3025          * but not DB or DATA records.  Those must have already been deleted
3026          * by the normal truncation mechanic.
3027          */
3028         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
3029                 RB_EMPTY(&ip->rec_tree)  &&
3030             (ip->sync_flags & HAMMER_INODE_DELETING) &&
3031             (ip->flags & HAMMER_INODE_DELETED) == 0) {
3032                 int count1 = 0;
3033
3034                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
3035                 if (error == 0) {
3036                         ip->flags |= HAMMER_INODE_DELETED;
3037                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
3038                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3039                         KKASSERT(RB_EMPTY(&ip->rec_tree));
3040
3041                         /*
3042                          * Set delete_tid in both the frontend and backend
3043                          * copy of the inode record.  The DELETED flag handles
3044                          * this, do not set DDIRTY.
3045                          */
3046                         ip->ino_leaf.base.delete_tid = trans->tid;
3047                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
3048                         ip->ino_leaf.delete_ts = trans->time32;
3049                         ip->sync_ino_leaf.delete_ts = trans->time32;
3050
3051
3052                         /*
3053                          * Adjust the inode count in the volume header
3054                          */
3055                         hammer_sync_lock_sh(trans);
3056                         if (ip->flags & HAMMER_INODE_ONDISK) {
3057                                 hammer_modify_volume_field(trans,
3058                                                            trans->rootvol,
3059                                                            vol0_stat_inodes);
3060                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
3061                                 hammer_modify_volume_done(trans->rootvol);
3062                         }
3063                         hammer_sync_unlock(trans);
3064                 }
3065         }
3066
3067         if (error)
3068                 goto done;
3069         ip->sync_flags &= ~HAMMER_INODE_BUFS;
3070
3071 defer_buffer_flush:
3072         /*
3073          * Now update the inode's on-disk inode-data and/or on-disk record.
3074          * DELETED and ONDISK are managed only in ip->flags.
3075          *
3076          * In the case of a defered buffer flush we still update the on-disk
3077          * inode to satisfy visibility requirements if there happen to be
3078          * directory dependancies.
3079          */
3080         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
3081         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3082                 /*
3083                  * If deleted and on-disk, don't set any additional flags.
3084                  * the delete flag takes care of things.
3085                  *
3086                  * Clear flags which may have been set by the frontend.
3087                  */
3088                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3089                                     HAMMER_INODE_SDIRTY |
3090                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3091                                     HAMMER_INODE_DELETING);
3092                 break;
3093         case HAMMER_INODE_DELETED:
3094                 /*
3095                  * Take care of the case where a deleted inode was never
3096                  * flushed to the disk in the first place.
3097                  *
3098                  * Clear flags which may have been set by the frontend.
3099                  */
3100                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3101                                     HAMMER_INODE_SDIRTY |
3102                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3103                                     HAMMER_INODE_DELETING);
3104                 while (RB_ROOT(&ip->rec_tree)) {
3105                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
3106                         hammer_ref(&record->lock);
3107                         KKASSERT(hammer_oneref(&record->lock));
3108                         record->flags |= HAMMER_RECF_DELETED_BE;
3109                         ++record->ip->rec_generation;
3110                         hammer_rel_mem_record(record);
3111                 }
3112                 break;
3113         case HAMMER_INODE_ONDISK:
3114                 /*
3115                  * If already on-disk, do not set any additional flags.
3116                  */
3117                 break;
3118         default:
3119                 /*
3120                  * If not on-disk and not deleted, set DDIRTY to force
3121                  * an initial record to be written.
3122                  *
3123                  * Also set the create_tid in both the frontend and backend
3124                  * copy of the inode record.
3125                  */
3126                 ip->ino_leaf.base.create_tid = trans->tid;
3127                 ip->ino_leaf.create_ts = trans->time32;
3128                 ip->sync_ino_leaf.base.create_tid = trans->tid;
3129                 ip->sync_ino_leaf.create_ts = trans->time32;
3130                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
3131                 break;
3132         }
3133
3134         /*
3135          * If DDIRTY or SDIRTY is set, write out a new record.
3136          * If the inode is already on-disk the old record is marked as
3137          * deleted.
3138          *
3139          * If DELETED is set hammer_update_inode() will delete the existing
3140          * record without writing out a new one.
3141          *
3142          * If *ONLY* the ITIMES flag is set we can update the record in-place.
3143          */
3144         if (ip->flags & HAMMER_INODE_DELETED) {
3145                 error = hammer_update_inode(&cursor, ip);
3146         } else
3147         if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3148             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3149                 error = hammer_update_itimes(&cursor, ip);
3150         } else
3151         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3152                               HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3153                 error = hammer_update_inode(&cursor, ip);
3154         }
3155 done:
3156         if (error) {
3157                 hammer_critical_error(ip->hmp, ip, error,
3158                                       "while syncing inode");
3159         }
3160         hammer_done_cursor(&cursor);
3161         return(error);
3162 }
3163
3164 /*
3165  * This routine is called when the OS is no longer actively referencing
3166  * the inode (but might still be keeping it cached), or when releasing
3167  * the last reference to an inode.
3168  *
3169  * At this point if the inode's nlinks count is zero we want to destroy
3170  * it, which may mean destroying it on-media too.
3171  */
3172 void
3173 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3174 {
3175         struct vnode *vp;
3176
3177         /*
3178          * Set the DELETING flag when the link count drops to 0 and the
3179          * OS no longer has any opens on the inode.
3180          *
3181          * The backend will clear DELETING (a mod flag) and set DELETED
3182          * (a state flag) when it is actually able to perform the
3183          * operation.
3184          *
3185          * Don't reflag the deletion if the flusher is currently syncing
3186          * one that was already flagged.  A previously set DELETING flag
3187          * may bounce around flags and sync_flags until the operation is
3188          * completely done.
3189          *
3190          * Do not attempt to modify a snapshot inode (one set to read-only).
3191          */
3192         if (ip->ino_data.nlinks == 0 &&
3193             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3194                 ip->flags |= HAMMER_INODE_DELETING;
3195                 ip->flags |= HAMMER_INODE_TRUNCATED;
3196                 ip->trunc_off = 0;
3197                 vp = NULL;
3198                 if (getvp) {
3199                         if (hammer_get_vnode(ip, &vp) != 0)
3200                                 return;
3201                 }
3202
3203                 /*
3204                  * Final cleanup
3205                  */
3206                 if (ip->vp)
3207                         nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0);
3208                 if (getvp)
3209                         vput(vp);
3210         }
3211 }
3212
3213 /*
3214  * After potentially resolving a dependancy the inode is tested
3215  * to determine whether it needs to be reflushed.
3216  */
3217 void
3218 hammer_test_inode(hammer_inode_t ip)
3219 {
3220         if (ip->flags & HAMMER_INODE_REFLUSH) {
3221                 ip->flags &= ~HAMMER_INODE_REFLUSH;
3222                 hammer_ref(&ip->lock);
3223                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
3224                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
3225                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3226                 } else {
3227                         hammer_flush_inode(ip, 0);
3228                 }
3229                 hammer_rel_inode(ip, 0);
3230         }
3231 }
3232
3233 /*
3234  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
3235  * reassociated with a vp or just before it gets freed.
3236  *
3237  * Pipeline wakeups to threads blocked due to an excessive number of
3238  * detached inodes.  This typically occurs when atime updates accumulate
3239  * while scanning a directory tree.
3240  */
3241 static void
3242 hammer_inode_wakereclaims(hammer_inode_t ip)
3243 {
3244         struct hammer_reclaim *reclaim;
3245         hammer_mount_t hmp = ip->hmp;
3246
3247         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3248                 return;
3249
3250         --hammer_count_reclaims;
3251         --hmp->count_reclaims;
3252         ip->flags &= ~HAMMER_INODE_RECLAIM;
3253
3254         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3255                 KKASSERT(reclaim->count > 0);
3256                 if (--reclaim->count == 0) {
3257                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3258                         wakeup(reclaim);
3259                 }
3260         }
3261 }
3262
3263 /*
3264  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3265  * inodes build up before we start blocking.  This routine is called
3266  * if a new inode is created or an inode is loaded from media.
3267  *
3268  * When we block we don't care *which* inode has finished reclaiming,
3269  * as long as one does.
3270  *
3271  * The reclaim pipeline is primarily governed by the auto-flush which is
3272  * 1/4 hammer_limit_reclaims.  We don't want to block if the count is
3273  * less than 1/2 hammer_limit_reclaims.  From 1/2 to full count is
3274  * dynamically governed.
3275  */
3276 void
3277 hammer_inode_waitreclaims(hammer_transaction_t trans)
3278 {
3279         hammer_mount_t hmp = trans->hmp;
3280         struct hammer_reclaim reclaim;
3281         int lower_limit;
3282
3283         /*
3284          * Track inode load, delay if the number of reclaiming inodes is
3285          * between 2/4 and 4/4 hammer_limit_reclaims, depending.
3286          */
3287         if (curthread->td_proc) {
3288                 struct hammer_inostats *stats;
3289
3290                 stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3291                 ++stats->count;
3292
3293                 if (stats->count > hammer_limit_reclaims / 2)
3294                         stats->count = hammer_limit_reclaims / 2;
3295                 lower_limit = hammer_limit_reclaims - stats->count;
3296                 if (hammer_debug_general & 0x10000) {
3297                         kprintf("pid %5d limit %d\n",
3298                                 (int)curthread->td_proc->p_pid, lower_limit);
3299                 }
3300         } else {
3301                 lower_limit = hammer_limit_reclaims * 3 / 4;
3302         }
3303         if (hmp->count_reclaims >= lower_limit) {
3304                 reclaim.count = 1;
3305                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3306                 tsleep(&reclaim, 0, "hmrrcm", hz);
3307                 if (reclaim.count > 0)
3308                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3309         }
3310 }
3311
3312 /*
3313  * Keep track of reclaim statistics on a per-pid basis using a loose
3314  * 4-way set associative hash table.  Collisions inherit the count of
3315  * the previous entry.
3316  *
3317  * NOTE: We want to be careful here to limit the chain size.  If the chain
3318  *       size is too large a pid will spread its stats out over too many
3319  *       entries under certain types of heavy filesystem activity and
3320  *       wind up not delaying long enough.
3321  */
3322 static
3323 struct hammer_inostats *
3324 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3325 {
3326         struct hammer_inostats *stats;
3327         int delta;
3328         int chain;
3329         static volatile int iterator;   /* we don't care about MP races */
3330
3331         /*
3332          * Chain up to 4 times to find our entry.
3333          */
3334         for (chain = 0; chain < 4; ++chain) {
3335                 stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3336                 if (stats->pid == pid)
3337                         break;
3338         }
3339
3340         /*
3341          * Replace one of the four chaining entries with our new entry.
3342          */
3343         if (chain == 4) {
3344                 stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3345                                        HAMMER_INOSTATS_HMASK];
3346                 stats->pid = pid;
3347         }
3348
3349         /*
3350          * Decay the entry
3351          */
3352         if (stats->count && stats->ltick != ticks) {
3353                 delta = ticks - stats->ltick;
3354                 stats->ltick = ticks;
3355                 if (delta <= 0 || delta > hz * 60)
3356                         stats->count = 0;
3357                 else
3358                         stats->count = stats->count * hz / (hz + delta);
3359         }
3360         if (hammer_debug_general & 0x10000)
3361                 kprintf("pid %5d stats %d\n", (int)pid, stats->count);
3362         return (stats);
3363 }
3364
3365 #if 0
3366
3367 /*
3368  * XXX not used, doesn't work very well due to the large batching nature
3369  * of flushes.
3370  *
3371  * A larger then normal backlog of inodes is sitting in the flusher,
3372  * enforce a general slowdown to let it catch up.  This routine is only
3373  * called on completion of a non-flusher-related transaction which
3374  * performed B-Tree node I/O.
3375  *
3376  * It is possible for the flusher to stall in a continuous load.
3377  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3378  * If the flusher is unable to catch up the inode count can bloat until
3379  * we run out of kvm.
3380  *
3381  * This is a bit of a hack.
3382  */
3383 void
3384 hammer_inode_waithard(hammer_mount_t hmp)
3385 {
3386         /*
3387          * Hysteresis.
3388          */
3389         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3390                 if (hmp->count_reclaims < hammer_limit_reclaims / 2 &&
3391                     hmp->count_iqueued < hmp->count_inodes / 20) {
3392                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3393                         return;
3394                 }
3395         } else {
3396                 if (hmp->count_reclaims < hammer_limit_reclaims ||
3397                     hmp->count_iqueued < hmp->count_inodes / 10) {
3398                         return;
3399                 }
3400                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3401         }
3402
3403         /*
3404          * Block for one flush cycle.
3405          */
3406         hammer_flusher_wait_next(hmp);
3407 }
3408
3409 #endif