sys/vfs/hammer/hammer_inode.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.114 2008/09/24 00:53:51 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38 #include <vm/vm_extern.h>
  39 #include <sys/buf.h>
  40 #include <sys/buf2.h>
  41
  42 static int      hammer_unload_inode(struct hammer_inode *ip);
  43 static void     hammer_free_inode(hammer_inode_t ip);
  44 static void     hammer_flush_inode_core(hammer_inode_t ip,
  45                                         hammer_flush_group_t flg, int flags);
  46 static int      hammer_setup_child_callback(hammer_record_t rec, void *data);
  47 #if 0
  48 static int      hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
  49 #endif
  50 static int      hammer_setup_parent_inodes(hammer_inode_t ip,
  51                                         hammer_flush_group_t flg);
  52 static int      hammer_setup_parent_inodes_helper(hammer_record_t record,
  53                                         hammer_flush_group_t flg);
  54 static void     hammer_inode_wakereclaims(hammer_inode_t ip);
  55
  56 #ifdef DEBUG_TRUNCATE
  57 extern struct hammer_inode *HammerTruncIp;
  58 #endif
  59
  60 /*
  61  * RB-Tree support for inode structures
  62  */
  63 int
  64 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  65 {
  66         if (ip1->obj_localization < ip2->obj_localization)
  67                 return(-1);
  68         if (ip1->obj_localization > ip2->obj_localization)
  69                 return(1);
  70         if (ip1->obj_id < ip2->obj_id)
  71                 return(-1);
  72         if (ip1->obj_id > ip2->obj_id)
  73                 return(1);
  74         if (ip1->obj_asof < ip2->obj_asof)
  75                 return(-1);
  76         if (ip1->obj_asof > ip2->obj_asof)
  77                 return(1);
  78         return(0);
  79 }
  80
  81 /*
  82  * RB-Tree support for inode structures / special LOOKUP_INFO
  83  */
  84 static int
  85 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  86 {
  87         if (info->obj_localization < ip->obj_localization)
  88                 return(-1);
  89         if (info->obj_localization > ip->obj_localization)
  90                 return(1);
  91         if (info->obj_id < ip->obj_id)
  92                 return(-1);
  93         if (info->obj_id > ip->obj_id)
  94                 return(1);
  95         if (info->obj_asof < ip->obj_asof)
  96                 return(-1);
  97         if (info->obj_asof > ip->obj_asof)
  98                 return(1);
  99         return(0);
 100 }
 101
 102 /*
 103  * Used by hammer_scan_inode_snapshots() to locate all of an object's
 104  * snapshots.  Note that the asof field is not tested, which we can get
 105  * away with because it is the lowest-priority field.
 106  */
 107 static int
 108 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
 109 {
 110         hammer_inode_info_t info = data;
 111
 112         if (ip->obj_localization > info->obj_localization)
 113                 return(1);
 114         if (ip->obj_localization < info->obj_localization)
 115                 return(-1);
 116         if (ip->obj_id > info->obj_id)
 117                 return(1);
 118         if (ip->obj_id < info->obj_id)
 119                 return(-1);
 120         return(0);
 121 }
 122
 123 /*
 124  * Used by hammer_unload_pseudofs() to locate all inodes associated with
 125  * a particular PFS.
 126  */
 127 static int
 128 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
 129 {
 130         u_int32_t localization = *(u_int32_t *)data;
 131         if (ip->obj_localization > localization)
 132                 return(1);
 133         if (ip->obj_localization < localization)
 134                 return(-1);
 135         return(0);
 136 }
 137
 138 /*
 139  * RB-Tree support for pseudofs structures
 140  */
 141 static int
 142 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
 143 {
 144         if (p1->localization < p2->localization)
 145                 return(-1);
 146         if (p1->localization > p2->localization)
 147                 return(1);
 148         return(0);
 149 }
 150
 151
 152 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 153 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 154                 hammer_inode_info_cmp, hammer_inode_info_t);
 155 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 156              hammer_pfs_rb_compare, u_int32_t, localization);
 157
 158 /*
 159  * The kernel is not actively referencing this vnode but is still holding
 160  * it cached.
 161  *
 162  * This is called from the frontend.
 163  */
 164 int
 165 hammer_vop_inactive(struct vop_inactive_args *ap)
 166 {
 167         struct hammer_inode *ip = VTOI(ap->a_vp);
 168
 169         /*
 170          * Degenerate case
 171          */
 172         if (ip == NULL) {
 173                 vrecycle(ap->a_vp);
 174                 return(0);
 175         }
 176
 177         /*
 178          * If the inode no longer has visibility in the filesystem try to
 179          * recycle it immediately, even if the inode is dirty.  Recycling
 180          * it quickly allows the system to reclaim buffer cache and VM
 181          * resources which can matter a lot in a heavily loaded system.
 182          *
 183          * This can deadlock in vfsync() if we aren't careful.
 184          *
 185          * Do not queue the inode to the flusher if we still have visibility,
 186          * otherwise namespace calls such as chmod will unnecessarily generate
 187          * multiple inode updates.
 188          */
 189         hammer_inode_unloadable_check(ip, 0);
 190         if (ip->ino_data.nlinks == 0) {
 191                 if (ip->flags & HAMMER_INODE_MODMASK)
 192                         hammer_flush_inode(ip, 0);
 193                 vrecycle(ap->a_vp);
 194         }
 195         return(0);
 196 }
 197
 198 /*
 199  * Release the vnode association.  This is typically (but not always)
 200  * the last reference on the inode.
 201  *
 202  * Once the association is lost we are on our own with regards to
 203  * flushing the inode.
 204  */
 205 int
 206 hammer_vop_reclaim(struct vop_reclaim_args *ap)
 207 {
 208         struct hammer_inode *ip;
 209         hammer_mount_t hmp;
 210         struct vnode *vp;
 211
 212         vp = ap->a_vp;
 213
 214         if ((ip = vp->v_data) != NULL) {
 215                 hmp = ip->hmp;
 216                 vp->v_data = NULL;
 217                 ip->vp = NULL;
 218
 219                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
 220                         ++hammer_count_reclaiming;
 221                         ++hmp->inode_reclaims;
 222                         ip->flags |= HAMMER_INODE_RECLAIM;
 223                 }
 224                 hammer_rel_inode(ip, 1);
 225         }
 226         return(0);
 227 }
 228
 229 /*
 230  * Return a locked vnode for the specified inode.  The inode must be
 231  * referenced but NOT LOCKED on entry and will remain referenced on
 232  * return.
 233  *
 234  * Called from the frontend.
 235  */
 236 int
 237 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 238 {
 239         hammer_mount_t hmp;
 240         struct vnode *vp;
 241         int error = 0;
 242         u_int8_t obj_type;
 243
 244         hmp = ip->hmp;
 245
 246         for (;;) {
 247                 if ((vp = ip->vp) == NULL) {
 248                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
 249                         if (error)
 250                                 break;
 251                         hammer_lock_ex(&ip->lock);
 252                         if (ip->vp != NULL) {
 253                                 hammer_unlock(&ip->lock);
 254                                 vp->v_type = VBAD;
 255                                 vx_put(vp);
 256                                 continue;
 257                         }
 258                         hammer_ref(&ip->lock);
 259                         vp = *vpp;
 260                         ip->vp = vp;
 261
 262                         obj_type = ip->ino_data.obj_type;
 263                         vp->v_type = hammer_get_vnode_type(obj_type);
 264
 265                         hammer_inode_wakereclaims(ip);
 266
 267                         switch(ip->ino_data.obj_type) {
 268                         case HAMMER_OBJTYPE_CDEV:
 269                         case HAMMER_OBJTYPE_BDEV:
 270                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
 271                                 addaliasu(vp, ip->ino_data.rmajor,
 272                                           ip->ino_data.rminor);
 273                                 break;
 274                         case HAMMER_OBJTYPE_FIFO:
 275                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
 276                                 break;
 277                         default:
 278                                 break;
 279                         }
 280
 281                         /*
 282                          * Only mark as the root vnode if the ip is not
 283                          * historical, otherwise the VFS cache will get
 284                          * confused.  The other half of the special handling
 285                          * is in hammer_vop_nlookupdotdot().
 286                          *
 287                          * Pseudo-filesystem roots can be accessed via
 288                          * non-root filesystem paths and setting VROOT may
 289                          * confuse the namecache.  Set VPFSROOT instead.
 290                          */
 291                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 292                             ip->obj_asof == hmp->asof) {
 293                                 if (ip->obj_localization == 0)
 294                                         vp->v_flag |= VROOT;
 295                                 else
 296                                         vp->v_flag |= VPFSROOT;
 297                         }
 298
 299                         vp->v_data = (void *)ip;
 300                         /* vnode locked by getnewvnode() */
 301                         /* make related vnode dirty if inode dirty? */
 302                         hammer_unlock(&ip->lock);
 303                         if (vp->v_type == VREG)
 304                                 vinitvmio(vp, ip->ino_data.size);
 305                         break;
 306                 }
 307
 308                 /*
 309                  * loop if the vget fails (aka races), or if the vp
 310                  * no longer matches ip->vp.
 311                  */
 312                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 313                         if (vp == ip->vp)
 314                                 break;
 315                         vput(vp);
 316                 }
 317         }
 318         *vpp = vp;
 319         return(error);
 320 }
 321
 322 /*
 323  * Locate all copies of the inode for obj_id compatible with the specified
 324  * asof, reference, and issue the related call-back.  This routine is used
 325  * for direct-io invalidation and does not create any new inodes.
 326  */
 327 void
 328 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
 329                             int (*callback)(hammer_inode_t ip, void *data),
 330                             void *data)
 331 {
 332         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
 333                                    hammer_inode_info_cmp_all_history,
 334                                    callback, iinfo);
 335 }
 336
 337 /*
 338  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 339  * do not attach or detach the related vnode (use hammer_get_vnode() for
 340  * that).
 341  *
 342  * The flags argument is only applied for newly created inodes, and only
 343  * certain flags are inherited.
 344  *
 345  * Called from the frontend.
 346  */
 347 struct hammer_inode *
 348 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
 349                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 350                  int flags, int *errorp)
 351 {
 352         hammer_mount_t hmp = trans->hmp;
 353         struct hammer_inode_info iinfo;
 354         struct hammer_cursor cursor;
 355         struct hammer_inode *ip;
 356
 357
 358         /*
 359          * Determine if we already have an inode cached.  If we do then
 360          * we are golden.
 361          */
 362         iinfo.obj_id = obj_id;
 363         iinfo.obj_asof = asof;
 364         iinfo.obj_localization = localization;
 365 loop:
 366         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 367         if (ip) {
 368                 hammer_ref(&ip->lock);
 369                 *errorp = 0;
 370                 return(ip);
 371         }
 372
 373         /*
 374          * Allocate a new inode structure and deal with races later.
 375          */
 376         ip = kmalloc(sizeof(*ip), M_HAMMER_INO, M_WAITOK|M_ZERO);
 377         ++hammer_count_inodes;
 378         ++hmp->count_inodes;
 379         ip->obj_id = obj_id;
 380         ip->obj_asof = iinfo.obj_asof;
 381         ip->obj_localization = localization;
 382         ip->hmp = hmp;
 383         ip->flags = flags & HAMMER_INODE_RO;
 384         ip->cache[0].ip = ip;
 385         ip->cache[1].ip = ip;
 386         if (hmp->ronly)
 387                 ip->flags |= HAMMER_INODE_RO;
 388         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 389                 0x7FFFFFFFFFFFFFFFLL;
 390         RB_INIT(&ip->rec_tree);
 391         TAILQ_INIT(&ip->target_list);
 392         hammer_ref(&ip->lock);
 393
 394         /*
 395          * Locate the on-disk inode.  If this is a PFS root we always
 396          * access the current version of the root inode and (if it is not
 397          * a master) always access information under it with a snapshot
 398          * TID.
 399          */
 400 retry:
 401         hammer_init_cursor(trans, &cursor, (dip ? &dip->cache[0] : NULL), NULL);
 402         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
 403         cursor.key_beg.obj_id = ip->obj_id;
 404         cursor.key_beg.key = 0;
 405         cursor.key_beg.create_tid = 0;
 406         cursor.key_beg.delete_tid = 0;
 407         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 408         cursor.key_beg.obj_type = 0;
 409
 410         cursor.asof = iinfo.obj_asof;
 411         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 412                        HAMMER_CURSOR_ASOF;
 413
 414         *errorp = hammer_btree_lookup(&cursor);
 415         if (*errorp == EDEADLK) {
 416                 hammer_done_cursor(&cursor);
 417                 goto retry;
 418         }
 419
 420         /*
 421          * On success the B-Tree lookup will hold the appropriate
 422          * buffer cache buffers and provide a pointer to the requested
 423          * information.  Copy the information to the in-memory inode
 424          * and cache the B-Tree node to improve future operations.
 425          */
 426         if (*errorp == 0) {
 427                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 428                 ip->ino_data = cursor.data->inode;
 429
 430                 /*
 431                  * cache[0] tries to cache the location of the object inode.
 432                  * The assumption is that it is near the directory inode.
 433                  *
 434                  * cache[1] tries to cache the location of the object data.
 435                  * The assumption is that it is near the directory data.
 436                  */
 437                 hammer_cache_node(&ip->cache[0], cursor.node);
 438                 if (dip && dip->cache[1].node)
 439                         hammer_cache_node(&ip->cache[1], dip->cache[1].node);
 440
 441                 /*
 442                  * The file should not contain any data past the file size
 443                  * stored in the inode.  Setting save_trunc_off to the
 444                  * file size instead of max reduces B-Tree lookup overheads
 445                  * on append by allowing the flusher to avoid checking for
 446                  * record overwrites.
 447                  */
 448                 ip->save_trunc_off = ip->ino_data.size;
 449
 450                 /*
 451                  * Locate and assign the pseudofs management structure to
 452                  * the inode.
 453                  */
 454                 if (dip && dip->obj_localization == ip->obj_localization) {
 455                         ip->pfsm = dip->pfsm;
 456                         hammer_ref(&ip->pfsm->lock);
 457                 } else {
 458                         ip->pfsm = hammer_load_pseudofs(trans,
 459                                                         ip->obj_localization,
 460                                                         errorp);
 461                         *errorp = 0;    /* ignore ENOENT */
 462                 }
 463         }
 464
 465         /*
 466          * The inode is placed on the red-black tree and will be synced to
 467          * the media when flushed or by the filesystem sync.  If this races
 468          * another instantiation/lookup the insertion will fail.
 469          */
 470         if (*errorp == 0) {
 471                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 472                         hammer_free_inode(ip);
 473                         hammer_done_cursor(&cursor);
 474                         goto loop;
 475                 }
 476                 ip->flags |= HAMMER_INODE_ONDISK;
 477         } else {
 478                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 479                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 480                         --hmp->rsv_inodes;
 481                 }
 482
 483                 hammer_free_inode(ip);
 484                 ip = NULL;
 485         }
 486         hammer_done_cursor(&cursor);
 487         trans->flags |= HAMMER_TRANSF_NEWINODE;
 488         return (ip);
 489 }
 490
 491 /*
 492  * Create a new filesystem object, returning the inode in *ipp.  The
 493  * returned inode will be referenced.  The inode is created in-memory.
 494  *
 495  * If pfsm is non-NULL the caller wishes to create the root inode for
 496  * a master PFS.
 497  */
 498 int
 499 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 500                     struct ucred *cred, hammer_inode_t dip,
 501                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
 502 {
 503         hammer_mount_t hmp;
 504         hammer_inode_t ip;
 505         uid_t xuid;
 506         int error;
 507
 508         hmp = trans->hmp;
 509
 510         ip = kmalloc(sizeof(*ip), M_HAMMER_INO, M_WAITOK|M_ZERO);
 511         ++hammer_count_inodes;
 512         ++hmp->count_inodes;
 513
 514         if (pfsm) {
 515                 KKASSERT(pfsm->localization != 0);
 516                 ip->obj_id = HAMMER_OBJID_ROOT;
 517                 ip->obj_localization = pfsm->localization;
 518         } else {
 519                 KKASSERT(dip != NULL);
 520                 ip->obj_id = hammer_alloc_objid(hmp, dip);
 521                 ip->obj_localization = dip->obj_localization;
 522         }
 523
 524         KKASSERT(ip->obj_id != 0);
 525         ip->obj_asof = hmp->asof;
 526         ip->hmp = hmp;
 527         ip->flush_state = HAMMER_FST_IDLE;
 528         ip->flags = HAMMER_INODE_DDIRTY |
 529                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
 530         ip->cache[0].ip = ip;
 531         ip->cache[1].ip = ip;
 532
 533         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 534         /* ip->save_trunc_off = 0; (already zero) */
 535         RB_INIT(&ip->rec_tree);
 536         TAILQ_INIT(&ip->target_list);
 537
 538         ip->ino_data.atime = trans->time;
 539         ip->ino_data.mtime = trans->time;
 540         ip->ino_data.size = 0;
 541         ip->ino_data.nlinks = 0;
 542
 543         /*
 544          * A nohistory designator on the parent directory is inherited by
 545          * the child.  We will do this even for pseudo-fs creation... the
 546          * sysad can turn it off.
 547          */
 548         if (dip) {
 549                 ip->ino_data.uflags = dip->ino_data.uflags &
 550                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 551         }
 552
 553         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 554         ip->ino_leaf.base.localization = ip->obj_localization +
 555                                          HAMMER_LOCALIZE_INODE;
 556         ip->ino_leaf.base.obj_id = ip->obj_id;
 557         ip->ino_leaf.base.key = 0;
 558         ip->ino_leaf.base.create_tid = 0;
 559         ip->ino_leaf.base.delete_tid = 0;
 560         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 561         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 562
 563         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 564         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 565         ip->ino_data.mode = vap->va_mode;
 566         ip->ino_data.ctime = trans->time;
 567
 568         /*
 569          * Setup the ".." pointer.  This only needs to be done for directories
 570          * but we do it for all objects as a recovery aid.
 571          */
 572         if (dip)
 573                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
 574 #if 0
 575         /*
 576          * The parent_obj_localization field only applies to pseudo-fs roots.
 577          * XXX this is no longer applicable, PFSs are no longer directly
 578          * tied into the parent's directory structure.
 579          */
 580         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
 581             ip->obj_id == HAMMER_OBJID_ROOT) {
 582                 ip->ino_data.ext.obj.parent_obj_localization =
 583                                                 dip->obj_localization;
 584         }
 585 #endif
 586
 587         switch(ip->ino_leaf.base.obj_type) {
 588         case HAMMER_OBJTYPE_CDEV:
 589         case HAMMER_OBJTYPE_BDEV:
 590                 ip->ino_data.rmajor = vap->va_rmajor;
 591                 ip->ino_data.rminor = vap->va_rminor;
 592                 break;
 593         default:
 594                 break;
 595         }
 596
 597         /*
 598          * Calculate default uid/gid and overwrite with information from
 599          * the vap.
 600          */
 601         if (dip) {
 602                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 603                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
 604                                              xuid, cred, &vap->va_mode);
 605         } else {
 606                 xuid = 0;
 607         }
 608         ip->ino_data.mode = vap->va_mode;
 609
 610         if (vap->va_vaflags & VA_UID_UUID_VALID)
 611                 ip->ino_data.uid = vap->va_uid_uuid;
 612         else if (vap->va_uid != (uid_t)VNOVAL)
 613                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 614         else
 615                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 616
 617         if (vap->va_vaflags & VA_GID_UUID_VALID)
 618                 ip->ino_data.gid = vap->va_gid_uuid;
 619         else if (vap->va_gid != (gid_t)VNOVAL)
 620                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 621         else if (dip)
 622                 ip->ino_data.gid = dip->ino_data.gid;
 623
 624         hammer_ref(&ip->lock);
 625
 626         if (pfsm) {
 627                 ip->pfsm = pfsm;
 628                 hammer_ref(&pfsm->lock);
 629                 error = 0;
 630         } else if (dip->obj_localization == ip->obj_localization) {
 631                 ip->pfsm = dip->pfsm;
 632                 hammer_ref(&ip->pfsm->lock);
 633                 error = 0;
 634         } else {
 635                 ip->pfsm = hammer_load_pseudofs(trans,
 636                                                 ip->obj_localization,
 637                                                 &error);
 638                 error = 0;      /* ignore ENOENT */
 639         }
 640
 641         if (error) {
 642                 hammer_free_inode(ip);
 643                 ip = NULL;
 644         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 645                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 646                 /* not reached */
 647                 hammer_free_inode(ip);
 648         }
 649         *ipp = ip;
 650         return(error);
 651 }
 652
 653 /*
 654  * Final cleanup / freeing of an inode structure
 655  */
 656 static void
 657 hammer_free_inode(hammer_inode_t ip)
 658 {
 659         KKASSERT(ip->lock.refs == 1);
 660         hammer_uncache_node(&ip->cache[0]);
 661         hammer_uncache_node(&ip->cache[1]);
 662         hammer_inode_wakereclaims(ip);
 663         if (ip->objid_cache)
 664                 hammer_clear_objid(ip);
 665         --hammer_count_inodes;
 666         --ip->hmp->count_inodes;
 667         if (ip->pfsm) {
 668                 hammer_rel_pseudofs(ip->hmp, ip->pfsm);
 669                 ip->pfsm = NULL;
 670         }
 671         kfree(ip, M_HAMMER_INO);
 672         ip = NULL;
 673 }
 674
 675 /*
 676  * Retrieve pseudo-fs data.  NULL will never be returned.
 677  *
 678  * If an error occurs *errorp will be set and a default template is returned,
 679  * otherwise *errorp is set to 0.  Typically when an error occurs it will
 680  * be ENOENT.
 681  */
 682 hammer_pseudofs_inmem_t
 683 hammer_load_pseudofs(hammer_transaction_t trans,
 684                      u_int32_t localization, int *errorp)
 685 {
 686         hammer_mount_t hmp = trans->hmp;
 687         hammer_inode_t ip;
 688         hammer_pseudofs_inmem_t pfsm;
 689         struct hammer_cursor cursor;
 690         int bytes;
 691
 692 retry:
 693         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
 694         if (pfsm) {
 695                 hammer_ref(&pfsm->lock);
 696                 *errorp = 0;
 697                 return(pfsm);
 698         }
 699
 700         /*
 701          * PFS records are stored in the root inode (not the PFS root inode,
 702          * but the real root).  Avoid an infinite recursion if loading
 703          * the PFS for the real root.
 704          */
 705         if (localization) {
 706                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
 707                                       HAMMER_MAX_TID,
 708                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
 709         } else {
 710                 ip = NULL;
 711         }
 712
 713         pfsm = kmalloc(sizeof(*pfsm), M_HAMMER, M_WAITOK | M_ZERO);
 714         pfsm->localization = localization;
 715         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
 716         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
 717
 718         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
 719         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
 720                                       HAMMER_LOCALIZE_MISC;
 721         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 722         cursor.key_beg.create_tid = 0;
 723         cursor.key_beg.delete_tid = 0;
 724         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 725         cursor.key_beg.obj_type = 0;
 726         cursor.key_beg.key = localization;
 727         cursor.asof = HAMMER_MAX_TID;
 728         cursor.flags |= HAMMER_CURSOR_ASOF;
 729
 730         if (ip)
 731                 *errorp = hammer_ip_lookup(&cursor);
 732         else
 733                 *errorp = hammer_btree_lookup(&cursor);
 734         if (*errorp == 0) {
 735                 *errorp = hammer_ip_resolve_data(&cursor);
 736                 if (*errorp == 0) {
 737                         if (cursor.data->pfsd.mirror_flags &
 738                             HAMMER_PFSD_DELETED) {
 739                                 *errorp = ENOENT;
 740                         } else {
 741                                 bytes = cursor.leaf->data_len;
 742                                 if (bytes > sizeof(pfsm->pfsd))
 743                                         bytes = sizeof(pfsm->pfsd);
 744                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
 745                         }
 746                 }
 747         }
 748         hammer_done_cursor(&cursor);
 749
 750         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 751         hammer_ref(&pfsm->lock);
 752         if (ip)
 753                 hammer_rel_inode(ip, 0);
 754         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
 755                 kfree(pfsm, M_HAMMER);
 756                 goto retry;
 757         }
 758         return(pfsm);
 759 }
 760
 761 /*
 762  * Store pseudo-fs data.  The backend will automatically delete any prior
 763  * on-disk pseudo-fs data but we have to delete in-memory versions.
 764  */
 765 int
 766 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
 767 {
 768         struct hammer_cursor cursor;
 769         hammer_record_t record;
 770         hammer_inode_t ip;
 771         int error;
 772
 773         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 774                               HAMMER_DEF_LOCALIZATION, 0, &error);
 775 retry:
 776         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 777         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
 778         cursor.key_beg.localization = ip->obj_localization +
 779                                       HAMMER_LOCALIZE_MISC;
 780         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 781         cursor.key_beg.create_tid = 0;
 782         cursor.key_beg.delete_tid = 0;
 783         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 784         cursor.key_beg.obj_type = 0;
 785         cursor.key_beg.key = pfsm->localization;
 786         cursor.asof = HAMMER_MAX_TID;
 787         cursor.flags |= HAMMER_CURSOR_ASOF;
 788
 789         error = hammer_ip_lookup(&cursor);
 790         if (error == 0 && hammer_cursor_inmem(&cursor)) {
 791                 record = cursor.iprec;
 792                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 793                         KKASSERT(cursor.deadlk_rec == NULL);
 794                         hammer_ref(&record->lock);
 795                         cursor.deadlk_rec = record;
 796                         error = EDEADLK;
 797                 } else {
 798                         record->flags |= HAMMER_RECF_DELETED_FE;
 799                         error = 0;
 800                 }
 801         }
 802         if (error == 0 || error == ENOENT) {
 803                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
 804                 record->type = HAMMER_MEM_RECORD_GENERAL;
 805
 806                 record->leaf.base.localization = ip->obj_localization +
 807                                                  HAMMER_LOCALIZE_MISC;
 808                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
 809                 record->leaf.base.key = pfsm->localization;
 810                 record->leaf.data_len = sizeof(pfsm->pfsd);
 811                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
 812                 error = hammer_ip_add_record(trans, record);
 813         }
 814         hammer_done_cursor(&cursor);
 815         if (error == EDEADLK)
 816                 goto retry;
 817         hammer_rel_inode(ip, 0);
 818         return(error);
 819 }
 820
 821 /*
 822  * Create a root directory for a PFS if one does not alredy exist.
 823  *
 824  * The PFS root stands alone so we must also bump the nlinks count
 825  * to prevent it from being destroyed on release.
 826  */
 827 int
 828 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
 829                        hammer_pseudofs_inmem_t pfsm)
 830 {
 831         hammer_inode_t ip;
 832         struct vattr vap;
 833         int error;
 834
 835         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 836                               pfsm->localization, 0, &error);
 837         if (ip == NULL) {
 838                 vattr_null(&vap);
 839                 vap.va_mode = 0755;
 840                 vap.va_type = VDIR;
 841                 error = hammer_create_inode(trans, &vap, cred, NULL, pfsm, &ip);
 842                 if (error == 0) {
 843                         ++ip->ino_data.nlinks;
 844                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 845                 }
 846         }
 847         if (ip)
 848                 hammer_rel_inode(ip, 0);
 849         return(error);
 850 }
 851
 852 /*
 853  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
 854  * if we are unable to disassociate all the inodes.
 855  */
 856 static
 857 int
 858 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
 859 {
 860         int res;
 861
 862         hammer_ref(&ip->lock);
 863         if (ip->lock.refs == 2 && ip->vp)
 864                 vclean_unlocked(ip->vp);
 865         if (ip->lock.refs == 1 && ip->vp == NULL)
 866                 res = 0;
 867         else
 868                 res = -1;       /* stop, someone is using the inode */
 869         hammer_rel_inode(ip, 0);
 870         return(res);
 871 }
 872
 873 int
 874 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
 875 {
 876         int res;
 877         int try;
 878
 879         for (try = res = 0; try < 4; ++try) {
 880                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
 881                                            hammer_inode_pfs_cmp,
 882                                            hammer_unload_pseudofs_callback,
 883                                            &localization);
 884                 if (res == 0 && try > 1)
 885                         break;
 886                 hammer_flusher_sync(trans->hmp);
 887         }
 888         if (res != 0)
 889                 res = ENOTEMPTY;
 890         return(res);
 891 }
 892
 893
 894 /*
 895  * Release a reference on a PFS
 896  */
 897 void
 898 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
 899 {
 900         hammer_unref(&pfsm->lock);
 901         if (pfsm->lock.refs == 0) {
 902                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
 903                 kfree(pfsm, M_HAMMER);
 904         }
 905 }
 906
 907 /*
 908  * Called by hammer_sync_inode().
 909  */
 910 static int
 911 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 912 {
 913         hammer_transaction_t trans = cursor->trans;
 914         hammer_record_t record;
 915         int error;
 916         int redirty;
 917
 918 retry:
 919         error = 0;
 920
 921         /*
 922          * If the inode has a presence on-disk then locate it and mark
 923          * it deleted, setting DELONDISK.
 924          *
 925          * The record may or may not be physically deleted, depending on
 926          * the retention policy.
 927          */
 928         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 929             HAMMER_INODE_ONDISK) {
 930                 hammer_normalize_cursor(cursor);
 931                 cursor->key_beg.localization = ip->obj_localization +
 932                                                HAMMER_LOCALIZE_INODE;
 933                 cursor->key_beg.obj_id = ip->obj_id;
 934                 cursor->key_beg.key = 0;
 935                 cursor->key_beg.create_tid = 0;
 936                 cursor->key_beg.delete_tid = 0;
 937                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 938                 cursor->key_beg.obj_type = 0;
 939                 cursor->asof = ip->obj_asof;
 940                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 941                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 942                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 943
 944                 error = hammer_btree_lookup(cursor);
 945                 if (hammer_debug_inode)
 946                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 947
 948                 if (error == 0) {
 949                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
 950                         if (hammer_debug_inode)
 951                                 kprintf(" error %d\n", error);
 952                         if (error == 0) {
 953                                 ip->flags |= HAMMER_INODE_DELONDISK;
 954                         }
 955                         if (cursor->node)
 956                                 hammer_cache_node(&ip->cache[0], cursor->node);
 957                 }
 958                 if (error == EDEADLK) {
 959                         hammer_done_cursor(cursor);
 960                         error = hammer_init_cursor(trans, cursor,
 961                                                    &ip->cache[0], ip);
 962                         if (hammer_debug_inode)
 963                                 kprintf("IPDED %p %d\n", ip, error);
 964                         if (error == 0)
 965                                 goto retry;
 966                 }
 967         }
 968
 969         /*
 970          * Ok, write out the initial record or a new record (after deleting
 971          * the old one), unless the DELETED flag is set.  This routine will
 972          * clear DELONDISK if it writes out a record.
 973          *
 974          * Update our inode statistics if this is the first application of
 975          * the inode on-disk.
 976          */
 977         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 978                 /*
 979                  * Generate a record and write it to the media.  We clean-up
 980                  * the state before releasing so we do not have to set-up
 981                  * a flush_group.
 982                  */
 983                 record = hammer_alloc_mem_record(ip, 0);
 984                 record->type = HAMMER_MEM_RECORD_INODE;
 985                 record->flush_state = HAMMER_FST_FLUSH;
 986                 record->leaf = ip->sync_ino_leaf;
 987                 record->leaf.base.create_tid = trans->tid;
 988                 record->leaf.data_len = sizeof(ip->sync_ino_data);
 989                 record->leaf.create_ts = trans->time32;
 990                 record->data = (void *)&ip->sync_ino_data;
 991                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
 992
 993                 /*
 994                  * If this flag is set we cannot sync the new file size
 995                  * because we haven't finished related truncations.  The
 996                  * inode will be flushed in another flush group to finish
 997                  * the job.
 998                  */
 999                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1000                     ip->sync_ino_data.size != ip->ino_data.size) {
1001                         redirty = 1;
1002                         ip->sync_ino_data.size = ip->ino_data.size;
1003                 } else {
1004                         redirty = 0;
1005                 }
1006
1007                 for (;;) {
1008                         error = hammer_ip_sync_record_cursor(cursor, record);
1009                         if (hammer_debug_inode)
1010                                 kprintf("GENREC %p rec %08x %d\n",
1011                                         ip, record->flags, error);
1012                         if (error != EDEADLK)
1013                                 break;
1014                         hammer_done_cursor(cursor);
1015                         error = hammer_init_cursor(trans, cursor,
1016                                                    &ip->cache[0], ip);
1017                         if (hammer_debug_inode)
1018                                 kprintf("GENREC reinit %d\n", error);
1019                         if (error)
1020                                 break;
1021                 }
1022
1023                 /*
1024                  * The record isn't managed by the inode's record tree,
1025                  * destroy it whether we succeed or fail.
1026                  */
1027                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1028                 record->flags |= HAMMER_RECF_DELETED_FE | HAMMER_RECF_COMMITTED;
1029                 record->flush_state = HAMMER_FST_IDLE;
1030                 hammer_rel_mem_record(record);
1031
1032                 /*
1033                  * Finish up.
1034                  */
1035                 if (error == 0) {
1036                         if (hammer_debug_inode)
1037                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1038                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1039                                             HAMMER_INODE_ATIME |
1040                                             HAMMER_INODE_MTIME);
1041                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1042                         if (redirty)
1043                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1044
1045                         /*
1046                          * Root volume count of inodes
1047                          */
1048                         hammer_sync_lock_sh(trans);
1049                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1050                                 hammer_modify_volume_field(trans,
1051                                                            trans->rootvol,
1052                                                            vol0_stat_inodes);
1053                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1054                                 hammer_modify_volume_done(trans->rootvol);
1055                                 ip->flags |= HAMMER_INODE_ONDISK;
1056                                 if (hammer_debug_inode)
1057                                         kprintf("NOWONDISK %p\n", ip);
1058                         }
1059                         hammer_sync_unlock(trans);
1060                 }
1061         }
1062
1063         /*
1064          * If the inode has been destroyed, clean out any left-over flags
1065          * that may have been set by the frontend.
1066          */
1067         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1068                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1069                                     HAMMER_INODE_ATIME |
1070                                     HAMMER_INODE_MTIME);
1071         }
1072         return(error);
1073 }
1074
1075 /*
1076  * Update only the itimes fields.
1077  *
1078  * ATIME can be updated without generating any UNDO.  MTIME is updated
1079  * with UNDO so it is guaranteed to be synchronized properly in case of
1080  * a crash.
1081  *
1082  * Neither field is included in the B-Tree leaf element's CRC, which is how
1083  * we can get away with updating ATIME the way we do.
1084  */
1085 static int
1086 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1087 {
1088         hammer_transaction_t trans = cursor->trans;
1089         int error;
1090
1091 retry:
1092         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1093             HAMMER_INODE_ONDISK) {
1094                 return(0);
1095         }
1096
1097         hammer_normalize_cursor(cursor);
1098         cursor->key_beg.localization = ip->obj_localization +
1099                                        HAMMER_LOCALIZE_INODE;
1100         cursor->key_beg.obj_id = ip->obj_id;
1101         cursor->key_beg.key = 0;
1102         cursor->key_beg.create_tid = 0;
1103         cursor->key_beg.delete_tid = 0;
1104         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1105         cursor->key_beg.obj_type = 0;
1106         cursor->asof = ip->obj_asof;
1107         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1108         cursor->flags |= HAMMER_CURSOR_ASOF;
1109         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1110         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1111         cursor->flags |= HAMMER_CURSOR_BACKEND;
1112
1113         error = hammer_btree_lookup(cursor);
1114         if (error == 0) {
1115                 hammer_cache_node(&ip->cache[0], cursor->node);
1116                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1117                         /*
1118                          * Updating MTIME requires an UNDO.  Just cover
1119                          * both atime and mtime.
1120                          */
1121                         hammer_sync_lock_sh(trans);
1122                         hammer_modify_buffer(trans, cursor->data_buffer,
1123                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1124                                      HAMMER_ITIMES_BYTES);
1125                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1126                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1127                         hammer_modify_buffer_done(cursor->data_buffer);
1128                         hammer_sync_unlock(trans);
1129                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1130                         /*
1131                          * Updating atime only can be done in-place with
1132                          * no UNDO.
1133                          */
1134                         hammer_sync_lock_sh(trans);
1135                         hammer_modify_buffer(trans, cursor->data_buffer,
1136                                              NULL, 0);
1137                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1138                         hammer_modify_buffer_done(cursor->data_buffer);
1139                         hammer_sync_unlock(trans);
1140                 }
1141                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1142         }
1143         if (error == EDEADLK) {
1144                 hammer_done_cursor(cursor);
1145                 error = hammer_init_cursor(trans, cursor,
1146                                            &ip->cache[0], ip);
1147                 if (error == 0)
1148                         goto retry;
1149         }
1150         return(error);
1151 }
1152
1153 /*
1154  * Release a reference on an inode, flush as requested.
1155  *
1156  * On the last reference we queue the inode to the flusher for its final
1157  * disposition.
1158  */
1159 void
1160 hammer_rel_inode(struct hammer_inode *ip, int flush)
1161 {
1162         /*hammer_mount_t hmp = ip->hmp;*/
1163
1164         /*
1165          * Handle disposition when dropping the last ref.
1166          */
1167         for (;;) {
1168                 if (ip->lock.refs == 1) {
1169                         /*
1170                          * Determine whether on-disk action is needed for
1171                          * the inode's final disposition.
1172                          */
1173                         KKASSERT(ip->vp == NULL);
1174                         hammer_inode_unloadable_check(ip, 0);
1175                         if (ip->flags & HAMMER_INODE_MODMASK) {
1176                                 hammer_flush_inode(ip, 0);
1177                         } else if (ip->lock.refs == 1) {
1178                                 hammer_unload_inode(ip);
1179                                 break;
1180                         }
1181                 } else {
1182                         if (flush)
1183                                 hammer_flush_inode(ip, 0);
1184
1185                         /*
1186                          * The inode still has multiple refs, try to drop
1187                          * one ref.
1188                          */
1189                         KKASSERT(ip->lock.refs >= 1);
1190                         if (ip->lock.refs > 1) {
1191                                 hammer_unref(&ip->lock);
1192                                 break;
1193                         }
1194                 }
1195         }
1196 }
1197
1198 /*
1199  * Unload and destroy the specified inode.  Must be called with one remaining
1200  * reference.  The reference is disposed of.
1201  *
1202  * The inode must be completely clean.
1203  */
1204 static int
1205 hammer_unload_inode(struct hammer_inode *ip)
1206 {
1207         hammer_mount_t hmp = ip->hmp;
1208
1209         KASSERT(ip->lock.refs == 1,
1210                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1211         KKASSERT(ip->vp == NULL);
1212         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1213         KKASSERT(ip->cursor_ip_refs == 0);
1214         KKASSERT(ip->lock.lockcount == 0);
1215         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1216
1217         KKASSERT(RB_EMPTY(&ip->rec_tree));
1218         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1219
1220         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1221
1222         hammer_free_inode(ip);
1223         return(0);
1224 }
1225
1226 /*
1227  * Called during unmounting if a critical error occured.  The in-memory
1228  * inode and all related structures are destroyed.
1229  *
1230  * If a critical error did not occur the unmount code calls the standard
1231  * release and asserts that the inode is gone.
1232  */
1233 int
1234 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1235 {
1236         hammer_record_t rec;
1237
1238         /*
1239          * Get rid of the inodes in-memory records, regardless of their
1240          * state, and clear the mod-mask.
1241          */
1242         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1243                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1244                 rec->target_ip = NULL;
1245                 if (rec->flush_state == HAMMER_FST_SETUP)
1246                         rec->flush_state = HAMMER_FST_IDLE;
1247         }
1248         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1249                 if (rec->flush_state == HAMMER_FST_FLUSH)
1250                         --rec->flush_group->refs;
1251                 else
1252                         hammer_ref(&rec->lock);
1253                 KKASSERT(rec->lock.refs == 1);
1254                 rec->flush_state = HAMMER_FST_IDLE;
1255                 rec->flush_group = NULL;
1256                 rec->flags |= HAMMER_RECF_DELETED_FE;
1257                 rec->flags |= HAMMER_RECF_DELETED_BE;
1258                 hammer_rel_mem_record(rec);
1259         }
1260         ip->flags &= ~HAMMER_INODE_MODMASK;
1261         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1262         KKASSERT(ip->vp == NULL);
1263
1264         /*
1265          * Remove the inode from any flush group, force it idle.  FLUSH
1266          * and SETUP states have an inode ref.
1267          */
1268         switch(ip->flush_state) {
1269         case HAMMER_FST_FLUSH:
1270                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1271                 --ip->flush_group->refs;
1272                 ip->flush_group = NULL;
1273                 /* fall through */
1274         case HAMMER_FST_SETUP:
1275                 hammer_unref(&ip->lock);
1276                 ip->flush_state = HAMMER_FST_IDLE;
1277                 /* fall through */
1278         case HAMMER_FST_IDLE:
1279                 break;
1280         }
1281
1282         /*
1283          * There shouldn't be any associated vnode.  The unload needs at
1284          * least one ref, if we do have a vp steal its ip ref.
1285          */
1286         if (ip->vp) {
1287                 kprintf("hammer_destroy_inode_callback: Unexpected "
1288                         "vnode association ip %p vp %p\n", ip, ip->vp);
1289                 ip->vp->v_data = NULL;
1290                 ip->vp = NULL;
1291         } else {
1292                 hammer_ref(&ip->lock);
1293         }
1294         hammer_unload_inode(ip);
1295         return(0);
1296 }
1297
1298 /*
1299  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1300  * the read-only flag for cached inodes.
1301  *
1302  * This routine is called from a RB_SCAN().
1303  */
1304 int
1305 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1306 {
1307         hammer_mount_t hmp = ip->hmp;
1308
1309         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1310                 ip->flags |= HAMMER_INODE_RO;
1311         else
1312                 ip->flags &= ~HAMMER_INODE_RO;
1313         return(0);
1314 }
1315
1316 /*
1317  * A transaction has modified an inode, requiring updates as specified by
1318  * the passed flags.
1319  *
1320  * HAMMER_INODE_DDIRTY: Inode data has been updated
1321  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1322  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1323  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1324  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1325  */
1326 void
1327 hammer_modify_inode(hammer_inode_t ip, int flags)
1328 {
1329         /*
1330          * ronly of 0 or 2 does not trigger assertion.
1331          * 2 is a special error state
1332          */
1333         KKASSERT(ip->hmp->ronly != 1 ||
1334                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1335                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1336                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1337         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1338                 ip->flags |= HAMMER_INODE_RSV_INODES;
1339                 ++ip->hmp->rsv_inodes;
1340         }
1341
1342         ip->flags |= flags;
1343 }
1344
1345 /*
1346  * Request that an inode be flushed.  This whole mess cannot block and may
1347  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1348  * actively flush the inode until the flush can be done.
1349  *
1350  * The inode may already be flushing, or may be in a setup state.  We can
1351  * place the inode in a flushing state if it is currently idle and flag it
1352  * to reflush if it is currently flushing.
1353  *
1354  * Upon return if the inode could not be flushed due to a setup
1355  * dependancy, then it will be automatically flushed when the dependancy
1356  * is satisfied.
1357  */
1358 void
1359 hammer_flush_inode(hammer_inode_t ip, int flags)
1360 {
1361         hammer_mount_t hmp;
1362         hammer_flush_group_t flg;
1363         int good;
1364
1365         /*
1366          * next_flush_group is the first flush group we can place the inode
1367          * in.  It may be NULL.  If it becomes full we append a new flush
1368          * group and make that the next_flush_group.
1369          */
1370         hmp = ip->hmp;
1371         while ((flg = hmp->next_flush_group) != NULL) {
1372                 KKASSERT(flg->running == 0);
1373                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1374                         break;
1375                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1376                 hammer_flusher_async(ip->hmp, flg);
1377         }
1378         if (flg == NULL) {
1379                 flg = kmalloc(sizeof(*flg), M_HAMMER, M_WAITOK|M_ZERO);
1380                 hmp->next_flush_group = flg;
1381                 TAILQ_INIT(&flg->flush_list);
1382                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1383         }
1384
1385         /*
1386          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1387          * state we have to put it back into an IDLE state so we can
1388          * drop the extra ref.
1389          *
1390          * If we have a parent dependancy we must still fall through
1391          * so we can run it.
1392          */
1393         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1394                 if (ip->flush_state == HAMMER_FST_SETUP &&
1395                     TAILQ_EMPTY(&ip->target_list)) {
1396                         ip->flush_state = HAMMER_FST_IDLE;
1397                         hammer_rel_inode(ip, 0);
1398                 }
1399                 if (ip->flush_state == HAMMER_FST_IDLE)
1400                         return;
1401         }
1402
1403         /*
1404          * Our flush action will depend on the current state.
1405          */
1406         switch(ip->flush_state) {
1407         case HAMMER_FST_IDLE:
1408                 /*
1409                  * We have no dependancies and can flush immediately.  Some
1410                  * our children may not be flushable so we have to re-test
1411                  * with that additional knowledge.
1412                  */
1413                 hammer_flush_inode_core(ip, flg, flags);
1414                 break;
1415         case HAMMER_FST_SETUP:
1416                 /*
1417                  * Recurse upwards through dependancies via target_list
1418                  * and start their flusher actions going if possible.
1419                  *
1420                  * 'good' is our connectivity.  -1 means we have none and
1421                  * can't flush, 0 means there weren't any dependancies, and
1422                  * 1 means we have good connectivity.
1423                  */
1424                 good = hammer_setup_parent_inodes(ip, flg);
1425
1426                 if (good >= 0) {
1427                         /*
1428                          * We can continue if good >= 0.  Determine how
1429                          * many records under our inode can be flushed (and
1430                          * mark them).
1431                          */
1432                         hammer_flush_inode_core(ip, flg, flags);
1433                 } else {
1434                         /*
1435                          * Parent has no connectivity, tell it to flush
1436                          * us as soon as it does.
1437                          *
1438                          * The REFLUSH flag is also needed to trigger
1439                          * dependancy wakeups.
1440                          */
1441                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1442                                      HAMMER_INODE_REFLUSH;
1443                         if (flags & HAMMER_FLUSH_SIGNAL) {
1444                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1445                                 hammer_flusher_async(ip->hmp, flg);
1446                         }
1447                 }
1448                 break;
1449         case HAMMER_FST_FLUSH:
1450                 /*
1451                  * We are already flushing, flag the inode to reflush
1452                  * if needed after it completes its current flush.
1453                  *
1454                  * The REFLUSH flag is also needed to trigger
1455                  * dependancy wakeups.
1456                  */
1457                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1458                         ip->flags |= HAMMER_INODE_REFLUSH;
1459                 if (flags & HAMMER_FLUSH_SIGNAL) {
1460                         ip->flags |= HAMMER_INODE_RESIGNAL;
1461                         hammer_flusher_async(ip->hmp, flg);
1462                 }
1463                 break;
1464         }
1465 }
1466
1467 /*
1468  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1469  * ip which reference our ip.
1470  *
1471  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1472  *     so for now do not ref/deref the structures.  Note that if we use the
1473  *     ref/rel code later, the rel CAN block.
1474  */
1475 static int
1476 hammer_setup_parent_inodes(hammer_inode_t ip, hammer_flush_group_t flg)
1477 {
1478         hammer_record_t depend;
1479         int good;
1480         int r;
1481
1482         good = 0;
1483         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1484                 r = hammer_setup_parent_inodes_helper(depend, flg);
1485                 KKASSERT(depend->target_ip == ip);
1486                 if (r < 0 && good == 0)
1487                         good = -1;
1488                 if (r > 0)
1489                         good = 1;
1490         }
1491         return(good);
1492 }
1493
1494 /*
1495  * This helper function takes a record representing the dependancy between
1496  * the parent inode and child inode.
1497  *
1498  * record->ip           = parent inode
1499  * record->target_ip    = child inode
1500  *
1501  * We are asked to recurse upwards and convert the record from SETUP
1502  * to FLUSH if possible.
1503  *
1504  * Return 1 if the record gives us connectivity
1505  *
1506  * Return 0 if the record is not relevant
1507  *
1508  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1509  */
1510 static int
1511 hammer_setup_parent_inodes_helper(hammer_record_t record,
1512                                   hammer_flush_group_t flg)
1513 {
1514         hammer_mount_t hmp;
1515         hammer_inode_t pip;
1516         int good;
1517
1518         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1519         pip = record->ip;
1520         hmp = pip->hmp;
1521
1522         /*
1523          * If the record is already flushing, is it in our flush group?
1524          *
1525          * If it is in our flush group but it is a general record or a
1526          * delete-on-disk, it does not improve our connectivity (return 0),
1527          * and if the target inode is not trying to destroy itself we can't
1528          * allow the operation yet anyway (the second return -1).
1529          */
1530         if (record->flush_state == HAMMER_FST_FLUSH) {
1531                 /*
1532                  * If not in our flush group ask the parent to reflush
1533                  * us as soon as possible.
1534                  */
1535                 if (record->flush_group != flg) {
1536                         pip->flags |= HAMMER_INODE_REFLUSH;
1537                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1538                         return(-1);
1539                 }
1540
1541                 /*
1542                  * If in our flush group everything is already set up,
1543                  * just return whether the record will improve our
1544                  * visibility or not.
1545                  */
1546                 if (record->type == HAMMER_MEM_RECORD_ADD)
1547                         return(1);
1548                 return(0);
1549         }
1550
1551         /*
1552          * It must be a setup record.  Try to resolve the setup dependancies
1553          * by recursing upwards so we can place ip on the flush list.
1554          */
1555         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1556
1557         good = hammer_setup_parent_inodes(pip, flg);
1558
1559         /*
1560          * If good < 0 the parent has no connectivity and we cannot safely
1561          * flush the directory entry, which also means we can't flush our
1562          * ip.  Flag the parent and us for downward recursion once the
1563          * parent's connectivity is resolved.
1564          */
1565         if (good < 0) {
1566                 /* pip->flags |= HAMMER_INODE_CONN_DOWN; set by recursion */
1567                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1568                 return(good);
1569         }
1570
1571         /*
1572          * We are go, place the parent inode in a flushing state so we can
1573          * place its record in a flushing state.  Note that the parent
1574          * may already be flushing.  The record must be in the same flush
1575          * group as the parent.
1576          */
1577         if (pip->flush_state != HAMMER_FST_FLUSH)
1578                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1579         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1580         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1581
1582 #if 0
1583         if (record->type == HAMMER_MEM_RECORD_DEL &&
1584             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1585                 /*
1586                  * Regardless of flushing state we cannot sync this path if the
1587                  * record represents a delete-on-disk but the target inode
1588                  * is not ready to sync its own deletion.
1589                  *
1590                  * XXX need to count effective nlinks to determine whether
1591                  * the flush is ok, otherwise removing a hardlink will
1592                  * just leave the DEL record to rot.
1593                  */
1594                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1595                 return(-1);
1596         } else
1597 #endif
1598         if (pip->flush_group == flg) {
1599                 /*
1600                  * Because we have not calculated nlinks yet we can just
1601                  * set records to the flush state if the parent is in
1602                  * the same flush group as we are.
1603                  */
1604                 record->flush_state = HAMMER_FST_FLUSH;
1605                 record->flush_group = flg;
1606                 ++record->flush_group->refs;
1607                 hammer_ref(&record->lock);
1608
1609                 /*
1610                  * A general directory-add contributes to our visibility.
1611                  *
1612                  * Otherwise it is probably a directory-delete or
1613                  * delete-on-disk record and does not contribute to our
1614                  * visbility (but we can still flush it).
1615                  */
1616                 if (record->type == HAMMER_MEM_RECORD_ADD)
1617                         return(1);
1618                 return(0);
1619         } else {
1620                 /*
1621                  * If the parent is not in our flush group we cannot
1622                  * flush this record yet, there is no visibility.
1623                  * We tell the parent to reflush and mark ourselves
1624                  * so the parent knows it should flush us too.
1625                  */
1626                 pip->flags |= HAMMER_INODE_REFLUSH;
1627                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1628                 return(-1);
1629         }
1630 }
1631
1632 /*
1633  * This is the core routine placing an inode into the FST_FLUSH state.
1634  */
1635 static void
1636 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1637 {
1638         int go_count;
1639
1640         /*
1641          * Set flush state and prevent the flusher from cycling into
1642          * the next flush group.  Do not place the ip on the list yet.
1643          * Inodes not in the idle state get an extra reference.
1644          */
1645         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1646         if (ip->flush_state == HAMMER_FST_IDLE)
1647                 hammer_ref(&ip->lock);
1648         ip->flush_state = HAMMER_FST_FLUSH;
1649         ip->flush_group = flg;
1650         ++ip->hmp->flusher.group_lock;
1651         ++ip->hmp->count_iqueued;
1652         ++hammer_count_iqueued;
1653         ++flg->total_count;
1654
1655         /*
1656          * If the flush group reaches the autoflush limit we want to signal
1657          * the flusher.  This is particularly important for remove()s.
1658          */
1659         if (flg->total_count == hammer_autoflush)
1660                 flags |= HAMMER_FLUSH_SIGNAL;
1661
1662         /*
1663          * We need to be able to vfsync/truncate from the backend.
1664          */
1665         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1666         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1667                 ip->flags |= HAMMER_INODE_VHELD;
1668                 vref(ip->vp);
1669         }
1670
1671         /*
1672          * Figure out how many in-memory records we can actually flush
1673          * (not including inode meta-data, buffers, etc).
1674          */
1675         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1676         if (flags & HAMMER_FLUSH_RECURSION) {
1677                 /*
1678                  * If this is a upwards recursion we do not want to
1679                  * recurse down again!
1680                  */
1681                 go_count = 1;
1682 #if 0
1683         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1684                 /*
1685                  * No new records are added if we must complete a flush
1686                  * from a previous cycle, but we do have to move the records
1687                  * from the previous cycle to the current one.
1688                  */
1689 #if 0
1690                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1691                                    hammer_syncgrp_child_callback, NULL);
1692 #endif
1693                 go_count = 1;
1694 #endif
1695         } else {
1696                 /*
1697                  * Normal flush, scan records and bring them into the flush.
1698                  * Directory adds and deletes are usually skipped (they are
1699                  * grouped with the related inode rather then with the
1700                  * directory).
1701                  *
1702                  * go_count can be negative, which means the scan aborted
1703                  * due to the flush group being over-full and we should
1704                  * flush what we have.
1705                  */
1706                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1707                                    hammer_setup_child_callback, NULL);
1708         }
1709
1710         /*
1711          * This is a more involved test that includes go_count.  If we
1712          * can't flush, flag the inode and return.  If go_count is 0 we
1713          * were are unable to flush any records in our rec_tree and
1714          * must ignore the XDIRTY flag.
1715          */
1716         if (go_count == 0) {
1717                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1718                         --ip->hmp->count_iqueued;
1719                         --hammer_count_iqueued;
1720
1721                         --flg->total_count;
1722                         ip->flush_state = HAMMER_FST_SETUP;
1723                         ip->flush_group = NULL;
1724                         if (ip->flags & HAMMER_INODE_VHELD) {
1725                                 ip->flags &= ~HAMMER_INODE_VHELD;
1726                                 vrele(ip->vp);
1727                         }
1728
1729                         /*
1730                          * REFLUSH is needed to trigger dependancy wakeups
1731                          * when an inode is in SETUP.
1732                          */
1733                         ip->flags |= HAMMER_INODE_REFLUSH;
1734                         if (flags & HAMMER_FLUSH_SIGNAL) {
1735                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1736                                 hammer_flusher_async(ip->hmp, flg);
1737                         }
1738                         if (--ip->hmp->flusher.group_lock == 0)
1739                                 wakeup(&ip->hmp->flusher.group_lock);
1740                         return;
1741                 }
1742         }
1743
1744         /*
1745          * Snapshot the state of the inode for the backend flusher.
1746          *
1747          * We continue to retain save_trunc_off even when all truncations
1748          * have been resolved as an optimization to determine if we can
1749          * skip the B-Tree lookup for overwrite deletions.
1750          *
1751          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1752          * and stays in ip->flags.  Once set, it stays set until the
1753          * inode is destroyed.
1754          */
1755         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1756                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1757                 ip->sync_trunc_off = ip->trunc_off;
1758                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1759                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1760                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1761
1762                 /*
1763                  * The save_trunc_off used to cache whether the B-Tree
1764                  * holds any records past that point is not used until
1765                  * after the truncation has succeeded, so we can safely
1766                  * set it now.
1767                  */
1768                 if (ip->save_trunc_off > ip->sync_trunc_off)
1769                         ip->save_trunc_off = ip->sync_trunc_off;
1770         }
1771         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1772                            ~HAMMER_INODE_TRUNCATED);
1773         ip->sync_ino_leaf = ip->ino_leaf;
1774         ip->sync_ino_data = ip->ino_data;
1775         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1776 #ifdef DEBUG_TRUNCATE
1777         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1778                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1779 #endif
1780
1781         /*
1782          * The flusher list inherits our inode and reference.
1783          */
1784         KKASSERT(flg->running == 0);
1785         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
1786         if (--ip->hmp->flusher.group_lock == 0)
1787                 wakeup(&ip->hmp->flusher.group_lock);
1788
1789         if (flags & HAMMER_FLUSH_SIGNAL) {
1790                 hammer_flusher_async(ip->hmp, flg);
1791         }
1792 }
1793
1794 /*
1795  * Callback for scan of ip->rec_tree.  Try to include each record in our
1796  * flush.  ip->flush_group has been set but the inode has not yet been
1797  * moved into a flushing state.
1798  *
1799  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1800  * both inodes.
1801  *
1802  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1803  * the caller from shortcutting the flush.
1804  */
1805 static int
1806 hammer_setup_child_callback(hammer_record_t rec, void *data)
1807 {
1808         hammer_flush_group_t flg;
1809         hammer_inode_t target_ip;
1810         hammer_inode_t ip;
1811         int r;
1812
1813         /*
1814          * Deleted records are ignored.  Note that the flush detects deleted
1815          * front-end records at multiple points to deal with races.  This is
1816          * just the first line of defense.  The only time DELETED_FE cannot
1817          * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1818          *
1819          * Don't get confused between record deletion and, say, directory
1820          * entry deletion.  The deletion of a directory entry that is on
1821          * the media has nothing to do with the record deletion flags.
1822          */
1823         if (rec->flags & (HAMMER_RECF_DELETED_FE|HAMMER_RECF_DELETED_BE)) {
1824                 if (rec->flush_state == HAMMER_FST_FLUSH) {
1825                         KKASSERT(rec->flush_group == rec->ip->flush_group);
1826                         r = 1;
1827                 } else {
1828                         r = 0;
1829                 }
1830                 return(r);
1831         }
1832
1833         /*
1834          * If the record is in an idle state it has no dependancies and
1835          * can be flushed.
1836          */
1837         ip = rec->ip;
1838         flg = ip->flush_group;
1839         r = 0;
1840
1841         switch(rec->flush_state) {
1842         case HAMMER_FST_IDLE:
1843                 /*
1844                  * The record has no setup dependancy, we can flush it.
1845                  */
1846                 KKASSERT(rec->target_ip == NULL);
1847                 rec->flush_state = HAMMER_FST_FLUSH;
1848                 rec->flush_group = flg;
1849                 ++flg->refs;
1850                 hammer_ref(&rec->lock);
1851                 r = 1;
1852                 break;
1853         case HAMMER_FST_SETUP:
1854                 /*
1855                  * The record has a setup dependancy.  These are typically
1856                  * directory entry adds and deletes.  Such entries will be
1857                  * flushed when their inodes are flushed so we do not
1858                  * usually have to add them to the flush here.  However,
1859                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
1860                  * it is asking us to flush this record (and it).
1861                  */
1862                 target_ip = rec->target_ip;
1863                 KKASSERT(target_ip != NULL);
1864                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1865
1866                 /*
1867                  * If the target IP is already flushing in our group
1868                  * we could associate the record, but target_ip has
1869                  * already synced ino_data to sync_ino_data and we
1870                  * would also have to adjust nlinks.   Plus there are
1871                  * ordering issues for adds and deletes.
1872                  *
1873                  * Reflush downward if this is an ADD, and upward if
1874                  * this is a DEL.
1875                  */
1876                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1877                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
1878                                 ip->flags |= HAMMER_INODE_REFLUSH;
1879                         else
1880                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1881                         break;
1882                 }
1883
1884                 /*
1885                  * Target IP is not yet flushing.  This can get complex
1886                  * because we have to be careful about the recursion.
1887                  *
1888                  * Directories create an issue for us in that if a flush
1889                  * of a directory is requested the expectation is to flush
1890                  * any pending directory entries, but this will cause the
1891                  * related inodes to recursively flush as well.  We can't
1892                  * really defer the operation so just get as many as we
1893                  * can and
1894                  */
1895 #if 0
1896                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
1897                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
1898                         /*
1899                          * We aren't reclaiming and the target ip was not
1900                          * previously prevented from flushing due to this
1901                          * record dependancy.  Do not flush this record.
1902                          */
1903                         /*r = 0;*/
1904                 } else
1905 #endif
1906                 if (flg->total_count + flg->refs >
1907                            ip->hmp->undo_rec_limit) {
1908                         /*
1909                          * Our flush group is over-full and we risk blowing
1910                          * out the UNDO FIFO.  Stop the scan, flush what we
1911                          * have, then reflush the directory.
1912                          *
1913                          * The directory may be forced through multiple
1914                          * flush groups before it can be completely
1915                          * flushed.
1916                          */
1917                         ip->flags |= HAMMER_INODE_RESIGNAL |
1918                                      HAMMER_INODE_REFLUSH;
1919                         r = -1;
1920                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1921                         /*
1922                          * If the target IP is not flushing we can force
1923                          * it to flush, even if it is unable to write out
1924                          * any of its own records we have at least one in
1925                          * hand that we CAN deal with.
1926                          */
1927                         rec->flush_state = HAMMER_FST_FLUSH;
1928                         rec->flush_group = flg;
1929                         ++flg->refs;
1930                         hammer_ref(&rec->lock);
1931                         hammer_flush_inode_core(target_ip, flg,
1932                                                 HAMMER_FLUSH_RECURSION);
1933                         r = 1;
1934                 } else {
1935                         /*
1936                          * General or delete-on-disk record.
1937                          *
1938                          * XXX this needs help.  If a delete-on-disk we could
1939                          * disconnect the target.  If the target has its own
1940                          * dependancies they really need to be flushed.
1941                          *
1942                          * XXX
1943                          */
1944                         rec->flush_state = HAMMER_FST_FLUSH;
1945                         rec->flush_group = flg;
1946                         ++flg->refs;
1947                         hammer_ref(&rec->lock);
1948                         hammer_flush_inode_core(target_ip, flg,
1949                                                 HAMMER_FLUSH_RECURSION);
1950                         r = 1;
1951                 }
1952                 break;
1953         case HAMMER_FST_FLUSH:
1954                 /*
1955                  * The flush_group should already match.
1956                  */
1957                 KKASSERT(rec->flush_group == flg);
1958                 r = 1;
1959                 break;
1960         }
1961         return(r);
1962 }
1963
1964 #if 0
1965 /*
1966  * This version just moves records already in a flush state to the new
1967  * flush group and that is it.
1968  */
1969 static int
1970 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
1971 {
1972         hammer_inode_t ip = rec->ip;
1973
1974         switch(rec->flush_state) {
1975         case HAMMER_FST_FLUSH:
1976                 KKASSERT(rec->flush_group == ip->flush_group);
1977                 break;
1978         default:
1979                 break;
1980         }
1981         return(0);
1982 }
1983 #endif
1984
1985 /*
1986  * Wait for a previously queued flush to complete.
1987  *
1988  * If a critical error occured we don't try to wait.
1989  */
1990 void
1991 hammer_wait_inode(hammer_inode_t ip)
1992 {
1993         hammer_flush_group_t flg;
1994
1995         flg = NULL;
1996         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
1997                 while (ip->flush_state != HAMMER_FST_IDLE &&
1998                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
1999                         if (ip->flush_state == HAMMER_FST_SETUP)
2000                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2001                         if (ip->flush_state != HAMMER_FST_IDLE) {
2002                                 ip->flags |= HAMMER_INODE_FLUSHW;
2003                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2004                         }
2005                 }
2006         }
2007 }
2008
2009 /*
2010  * Called by the backend code when a flush has been completed.
2011  * The inode has already been removed from the flush list.
2012  *
2013  * A pipelined flush can occur, in which case we must re-enter the
2014  * inode on the list and re-copy its fields.
2015  */
2016 void
2017 hammer_flush_inode_done(hammer_inode_t ip, int error)
2018 {
2019         hammer_mount_t hmp;
2020         int dorel;
2021
2022         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2023
2024         hmp = ip->hmp;
2025
2026         /*
2027          * Merge left-over flags back into the frontend and fix the state.
2028          * Incomplete truncations are retained by the backend.
2029          */
2030         ip->error = error;
2031         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2032         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2033
2034         /*
2035          * The backend may have adjusted nlinks, so if the adjusted nlinks
2036          * does not match the fronttend set the frontend's RDIRTY flag again.
2037          */
2038         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2039                 ip->flags |= HAMMER_INODE_DDIRTY;
2040
2041         /*
2042          * Fix up the dirty buffer status.
2043          */
2044         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2045                 ip->flags |= HAMMER_INODE_BUFS;
2046         }
2047
2048         /*
2049          * Re-set the XDIRTY flag if some of the inode's in-memory records
2050          * could not be flushed.
2051          */
2052         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2053                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2054                  (!RB_EMPTY(&ip->rec_tree) &&
2055                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2056
2057         /*
2058          * Do not lose track of inodes which no longer have vnode
2059          * assocations, otherwise they may never get flushed again.
2060          *
2061          * The reflush flag can be set superfluously, causing extra pain
2062          * for no reason.  If the inode is no longer modified it no longer
2063          * needs to be flushed.
2064          */
2065         if (ip->flags & HAMMER_INODE_MODMASK) {
2066                 if (ip->vp == NULL)
2067                         ip->flags |= HAMMER_INODE_REFLUSH;
2068         } else {
2069                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2070         }
2071
2072         /*
2073          * Adjust the flush state.
2074          */
2075         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2076                 /*
2077                  * We were unable to flush out all our records, leave the
2078                  * inode in a flush state and in the current flush group.
2079                  * The flush group will be re-run.
2080                  *
2081                  * This occurs if the UNDO block gets too full or there is
2082                  * too much dirty meta-data and allows the flusher to
2083                  * finalize the UNDO block and then re-flush.
2084                  */
2085                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2086                 dorel = 0;
2087         } else {
2088                 /*
2089                  * Remove from the flush_group
2090                  */
2091                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2092                 ip->flush_group = NULL;
2093
2094                 /*
2095                  * Clean up the vnode ref and tracking counts.
2096                  */
2097                 if (ip->flags & HAMMER_INODE_VHELD) {
2098                         ip->flags &= ~HAMMER_INODE_VHELD;
2099                         vrele(ip->vp);
2100                 }
2101                 --hmp->count_iqueued;
2102                 --hammer_count_iqueued;
2103
2104                 /*
2105                  * And adjust the state.
2106                  */
2107                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2108                         ip->flush_state = HAMMER_FST_IDLE;
2109                         dorel = 1;
2110                 } else {
2111                         ip->flush_state = HAMMER_FST_SETUP;
2112                         dorel = 0;
2113                 }
2114
2115                 /*
2116                  * If the frontend is waiting for a flush to complete,
2117                  * wake it up.
2118                  */
2119                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2120                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2121                         wakeup(&ip->flags);
2122                 }
2123
2124                 /*
2125                  * If the frontend made more changes and requested another
2126                  * flush, then try to get it running.
2127                  *
2128                  * Reflushes are aborted when the inode is errored out.
2129                  */
2130                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2131                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2132                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2133                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2134                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2135                         } else {
2136                                 hammer_flush_inode(ip, 0);
2137                         }
2138                 }
2139         }
2140
2141         /*
2142          * If we have no parent dependancies we can clear CONN_DOWN
2143          */
2144         if (TAILQ_EMPTY(&ip->target_list))
2145                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2146
2147         /*
2148          * If the inode is now clean drop the space reservation.
2149          */
2150         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2151             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2152                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2153                 --hmp->rsv_inodes;
2154         }
2155
2156         if (dorel)
2157                 hammer_rel_inode(ip, 0);
2158 }
2159
2160 /*
2161  * Called from hammer_sync_inode() to synchronize in-memory records
2162  * to the media.
2163  */
2164 static int
2165 hammer_sync_record_callback(hammer_record_t record, void *data)
2166 {
2167         hammer_cursor_t cursor = data;
2168         hammer_transaction_t trans = cursor->trans;
2169         hammer_mount_t hmp = trans->hmp;
2170         int error;
2171
2172         /*
2173          * Skip records that do not belong to the current flush.
2174          */
2175         ++hammer_stats_record_iterations;
2176         if (record->flush_state != HAMMER_FST_FLUSH)
2177                 return(0);
2178
2179 #if 1
2180         if (record->flush_group != record->ip->flush_group) {
2181                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2182                 Debugger("blah2");
2183                 return(0);
2184         }
2185 #endif
2186         KKASSERT(record->flush_group == record->ip->flush_group);
2187
2188         /*
2189          * Interlock the record using the BE flag.  Once BE is set the
2190          * frontend cannot change the state of FE.
2191          *
2192          * NOTE: If FE is set prior to us setting BE we still sync the
2193          * record out, but the flush completion code converts it to
2194          * a delete-on-disk record instead of destroying it.
2195          */
2196         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2197         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2198
2199         /*
2200          * The backend may have already disposed of the record.
2201          */
2202         if (record->flags & HAMMER_RECF_DELETED_BE) {
2203                 error = 0;
2204                 goto done;
2205         }
2206
2207         /*
2208          * If the whole inode is being deleting all on-disk records will
2209          * be deleted very soon, we can't sync any new records to disk
2210          * because they will be deleted in the same transaction they were
2211          * created in (delete_tid == create_tid), which will assert.
2212          *
2213          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2214          * that we currently panic on.
2215          */
2216         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2217                 switch(record->type) {
2218                 case HAMMER_MEM_RECORD_DATA:
2219                         /*
2220                          * We don't have to do anything, if the record was
2221                          * committed the space will have been accounted for
2222                          * in the blockmap.
2223                          */
2224                         /* fall through */
2225                 case HAMMER_MEM_RECORD_GENERAL:
2226                         record->flags |= HAMMER_RECF_DELETED_FE;
2227                         record->flags |= HAMMER_RECF_DELETED_BE;
2228                         error = 0;
2229                         goto done;
2230                 case HAMMER_MEM_RECORD_ADD:
2231                         panic("hammer_sync_record_callback: illegal add "
2232                               "during inode deletion record %p", record);
2233                         break; /* NOT REACHED */
2234                 case HAMMER_MEM_RECORD_INODE:
2235                         panic("hammer_sync_record_callback: attempt to "
2236                               "sync inode record %p?", record);
2237                         break; /* NOT REACHED */
2238                 case HAMMER_MEM_RECORD_DEL:
2239                         /*
2240                          * Follow through and issue the on-disk deletion
2241                          */
2242                         break;
2243                 }
2244         }
2245
2246         /*
2247          * If DELETED_FE is set special handling is needed for directory
2248          * entries.  Dependant pieces related to the directory entry may
2249          * have already been synced to disk.  If this occurs we have to
2250          * sync the directory entry and then change the in-memory record
2251          * from an ADD to a DELETE to cover the fact that it's been
2252          * deleted by the frontend.
2253          *
2254          * A directory delete covering record (MEM_RECORD_DEL) can never
2255          * be deleted by the frontend.
2256          *
2257          * Any other record type (aka DATA) can be deleted by the frontend.
2258          * XXX At the moment the flusher must skip it because there may
2259          * be another data record in the flush group for the same block,
2260          * meaning that some frontend data changes can leak into the backend's
2261          * synchronization point.
2262          */
2263         if (record->flags & HAMMER_RECF_DELETED_FE) {
2264                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2265                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2266                 } else {
2267                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2268                         record->flags |= HAMMER_RECF_DELETED_BE;
2269                         error = 0;
2270                         goto done;
2271                 }
2272         }
2273
2274         /*
2275          * Assign the create_tid for new records.  Deletions already
2276          * have the record's entire key properly set up.
2277          */
2278         if (record->type != HAMMER_MEM_RECORD_DEL)
2279                 record->leaf.base.create_tid = trans->tid;
2280                 record->leaf.create_ts = trans->time32;
2281         for (;;) {
2282                 error = hammer_ip_sync_record_cursor(cursor, record);
2283                 if (error != EDEADLK)
2284                         break;
2285                 hammer_done_cursor(cursor);
2286                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2287                                            record->ip);
2288                 if (error)
2289                         break;
2290         }
2291         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2292
2293         if (error)
2294                 error = -error;
2295 done:
2296         hammer_flush_record_done(record, error);
2297
2298         /*
2299          * Do partial finalization if we have built up too many dirty
2300          * buffers.  Otherwise a buffer cache deadlock can occur when
2301          * doing things like creating tens of thousands of tiny files.
2302          *
2303          * We must release our cursor lock to avoid a 3-way deadlock
2304          * due to the exclusive sync lock the finalizer must get.
2305          */
2306         if (hammer_flusher_meta_limit(hmp)) {
2307                 hammer_unlock_cursor(cursor, 0);
2308                 hammer_flusher_finalize(trans, 0);
2309                 hammer_lock_cursor(cursor, 0);
2310         }
2311
2312         return(error);
2313 }
2314
2315 /*
2316  * Backend function called by the flusher to sync an inode to media.
2317  */
2318 int
2319 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2320 {
2321         struct hammer_cursor cursor;
2322         hammer_node_t tmp_node;
2323         hammer_record_t depend;
2324         hammer_record_t next;
2325         int error, tmp_error;
2326         u_int64_t nlinks;
2327
2328         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2329                 return(0);
2330
2331         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2332         if (error)
2333                 goto done;
2334
2335         /*
2336          * Any directory records referencing this inode which are not in
2337          * our current flush group must adjust our nlink count for the
2338          * purposes of synchronization to disk.
2339          *
2340          * Records which are in our flush group can be unlinked from our
2341          * inode now, potentially allowing the inode to be physically
2342          * deleted.
2343          *
2344          * This cannot block.
2345          */
2346         nlinks = ip->ino_data.nlinks;
2347         next = TAILQ_FIRST(&ip->target_list);
2348         while ((depend = next) != NULL) {
2349                 next = TAILQ_NEXT(depend, target_entry);
2350                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2351                     depend->flush_group == ip->flush_group) {
2352                         /*
2353                          * If this is an ADD that was deleted by the frontend
2354                          * the frontend nlinks count will have already been
2355                          * decremented, but the backend is going to sync its
2356                          * directory entry and must account for it.  The
2357                          * record will be converted to a delete-on-disk when
2358                          * it gets synced.
2359                          *
2360                          * If the ADD was not deleted by the frontend we
2361                          * can remove the dependancy from our target_list.
2362                          */
2363                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2364                                 ++nlinks;
2365                         } else {
2366                                 TAILQ_REMOVE(&ip->target_list, depend,
2367                                              target_entry);
2368                                 depend->target_ip = NULL;
2369                         }
2370                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2371                         /*
2372                          * Not part of our flush group
2373                          */
2374                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2375                         switch(depend->type) {
2376                         case HAMMER_MEM_RECORD_ADD:
2377                                 --nlinks;
2378                                 break;
2379                         case HAMMER_MEM_RECORD_DEL:
2380                                 ++nlinks;
2381                                 break;
2382                         default:
2383                                 break;
2384                         }
2385                 }
2386         }
2387
2388         /*
2389          * Set dirty if we had to modify the link count.
2390          */
2391         if (ip->sync_ino_data.nlinks != nlinks) {
2392                 KKASSERT((int64_t)nlinks >= 0);
2393                 ip->sync_ino_data.nlinks = nlinks;
2394                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2395         }
2396
2397         /*
2398          * If there is a trunction queued destroy any data past the (aligned)
2399          * truncation point.  Userland will have dealt with the buffer
2400          * containing the truncation point for us.
2401          *
2402          * We don't flush pending frontend data buffers until after we've
2403          * dealt with the truncation.
2404          */
2405         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2406                 /*
2407                  * Interlock trunc_off.  The VOP front-end may continue to
2408                  * make adjustments to it while we are blocked.
2409                  */
2410                 off_t trunc_off;
2411                 off_t aligned_trunc_off;
2412                 int blkmask;
2413
2414                 trunc_off = ip->sync_trunc_off;
2415                 blkmask = hammer_blocksize(trunc_off) - 1;
2416                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2417
2418                 /*
2419                  * Delete any whole blocks on-media.  The front-end has
2420                  * already cleaned out any partial block and made it
2421                  * pending.  The front-end may have updated trunc_off
2422                  * while we were blocked so we only use sync_trunc_off.
2423                  *
2424                  * This operation can blow out the buffer cache, EWOULDBLOCK
2425                  * means we were unable to complete the deletion.  The
2426                  * deletion will update sync_trunc_off in that case.
2427                  */
2428                 error = hammer_ip_delete_range(&cursor, ip,
2429                                                 aligned_trunc_off,
2430                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2431                 if (error == EWOULDBLOCK) {
2432                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2433                         error = 0;
2434                         goto defer_buffer_flush;
2435                 }
2436
2437                 if (error)
2438                         goto done;
2439
2440                 /*
2441                  * Clear the truncation flag on the backend after we have
2442                  * complete the deletions.  Backend data is now good again
2443                  * (including new records we are about to sync, below).
2444                  *
2445                  * Leave sync_trunc_off intact.  As we write additional
2446                  * records the backend will update sync_trunc_off.  This
2447                  * tells the backend whether it can skip the overwrite
2448                  * test.  This should work properly even when the backend
2449                  * writes full blocks where the truncation point straddles
2450                  * the block because the comparison is against the base
2451                  * offset of the record.
2452                  */
2453                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2454                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2455         } else {
2456                 error = 0;
2457         }
2458
2459         /*
2460          * Now sync related records.  These will typically be directory
2461          * entries, records tracking direct-writes, or delete-on-disk records.
2462          */
2463         if (error == 0) {
2464                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2465                                     hammer_sync_record_callback, &cursor);
2466                 if (tmp_error < 0)
2467                         tmp_error = -error;
2468                 if (tmp_error)
2469                         error = tmp_error;
2470         }
2471         hammer_cache_node(&ip->cache[1], cursor.node);
2472
2473         /*
2474          * Re-seek for inode update, assuming our cache hasn't been ripped
2475          * out from under us.
2476          */
2477         if (error == 0) {
2478                 tmp_node = hammer_ref_node_safe(ip->hmp, &ip->cache[0], &error);
2479                 if (tmp_node) {
2480                         hammer_cursor_downgrade(&cursor);
2481                         hammer_lock_sh(&tmp_node->lock);
2482                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2483                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2484                         hammer_unlock(&tmp_node->lock);
2485                         hammer_rel_node(tmp_node);
2486                 }
2487                 error = 0;
2488         }
2489
2490         /*
2491          * If we are deleting the inode the frontend had better not have
2492          * any active references on elements making up the inode.
2493          *
2494          * The call to hammer_ip_delete_clean() cleans up auxillary records
2495          * but not DB or DATA records.  Those must have already been deleted
2496          * by the normal truncation mechanic.
2497          */
2498         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2499                 RB_EMPTY(&ip->rec_tree)  &&
2500             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2501             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2502                 int count1 = 0;
2503
2504                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2505                 if (error == 0) {
2506                         ip->flags |= HAMMER_INODE_DELETED;
2507                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2508                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2509                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2510
2511                         /*
2512                          * Set delete_tid in both the frontend and backend
2513                          * copy of the inode record.  The DELETED flag handles
2514                          * this, do not set RDIRTY.
2515                          */
2516                         ip->ino_leaf.base.delete_tid = trans->tid;
2517                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2518                         ip->ino_leaf.delete_ts = trans->time32;
2519                         ip->sync_ino_leaf.delete_ts = trans->time32;
2520
2521
2522                         /*
2523                          * Adjust the inode count in the volume header
2524                          */
2525                         hammer_sync_lock_sh(trans);
2526                         if (ip->flags & HAMMER_INODE_ONDISK) {
2527                                 hammer_modify_volume_field(trans,
2528                                                            trans->rootvol,
2529                                                            vol0_stat_inodes);
2530                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2531                                 hammer_modify_volume_done(trans->rootvol);
2532                         }
2533                         hammer_sync_unlock(trans);
2534                 }
2535         }
2536
2537         if (error)
2538                 goto done;
2539         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2540
2541 defer_buffer_flush:
2542         /*
2543          * Now update the inode's on-disk inode-data and/or on-disk record.
2544          * DELETED and ONDISK are managed only in ip->flags.
2545          *
2546          * In the case of a defered buffer flush we still update the on-disk
2547          * inode to satisfy visibility requirements if there happen to be
2548          * directory dependancies.
2549          */
2550         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2551         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2552                 /*
2553                  * If deleted and on-disk, don't set any additional flags.
2554                  * the delete flag takes care of things.
2555                  *
2556                  * Clear flags which may have been set by the frontend.
2557                  */
2558                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2559                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2560                                     HAMMER_INODE_DELETING);
2561                 break;
2562         case HAMMER_INODE_DELETED:
2563                 /*
2564                  * Take care of the case where a deleted inode was never
2565                  * flushed to the disk in the first place.
2566                  *
2567                  * Clear flags which may have been set by the frontend.
2568                  */
2569                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2570                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2571                                     HAMMER_INODE_DELETING);
2572                 while (RB_ROOT(&ip->rec_tree)) {
2573                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2574                         hammer_ref(&record->lock);
2575                         KKASSERT(record->lock.refs == 1);
2576                         record->flags |= HAMMER_RECF_DELETED_FE;
2577                         record->flags |= HAMMER_RECF_DELETED_BE;
2578                         hammer_rel_mem_record(record);
2579                 }
2580                 break;
2581         case HAMMER_INODE_ONDISK:
2582                 /*
2583                  * If already on-disk, do not set any additional flags.
2584                  */
2585                 break;
2586         default:
2587                 /*
2588                  * If not on-disk and not deleted, set DDIRTY to force
2589                  * an initial record to be written.
2590                  *
2591                  * Also set the create_tid in both the frontend and backend
2592                  * copy of the inode record.
2593                  */
2594                 ip->ino_leaf.base.create_tid = trans->tid;
2595                 ip->ino_leaf.create_ts = trans->time32;
2596                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2597                 ip->sync_ino_leaf.create_ts = trans->time32;
2598                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2599                 break;
2600         }
2601
2602         /*
2603          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2604          * is already on-disk the old record is marked as deleted.
2605          *
2606          * If DELETED is set hammer_update_inode() will delete the existing
2607          * record without writing out a new one.
2608          *
2609          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2610          */
2611         if (ip->flags & HAMMER_INODE_DELETED) {
2612                 error = hammer_update_inode(&cursor, ip);
2613         } else
2614         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2615             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2616                 error = hammer_update_itimes(&cursor, ip);
2617         } else
2618         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2619                 error = hammer_update_inode(&cursor, ip);
2620         }
2621 done:
2622         if (error) {
2623                 hammer_critical_error(ip->hmp, ip, error,
2624                                       "while syncing inode");
2625         }
2626         hammer_done_cursor(&cursor);
2627         return(error);
2628 }
2629
2630 /*
2631  * This routine is called when the OS is no longer actively referencing
2632  * the inode (but might still be keeping it cached), or when releasing
2633  * the last reference to an inode.
2634  *
2635  * At this point if the inode's nlinks count is zero we want to destroy
2636  * it, which may mean destroying it on-media too.
2637  */
2638 void
2639 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2640 {
2641         struct vnode *vp;
2642
2643         /*
2644          * Set the DELETING flag when the link count drops to 0 and the
2645          * OS no longer has any opens on the inode.
2646          *
2647          * The backend will clear DELETING (a mod flag) and set DELETED
2648          * (a state flag) when it is actually able to perform the
2649          * operation.
2650          *
2651          * Don't reflag the deletion if the flusher is currently syncing
2652          * one that was already flagged.  A previously set DELETING flag
2653          * may bounce around flags and sync_flags until the operation is
2654          * completely done.
2655          */
2656         if (ip->ino_data.nlinks == 0 &&
2657             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2658                 ip->flags |= HAMMER_INODE_DELETING;
2659                 ip->flags |= HAMMER_INODE_TRUNCATED;
2660                 ip->trunc_off = 0;
2661                 vp = NULL;
2662                 if (getvp) {
2663                         if (hammer_get_vnode(ip, &vp) != 0)
2664                                 return;
2665                 }
2666
2667                 /*
2668                  * Final cleanup
2669                  */
2670                 if (ip->vp) {
2671                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2672                         vnode_pager_setsize(ip->vp, 0);
2673                 }
2674                 if (getvp) {
2675                         vput(vp);
2676                 }
2677         }
2678 }
2679
2680 /*
2681  * After potentially resolving a dependancy the inode is tested
2682  * to determine whether it needs to be reflushed.
2683  */
2684 void
2685 hammer_test_inode(hammer_inode_t ip)
2686 {
2687         if (ip->flags & HAMMER_INODE_REFLUSH) {
2688                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2689                 hammer_ref(&ip->lock);
2690                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2691                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2692                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2693                 } else {
2694                         hammer_flush_inode(ip, 0);
2695                 }
2696                 hammer_rel_inode(ip, 0);
2697         }
2698 }
2699
2700 /*
2701  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2702  * reassociated with a vp or just before it gets freed.
2703  *
2704  * Wakeup one thread blocked waiting on reclaims to complete.  Note that
2705  * the inode the thread is waiting on behalf of is a different inode then
2706  * the inode we are called with.  This is to create a pipeline.
2707  */
2708 static void
2709 hammer_inode_wakereclaims(hammer_inode_t ip)
2710 {
2711         struct hammer_reclaim *reclaim;
2712         hammer_mount_t hmp = ip->hmp;
2713
2714         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2715                 return;
2716
2717         --hammer_count_reclaiming;
2718         --hmp->inode_reclaims;
2719         ip->flags &= ~HAMMER_INODE_RECLAIM;
2720
2721         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
2722                 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2723                 reclaim->okydoky = 1;
2724                 wakeup(reclaim);
2725         }
2726 }
2727
2728 /*
2729  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
2730  * inodes build up before we start blocking.
2731  *
2732  * When we block we don't care *which* inode has finished reclaiming,
2733  * as lone as one does.  This is somewhat heuristical... we also put a
2734  * cap on how long we are willing to wait.
2735  */
2736 void
2737 hammer_inode_waitreclaims(hammer_mount_t hmp)
2738 {
2739         struct hammer_reclaim reclaim;
2740         int delay;
2741
2742         if (hmp->inode_reclaims > HAMMER_RECLAIM_WAIT) {
2743                 reclaim.okydoky = 0;
2744                 TAILQ_INSERT_TAIL(&hmp->reclaim_list,
2745                                   &reclaim, entry);
2746         } else {
2747                 reclaim.okydoky = 1;
2748         }
2749
2750         if (reclaim.okydoky == 0) {
2751                 delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2752                         (HAMMER_RECLAIM_WAIT * 5);
2753                 if (delay >= 0)
2754                         tsleep(&reclaim, 0, "hmrrcm", delay + 1);
2755                 if (reclaim.okydoky == 0)
2756                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2757         }
2758 }
2759