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 1.100 2008/07/12 02:47:39 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, int flags);
  45 static int      hammer_setup_child_callback(hammer_record_t rec, void *data);
  46 static int      hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
  47 static int      hammer_setup_parent_inodes(hammer_inode_t ip);
  48 static int      hammer_setup_parent_inodes_helper(hammer_record_t record);
  49 static void     hammer_inode_wakereclaims(hammer_inode_t ip);
  50
  51 #ifdef DEBUG_TRUNCATE
  52 extern struct hammer_inode *HammerTruncIp;
  53 #endif
  54
  55 /*
  56  * RB-Tree support for inode structures
  57  */
  58 int
  59 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  60 {
  61         if (ip1->obj_localization < ip2->obj_localization)
  62                 return(-1);
  63         if (ip1->obj_localization > ip2->obj_localization)
  64                 return(1);
  65         if (ip1->obj_id < ip2->obj_id)
  66                 return(-1);
  67         if (ip1->obj_id > ip2->obj_id)
  68                 return(1);
  69         if (ip1->obj_asof < ip2->obj_asof)
  70                 return(-1);
  71         if (ip1->obj_asof > ip2->obj_asof)
  72                 return(1);
  73         return(0);
  74 }
  75
  76 /*
  77  * RB-Tree support for inode structures / special LOOKUP_INFO
  78  */
  79 static int
  80 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  81 {
  82         if (info->obj_localization < ip->obj_localization)
  83                 return(-1);
  84         if (info->obj_localization > ip->obj_localization)
  85                 return(1);
  86         if (info->obj_id < ip->obj_id)
  87                 return(-1);
  88         if (info->obj_id > ip->obj_id)
  89                 return(1);
  90         if (info->obj_asof < ip->obj_asof)
  91                 return(-1);
  92         if (info->obj_asof > ip->obj_asof)
  93                 return(1);
  94         return(0);
  95 }
  96
  97 /*
  98  * Used by hammer_scan_inode_snapshots() to locate all of an object's
  99  * snapshots.  Note that the asof field is not tested, which we can get
 100  * away with because it is the lowest-priority field.
 101  */
 102 static int
 103 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
 104 {
 105         hammer_inode_info_t info = data;
 106
 107         if (ip->obj_localization > info->obj_localization)
 108                 return(1);
 109         if (ip->obj_localization < info->obj_localization)
 110                 return(-1);
 111         if (ip->obj_id > info->obj_id)
 112                 return(1);
 113         if (ip->obj_id < info->obj_id)
 114                 return(-1);
 115         return(0);
 116 }
 117
 118 /*
 119  * Used by hammer_unload_pseudofs() to locate all inodes associated with
 120  * a particular PFS.
 121  */
 122 static int
 123 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
 124 {
 125         u_int32_t localization = *(u_int32_t *)data;
 126         if (ip->obj_localization > localization)
 127                 return(1);
 128         if (ip->obj_localization < localization)
 129                 return(-1);
 130         return(0);
 131 }
 132
 133 /*
 134  * RB-Tree support for pseudofs structures
 135  */
 136 static int
 137 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
 138 {
 139         if (p1->localization < p2->localization)
 140                 return(-1);
 141         if (p1->localization > p2->localization)
 142                 return(1);
 143         return(0);
 144 }
 145
 146
 147 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 148 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 149                 hammer_inode_info_cmp, hammer_inode_info_t);
 150 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 151              hammer_pfs_rb_compare, u_int32_t, localization);
 152
 153 /*
 154  * The kernel is not actively referencing this vnode but is still holding
 155  * it cached.
 156  *
 157  * This is called from the frontend.
 158  */
 159 int
 160 hammer_vop_inactive(struct vop_inactive_args *ap)
 161 {
 162         struct hammer_inode *ip = VTOI(ap->a_vp);
 163
 164         /*
 165          * Degenerate case
 166          */
 167         if (ip == NULL) {
 168                 vrecycle(ap->a_vp);
 169                 return(0);
 170         }
 171
 172         /*
 173          * If the inode no longer has visibility in the filesystem try to
 174          * recycle it immediately, even if the inode is dirty.  Recycling
 175          * it quickly allows the system to reclaim buffer cache and VM
 176          * resources which can matter a lot in a heavily loaded system.
 177          *
 178          * This can deadlock in vfsync() if we aren't careful.
 179          *
 180          * Do not queue the inode to the flusher if we still have visibility,
 181          * otherwise namespace calls such as chmod will unnecessarily generate
 182          * multiple inode updates.
 183          */
 184         hammer_inode_unloadable_check(ip, 0);
 185         if (ip->ino_data.nlinks == 0) {
 186                 if (ip->flags & HAMMER_INODE_MODMASK)
 187                         hammer_flush_inode(ip, 0);
 188                 vrecycle(ap->a_vp);
 189         }
 190         return(0);
 191 }
 192
 193 /*
 194  * Release the vnode association.  This is typically (but not always)
 195  * the last reference on the inode.
 196  *
 197  * Once the association is lost we are on our own with regards to
 198  * flushing the inode.
 199  */
 200 int
 201 hammer_vop_reclaim(struct vop_reclaim_args *ap)
 202 {
 203         struct hammer_inode *ip;
 204         hammer_mount_t hmp;
 205         struct vnode *vp;
 206
 207         vp = ap->a_vp;
 208
 209         if ((ip = vp->v_data) != NULL) {
 210                 hmp = ip->hmp;
 211                 vp->v_data = NULL;
 212                 ip->vp = NULL;
 213
 214                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
 215                         ++hammer_count_reclaiming;
 216                         ++hmp->inode_reclaims;
 217                         ip->flags |= HAMMER_INODE_RECLAIM;
 218                         if (hmp->inode_reclaims > HAMMER_RECLAIM_FLUSH &&
 219                             (hmp->inode_reclaims & 255) == 0) {
 220                                 hammer_flusher_async(hmp);
 221                         }
 222                 }
 223                 hammer_rel_inode(ip, 1);
 224         }
 225         return(0);
 226 }
 227
 228 /*
 229  * Return a locked vnode for the specified inode.  The inode must be
 230  * referenced but NOT LOCKED on entry and will remain referenced on
 231  * return.
 232  *
 233  * Called from the frontend.
 234  */
 235 int
 236 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 237 {
 238         hammer_mount_t hmp;
 239         struct vnode *vp;
 240         int error = 0;
 241         u_int8_t obj_type;
 242
 243         hmp = ip->hmp;
 244
 245         for (;;) {
 246                 if ((vp = ip->vp) == NULL) {
 247                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
 248                         if (error)
 249                                 break;
 250                         hammer_lock_ex(&ip->lock);
 251                         if (ip->vp != NULL) {
 252                                 hammer_unlock(&ip->lock);
 253                                 vp->v_type = VBAD;
 254                                 vx_put(vp);
 255                                 continue;
 256                         }
 257                         hammer_ref(&ip->lock);
 258                         vp = *vpp;
 259                         ip->vp = vp;
 260
 261                         obj_type = ip->ino_data.obj_type;
 262                         vp->v_type = hammer_get_vnode_type(obj_type);
 263
 264                         hammer_inode_wakereclaims(ip);
 265
 266                         switch(ip->ino_data.obj_type) {
 267                         case HAMMER_OBJTYPE_CDEV:
 268                         case HAMMER_OBJTYPE_BDEV:
 269                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
 270                                 addaliasu(vp, ip->ino_data.rmajor,
 271                                           ip->ino_data.rminor);
 272                                 break;
 273                         case HAMMER_OBJTYPE_FIFO:
 274                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
 275                                 break;
 276                         default:
 277                                 break;
 278                         }
 279
 280                         /*
 281                          * Only mark as the root vnode if the ip is not
 282                          * historical, otherwise the VFS cache will get
 283                          * confused.  The other half of the special handling
 284                          * is in hammer_vop_nlookupdotdot().
 285                          *
 286                          * Pseudo-filesystem roots also do not count.
 287                          */
 288                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 289                             ip->obj_asof == hmp->asof &&
 290                             ip->obj_localization == 0) {
 291                                 vp->v_flag |= VROOT;
 292                         }
 293
 294                         vp->v_data = (void *)ip;
 295                         /* vnode locked by getnewvnode() */
 296                         /* make related vnode dirty if inode dirty? */
 297                         hammer_unlock(&ip->lock);
 298                         if (vp->v_type == VREG)
 299                                 vinitvmio(vp, ip->ino_data.size);
 300                         break;
 301                 }
 302
 303                 /*
 304                  * loop if the vget fails (aka races), or if the vp
 305                  * no longer matches ip->vp.
 306                  */
 307                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 308                         if (vp == ip->vp)
 309                                 break;
 310                         vput(vp);
 311                 }
 312         }
 313         *vpp = vp;
 314         return(error);
 315 }
 316
 317 /*
 318  * Locate all copies of the inode for obj_id compatible with the specified
 319  * asof, reference, and issue the related call-back.  This routine is used
 320  * for direct-io invalidation and does not create any new inodes.
 321  */
 322 void
 323 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
 324                             int (*callback)(hammer_inode_t ip, void *data),
 325                             void *data)
 326 {
 327         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
 328                                    hammer_inode_info_cmp_all_history,
 329                                    callback, iinfo);
 330 }
 331
 332 /*
 333  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 334  * do not attach or detach the related vnode (use hammer_get_vnode() for
 335  * that).
 336  *
 337  * The flags argument is only applied for newly created inodes, and only
 338  * certain flags are inherited.
 339  *
 340  * Called from the frontend.
 341  */
 342 struct hammer_inode *
 343 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
 344                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
 345                  int flags, int *errorp)
 346 {
 347         hammer_mount_t hmp = trans->hmp;
 348         struct hammer_inode_info iinfo;
 349         struct hammer_cursor cursor;
 350         struct hammer_inode *ip;
 351
 352
 353         /*
 354          * Determine if we already have an inode cached.  If we do then
 355          * we are golden.
 356          */
 357         iinfo.obj_id = obj_id;
 358         iinfo.obj_asof = asof;
 359         iinfo.obj_localization = localization;
 360 loop:
 361         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 362         if (ip) {
 363                 hammer_ref(&ip->lock);
 364                 *errorp = 0;
 365                 return(ip);
 366         }
 367
 368         /*
 369          * Allocate a new inode structure and deal with races later.
 370          */
 371         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 372         ++hammer_count_inodes;
 373         ++hmp->count_inodes;
 374         ip->obj_id = obj_id;
 375         ip->obj_asof = iinfo.obj_asof;
 376         ip->obj_localization = localization;
 377         ip->hmp = hmp;
 378         ip->flags = flags & HAMMER_INODE_RO;
 379         ip->cache[0].ip = ip;
 380         ip->cache[1].ip = ip;
 381         if (hmp->ronly)
 382                 ip->flags |= HAMMER_INODE_RO;
 383         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 384                 0x7FFFFFFFFFFFFFFFLL;
 385         RB_INIT(&ip->rec_tree);
 386         TAILQ_INIT(&ip->target_list);
 387         hammer_ref(&ip->lock);
 388
 389         /*
 390          * Locate the on-disk inode.  If this is a PFS root we always
 391          * access the current version of the root inode and (if it is not
 392          * a master) always access information under it with a snapshot
 393          * TID.
 394          */
 395 retry:
 396         hammer_init_cursor(trans, &cursor, (dip ? &dip->cache[0] : NULL), NULL);
 397         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
 398         cursor.key_beg.obj_id = ip->obj_id;
 399         cursor.key_beg.key = 0;
 400         cursor.key_beg.create_tid = 0;
 401         cursor.key_beg.delete_tid = 0;
 402         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 403         cursor.key_beg.obj_type = 0;
 404
 405         cursor.asof = iinfo.obj_asof;
 406         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 407                        HAMMER_CURSOR_ASOF;
 408
 409         *errorp = hammer_btree_lookup(&cursor);
 410         if (*errorp == EDEADLK) {
 411                 hammer_done_cursor(&cursor);
 412                 goto retry;
 413         }
 414
 415         /*
 416          * On success the B-Tree lookup will hold the appropriate
 417          * buffer cache buffers and provide a pointer to the requested
 418          * information.  Copy the information to the in-memory inode
 419          * and cache the B-Tree node to improve future operations.
 420          */
 421         if (*errorp == 0) {
 422                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 423                 ip->ino_data = cursor.data->inode;
 424
 425                 /*
 426                  * cache[0] tries to cache the location of the object inode.
 427                  * The assumption is that it is near the directory inode.
 428                  *
 429                  * cache[1] tries to cache the location of the object data.
 430                  * The assumption is that it is near the directory data.
 431                  */
 432                 hammer_cache_node(&ip->cache[0], cursor.node);
 433                 if (dip && dip->cache[1].node)
 434                         hammer_cache_node(&ip->cache[1], dip->cache[1].node);
 435
 436                 /*
 437                  * The file should not contain any data past the file size
 438                  * stored in the inode.  Setting save_trunc_off to the
 439                  * file size instead of max reduces B-Tree lookup overheads
 440                  * on append by allowing the flusher to avoid checking for
 441                  * record overwrites.
 442                  */
 443                 ip->save_trunc_off = ip->ino_data.size;
 444
 445                 /*
 446                  * Locate and assign the pseudofs management structure to
 447                  * the inode.
 448                  */
 449                 if (dip && dip->obj_localization == ip->obj_localization) {
 450                         ip->pfsm = dip->pfsm;
 451                         hammer_ref(&ip->pfsm->lock);
 452                 } else {
 453                         ip->pfsm = hammer_load_pseudofs(trans,
 454                                                         ip->obj_localization,
 455                                                         errorp);
 456                         *errorp = 0;    /* ignore ENOENT */
 457                 }
 458         }
 459
 460         /*
 461          * The inode is placed on the red-black tree and will be synced to
 462          * the media when flushed or by the filesystem sync.  If this races
 463          * another instantiation/lookup the insertion will fail.
 464          */
 465         if (*errorp == 0) {
 466                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 467                         hammer_free_inode(ip);
 468                         hammer_done_cursor(&cursor);
 469                         goto loop;
 470                 }
 471                 ip->flags |= HAMMER_INODE_ONDISK;
 472         } else {
 473                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 474                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 475                         --hmp->rsv_inodes;
 476                 }
 477
 478                 hammer_free_inode(ip);
 479                 ip = NULL;
 480         }
 481         hammer_done_cursor(&cursor);
 482         return (ip);
 483 }
 484
 485 /*
 486  * Create a new filesystem object, returning the inode in *ipp.  The
 487  * returned inode will be referenced.  The inode is created in-memory.
 488  *
 489  * If pfsm is non-NULL the caller wishes to create the root inode for
 490  * a master PFS.
 491  */
 492 int
 493 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 494                     struct ucred *cred, hammer_inode_t dip,
 495                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
 496 {
 497         hammer_mount_t hmp;
 498         hammer_inode_t ip;
 499         uid_t xuid;
 500         int error;
 501
 502         hmp = trans->hmp;
 503
 504         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 505         ++hammer_count_inodes;
 506         ++hmp->count_inodes;
 507
 508         if (pfsm) {
 509                 KKASSERT(pfsm->localization != 0);
 510                 ip->obj_id = HAMMER_OBJID_ROOT;
 511                 ip->obj_localization = pfsm->localization;
 512         } else {
 513                 KKASSERT(dip != NULL);
 514                 ip->obj_id = hammer_alloc_objid(hmp, dip);
 515                 ip->obj_localization = dip->obj_localization;
 516         }
 517
 518         KKASSERT(ip->obj_id != 0);
 519         ip->obj_asof = hmp->asof;
 520         ip->hmp = hmp;
 521         ip->flush_state = HAMMER_FST_IDLE;
 522         ip->flags = HAMMER_INODE_DDIRTY |
 523                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
 524         ip->cache[0].ip = ip;
 525         ip->cache[1].ip = ip;
 526
 527         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 528         /* ip->save_trunc_off = 0; (already zero) */
 529         RB_INIT(&ip->rec_tree);
 530         TAILQ_INIT(&ip->target_list);
 531
 532         ip->ino_data.atime = trans->time;
 533         ip->ino_data.mtime = trans->time;
 534         ip->ino_data.size = 0;
 535         ip->ino_data.nlinks = 0;
 536
 537         /*
 538          * A nohistory designator on the parent directory is inherited by
 539          * the child.  We will do this even for pseudo-fs creation... the
 540          * sysad can turn it off.
 541          */
 542         if (dip) {
 543                 ip->ino_data.uflags = dip->ino_data.uflags &
 544                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 545         }
 546
 547         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 548         ip->ino_leaf.base.localization = ip->obj_localization +
 549                                          HAMMER_LOCALIZE_INODE;
 550         ip->ino_leaf.base.obj_id = ip->obj_id;
 551         ip->ino_leaf.base.key = 0;
 552         ip->ino_leaf.base.create_tid = 0;
 553         ip->ino_leaf.base.delete_tid = 0;
 554         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 555         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 556
 557         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 558         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 559         ip->ino_data.mode = vap->va_mode;
 560         ip->ino_data.ctime = trans->time;
 561
 562         /*
 563          * Setup the ".." pointer.  This only needs to be done for directories
 564          * but we do it for all objects as a recovery aid.
 565          */
 566         if (dip)
 567                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
 568 #if 0
 569         /*
 570          * The parent_obj_localization field only applies to pseudo-fs roots.
 571          * XXX this is no longer applicable, PFSs are no longer directly
 572          * tied into the parent's directory structure.
 573          */
 574         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
 575             ip->obj_id == HAMMER_OBJID_ROOT) {
 576                 ip->ino_data.ext.obj.parent_obj_localization =
 577                                                 dip->obj_localization;
 578         }
 579 #endif
 580
 581         switch(ip->ino_leaf.base.obj_type) {
 582         case HAMMER_OBJTYPE_CDEV:
 583         case HAMMER_OBJTYPE_BDEV:
 584                 ip->ino_data.rmajor = vap->va_rmajor;
 585                 ip->ino_data.rminor = vap->va_rminor;
 586                 break;
 587         default:
 588                 break;
 589         }
 590
 591         /*
 592          * Calculate default uid/gid and overwrite with information from
 593          * the vap.
 594          */
 595         if (dip) {
 596                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 597                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
 598                                              xuid, cred, &vap->va_mode);
 599         } else {
 600                 xuid = 0;
 601         }
 602         ip->ino_data.mode = vap->va_mode;
 603
 604         if (vap->va_vaflags & VA_UID_UUID_VALID)
 605                 ip->ino_data.uid = vap->va_uid_uuid;
 606         else if (vap->va_uid != (uid_t)VNOVAL)
 607                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 608         else
 609                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 610
 611         if (vap->va_vaflags & VA_GID_UUID_VALID)
 612                 ip->ino_data.gid = vap->va_gid_uuid;
 613         else if (vap->va_gid != (gid_t)VNOVAL)
 614                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 615         else if (dip)
 616                 ip->ino_data.gid = dip->ino_data.gid;
 617
 618         hammer_ref(&ip->lock);
 619
 620         if (pfsm) {
 621                 ip->pfsm = pfsm;
 622                 hammer_ref(&pfsm->lock);
 623                 error = 0;
 624         } else if (dip->obj_localization == ip->obj_localization) {
 625                 ip->pfsm = dip->pfsm;
 626                 hammer_ref(&ip->pfsm->lock);
 627                 error = 0;
 628         } else {
 629                 ip->pfsm = hammer_load_pseudofs(trans,
 630                                                 ip->obj_localization,
 631                                                 &error);
 632                 error = 0;      /* ignore ENOENT */
 633         }
 634
 635         if (error) {
 636                 hammer_free_inode(ip);
 637                 ip = NULL;
 638         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 639                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 640                 /* not reached */
 641                 hammer_free_inode(ip);
 642         }
 643         *ipp = ip;
 644         return(error);
 645 }
 646
 647 /*
 648  * Final cleanup / freeing of an inode structure
 649  */
 650 static void
 651 hammer_free_inode(hammer_inode_t ip)
 652 {
 653         KKASSERT(ip->lock.refs == 1);
 654         hammer_uncache_node(&ip->cache[0]);
 655         hammer_uncache_node(&ip->cache[1]);
 656         hammer_inode_wakereclaims(ip);
 657         if (ip->objid_cache)
 658                 hammer_clear_objid(ip);
 659         --hammer_count_inodes;
 660         --ip->hmp->count_inodes;
 661         if (ip->pfsm) {
 662                 hammer_rel_pseudofs(ip->hmp, ip->pfsm);
 663                 ip->pfsm = NULL;
 664         }
 665         kfree(ip, M_HAMMER);
 666         ip = NULL;
 667 }
 668
 669 /*
 670  * Retrieve pseudo-fs data.  NULL will never be returned.
 671  *
 672  * If an error occurs *errorp will be set and a default template is returned,
 673  * otherwise *errorp is set to 0.  Typically when an error occurs it will
 674  * be ENOENT.
 675  */
 676 hammer_pseudofs_inmem_t
 677 hammer_load_pseudofs(hammer_transaction_t trans,
 678                      u_int32_t localization, int *errorp)
 679 {
 680         hammer_mount_t hmp = trans->hmp;
 681         hammer_inode_t ip;
 682         hammer_pseudofs_inmem_t pfsm;
 683         struct hammer_cursor cursor;
 684         int bytes;
 685
 686 retry:
 687         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
 688         if (pfsm) {
 689                 hammer_ref(&pfsm->lock);
 690                 *errorp = 0;
 691                 return(pfsm);
 692         }
 693
 694         /*
 695          * PFS records are stored in the root inode (not the PFS root inode,
 696          * but the real root).  Avoid an infinite recursion if loading
 697          * the PFS for the real root.
 698          */
 699         if (localization) {
 700                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
 701                                       HAMMER_MAX_TID,
 702                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
 703         } else {
 704                 ip = NULL;
 705         }
 706
 707         pfsm = kmalloc(sizeof(*pfsm), M_HAMMER, M_WAITOK | M_ZERO);
 708         pfsm->localization = localization;
 709         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
 710         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
 711
 712         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
 713         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
 714                                       HAMMER_LOCALIZE_MISC;
 715         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 716         cursor.key_beg.create_tid = 0;
 717         cursor.key_beg.delete_tid = 0;
 718         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 719         cursor.key_beg.obj_type = 0;
 720         cursor.key_beg.key = localization;
 721         cursor.asof = HAMMER_MAX_TID;
 722         cursor.flags |= HAMMER_CURSOR_ASOF;
 723
 724         if (ip)
 725                 *errorp = hammer_ip_lookup(&cursor);
 726         else
 727                 *errorp = hammer_btree_lookup(&cursor);
 728         if (*errorp == 0) {
 729                 *errorp = hammer_ip_resolve_data(&cursor);
 730                 if (*errorp == 0) {
 731                         if (cursor.data->pfsd.mirror_flags &
 732                             HAMMER_PFSD_DELETED) {
 733                                 *errorp = ENOENT;
 734                         } else {
 735                                 bytes = cursor.leaf->data_len;
 736                                 if (bytes > sizeof(pfsm->pfsd))
 737                                         bytes = sizeof(pfsm->pfsd);
 738                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
 739                         }
 740                 }
 741         }
 742         hammer_done_cursor(&cursor);
 743
 744         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 745         hammer_ref(&pfsm->lock);
 746         if (ip)
 747                 hammer_rel_inode(ip, 0);
 748         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
 749                 kfree(pfsm, M_HAMMER);
 750                 goto retry;
 751         }
 752         return(pfsm);
 753 }
 754
 755 /*
 756  * Store pseudo-fs data.  The backend will automatically delete any prior
 757  * on-disk pseudo-fs data but we have to delete in-memory versions.
 758  */
 759 int
 760 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
 761 {
 762         struct hammer_cursor cursor;
 763         hammer_record_t record;
 764         hammer_inode_t ip;
 765         int error;
 766
 767         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 768                               HAMMER_DEF_LOCALIZATION, 0, &error);
 769 retry:
 770         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
 771         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
 772         cursor.key_beg.localization = ip->obj_localization +
 773                                       HAMMER_LOCALIZE_MISC;
 774         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
 775         cursor.key_beg.create_tid = 0;
 776         cursor.key_beg.delete_tid = 0;
 777         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
 778         cursor.key_beg.obj_type = 0;
 779         cursor.key_beg.key = pfsm->localization;
 780         cursor.asof = HAMMER_MAX_TID;
 781         cursor.flags |= HAMMER_CURSOR_ASOF;
 782
 783         error = hammer_ip_lookup(&cursor);
 784         if (error == 0 && hammer_cursor_inmem(&cursor)) {
 785                 record = cursor.iprec;
 786                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 787                         KKASSERT(cursor.deadlk_rec == NULL);
 788                         hammer_ref(&record->lock);
 789                         cursor.deadlk_rec = record;
 790                         error = EDEADLK;
 791                 } else {
 792                         record->flags |= HAMMER_RECF_DELETED_FE;
 793                         error = 0;
 794                 }
 795         }
 796         if (error == 0 || error == ENOENT) {
 797                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
 798                 record->type = HAMMER_MEM_RECORD_GENERAL;
 799
 800                 record->leaf.base.localization = ip->obj_localization +
 801                                                  HAMMER_LOCALIZE_MISC;
 802                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
 803                 record->leaf.base.key = pfsm->localization;
 804                 record->leaf.data_len = sizeof(pfsm->pfsd);
 805                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
 806                 error = hammer_ip_add_record(trans, record);
 807         }
 808         hammer_done_cursor(&cursor);
 809         if (error == EDEADLK)
 810                 goto retry;
 811         hammer_rel_inode(ip, 0);
 812         return(error);
 813 }
 814
 815 /*
 816  * Create a root directory for a PFS if one does not alredy exist.
 817  *
 818  * The PFS root stands alone so we must also bump the nlinks count
 819  * to prevent it from being destroyed on release.
 820  */
 821 int
 822 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
 823                        hammer_pseudofs_inmem_t pfsm)
 824 {
 825         hammer_inode_t ip;
 826         struct vattr vap;
 827         int error;
 828
 829         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
 830                               pfsm->localization, 0, &error);
 831         if (ip == NULL) {
 832                 vattr_null(&vap);
 833                 vap.va_mode = 0755;
 834                 vap.va_type = VDIR;
 835                 error = hammer_create_inode(trans, &vap, cred, NULL, pfsm, &ip);
 836                 if (error == 0) {
 837                         ++ip->ino_data.nlinks;
 838                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 839                 }
 840         }
 841         if (ip)
 842                 hammer_rel_inode(ip, 0);
 843         return(error);
 844 }
 845
 846 /*
 847  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
 848  * if we are unable to disassociate all the inodes.
 849  */
 850 static
 851 int
 852 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
 853 {
 854         int res;
 855
 856         hammer_ref(&ip->lock);
 857         if (ip->lock.refs == 2 && ip->vp)
 858                 vclean_unlocked(ip->vp);
 859         if (ip->lock.refs == 1 && ip->vp == NULL)
 860                 res = 0;
 861         else
 862                 res = -1;       /* stop, someone is using the inode */
 863         hammer_rel_inode(ip, 0);
 864         return(res);
 865 }
 866
 867 int
 868 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
 869 {
 870         int res;
 871         int try;
 872
 873         for (try = res = 0; try < 4; ++try) {
 874                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
 875                                            hammer_inode_pfs_cmp,
 876                                            hammer_unload_pseudofs_callback,
 877                                            &localization);
 878                 if (res == 0 && try > 1)
 879                         break;
 880                 hammer_flusher_sync(trans->hmp);
 881         }
 882         if (res != 0)
 883                 res = ENOTEMPTY;
 884         return(res);
 885 }
 886
 887
 888 /*
 889  * Release a reference on a PFS
 890  */
 891 void
 892 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
 893 {
 894         hammer_unref(&pfsm->lock);
 895         if (pfsm->lock.refs == 0) {
 896                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
 897                 kfree(pfsm, M_HAMMER);
 898         }
 899 }
 900
 901 /*
 902  * Called by hammer_sync_inode().
 903  */
 904 static int
 905 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 906 {
 907         hammer_transaction_t trans = cursor->trans;
 908         hammer_record_t record;
 909         int error;
 910         int redirty;
 911
 912 retry:
 913         error = 0;
 914
 915         /*
 916          * If the inode has a presence on-disk then locate it and mark
 917          * it deleted, setting DELONDISK.
 918          *
 919          * The record may or may not be physically deleted, depending on
 920          * the retention policy.
 921          */
 922         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 923             HAMMER_INODE_ONDISK) {
 924                 hammer_normalize_cursor(cursor);
 925                 cursor->key_beg.localization = ip->obj_localization +
 926                                                HAMMER_LOCALIZE_INODE;
 927                 cursor->key_beg.obj_id = ip->obj_id;
 928                 cursor->key_beg.key = 0;
 929                 cursor->key_beg.create_tid = 0;
 930                 cursor->key_beg.delete_tid = 0;
 931                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 932                 cursor->key_beg.obj_type = 0;
 933                 cursor->asof = ip->obj_asof;
 934                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 935                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 936                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 937
 938                 error = hammer_btree_lookup(cursor);
 939                 if (hammer_debug_inode)
 940                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 941                 if (error) {
 942                         kprintf("error %d\n", error);
 943                         Debugger("hammer_update_inode");
 944                 }
 945
 946                 if (error == 0) {
 947                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
 948                         if (hammer_debug_inode)
 949                                 kprintf(" error %d\n", error);
 950                         if (error && error != EDEADLK) {
 951                                 kprintf("error %d\n", error);
 952                                 Debugger("hammer_update_inode2");
 953                         }
 954                         if (error == 0) {
 955                                 ip->flags |= HAMMER_INODE_DELONDISK;
 956                         }
 957                         if (cursor->node)
 958                                 hammer_cache_node(&ip->cache[0], cursor->node);
 959                 }
 960                 if (error == EDEADLK) {
 961                         hammer_done_cursor(cursor);
 962                         error = hammer_init_cursor(trans, cursor,
 963                                                    &ip->cache[0], ip);
 964                         if (hammer_debug_inode)
 965                                 kprintf("IPDED %p %d\n", ip, error);
 966                         if (error == 0)
 967                                 goto retry;
 968                 }
 969         }
 970
 971         /*
 972          * Ok, write out the initial record or a new record (after deleting
 973          * the old one), unless the DELETED flag is set.  This routine will
 974          * clear DELONDISK if it writes out a record.
 975          *
 976          * Update our inode statistics if this is the first application of
 977          * the inode on-disk.
 978          */
 979         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 980                 /*
 981                  * Generate a record and write it to the media
 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                 if (error) {
1023                         kprintf("error %d\n", error);
1024                         Debugger("hammer_update_inode3");
1025                 }
1026
1027                 /*
1028                  * The record isn't managed by the inode's record tree,
1029                  * destroy it whether we succeed or fail.
1030                  */
1031                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1032                 record->flags |= HAMMER_RECF_DELETED_FE;
1033                 record->flush_state = HAMMER_FST_IDLE;
1034                 hammer_rel_mem_record(record);
1035
1036                 /*
1037                  * Finish up.
1038                  */
1039                 if (error == 0) {
1040                         if (hammer_debug_inode)
1041                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1042                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1043                                             HAMMER_INODE_ATIME |
1044                                             HAMMER_INODE_MTIME);
1045                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1046                         if (redirty)
1047                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1048
1049                         /*
1050                          * Root volume count of inodes
1051                          */
1052                         hammer_sync_lock_sh(trans);
1053                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1054                                 hammer_modify_volume_field(trans,
1055                                                            trans->rootvol,
1056                                                            vol0_stat_inodes);
1057                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1058                                 hammer_modify_volume_done(trans->rootvol);
1059                                 ip->flags |= HAMMER_INODE_ONDISK;
1060                                 if (hammer_debug_inode)
1061                                         kprintf("NOWONDISK %p\n", ip);
1062                         }
1063                         hammer_sync_unlock(trans);
1064                 }
1065         }
1066
1067         /*
1068          * If the inode has been destroyed, clean out any left-over flags
1069          * that may have been set by the frontend.
1070          */
1071         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1072                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1073                                     HAMMER_INODE_ATIME |
1074                                     HAMMER_INODE_MTIME);
1075         }
1076         return(error);
1077 }
1078
1079 /*
1080  * Update only the itimes fields.
1081  *
1082  * ATIME can be updated without generating any UNDO.  MTIME is updated
1083  * with UNDO so it is guaranteed to be synchronized properly in case of
1084  * a crash.
1085  *
1086  * Neither field is included in the B-Tree leaf element's CRC, which is how
1087  * we can get away with updating ATIME the way we do.
1088  */
1089 static int
1090 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1091 {
1092         hammer_transaction_t trans = cursor->trans;
1093         int error;
1094
1095 retry:
1096         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1097             HAMMER_INODE_ONDISK) {
1098                 return(0);
1099         }
1100
1101         hammer_normalize_cursor(cursor);
1102         cursor->key_beg.localization = ip->obj_localization +
1103                                        HAMMER_LOCALIZE_INODE;
1104         cursor->key_beg.obj_id = ip->obj_id;
1105         cursor->key_beg.key = 0;
1106         cursor->key_beg.create_tid = 0;
1107         cursor->key_beg.delete_tid = 0;
1108         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1109         cursor->key_beg.obj_type = 0;
1110         cursor->asof = ip->obj_asof;
1111         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1112         cursor->flags |= HAMMER_CURSOR_ASOF;
1113         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1114         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1115         cursor->flags |= HAMMER_CURSOR_BACKEND;
1116
1117         error = hammer_btree_lookup(cursor);
1118         if (error) {
1119                 kprintf("error %d\n", error);
1120                 Debugger("hammer_update_itimes1");
1121         }
1122         if (error == 0) {
1123                 hammer_cache_node(&ip->cache[0], cursor->node);
1124                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1125                         /*
1126                          * Updating MTIME requires an UNDO.  Just cover
1127                          * both atime and mtime.
1128                          */
1129                         hammer_sync_lock_sh(trans);
1130                         hammer_modify_buffer(trans, cursor->data_buffer,
1131                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1132                                      HAMMER_ITIMES_BYTES);
1133                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1134                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1135                         hammer_modify_buffer_done(cursor->data_buffer);
1136                         hammer_sync_unlock(trans);
1137                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1138                         /*
1139                          * Updating atime only can be done in-place with
1140                          * no UNDO.
1141                          */
1142                         hammer_sync_lock_sh(trans);
1143                         hammer_modify_buffer(trans, cursor->data_buffer,
1144                                              NULL, 0);
1145                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1146                         hammer_modify_buffer_done(cursor->data_buffer);
1147                         hammer_sync_unlock(trans);
1148                 }
1149                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1150         }
1151         if (error == EDEADLK) {
1152                 hammer_done_cursor(cursor);
1153                 error = hammer_init_cursor(trans, cursor,
1154                                            &ip->cache[0], ip);
1155                 if (error == 0)
1156                         goto retry;
1157         }
1158         return(error);
1159 }
1160
1161 /*
1162  * Release a reference on an inode, flush as requested.
1163  *
1164  * On the last reference we queue the inode to the flusher for its final
1165  * disposition.
1166  */
1167 void
1168 hammer_rel_inode(struct hammer_inode *ip, int flush)
1169 {
1170         hammer_mount_t hmp = ip->hmp;
1171
1172         /*
1173          * Handle disposition when dropping the last ref.
1174          */
1175         for (;;) {
1176                 if (ip->lock.refs == 1) {
1177                         /*
1178                          * Determine whether on-disk action is needed for
1179                          * the inode's final disposition.
1180                          */
1181                         KKASSERT(ip->vp == NULL);
1182                         hammer_inode_unloadable_check(ip, 0);
1183                         if (ip->flags & HAMMER_INODE_MODMASK) {
1184                                 if (hmp->rsv_inodes > desiredvnodes) {
1185                                         hammer_flush_inode(ip,
1186                                                            HAMMER_FLUSH_SIGNAL);
1187                                 } else {
1188                                         hammer_flush_inode(ip, 0);
1189                                 }
1190                         } else if (ip->lock.refs == 1) {
1191                                 hammer_unload_inode(ip);
1192                                 break;
1193                         }
1194                 } else {
1195                         if (flush)
1196                                 hammer_flush_inode(ip, 0);
1197
1198                         /*
1199                          * The inode still has multiple refs, try to drop
1200                          * one ref.
1201                          */
1202                         KKASSERT(ip->lock.refs >= 1);
1203                         if (ip->lock.refs > 1) {
1204                                 hammer_unref(&ip->lock);
1205                                 break;
1206                         }
1207                 }
1208         }
1209 }
1210
1211 /*
1212  * Unload and destroy the specified inode.  Must be called with one remaining
1213  * reference.  The reference is disposed of.
1214  *
1215  * This can only be called in the context of the flusher.
1216  */
1217 static int
1218 hammer_unload_inode(struct hammer_inode *ip)
1219 {
1220         hammer_mount_t hmp = ip->hmp;
1221
1222         KASSERT(ip->lock.refs == 1,
1223                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1224         KKASSERT(ip->vp == NULL);
1225         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1226         KKASSERT(ip->cursor_ip_refs == 0);
1227         KKASSERT(ip->lock.lockcount == 0);
1228         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1229
1230         KKASSERT(RB_EMPTY(&ip->rec_tree));
1231         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1232
1233         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1234
1235         hammer_free_inode(ip);
1236         return(0);
1237 }
1238
1239 /*
1240  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1241  * the read-only flag for cached inodes.
1242  *
1243  * This routine is called from a RB_SCAN().
1244  */
1245 int
1246 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1247 {
1248         hammer_mount_t hmp = ip->hmp;
1249
1250         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1251                 ip->flags |= HAMMER_INODE_RO;
1252         else
1253                 ip->flags &= ~HAMMER_INODE_RO;
1254         return(0);
1255 }
1256
1257 /*
1258  * A transaction has modified an inode, requiring updates as specified by
1259  * the passed flags.
1260  *
1261  * HAMMER_INODE_DDIRTY: Inode data has been updated
1262  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1263  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1264  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1265  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1266  */
1267 void
1268 hammer_modify_inode(hammer_inode_t ip, int flags)
1269 {
1270         KKASSERT(ip->hmp->ronly == 0 ||
1271                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1272                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1273                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1274         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1275                 ip->flags |= HAMMER_INODE_RSV_INODES;
1276                 ++ip->hmp->rsv_inodes;
1277         }
1278
1279         ip->flags |= flags;
1280 }
1281
1282 /*
1283  * Request that an inode be flushed.  This whole mess cannot block and may
1284  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1285  * actively flush the inode until the flush can be done.
1286  *
1287  * The inode may already be flushing, or may be in a setup state.  We can
1288  * place the inode in a flushing state if it is currently idle and flag it
1289  * to reflush if it is currently flushing.
1290  *
1291  * If the HAMMER_FLUSH_SYNCHRONOUS flag is specified we will attempt to
1292  * flush the indoe synchronously using the caller's context.
1293  */
1294 void
1295 hammer_flush_inode(hammer_inode_t ip, int flags)
1296 {
1297         int good;
1298
1299         /*
1300          * Trivial 'nothing to flush' case.  If the inode is ina SETUP
1301          * state we have to put it back into an IDLE state so we can
1302          * drop the extra ref.
1303          */
1304         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1305                 if (ip->flush_state == HAMMER_FST_SETUP) {
1306                         ip->flush_state = HAMMER_FST_IDLE;
1307                         hammer_rel_inode(ip, 0);
1308                 }
1309                 return;
1310         }
1311
1312         /*
1313          * Our flush action will depend on the current state.
1314          */
1315         switch(ip->flush_state) {
1316         case HAMMER_FST_IDLE:
1317                 /*
1318                  * We have no dependancies and can flush immediately.  Some
1319                  * our children may not be flushable so we have to re-test
1320                  * with that additional knowledge.
1321                  */
1322                 hammer_flush_inode_core(ip, flags);
1323                 break;
1324         case HAMMER_FST_SETUP:
1325                 /*
1326                  * Recurse upwards through dependancies via target_list
1327                  * and start their flusher actions going if possible.
1328                  *
1329                  * 'good' is our connectivity.  -1 means we have none and
1330                  * can't flush, 0 means there weren't any dependancies, and
1331                  * 1 means we have good connectivity.
1332                  */
1333                 good = hammer_setup_parent_inodes(ip);
1334
1335                 /*
1336                  * We can continue if good >= 0.  Determine how many records
1337                  * under our inode can be flushed (and mark them).
1338                  */
1339                 if (good >= 0) {
1340                         hammer_flush_inode_core(ip, flags);
1341                 } else {
1342                         ip->flags |= HAMMER_INODE_REFLUSH;
1343                         if (flags & HAMMER_FLUSH_SIGNAL) {
1344                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1345                                 hammer_flusher_async(ip->hmp);
1346                         }
1347                 }
1348                 break;
1349         default:
1350                 /*
1351                  * We are already flushing, flag the inode to reflush
1352                  * if needed after it completes its current flush.
1353                  */
1354                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1355                         ip->flags |= HAMMER_INODE_REFLUSH;
1356                 if (flags & HAMMER_FLUSH_SIGNAL) {
1357                         ip->flags |= HAMMER_INODE_RESIGNAL;
1358                         hammer_flusher_async(ip->hmp);
1359                 }
1360                 break;
1361         }
1362 }
1363
1364 /*
1365  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1366  * ip which reference our ip.
1367  *
1368  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1369  *     so for now do not ref/deref the structures.  Note that if we use the
1370  *     ref/rel code later, the rel CAN block.
1371  */
1372 static int
1373 hammer_setup_parent_inodes(hammer_inode_t ip)
1374 {
1375         hammer_record_t depend;
1376 #if 0
1377         hammer_record_t next;
1378         hammer_inode_t  pip;
1379 #endif
1380         int good;
1381         int r;
1382
1383         good = 0;
1384         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1385                 r = hammer_setup_parent_inodes_helper(depend);
1386                 KKASSERT(depend->target_ip == ip);
1387                 if (r < 0 && good == 0)
1388                         good = -1;
1389                 if (r > 0)
1390                         good = 1;
1391         }
1392         return(good);
1393
1394 #if 0
1395 retry:
1396         good = 0;
1397         next = TAILQ_FIRST(&ip->target_list);
1398         if (next) {
1399                 hammer_ref(&next->lock);
1400                 hammer_ref(&next->ip->lock);
1401         }
1402         while ((depend = next) != NULL) {
1403                 if (depend->target_ip == NULL) {
1404                         pip = depend->ip;
1405                         hammer_rel_mem_record(depend);
1406                         hammer_rel_inode(pip, 0);
1407                         goto retry;
1408                 }
1409                 KKASSERT(depend->target_ip == ip);
1410                 next = TAILQ_NEXT(depend, target_entry);
1411                 if (next) {
1412                         hammer_ref(&next->lock);
1413                         hammer_ref(&next->ip->lock);
1414                 }
1415                 r = hammer_setup_parent_inodes_helper(depend);
1416                 if (r < 0 && good == 0)
1417                         good = -1;
1418                 if (r > 0)
1419                         good = 1;
1420                 pip = depend->ip;
1421                 hammer_rel_mem_record(depend);
1422                 hammer_rel_inode(pip, 0);
1423         }
1424         return(good);
1425 #endif
1426 }
1427
1428 /*
1429  * This helper function takes a record representing the dependancy between
1430  * the parent inode and child inode.
1431  *
1432  * record->ip           = parent inode
1433  * record->target_ip    = child inode
1434  *
1435  * We are asked to recurse upwards and convert the record from SETUP
1436  * to FLUSH if possible.
1437  *
1438  * Return 1 if the record gives us connectivity
1439  *
1440  * Return 0 if the record is not relevant
1441  *
1442  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1443  */
1444 static int
1445 hammer_setup_parent_inodes_helper(hammer_record_t record)
1446 {
1447         hammer_mount_t hmp;
1448         hammer_inode_t pip;
1449         int good;
1450
1451         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1452         pip = record->ip;
1453         hmp = pip->hmp;
1454
1455         /*
1456          * If the record is already flushing, is it in our flush group?
1457          *
1458          * If it is in our flush group but it is a general record or a
1459          * delete-on-disk, it does not improve our connectivity (return 0),
1460          * and if the target inode is not trying to destroy itself we can't
1461          * allow the operation yet anyway (the second return -1).
1462          */
1463         if (record->flush_state == HAMMER_FST_FLUSH) {
1464                 if (record->flush_group != hmp->flusher.next) {
1465                         pip->flags |= HAMMER_INODE_REFLUSH;
1466                         return(-1);
1467                 }
1468                 if (record->type == HAMMER_MEM_RECORD_ADD)
1469                         return(1);
1470                 /* GENERAL or DEL */
1471                 return(0);
1472         }
1473
1474         /*
1475          * It must be a setup record.  Try to resolve the setup dependancies
1476          * by recursing upwards so we can place ip on the flush list.
1477          */
1478         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1479
1480         good = hammer_setup_parent_inodes(pip);
1481
1482         /*
1483          * We can't flush ip because it has no connectivity (XXX also check
1484          * nlinks for pre-existing connectivity!).  Flag it so any resolution
1485          * recurses back down.
1486          */
1487         if (good < 0) {
1488                 pip->flags |= HAMMER_INODE_REFLUSH;
1489                 return(good);
1490         }
1491
1492         /*
1493          * We are go, place the parent inode in a flushing state so we can
1494          * place its record in a flushing state.  Note that the parent
1495          * may already be flushing.  The record must be in the same flush
1496          * group as the parent.
1497          */
1498         if (pip->flush_state != HAMMER_FST_FLUSH)
1499                 hammer_flush_inode_core(pip, HAMMER_FLUSH_RECURSION);
1500         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1501         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1502
1503 #if 0
1504         if (record->type == HAMMER_MEM_RECORD_DEL &&
1505             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1506                 /*
1507                  * Regardless of flushing state we cannot sync this path if the
1508                  * record represents a delete-on-disk but the target inode
1509                  * is not ready to sync its own deletion.
1510                  *
1511                  * XXX need to count effective nlinks to determine whether
1512                  * the flush is ok, otherwise removing a hardlink will
1513                  * just leave the DEL record to rot.
1514                  */
1515                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1516                 return(-1);
1517         } else
1518 #endif
1519         if (pip->flush_group == pip->hmp->flusher.next) {
1520                 /*
1521                  * This is the record we wanted to synchronize.  If the
1522                  * record went into a flush state while we blocked it
1523                  * had better be in the correct flush group.
1524                  */
1525                 if (record->flush_state != HAMMER_FST_FLUSH) {
1526                         record->flush_state = HAMMER_FST_FLUSH;
1527                         record->flush_group = pip->flush_group;
1528                         hammer_ref(&record->lock);
1529                 } else {
1530                         KKASSERT(record->flush_group == pip->flush_group);
1531                 }
1532                 if (record->type == HAMMER_MEM_RECORD_ADD)
1533                         return(1);
1534
1535                 /*
1536                  * A general or delete-on-disk record does not contribute
1537                  * to our visibility.  We can still flush it, however.
1538                  */
1539                 return(0);
1540         } else {
1541                 /*
1542                  * We couldn't resolve the dependancies, request that the
1543                  * inode be flushed when the dependancies can be resolved.
1544                  */
1545                 pip->flags |= HAMMER_INODE_REFLUSH;
1546                 return(-1);
1547         }
1548 }
1549
1550 /*
1551  * This is the core routine placing an inode into the FST_FLUSH state.
1552  */
1553 static void
1554 hammer_flush_inode_core(hammer_inode_t ip, int flags)
1555 {
1556         int go_count;
1557
1558         /*
1559          * Set flush state and prevent the flusher from cycling into
1560          * the next flush group.  Do not place the ip on the list yet.
1561          * Inodes not in the idle state get an extra reference.
1562          */
1563         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1564         if (ip->flush_state == HAMMER_FST_IDLE)
1565                 hammer_ref(&ip->lock);
1566         ip->flush_state = HAMMER_FST_FLUSH;
1567         ip->flush_group = ip->hmp->flusher.next;
1568         ++ip->hmp->flusher.group_lock;
1569         ++ip->hmp->count_iqueued;
1570         ++hammer_count_iqueued;
1571
1572         /*
1573          * We need to be able to vfsync/truncate from the backend.
1574          */
1575         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1576         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1577                 ip->flags |= HAMMER_INODE_VHELD;
1578                 vref(ip->vp);
1579         }
1580
1581         /*
1582          * Figure out how many in-memory records we can actually flush
1583          * (not including inode meta-data, buffers, etc).
1584          *
1585          * Do not add new records to the flush if this is a recursion or
1586          * if we must still complete a flush from the previous flush cycle.
1587          */
1588         if (flags & HAMMER_FLUSH_RECURSION) {
1589                 go_count = 1;
1590         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1591                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1592                                    hammer_syncgrp_child_callback, NULL);
1593                 go_count = 1;
1594         } else {
1595                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1596                                    hammer_setup_child_callback, NULL);
1597         }
1598
1599         /*
1600          * This is a more involved test that includes go_count.  If we
1601          * can't flush, flag the inode and return.  If go_count is 0 we
1602          * were are unable to flush any records in our rec_tree and
1603          * must ignore the XDIRTY flag.
1604          */
1605         if (go_count == 0) {
1606                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1607                         ip->flags |= HAMMER_INODE_REFLUSH;
1608
1609                         --ip->hmp->count_iqueued;
1610                         --hammer_count_iqueued;
1611
1612                         ip->flush_state = HAMMER_FST_SETUP;
1613                         if (ip->flags & HAMMER_INODE_VHELD) {
1614                                 ip->flags &= ~HAMMER_INODE_VHELD;
1615                                 vrele(ip->vp);
1616                         }
1617                         if (flags & HAMMER_FLUSH_SIGNAL) {
1618                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1619                                 hammer_flusher_async(ip->hmp);
1620                         }
1621                         if (--ip->hmp->flusher.group_lock == 0)
1622                                 wakeup(&ip->hmp->flusher.group_lock);
1623                         return;
1624                 }
1625         }
1626
1627         /*
1628          * Snapshot the state of the inode for the backend flusher.
1629          *
1630          * We continue to retain save_trunc_off even when all truncations
1631          * have been resolved as an optimization to determine if we can
1632          * skip the B-Tree lookup for overwrite deletions.
1633          *
1634          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1635          * and stays in ip->flags.  Once set, it stays set until the
1636          * inode is destroyed.
1637          *
1638          * NOTE: If a truncation from a previous flush cycle had to be
1639          * continued into this one, the TRUNCATED flag will still be
1640          * set in sync_flags as will WOULDBLOCK.  When this occurs
1641          * we CANNOT safely integrate a new truncation from the front-end
1642          * because there may be data records in-memory assigned a flush
1643          * state from the previous cycle that are supposed to be flushed
1644          * before the next frontend truncation.
1645          */
1646         if ((ip->flags & (HAMMER_INODE_TRUNCATED | HAMMER_INODE_WOULDBLOCK)) ==
1647             HAMMER_INODE_TRUNCATED) {
1648                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1649                 ip->sync_trunc_off = ip->trunc_off;
1650                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1651                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1652                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1653
1654                 /*
1655                  * The save_trunc_off used to cache whether the B-Tree
1656                  * holds any records past that point is not used until
1657                  * after the truncation has succeeded, so we can safely
1658                  * set it now.
1659                  */
1660                 if (ip->save_trunc_off > ip->sync_trunc_off)
1661                         ip->save_trunc_off = ip->sync_trunc_off;
1662         }
1663         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1664                            ~HAMMER_INODE_TRUNCATED);
1665         ip->sync_ino_leaf = ip->ino_leaf;
1666         ip->sync_ino_data = ip->ino_data;
1667         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1668 #ifdef DEBUG_TRUNCATE
1669         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1670                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1671 #endif
1672
1673         /*
1674          * The flusher list inherits our inode and reference.
1675          */
1676         TAILQ_INSERT_TAIL(&ip->hmp->flush_list, ip, flush_entry);
1677         if (--ip->hmp->flusher.group_lock == 0)
1678                 wakeup(&ip->hmp->flusher.group_lock);
1679
1680         if (flags & HAMMER_FLUSH_SIGNAL) {
1681                 hammer_flusher_async(ip->hmp);
1682         }
1683 }
1684
1685 /*
1686  * Callback for scan of ip->rec_tree.  Try to include each record in our
1687  * flush.  ip->flush_group has been set but the inode has not yet been
1688  * moved into a flushing state.
1689  *
1690  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1691  * both inodes.
1692  *
1693  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1694  * the caller from shortcutting the flush.
1695  */
1696 static int
1697 hammer_setup_child_callback(hammer_record_t rec, void *data)
1698 {
1699         hammer_inode_t target_ip;
1700         hammer_inode_t ip;
1701         int r;
1702
1703         /*
1704          * Deleted records are ignored.  Note that the flush detects deleted
1705          * front-end records at multiple points to deal with races.  This is
1706          * just the first line of defense.  The only time DELETED_FE cannot
1707          * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1708          *
1709          * Don't get confused between record deletion and, say, directory
1710          * entry deletion.  The deletion of a directory entry that is on
1711          * the media has nothing to do with the record deletion flags.
1712          *
1713          * The flush_group for a record already in a flush state must
1714          * be updated.  This case can only occur if the inode deleting
1715          * too many records had to be moved to the next flush group.
1716          */
1717         if (rec->flags & (HAMMER_RECF_DELETED_FE|HAMMER_RECF_DELETED_BE)) {
1718                 if (rec->flush_state == HAMMER_FST_FLUSH) {
1719                         KKASSERT(rec->ip->flags & HAMMER_INODE_WOULDBLOCK);
1720                         rec->flush_group = rec->ip->flush_group;
1721                         r = 1;
1722                 } else {
1723                         r = 0;
1724                 }
1725                 return(r);
1726         }
1727
1728         /*
1729          * If the record is in an idle state it has no dependancies and
1730          * can be flushed.
1731          */
1732         ip = rec->ip;
1733         r = 0;
1734
1735         switch(rec->flush_state) {
1736         case HAMMER_FST_IDLE:
1737                 /*
1738                  * Record has no setup dependancy, we can flush it.
1739                  */
1740                 KKASSERT(rec->target_ip == NULL);
1741                 rec->flush_state = HAMMER_FST_FLUSH;
1742                 rec->flush_group = ip->flush_group;
1743                 hammer_ref(&rec->lock);
1744                 r = 1;
1745                 break;
1746         case HAMMER_FST_SETUP:
1747                 /*
1748                  * Record has a setup dependancy.  Try to include the
1749                  * target ip in the flush.
1750                  *
1751                  * We have to be careful here, if we do not do the right
1752                  * thing we can lose track of dirty inodes and the system
1753                  * will lockup trying to allocate buffers.
1754                  */
1755                 target_ip = rec->target_ip;
1756                 KKASSERT(target_ip != NULL);
1757                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1758                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1759                         /*
1760                          * If the target IP is already flushing in our group
1761                          * we are golden, otherwise make sure the target
1762                          * reflushes.
1763                          */
1764                         if (target_ip->flush_group == ip->flush_group) {
1765                                 rec->flush_state = HAMMER_FST_FLUSH;
1766                                 rec->flush_group = ip->flush_group;
1767                                 hammer_ref(&rec->lock);
1768                                 r = 1;
1769                         } else {
1770                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1771                         }
1772                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1773                         /*
1774                          * If the target IP is not flushing we can force
1775                          * it to flush, even if it is unable to write out
1776                          * any of its own records we have at least one in
1777                          * hand that we CAN deal with.
1778                          */
1779                         rec->flush_state = HAMMER_FST_FLUSH;
1780                         rec->flush_group = ip->flush_group;
1781                         hammer_ref(&rec->lock);
1782                         hammer_flush_inode_core(target_ip,
1783                                                 HAMMER_FLUSH_RECURSION);
1784                         r = 1;
1785                 } else {
1786                         /*
1787                          * General or delete-on-disk record.
1788                          *
1789                          * XXX this needs help.  If a delete-on-disk we could
1790                          * disconnect the target.  If the target has its own
1791                          * dependancies they really need to be flushed.
1792                          *
1793                          * XXX
1794                          */
1795                         rec->flush_state = HAMMER_FST_FLUSH;
1796                         rec->flush_group = ip->flush_group;
1797                         hammer_ref(&rec->lock);
1798                         hammer_flush_inode_core(target_ip,
1799                                                 HAMMER_FLUSH_RECURSION);
1800                         r = 1;
1801                 }
1802                 break;
1803         case HAMMER_FST_FLUSH:
1804                 /*
1805                  * If the WOULDBLOCK flag is set records may have been left
1806                  * over from a previous flush attempt and should be moved
1807                  * to the current flush group.  If it is not set then all
1808                  * such records had better have been flushed already or
1809                  * already associated with the current flush group.
1810                  */
1811                 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1812                         rec->flush_group = ip->flush_group;
1813                 } else {
1814                         KKASSERT(rec->flush_group == ip->flush_group);
1815                 }
1816                 r = 1;
1817                 break;
1818         }
1819         return(r);
1820 }
1821
1822 /*
1823  * This version just moves records already in a flush state to the new
1824  * flush group and that is it.
1825  */
1826 static int
1827 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
1828 {
1829         hammer_inode_t ip = rec->ip;
1830
1831         switch(rec->flush_state) {
1832         case HAMMER_FST_FLUSH:
1833                 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1834                         rec->flush_group = ip->flush_group;
1835                 } else {
1836                         KKASSERT(rec->flush_group == ip->flush_group);
1837                 }
1838                 break;
1839         default:
1840                 break;
1841         }
1842         return(0);
1843 }
1844
1845 /*
1846  * Wait for a previously queued flush to complete.  Not only do we need to
1847  * wait for the inode to sync out, we also may have to run the flusher again
1848  * to get it past the UNDO position pertaining to the flush so a crash does
1849  * not 'undo' our flush.
1850  */
1851 void
1852 hammer_wait_inode(hammer_inode_t ip)
1853 {
1854         hammer_mount_t hmp = ip->hmp;
1855         int sync_group;
1856         int waitcount;
1857
1858         sync_group = ip->flush_group;
1859         waitcount = (ip->flags & HAMMER_INODE_REFLUSH) ? 2 : 1;
1860
1861         if (ip->flush_state == HAMMER_FST_SETUP) {
1862                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1863         }
1864         /* XXX can we make this != FST_IDLE ? check SETUP depends */
1865         while (ip->flush_state == HAMMER_FST_FLUSH &&
1866                (ip->flush_group - sync_group) < waitcount) {
1867                 ip->flags |= HAMMER_INODE_FLUSHW;
1868                 tsleep(&ip->flags, 0, "hmrwin", 0);
1869         }
1870         while (hmp->flusher.done - sync_group < waitcount) {
1871                 kprintf("Y");
1872                 hammer_flusher_sync(hmp);
1873         }
1874 }
1875
1876 /*
1877  * Called by the backend code when a flush has been completed.
1878  * The inode has already been removed from the flush list.
1879  *
1880  * A pipelined flush can occur, in which case we must re-enter the
1881  * inode on the list and re-copy its fields.
1882  */
1883 void
1884 hammer_flush_inode_done(hammer_inode_t ip)
1885 {
1886         hammer_mount_t hmp;
1887         int dorel;
1888
1889         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1890
1891         hmp = ip->hmp;
1892
1893         /*
1894          * Merge left-over flags back into the frontend and fix the state.
1895          * Incomplete truncations are retained by the backend.
1896          */
1897         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
1898         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
1899
1900         /*
1901          * The backend may have adjusted nlinks, so if the adjusted nlinks
1902          * does not match the fronttend set the frontend's RDIRTY flag again.
1903          */
1904         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
1905                 ip->flags |= HAMMER_INODE_DDIRTY;
1906
1907         /*
1908          * Fix up the dirty buffer status.
1909          */
1910         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
1911                 ip->flags |= HAMMER_INODE_BUFS;
1912         }
1913
1914         /*
1915          * Re-set the XDIRTY flag if some of the inode's in-memory records
1916          * could not be flushed.
1917          */
1918         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
1919                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
1920                  (!RB_EMPTY(&ip->rec_tree) &&
1921                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
1922
1923         /*
1924          * Do not lose track of inodes which no longer have vnode
1925          * assocations, otherwise they may never get flushed again.
1926          */
1927         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1928                 ip->flags |= HAMMER_INODE_REFLUSH;
1929
1930         /*
1931          * Clean up the vnode ref
1932          */
1933         if (ip->flags & HAMMER_INODE_VHELD) {
1934                 ip->flags &= ~HAMMER_INODE_VHELD;
1935                 vrele(ip->vp);
1936         }
1937
1938         /*
1939          * Adjust flush_state.  The target state (idle or setup) shouldn't
1940          * be terribly important since we will reflush if we really need
1941          * to do anything.
1942          *
1943          * If the WOULDBLOCK flag is set we must re-flush immediately
1944          * to continue a potentially large deletion.  The flag also causes
1945          * the hammer_setup_child_callback() to move records in the old
1946          * flush group to the new one.
1947          */
1948         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1949                 ip->flush_state = HAMMER_FST_IDLE;
1950                 hammer_flush_inode_core(ip, HAMMER_FLUSH_SIGNAL);
1951                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
1952                 dorel = 1;
1953         } else if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
1954                 ip->flush_state = HAMMER_FST_IDLE;
1955                 dorel = 1;
1956         } else {
1957                 ip->flush_state = HAMMER_FST_SETUP;
1958                 dorel = 0;
1959         }
1960
1961         --hmp->count_iqueued;
1962         --hammer_count_iqueued;
1963
1964         /*
1965          * If the frontend made more changes and requested another flush,
1966          * then try to get it running.
1967          */
1968         if (ip->flags & HAMMER_INODE_REFLUSH) {
1969                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1970                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1971                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1972                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1973                 } else {
1974                         hammer_flush_inode(ip, 0);
1975                 }
1976         }
1977
1978         /*
1979          * If the inode is now clean drop the space reservation.
1980          */
1981         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1982             (ip->flags & HAMMER_INODE_RSV_INODES)) {
1983                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
1984                 --hmp->rsv_inodes;
1985         }
1986
1987         /*
1988          * Finally, if the frontend is waiting for a flush to complete,
1989          * wake it up.
1990          */
1991         if (ip->flush_state != HAMMER_FST_FLUSH) {
1992                 if (ip->flags & HAMMER_INODE_FLUSHW) {
1993                         ip->flags &= ~HAMMER_INODE_FLUSHW;
1994                         wakeup(&ip->flags);
1995                 }
1996         }
1997         if (dorel)
1998                 hammer_rel_inode(ip, 0);
1999 }
2000
2001 /*
2002  * Called from hammer_sync_inode() to synchronize in-memory records
2003  * to the media.
2004  */
2005 static int
2006 hammer_sync_record_callback(hammer_record_t record, void *data)
2007 {
2008         hammer_cursor_t cursor = data;
2009         hammer_transaction_t trans = cursor->trans;
2010         hammer_mount_t hmp = trans->hmp;
2011         int error;
2012
2013         /*
2014          * Skip records that do not belong to the current flush.
2015          */
2016         ++hammer_stats_record_iterations;
2017         if (record->flush_state != HAMMER_FST_FLUSH)
2018                 return(0);
2019
2020 #if 1
2021         if (record->flush_group != record->ip->flush_group) {
2022                 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2023                 Debugger("blah2");
2024                 return(0);
2025         }
2026 #endif
2027         KKASSERT(record->flush_group == record->ip->flush_group);
2028
2029         /*
2030          * Interlock the record using the BE flag.  Once BE is set the
2031          * frontend cannot change the state of FE.
2032          *
2033          * NOTE: If FE is set prior to us setting BE we still sync the
2034          * record out, but the flush completion code converts it to
2035          * a delete-on-disk record instead of destroying it.
2036          */
2037         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2038         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2039
2040         /*
2041          * The backend may have already disposed of the record.
2042          */
2043         if (record->flags & HAMMER_RECF_DELETED_BE) {
2044                 error = 0;
2045                 goto done;
2046         }
2047
2048         /*
2049          * If the whole inode is being deleting all on-disk records will
2050          * be deleted very soon, we can't sync any new records to disk
2051          * because they will be deleted in the same transaction they were
2052          * created in (delete_tid == create_tid), which will assert.
2053          *
2054          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2055          * that we currently panic on.
2056          */
2057         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2058                 switch(record->type) {
2059                 case HAMMER_MEM_RECORD_DATA:
2060                         /*
2061                          * We don't have to do anything, if the record was
2062                          * committed the space will have been accounted for
2063                          * in the blockmap.
2064                          */
2065                         /* fall through */
2066                 case HAMMER_MEM_RECORD_GENERAL:
2067                         record->flags |= HAMMER_RECF_DELETED_FE;
2068                         record->flags |= HAMMER_RECF_DELETED_BE;
2069                         error = 0;
2070                         goto done;
2071                 case HAMMER_MEM_RECORD_ADD:
2072                         panic("hammer_sync_record_callback: illegal add "
2073                               "during inode deletion record %p", record);
2074                         break; /* NOT REACHED */
2075                 case HAMMER_MEM_RECORD_INODE:
2076                         panic("hammer_sync_record_callback: attempt to "
2077                               "sync inode record %p?", record);
2078                         break; /* NOT REACHED */
2079                 case HAMMER_MEM_RECORD_DEL:
2080                         /*
2081                          * Follow through and issue the on-disk deletion
2082                          */
2083                         break;
2084                 }
2085         }
2086
2087         /*
2088          * If DELETED_FE is set special handling is needed for directory
2089          * entries.  Dependant pieces related to the directory entry may
2090          * have already been synced to disk.  If this occurs we have to
2091          * sync the directory entry and then change the in-memory record
2092          * from an ADD to a DELETE to cover the fact that it's been
2093          * deleted by the frontend.
2094          *
2095          * A directory delete covering record (MEM_RECORD_DEL) can never
2096          * be deleted by the frontend.
2097          *
2098          * Any other record type (aka DATA) can be deleted by the frontend.
2099          * XXX At the moment the flusher must skip it because there may
2100          * be another data record in the flush group for the same block,
2101          * meaning that some frontend data changes can leak into the backend's
2102          * synchronization point.
2103          */
2104         if (record->flags & HAMMER_RECF_DELETED_FE) {
2105                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2106                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2107                 } else {
2108                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2109                         record->flags |= HAMMER_RECF_DELETED_BE;
2110                         error = 0;
2111                         goto done;
2112                 }
2113         }
2114
2115         /*
2116          * Assign the create_tid for new records.  Deletions already
2117          * have the record's entire key properly set up.
2118          */
2119         if (record->type != HAMMER_MEM_RECORD_DEL)
2120                 record->leaf.base.create_tid = trans->tid;
2121                 record->leaf.create_ts = trans->time32;
2122         for (;;) {
2123                 error = hammer_ip_sync_record_cursor(cursor, record);
2124                 if (error != EDEADLK)
2125                         break;
2126                 hammer_done_cursor(cursor);
2127                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2128                                            record->ip);
2129                 if (error)
2130                         break;
2131         }
2132         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2133
2134         if (error) {
2135                 error = -error;
2136                 if (error != -ENOSPC) {
2137                         kprintf("hammer_sync_record_callback: sync failed rec "
2138                                 "%p, error %d\n", record, error);
2139                         Debugger("sync failed rec");
2140                 }
2141         }
2142 done:
2143         hammer_flush_record_done(record, error);
2144
2145         /*
2146          * Do partial finalization if we have built up too many dirty
2147          * buffers.  Otherwise a buffer cache deadlock can occur when
2148          * doing things like creating tens of thousands of tiny files.
2149          *
2150          * We must release our cursor lock to avoid a 3-way deadlock
2151          * due to the exclusive sync lock the finalizer must get.
2152          */
2153         if (hammer_flusher_meta_limit(hmp)) {
2154                 hammer_unlock_cursor(cursor, 0);
2155                 hammer_flusher_finalize(trans, 0);
2156                 hammer_lock_cursor(cursor, 0);
2157         }
2158
2159         return(error);
2160 }
2161
2162 /*
2163  * XXX error handling
2164  */
2165 int
2166 hammer_sync_inode(hammer_inode_t ip)
2167 {
2168         struct hammer_transaction trans;
2169         struct hammer_cursor cursor;
2170         hammer_node_t tmp_node;
2171         hammer_record_t depend;
2172         hammer_record_t next;
2173         int error, tmp_error;
2174         u_int64_t nlinks;
2175
2176         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2177                 return(0);
2178
2179         hammer_start_transaction_fls(&trans, ip->hmp);
2180         error = hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2181         if (error)
2182                 goto done;
2183
2184         /*
2185          * Any directory records referencing this inode which are not in
2186          * our current flush group must adjust our nlink count for the
2187          * purposes of synchronization to disk.
2188          *
2189          * Records which are in our flush group can be unlinked from our
2190          * inode now, potentially allowing the inode to be physically
2191          * deleted.
2192          *
2193          * This cannot block.
2194          */
2195         nlinks = ip->ino_data.nlinks;
2196         next = TAILQ_FIRST(&ip->target_list);
2197         while ((depend = next) != NULL) {
2198                 next = TAILQ_NEXT(depend, target_entry);
2199                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2200                     depend->flush_group == ip->hmp->flusher.act) {
2201                         /*
2202                          * If this is an ADD that was deleted by the frontend
2203                          * the frontend nlinks count will have already been
2204                          * decremented, but the backend is going to sync its
2205                          * directory entry and must account for it.  The
2206                          * record will be converted to a delete-on-disk when
2207                          * it gets synced.
2208                          *
2209                          * If the ADD was not deleted by the frontend we
2210                          * can remove the dependancy from our target_list.
2211                          */
2212                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2213                                 ++nlinks;
2214                         } else {
2215                                 TAILQ_REMOVE(&ip->target_list, depend,
2216                                              target_entry);
2217                                 depend->target_ip = NULL;
2218                         }
2219                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2220                         /*
2221                          * Not part of our flush group
2222                          */
2223                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2224                         switch(depend->type) {
2225                         case HAMMER_MEM_RECORD_ADD:
2226                                 --nlinks;
2227                                 break;
2228                         case HAMMER_MEM_RECORD_DEL:
2229                                 ++nlinks;
2230                                 break;
2231                         default:
2232                                 break;
2233                         }
2234                 }
2235         }
2236
2237         /*
2238          * Set dirty if we had to modify the link count.
2239          */
2240         if (ip->sync_ino_data.nlinks != nlinks) {
2241                 KKASSERT((int64_t)nlinks >= 0);
2242                 ip->sync_ino_data.nlinks = nlinks;
2243                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2244         }
2245
2246         /*
2247          * If there is a trunction queued destroy any data past the (aligned)
2248          * truncation point.  Userland will have dealt with the buffer
2249          * containing the truncation point for us.
2250          *
2251          * We don't flush pending frontend data buffers until after we've
2252          * dealt with the truncation.
2253          */
2254         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2255                 /*
2256                  * Interlock trunc_off.  The VOP front-end may continue to
2257                  * make adjustments to it while we are blocked.
2258                  */
2259                 off_t trunc_off;
2260                 off_t aligned_trunc_off;
2261                 int blkmask;
2262
2263                 trunc_off = ip->sync_trunc_off;
2264                 blkmask = hammer_blocksize(trunc_off) - 1;
2265                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2266
2267                 /*
2268                  * Delete any whole blocks on-media.  The front-end has
2269                  * already cleaned out any partial block and made it
2270                  * pending.  The front-end may have updated trunc_off
2271                  * while we were blocked so we only use sync_trunc_off.
2272                  *
2273                  * This operation can blow out the buffer cache, EWOULDBLOCK
2274                  * means we were unable to complete the deletion.  The
2275                  * deletion will update sync_trunc_off in that case.
2276                  */
2277                 error = hammer_ip_delete_range(&cursor, ip,
2278                                                 aligned_trunc_off,
2279                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2280                 if (error == EWOULDBLOCK) {
2281                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2282                         error = 0;
2283                         goto defer_buffer_flush;
2284                 }
2285
2286                 if (error)
2287                         Debugger("hammer_ip_delete_range errored");
2288
2289                 /*
2290                  * Clear the truncation flag on the backend after we have
2291                  * complete the deletions.  Backend data is now good again
2292                  * (including new records we are about to sync, below).
2293                  *
2294                  * Leave sync_trunc_off intact.  As we write additional
2295                  * records the backend will update sync_trunc_off.  This
2296                  * tells the backend whether it can skip the overwrite
2297                  * test.  This should work properly even when the backend
2298                  * writes full blocks where the truncation point straddles
2299                  * the block because the comparison is against the base
2300                  * offset of the record.
2301                  */
2302                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2303                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2304         } else {
2305                 error = 0;
2306         }
2307
2308         /*
2309          * Now sync related records.  These will typically be directory
2310          * entries, records tracking direct-writes, or delete-on-disk records.
2311          */
2312         if (error == 0) {
2313                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2314                                     hammer_sync_record_callback, &cursor);
2315                 if (tmp_error < 0)
2316                         tmp_error = -error;
2317                 if (tmp_error)
2318                         error = tmp_error;
2319         }
2320         hammer_cache_node(&ip->cache[1], cursor.node);
2321
2322         /*
2323          * Re-seek for inode update, assuming our cache hasn't been ripped
2324          * out from under us.
2325          */
2326         if (error == 0) {
2327                 tmp_node = hammer_ref_node_safe(ip->hmp, &ip->cache[0], &error);
2328                 if (tmp_node) {
2329                         hammer_cursor_downgrade(&cursor);
2330                         hammer_lock_sh(&tmp_node->lock);
2331                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2332                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2333                         hammer_unlock(&tmp_node->lock);
2334                         hammer_rel_node(tmp_node);
2335                 }
2336                 error = 0;
2337         }
2338
2339         /*
2340          * If we are deleting the inode the frontend had better not have
2341          * any active references on elements making up the inode.
2342          *
2343          * The call to hammer_ip_delete_clean() cleans up auxillary records
2344          * but not DB or DATA records.  Those must have already been deleted
2345          * by the normal truncation mechanic.
2346          */
2347         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2348                 RB_EMPTY(&ip->rec_tree)  &&
2349             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2350             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2351                 int count1 = 0;
2352
2353                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2354                 if (error == 0) {
2355                         ip->flags |= HAMMER_INODE_DELETED;
2356                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2357                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2358                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2359
2360                         /*
2361                          * Set delete_tid in both the frontend and backend
2362                          * copy of the inode record.  The DELETED flag handles
2363                          * this, do not set RDIRTY.
2364                          */
2365                         ip->ino_leaf.base.delete_tid = trans.tid;
2366                         ip->sync_ino_leaf.base.delete_tid = trans.tid;
2367                         ip->ino_leaf.delete_ts = trans.time32;
2368                         ip->sync_ino_leaf.delete_ts = trans.time32;
2369
2370
2371                         /*
2372                          * Adjust the inode count in the volume header
2373                          */
2374                         hammer_sync_lock_sh(&trans);
2375                         if (ip->flags & HAMMER_INODE_ONDISK) {
2376                                 hammer_modify_volume_field(&trans,
2377                                                            trans.rootvol,
2378                                                            vol0_stat_inodes);
2379                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2380                                 hammer_modify_volume_done(trans.rootvol);
2381                         }
2382                         hammer_sync_unlock(&trans);
2383                 } else {
2384                         Debugger("hammer_ip_delete_clean errored");
2385                 }
2386         }
2387
2388         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2389
2390         if (error)
2391                 Debugger("RB_SCAN errored");
2392
2393 defer_buffer_flush:
2394         /*
2395          * Now update the inode's on-disk inode-data and/or on-disk record.
2396          * DELETED and ONDISK are managed only in ip->flags.
2397          *
2398          * In the case of a defered buffer flush we still update the on-disk
2399          * inode to satisfy visibility requirements if there happen to be
2400          * directory dependancies.
2401          */
2402         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2403         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2404                 /*
2405                  * If deleted and on-disk, don't set any additional flags.
2406                  * the delete flag takes care of things.
2407                  *
2408                  * Clear flags which may have been set by the frontend.
2409                  */
2410                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2411                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2412                                     HAMMER_INODE_DELETING);
2413                 break;
2414         case HAMMER_INODE_DELETED:
2415                 /*
2416                  * Take care of the case where a deleted inode was never
2417                  * flushed to the disk in the first place.
2418                  *
2419                  * Clear flags which may have been set by the frontend.
2420                  */
2421                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2422                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2423                                     HAMMER_INODE_DELETING);
2424                 while (RB_ROOT(&ip->rec_tree)) {
2425                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2426                         hammer_ref(&record->lock);
2427                         KKASSERT(record->lock.refs == 1);
2428                         record->flags |= HAMMER_RECF_DELETED_FE;
2429                         record->flags |= HAMMER_RECF_DELETED_BE;
2430                         hammer_rel_mem_record(record);
2431                 }
2432                 break;
2433         case HAMMER_INODE_ONDISK:
2434                 /*
2435                  * If already on-disk, do not set any additional flags.
2436                  */
2437                 break;
2438         default:
2439                 /*
2440                  * If not on-disk and not deleted, set DDIRTY to force
2441                  * an initial record to be written.
2442                  *
2443                  * Also set the create_tid in both the frontend and backend
2444                  * copy of the inode record.
2445                  */
2446                 ip->ino_leaf.base.create_tid = trans.tid;
2447                 ip->ino_leaf.create_ts = trans.time32;
2448                 ip->sync_ino_leaf.base.create_tid = trans.tid;
2449                 ip->sync_ino_leaf.create_ts = trans.time32;
2450                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2451                 break;
2452         }
2453
2454         /*
2455          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2456          * is already on-disk the old record is marked as deleted.
2457          *
2458          * If DELETED is set hammer_update_inode() will delete the existing
2459          * record without writing out a new one.
2460          *
2461          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2462          */
2463         if (ip->flags & HAMMER_INODE_DELETED) {
2464                 error = hammer_update_inode(&cursor, ip);
2465         } else
2466         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2467             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2468                 error = hammer_update_itimes(&cursor, ip);
2469         } else
2470         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2471                 error = hammer_update_inode(&cursor, ip);
2472         }
2473         if (error)
2474                 Debugger("hammer_update_itimes/inode errored");
2475 done:
2476         /*
2477          * Save the TID we used to sync the inode with to make sure we
2478          * do not improperly reuse it.
2479          */
2480         hammer_done_cursor(&cursor);
2481         hammer_done_transaction(&trans);
2482         return(error);
2483 }
2484
2485 /*
2486  * This routine is called when the OS is no longer actively referencing
2487  * the inode (but might still be keeping it cached), or when releasing
2488  * the last reference to an inode.
2489  *
2490  * At this point if the inode's nlinks count is zero we want to destroy
2491  * it, which may mean destroying it on-media too.
2492  */
2493 void
2494 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2495 {
2496         struct vnode *vp;
2497
2498         /*
2499          * Set the DELETING flag when the link count drops to 0 and the
2500          * OS no longer has any opens on the inode.
2501          *
2502          * The backend will clear DELETING (a mod flag) and set DELETED
2503          * (a state flag) when it is actually able to perform the
2504          * operation.
2505          */
2506         if (ip->ino_data.nlinks == 0 &&
2507             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2508                 ip->flags |= HAMMER_INODE_DELETING;
2509                 ip->flags |= HAMMER_INODE_TRUNCATED;
2510                 ip->trunc_off = 0;
2511                 vp = NULL;
2512                 if (getvp) {
2513                         if (hammer_get_vnode(ip, &vp) != 0)
2514                                 return;
2515                 }
2516
2517                 /*
2518                  * Final cleanup
2519                  */
2520                 if (ip->vp) {
2521                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2522                         vnode_pager_setsize(ip->vp, 0);
2523                 }
2524                 if (getvp) {
2525                         vput(vp);
2526                 }
2527         }
2528 }
2529
2530 /*
2531  * Re-test an inode when a dependancy had gone away to see if we
2532  * can chain flush it.
2533  */
2534 void
2535 hammer_test_inode(hammer_inode_t ip)
2536 {
2537         if (ip->flags & HAMMER_INODE_REFLUSH) {
2538                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2539                 hammer_ref(&ip->lock);
2540                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2541                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2542                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2543                 } else {
2544                         hammer_flush_inode(ip, 0);
2545                 }
2546                 hammer_rel_inode(ip, 0);
2547         }
2548 }
2549
2550 /*
2551  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2552  * reassociated with a vp or just before it gets freed.
2553  *
2554  * Wakeup one thread blocked waiting on reclaims to complete.  Note that
2555  * the inode the thread is waiting on behalf of is a different inode then
2556  * the inode we are called with.  This is to create a pipeline.
2557  */
2558 static void
2559 hammer_inode_wakereclaims(hammer_inode_t ip)
2560 {
2561         struct hammer_reclaim *reclaim;
2562         hammer_mount_t hmp = ip->hmp;
2563
2564         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2565                 return;
2566
2567         --hammer_count_reclaiming;
2568         --hmp->inode_reclaims;
2569         ip->flags &= ~HAMMER_INODE_RECLAIM;
2570
2571         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
2572                 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2573                 reclaim->okydoky = 1;
2574                 wakeup(reclaim);
2575         }
2576 }
2577
2578 /*
2579  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
2580  * inodes build up before we start blocking.
2581  *
2582  * When we block we don't care *which* inode has finished reclaiming,
2583  * as lone as one does.  This is somewhat heuristical... we also put a
2584  * cap on how long we are willing to wait.
2585  */
2586 void
2587 hammer_inode_waitreclaims(hammer_mount_t hmp)
2588 {
2589         struct hammer_reclaim reclaim;
2590         int delay;
2591
2592         if (hmp->inode_reclaims > HAMMER_RECLAIM_WAIT) {
2593                 reclaim.okydoky = 0;
2594                 TAILQ_INSERT_TAIL(&hmp->reclaim_list,
2595                                   &reclaim, entry);
2596         } else {
2597                 reclaim.okydoky = 1;
2598         }
2599
2600         if (reclaim.okydoky == 0) {
2601                 delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2602                         HAMMER_RECLAIM_WAIT;
2603                 if (delay >= 0)
2604                         tsleep(&reclaim, 0, "hmrrcm", delay + 1);
2605                 if (reclaim.okydoky == 0)
2606                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2607         }
2608 }
2609