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 1.61 2008/05/22 04:14:01 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_flush_inode_core(hammer_inode_t ip, int flags);
  44 static int hammer_setup_child_callback(hammer_record_t rec, void *data);
  45 static int hammer_setup_parent_inodes(hammer_record_t record);
  46
  47 /*
  48  * The kernel is not actively referencing this vnode but is still holding
  49  * it cached.
  50  *
  51  * This is called from the frontend.
  52  */
  53 int
  54 hammer_vop_inactive(struct vop_inactive_args *ap)
  55 {
  56         struct hammer_inode *ip = VTOI(ap->a_vp);
  57
  58         /*
  59          * Degenerate case
  60          */
  61         if (ip == NULL) {
  62                 vrecycle(ap->a_vp);
  63                 return(0);
  64         }
  65
  66         /*
  67          * If the inode no longer has visibility in the filesystem and is
  68          * fairly clean, try to recycle it immediately.  This can deadlock
  69          * in vfsync() if we aren't careful.
  70          *
  71          * Do not queue the inode to the flusher if we still have visibility,
  72          * otherwise namespace calls such as chmod will unnecessarily generate
  73          * multiple inode updates.
  74          */
  75         hammer_inode_unloadable_check(ip, 0);
  76         if (ip->ino_data.nlinks == 0) {
  77                 if (ip->flags & HAMMER_INODE_MODMASK)
  78                         hammer_flush_inode(ip, 0);
  79                 else
  80                         vrecycle(ap->a_vp);
  81         }
  82         return(0);
  83 }
  84
  85 /*
  86  * Release the vnode association.  This is typically (but not always)
  87  * the last reference on the inode.
  88  *
  89  * Once the association is lost we are on our own with regards to
  90  * flushing the inode.
  91  */
  92 int
  93 hammer_vop_reclaim(struct vop_reclaim_args *ap)
  94 {
  95         struct hammer_inode *ip;
  96         struct vnode *vp;
  97
  98         vp = ap->a_vp;
  99
 100         if ((ip = vp->v_data) != NULL) {
 101                 vp->v_data = NULL;
 102                 ip->vp = NULL;
 103                 hammer_rel_inode(ip, 1);
 104         }
 105         return(0);
 106 }
 107
 108 /*
 109  * Return a locked vnode for the specified inode.  The inode must be
 110  * referenced but NOT LOCKED on entry and will remain referenced on
 111  * return.
 112  *
 113  * Called from the frontend.
 114  */
 115 int
 116 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 117 {
 118         struct vnode *vp;
 119         int error = 0;
 120
 121         for (;;) {
 122                 if ((vp = ip->vp) == NULL) {
 123                         error = getnewvnode(VT_HAMMER, ip->hmp->mp, vpp, 0, 0);
 124                         if (error)
 125                                 break;
 126                         hammer_lock_ex(&ip->lock);
 127                         if (ip->vp != NULL) {
 128                                 hammer_unlock(&ip->lock);
 129                                 vp->v_type = VBAD;
 130                                 vx_put(vp);
 131                                 continue;
 132                         }
 133                         hammer_ref(&ip->lock);
 134                         vp = *vpp;
 135                         ip->vp = vp;
 136                         vp->v_type =
 137                                 hammer_get_vnode_type(ip->ino_data.obj_type);
 138
 139                         switch(ip->ino_data.obj_type) {
 140                         case HAMMER_OBJTYPE_CDEV:
 141                         case HAMMER_OBJTYPE_BDEV:
 142                                 vp->v_ops = &ip->hmp->mp->mnt_vn_spec_ops;
 143                                 addaliasu(vp, ip->ino_data.rmajor,
 144                                           ip->ino_data.rminor);
 145                                 break;
 146                         case HAMMER_OBJTYPE_FIFO:
 147                                 vp->v_ops = &ip->hmp->mp->mnt_vn_fifo_ops;
 148                                 break;
 149                         default:
 150                                 break;
 151                         }
 152
 153                         /*
 154                          * Only mark as the root vnode if the ip is not
 155                          * historical, otherwise the VFS cache will get
 156                          * confused.  The other half of the special handling
 157                          * is in hammer_vop_nlookupdotdot().
 158                          */
 159                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 160                             ip->obj_asof == ip->hmp->asof) {
 161                                 vp->v_flag |= VROOT;
 162                         }
 163
 164                         vp->v_data = (void *)ip;
 165                         /* vnode locked by getnewvnode() */
 166                         /* make related vnode dirty if inode dirty? */
 167                         hammer_unlock(&ip->lock);
 168                         if (vp->v_type == VREG)
 169                                 vinitvmio(vp, ip->ino_data.size);
 170                         break;
 171                 }
 172
 173                 /*
 174                  * loop if the vget fails (aka races), or if the vp
 175                  * no longer matches ip->vp.
 176                  */
 177                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 178                         if (vp == ip->vp)
 179                                 break;
 180                         vput(vp);
 181                 }
 182         }
 183         *vpp = vp;
 184         return(error);
 185 }
 186
 187 /*
 188  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 189  * do not attach or detach the related vnode (use hammer_get_vnode() for
 190  * that).
 191  *
 192  * The flags argument is only applied for newly created inodes, and only
 193  * certain flags are inherited.
 194  *
 195  * Called from the frontend.
 196  */
 197 struct hammer_inode *
 198 hammer_get_inode(hammer_transaction_t trans, struct hammer_node **cache,
 199                  u_int64_t obj_id, hammer_tid_t asof, int flags, int *errorp)
 200 {
 201         hammer_mount_t hmp = trans->hmp;
 202         struct hammer_inode_info iinfo;
 203         struct hammer_cursor cursor;
 204         struct hammer_inode *ip;
 205
 206         /*
 207          * Determine if we already have an inode cached.  If we do then
 208          * we are golden.
 209          */
 210         iinfo.obj_id = obj_id;
 211         iinfo.obj_asof = asof;
 212 loop:
 213         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 214         if (ip) {
 215                 hammer_ref(&ip->lock);
 216                 *errorp = 0;
 217                 return(ip);
 218         }
 219
 220         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 221         ++hammer_count_inodes;
 222         ip->obj_id = obj_id;
 223         ip->obj_asof = iinfo.obj_asof;
 224         ip->hmp = hmp;
 225         ip->flags = flags & HAMMER_INODE_RO;
 226         if (hmp->ronly)
 227                 ip->flags |= HAMMER_INODE_RO;
 228         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 229         RB_INIT(&ip->rec_tree);
 230         TAILQ_INIT(&ip->bio_list);
 231         TAILQ_INIT(&ip->bio_alt_list);
 232         TAILQ_INIT(&ip->target_list);
 233
 234         /*
 235          * Locate the on-disk inode.
 236          */
 237 retry:
 238         hammer_init_cursor(trans, &cursor, cache, NULL);
 239         cursor.key_beg.localization = HAMMER_LOCALIZE_INODE;
 240         cursor.key_beg.obj_id = ip->obj_id;
 241         cursor.key_beg.key = 0;
 242         cursor.key_beg.create_tid = 0;
 243         cursor.key_beg.delete_tid = 0;
 244         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 245         cursor.key_beg.obj_type = 0;
 246         cursor.asof = iinfo.obj_asof;
 247         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 248                        HAMMER_CURSOR_ASOF;
 249
 250         *errorp = hammer_btree_lookup(&cursor);
 251         if (*errorp == EDEADLK) {
 252                 hammer_done_cursor(&cursor);
 253                 goto retry;
 254         }
 255
 256         /*
 257          * On success the B-Tree lookup will hold the appropriate
 258          * buffer cache buffers and provide a pointer to the requested
 259          * information.  Copy the information to the in-memory inode
 260          * and cache the B-Tree node to improve future operations.
 261          */
 262         if (*errorp == 0) {
 263                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 264                 ip->ino_data = cursor.data->inode;
 265                 hammer_cache_node(cursor.node, &ip->cache[0]);
 266                 if (cache)
 267                         hammer_cache_node(cursor.node, cache);
 268         }
 269
 270         /*
 271          * On success load the inode's record and data and insert the
 272          * inode into the B-Tree.  It is possible to race another lookup
 273          * insertion of the same inode so deal with that condition too.
 274          *
 275          * The cursor's locked node interlocks against others creating and
 276          * destroying ip while we were blocked.
 277          */
 278         if (*errorp == 0) {
 279                 hammer_ref(&ip->lock);
 280                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 281                         hammer_uncache_node(&ip->cache[0]);
 282                         hammer_uncache_node(&ip->cache[1]);
 283                         KKASSERT(ip->lock.refs == 1);
 284                         --hammer_count_inodes;
 285                         kfree(ip, M_HAMMER);
 286                         hammer_done_cursor(&cursor);
 287                         goto loop;
 288                 }
 289                 ip->flags |= HAMMER_INODE_ONDISK;
 290         } else {
 291                 /*
 292                  * Do not panic on read-only accesses which fail, particularly
 293                  * historical accesses where the snapshot might not have
 294                  * complete connectivity.
 295                  */
 296                 if ((flags & HAMMER_INODE_RO) == 0) {
 297                         kprintf("hammer_get_inode: failed ip %p obj_id %016llx cursor %p error %d\n",
 298                                 ip, ip->obj_id, &cursor, *errorp);
 299                         Debugger("x");
 300                 }
 301                 --hammer_count_inodes;
 302                 kfree(ip, M_HAMMER);
 303                 ip = NULL;
 304         }
 305         hammer_done_cursor(&cursor);
 306         return (ip);
 307 }
 308
 309 /*
 310  * Create a new filesystem object, returning the inode in *ipp.  The
 311  * returned inode will be referenced.
 312  *
 313  * The inode is created in-memory.
 314  */
 315 int
 316 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 317                     struct ucred *cred, hammer_inode_t dip,
 318                     struct hammer_inode **ipp)
 319 {
 320         hammer_mount_t hmp;
 321         hammer_inode_t ip;
 322         uid_t xuid;
 323
 324         hmp = trans->hmp;
 325         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 326         ++hammer_count_inodes;
 327         ip->obj_id = hammer_alloc_objid(trans, dip);
 328         KKASSERT(ip->obj_id != 0);
 329         ip->obj_asof = hmp->asof;
 330         ip->hmp = hmp;
 331         ip->flush_state = HAMMER_FST_IDLE;
 332         ip->flags = HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES;
 333
 334         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 335         RB_INIT(&ip->rec_tree);
 336         TAILQ_INIT(&ip->bio_list);
 337         TAILQ_INIT(&ip->bio_alt_list);
 338         TAILQ_INIT(&ip->target_list);
 339
 340         ip->ino_leaf.atime = trans->time;
 341         ip->ino_data.mtime = trans->time;
 342         ip->ino_data.size = 0;
 343         ip->ino_data.nlinks = 0;
 344         /* XXX */
 345         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 346         ip->ino_leaf.base.localization = HAMMER_LOCALIZE_INODE;
 347         ip->ino_leaf.base.obj_id = ip->obj_id;
 348         ip->ino_leaf.base.key = 0;
 349         ip->ino_leaf.base.create_tid = 0;
 350         ip->ino_leaf.base.delete_tid = 0;
 351         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 352         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 353
 354         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 355         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 356         ip->ino_data.mode = vap->va_mode;
 357         ip->ino_data.ctime = trans->time;
 358         ip->ino_data.parent_obj_id = (dip) ? dip->ino_leaf.base.obj_id : 0;
 359
 360         switch(ip->ino_leaf.base.obj_type) {
 361         case HAMMER_OBJTYPE_CDEV:
 362         case HAMMER_OBJTYPE_BDEV:
 363                 ip->ino_data.rmajor = vap->va_rmajor;
 364                 ip->ino_data.rminor = vap->va_rminor;
 365                 break;
 366         default:
 367                 break;
 368         }
 369
 370         /*
 371          * Calculate default uid/gid and overwrite with information from
 372          * the vap.
 373          */
 374         xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 375         ip->ino_data.gid = dip->ino_data.gid;
 376         xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode, xuid, cred,
 377                                      &vap->va_mode);
 378         ip->ino_data.mode = vap->va_mode;
 379
 380         if (vap->va_vaflags & VA_UID_UUID_VALID)
 381                 ip->ino_data.uid = vap->va_uid_uuid;
 382         else if (vap->va_uid != (uid_t)VNOVAL)
 383                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 384         if (vap->va_vaflags & VA_GID_UUID_VALID)
 385                 ip->ino_data.gid = vap->va_gid_uuid;
 386         else if (vap->va_gid != (gid_t)VNOVAL)
 387                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 388
 389         hammer_ref(&ip->lock);
 390         if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 391                 hammer_unref(&ip->lock);
 392                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 393         }
 394         *ipp = ip;
 395         return(0);
 396 }
 397
 398 /*
 399  * Called by hammer_sync_inode().
 400  */
 401 static int
 402 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 403 {
 404         hammer_transaction_t trans = cursor->trans;
 405         hammer_record_t record;
 406         int error;
 407
 408 retry:
 409         error = 0;
 410
 411         /*
 412          * If the inode has a presence on-disk then locate it and mark
 413          * it deleted, setting DELONDISK.
 414          *
 415          * The record may or may not be physically deleted, depending on
 416          * the retention policy.
 417          */
 418         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 419             HAMMER_INODE_ONDISK) {
 420                 hammer_normalize_cursor(cursor);
 421                 cursor->key_beg.localization = HAMMER_LOCALIZE_INODE;
 422                 cursor->key_beg.obj_id = ip->obj_id;
 423                 cursor->key_beg.key = 0;
 424                 cursor->key_beg.create_tid = 0;
 425                 cursor->key_beg.delete_tid = 0;
 426                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 427                 cursor->key_beg.obj_type = 0;
 428                 cursor->asof = ip->obj_asof;
 429                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 430                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 431                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 432
 433                 error = hammer_btree_lookup(cursor);
 434                 if (hammer_debug_inode)
 435                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 436                 if (error) {
 437                         kprintf("error %d\n", error);
 438                         Debugger("hammer_update_inode");
 439                 }
 440
 441                 if (error == 0) {
 442                         error = hammer_ip_delete_record(cursor, trans->tid);
 443                         if (hammer_debug_inode)
 444                                 kprintf(" error %d\n", error);
 445                         if (error && error != EDEADLK) {
 446                                 kprintf("error %d\n", error);
 447                                 Debugger("hammer_update_inode2");
 448                         }
 449                         if (error == 0) {
 450                                 ip->flags |= HAMMER_INODE_DELONDISK;
 451                         }
 452                         if (cursor->node)
 453                                 hammer_cache_node(cursor->node, &ip->cache[0]);
 454                 }
 455                 if (error == EDEADLK) {
 456                         hammer_done_cursor(cursor);
 457                         error = hammer_init_cursor(trans, cursor,
 458                                                    &ip->cache[0], ip);
 459                         if (hammer_debug_inode)
 460                                 kprintf("IPDED %p %d\n", ip, error);
 461                         if (error == 0)
 462                                 goto retry;
 463                 }
 464         }
 465
 466         /*
 467          * Ok, write out the initial record or a new record (after deleting
 468          * the old one), unless the DELETED flag is set.  This routine will
 469          * clear DELONDISK if it writes out a record.
 470          *
 471          * Update our inode statistics if this is the first application of
 472          * the inode on-disk.
 473          */
 474         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 475                 /*
 476                  * Generate a record and write it to the media
 477                  */
 478                 record = hammer_alloc_mem_record(ip, 0);
 479                 record->type = HAMMER_MEM_RECORD_INODE;
 480                 record->flush_state = HAMMER_FST_FLUSH;
 481                 record->leaf = ip->sync_ino_leaf;
 482                 record->leaf.base.create_tid = trans->tid;
 483                 record->leaf.data_len = sizeof(ip->sync_ino_data);
 484                 record->data = (void *)&ip->sync_ino_data;
 485                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
 486                 for (;;) {
 487                         error = hammer_ip_sync_record_cursor(cursor, record);
 488                         if (hammer_debug_inode)
 489                                 kprintf("GENREC %p rec %08x %d\n",
 490                                         ip, record->flags, error);
 491                         if (error != EDEADLK)
 492                                 break;
 493                         hammer_done_cursor(cursor);
 494                         error = hammer_init_cursor(trans, cursor,
 495                                                    &ip->cache[0], ip);
 496                         if (hammer_debug_inode)
 497                                 kprintf("GENREC reinit %d\n", error);
 498                         if (error)
 499                                 break;
 500                 }
 501                 if (error) {
 502                         kprintf("error %d\n", error);
 503                         Debugger("hammer_update_inode3");
 504                 }
 505
 506                 /*
 507                  * The record isn't managed by the inode's record tree,
 508                  * destroy it whether we succeed or fail.
 509                  */
 510                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 511                 record->flags |= HAMMER_RECF_DELETED_FE;
 512                 record->flush_state = HAMMER_FST_IDLE;
 513                 hammer_rel_mem_record(record);
 514
 515                 /*
 516                  * Finish up.
 517                  */
 518                 if (error == 0) {
 519                         if (hammer_debug_inode)
 520                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
 521                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
 522                                             HAMMER_INODE_ITIMES);
 523                         ip->flags &= ~HAMMER_INODE_DELONDISK;
 524
 525                         /*
 526                          * Root volume count of inodes
 527                          */
 528                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
 529                                 hammer_modify_volume_field(trans,
 530                                                            trans->rootvol,
 531                                                            vol0_stat_inodes);
 532                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
 533                                 hammer_modify_volume_done(trans->rootvol);
 534                                 ip->flags |= HAMMER_INODE_ONDISK;
 535                                 if (hammer_debug_inode)
 536                                         kprintf("NOWONDISK %p\n", ip);
 537                         }
 538                 }
 539         }
 540
 541         /*
 542          * If the inode has been destroyed, clean out any left-over flags
 543          * that may have been set by the frontend.
 544          */
 545         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
 546                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
 547                                     HAMMER_INODE_ITIMES);
 548         }
 549         return(error);
 550 }
 551
 552 /*
 553  * Update only the itimes fields.  This is done no-historically.  The
 554  * record is updated in-place on the disk.
 555  */
 556 static int
 557 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
 558 {
 559         hammer_transaction_t trans = cursor->trans;
 560         struct hammer_btree_leaf_elm *leaf;
 561         int error;
 562
 563 retry:
 564         error = 0;
 565         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 566             HAMMER_INODE_ONDISK) {
 567                 hammer_normalize_cursor(cursor);
 568                 cursor->key_beg.localization = HAMMER_LOCALIZE_INODE;
 569                 cursor->key_beg.obj_id = ip->obj_id;
 570                 cursor->key_beg.key = 0;
 571                 cursor->key_beg.create_tid = 0;
 572                 cursor->key_beg.delete_tid = 0;
 573                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 574                 cursor->key_beg.obj_type = 0;
 575                 cursor->asof = ip->obj_asof;
 576                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 577                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 578                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 579
 580                 error = hammer_btree_lookup(cursor);
 581                 if (error) {
 582                         kprintf("error %d\n", error);
 583                         Debugger("hammer_update_itimes1");
 584                 }
 585                 if (error == 0) {
 586                         /*
 587                          * Do not generate UNDO records for atime updates.
 588                          */
 589                         leaf = cursor->leaf;
 590                         hammer_modify_node(trans, cursor->node,
 591                                            &leaf->atime, sizeof(leaf->atime));
 592                         leaf->atime = ip->sync_ino_leaf.atime;
 593                         hammer_modify_node_done(cursor->node);
 594                         /*rec->ino_mtime = ip->sync_ino_rec.ino_mtime;*/
 595                         ip->sync_flags &= ~HAMMER_INODE_ITIMES;
 596                         /* XXX recalculate crc */
 597                         hammer_cache_node(cursor->node, &ip->cache[0]);
 598                 }
 599                 if (error == EDEADLK) {
 600                         hammer_done_cursor(cursor);
 601                         error = hammer_init_cursor(trans, cursor,
 602                                                    &ip->cache[0], ip);
 603                         if (error == 0)
 604                                 goto retry;
 605                 }
 606         }
 607         return(error);
 608 }
 609
 610 /*
 611  * Release a reference on an inode, flush as requested.
 612  *
 613  * On the last reference we queue the inode to the flusher for its final
 614  * disposition.
 615  */
 616 void
 617 hammer_rel_inode(struct hammer_inode *ip, int flush)
 618 {
 619         hammer_mount_t hmp = ip->hmp;
 620
 621         /*
 622          * Handle disposition when dropping the last ref.
 623          */
 624         for (;;) {
 625                 if (ip->lock.refs == 1) {
 626                         /*
 627                          * Determine whether on-disk action is needed for
 628                          * the inode's final disposition.
 629                          */
 630                         KKASSERT(ip->vp == NULL);
 631                         hammer_inode_unloadable_check(ip, 0);
 632                         if (ip->flags & HAMMER_INODE_MODMASK) {
 633                                 hammer_flush_inode(ip, 0);
 634                         } else if (ip->lock.refs == 1) {
 635                                 hammer_unload_inode(ip);
 636                                 break;
 637                         }
 638                 } else {
 639                         if (flush)
 640                                 hammer_flush_inode(ip, 0);
 641
 642                         /*
 643                          * The inode still has multiple refs, try to drop
 644                          * one ref.
 645                          */
 646                         KKASSERT(ip->lock.refs >= 1);
 647                         if (ip->lock.refs > 1) {
 648                                 hammer_unref(&ip->lock);
 649                                 break;
 650                         }
 651                 }
 652         }
 653
 654         /*
 655          * XXX bad hack until I add code to track inodes in SETUP.  We
 656          * can queue a lot of inodes to the syncer but if we don't wake
 657          * it up the undo sets will be too large or too many unflushed
 658          * records will build up and blow our malloc limit.
 659          */
 660         if (++hmp->reclaim_count > 256) {
 661                 hmp->reclaim_count = 0;
 662                 hammer_flusher_async(hmp);
 663         }
 664 }
 665
 666 /*
 667  * Unload and destroy the specified inode.  Must be called with one remaining
 668  * reference.  The reference is disposed of.
 669  *
 670  * This can only be called in the context of the flusher.
 671  */
 672 static int
 673 hammer_unload_inode(struct hammer_inode *ip)
 674 {
 675         KASSERT(ip->lock.refs == 1,
 676                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
 677         KKASSERT(ip->vp == NULL);
 678         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
 679         KKASSERT(ip->cursor_ip_refs == 0);
 680         KKASSERT(ip->lock.lockcount == 0);
 681         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
 682
 683         KKASSERT(RB_EMPTY(&ip->rec_tree));
 684         KKASSERT(TAILQ_EMPTY(&ip->target_list));
 685         KKASSERT(TAILQ_EMPTY(&ip->bio_list));
 686         KKASSERT(TAILQ_EMPTY(&ip->bio_alt_list));
 687
 688         RB_REMOVE(hammer_ino_rb_tree, &ip->hmp->rb_inos_root, ip);
 689
 690         hammer_uncache_node(&ip->cache[0]);
 691         hammer_uncache_node(&ip->cache[1]);
 692         if (ip->objid_cache)
 693                 hammer_clear_objid(ip);
 694         --hammer_count_inodes;
 695         kfree(ip, M_HAMMER);
 696
 697         return(0);
 698 }
 699
 700 /*
 701  * A transaction has modified an inode, requiring updates as specified by
 702  * the passed flags.
 703  *
 704  * HAMMER_INODE_DDIRTY: Inode data has been updated
 705  * HAMMER_INODE_XDIRTY: Dirty in-memory records
 706  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
 707  * HAMMER_INODE_DELETED: Inode record/data must be deleted
 708  * HAMMER_INODE_ITIMES: mtime/atime has been updated
 709  */
 710 void
 711 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
 712 {
 713         KKASSERT ((ip->flags & HAMMER_INODE_RO) == 0 ||
 714                   (flags & (HAMMER_INODE_DDIRTY |
 715                             HAMMER_INODE_XDIRTY | HAMMER_INODE_BUFS |
 716                             HAMMER_INODE_DELETED | HAMMER_INODE_ITIMES)) == 0);
 717
 718         ip->flags |= flags;
 719 }
 720
 721 /*
 722  * Request that an inode be flushed.  This whole mess cannot block and may
 723  * recurse.  Once requested HAMMER will attempt to actively flush it until
 724  * the flush can be done.
 725  *
 726  * The inode may already be flushing, or may be in a setup state.  We can
 727  * place the inode in a flushing state if it is currently idle and flag it
 728  * to reflush if it is currently flushing.
 729  */
 730 void
 731 hammer_flush_inode(hammer_inode_t ip, int flags)
 732 {
 733         hammer_record_t depend;
 734         int r, good;
 735
 736         /*
 737          * Trivial 'nothing to flush' case.  If the inode is ina SETUP
 738          * state we have to put it back into an IDLE state so we can
 739          * drop the extra ref.
 740          */
 741         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
 742                 if (ip->flush_state == HAMMER_FST_SETUP) {
 743                         ip->flush_state = HAMMER_FST_IDLE;
 744                         hammer_rel_inode(ip, 0);
 745                 }
 746                 return;
 747         }
 748
 749         /*
 750          * Our flush action will depend on the current state.
 751          */
 752         switch(ip->flush_state) {
 753         case HAMMER_FST_IDLE:
 754                 /*
 755                  * We have no dependancies and can flush immediately.  Some
 756                  * our children may not be flushable so we have to re-test
 757                  * with that additional knowledge.
 758                  */
 759                 hammer_flush_inode_core(ip, flags);
 760                 break;
 761         case HAMMER_FST_SETUP:
 762                 /*
 763                  * Recurse upwards through dependancies via target_list
 764                  * and start their flusher actions going if possible.
 765                  *
 766                  * 'good' is our connectivity.  -1 means we have none and
 767                  * can't flush, 0 means there weren't any dependancies, and
 768                  * 1 means we have good connectivity.
 769                  */
 770                 good = 0;
 771                 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 772                         r = hammer_setup_parent_inodes(depend);
 773                         if (r < 0 && good == 0)
 774                                 good = -1;
 775                         if (r > 0)
 776                                 good = 1;
 777                 }
 778
 779                 /*
 780                  * We can continue if good >= 0.  Determine how many records
 781                  * under our inode can be flushed (and mark them).
 782                  */
 783                 if (good >= 0) {
 784                         hammer_flush_inode_core(ip, flags);
 785                 } else {
 786                         ip->flags |= HAMMER_INODE_REFLUSH;
 787                         if (flags & HAMMER_FLUSH_SIGNAL) {
 788                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 789                                 hammer_flusher_async(ip->hmp);
 790                         }
 791                 }
 792                 break;
 793         default:
 794                 /*
 795                  * We are already flushing, flag the inode to reflush
 796                  * if needed after it completes its current flush.
 797                  */
 798                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
 799                         ip->flags |= HAMMER_INODE_REFLUSH;
 800                 if (flags & HAMMER_FLUSH_SIGNAL) {
 801                         ip->flags |= HAMMER_INODE_RESIGNAL;
 802                         hammer_flusher_async(ip->hmp);
 803                 }
 804                 break;
 805         }
 806 }
 807
 808 /*
 809  * We are asked to recurse upwards and convert the record from SETUP
 810  * to FLUSH if possible.  record->ip is a parent of the caller's inode,
 811  * and record->target_ip is the caller's inode.
 812  *
 813  * Return 1 if the record gives us connectivity
 814  *
 815  * Return 0 if the record is not relevant
 816  *
 817  * Return -1 if we can't resolve the dependancy and there is no connectivity.
 818  */
 819 static int
 820 hammer_setup_parent_inodes(hammer_record_t record)
 821 {
 822         hammer_mount_t hmp = record->ip->hmp;
 823         hammer_record_t depend;
 824         hammer_inode_t ip;
 825         int r, good;
 826
 827         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
 828         ip = record->ip;
 829
 830         /*
 831          * If the record is already flushing, is it in our flush group?
 832          *
 833          * If it is in our flush group but it is a general record or a
 834          * delete-on-disk, it does not improve our connectivity (return 0),
 835          * and if the target inode is not trying to destroy itself we can't
 836          * allow the operation yet anyway (the second return -1).
 837          */
 838         if (record->flush_state == HAMMER_FST_FLUSH) {
 839                 if (record->flush_group != hmp->flusher_next) {
 840                         ip->flags |= HAMMER_INODE_REFLUSH;
 841                         return(-1);
 842                 }
 843                 if (record->type == HAMMER_MEM_RECORD_ADD)
 844                         return(1);
 845                 /* GENERAL or DEL */
 846                 return(0);
 847         }
 848
 849         /*
 850          * It must be a setup record.  Try to resolve the setup dependancies
 851          * by recursing upwards so we can place ip on the flush list.
 852          */
 853         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
 854
 855         good = 0;
 856         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 857                 r = hammer_setup_parent_inodes(depend);
 858                 if (r < 0 && good == 0)
 859                         good = -1;
 860                 if (r > 0)
 861                         good = 1;
 862         }
 863
 864         /*
 865          * We can't flush ip because it has no connectivity (XXX also check
 866          * nlinks for pre-existing connectivity!).  Flag it so any resolution
 867          * recurses back down.
 868          */
 869         if (good < 0) {
 870                 ip->flags |= HAMMER_INODE_REFLUSH;
 871                 return(good);
 872         }
 873
 874         /*
 875          * We are go, place the parent inode in a flushing state so we can
 876          * place its record in a flushing state.  Note that the parent
 877          * may already be flushing.  The record must be in the same flush
 878          * group as the parent.
 879          */
 880         if (ip->flush_state != HAMMER_FST_FLUSH)
 881                 hammer_flush_inode_core(ip, HAMMER_FLUSH_RECURSION);
 882         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
 883         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
 884
 885 #if 0
 886         if (record->type == HAMMER_MEM_RECORD_DEL &&
 887             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
 888                 /*
 889                  * Regardless of flushing state we cannot sync this path if the
 890                  * record represents a delete-on-disk but the target inode
 891                  * is not ready to sync its own deletion.
 892                  *
 893                  * XXX need to count effective nlinks to determine whether
 894                  * the flush is ok, otherwise removing a hardlink will
 895                  * just leave the DEL record to rot.
 896                  */
 897                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
 898                 return(-1);
 899         } else
 900 #endif
 901         if (ip->flush_group == ip->hmp->flusher_next) {
 902                 /*
 903                  * This is the record we wanted to synchronize.
 904                  */
 905                 record->flush_state = HAMMER_FST_FLUSH;
 906                 record->flush_group = ip->flush_group;
 907                 hammer_ref(&record->lock);
 908                 if (record->type == HAMMER_MEM_RECORD_ADD)
 909                         return(1);
 910
 911                 /*
 912                  * A general or delete-on-disk record does not contribute
 913                  * to our visibility.  We can still flush it, however.
 914                  */
 915                 return(0);
 916         } else {
 917                 /*
 918                  * We couldn't resolve the dependancies, request that the
 919                  * inode be flushed when the dependancies can be resolved.
 920                  */
 921                 ip->flags |= HAMMER_INODE_REFLUSH;
 922                 return(-1);
 923         }
 924 }
 925
 926 /*
 927  * This is the core routine placing an inode into the FST_FLUSH state.
 928  */
 929 static void
 930 hammer_flush_inode_core(hammer_inode_t ip, int flags)
 931 {
 932         int go_count;
 933
 934         /*
 935          * Set flush state and prevent the flusher from cycling into
 936          * the next flush group.  Do not place the ip on the list yet.
 937          * Inodes not in the idle state get an extra reference.
 938          */
 939         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
 940         if (ip->flush_state == HAMMER_FST_IDLE)
 941                 hammer_ref(&ip->lock);
 942         ip->flush_state = HAMMER_FST_FLUSH;
 943         ip->flush_group = ip->hmp->flusher_next;
 944         ++ip->hmp->flusher_lock;
 945
 946         /*
 947          * We need to be able to vfsync/truncate from the backend.
 948          */
 949         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
 950         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
 951                 ip->flags |= HAMMER_INODE_VHELD;
 952                 vref(ip->vp);
 953         }
 954
 955         /*
 956          * Figure out how many in-memory records we can actually flush
 957          * (not including inode meta-data, buffers, etc).
 958          */
 959         if (flags & HAMMER_FLUSH_RECURSION) {
 960                 go_count = 1;
 961         } else {
 962                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
 963                                    hammer_setup_child_callback, NULL);
 964         }
 965
 966         /*
 967          * This is a more involved test that includes go_count.  If we
 968          * can't flush, flag the inode and return.  If go_count is 0 we
 969          * were are unable to flush any records in our rec_tree and
 970          * must ignore the XDIRTY flag.
 971          */
 972         if (go_count == 0) {
 973                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
 974                         ip->flags |= HAMMER_INODE_REFLUSH;
 975                         ip->flush_state = HAMMER_FST_SETUP;
 976                         if (ip->flags & HAMMER_INODE_VHELD) {
 977                                 ip->flags &= ~HAMMER_INODE_VHELD;
 978                                 vrele(ip->vp);
 979                         }
 980                         if (flags & HAMMER_FLUSH_SIGNAL) {
 981                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 982                                 hammer_flusher_async(ip->hmp);
 983                         }
 984                         if (--ip->hmp->flusher_lock == 0)
 985                                 wakeup(&ip->hmp->flusher_lock);
 986                         return;
 987                 }
 988         }
 989
 990         /*
 991          * Snapshot the state of the inode for the backend flusher.
 992          *
 993          * The truncation must be retained in the frontend until after
 994          * we've actually performed the record deletion.
 995          *
 996          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
 997          * and stays in ip->flags.  Once set, it stays set until the
 998          * inode is destroyed.
 999          */
1000         ip->sync_flags = (ip->flags & HAMMER_INODE_MODMASK);
1001         ip->sync_trunc_off = ip->trunc_off;
1002         ip->sync_ino_leaf = ip->ino_leaf;
1003         ip->sync_ino_data = ip->ino_data;
1004         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1005
1006         /*
1007          * The flusher list inherits our inode and reference.
1008          */
1009         TAILQ_INSERT_TAIL(&ip->hmp->flush_list, ip, flush_entry);
1010         if (--ip->hmp->flusher_lock == 0)
1011                 wakeup(&ip->hmp->flusher_lock);
1012
1013         if (flags & HAMMER_FLUSH_SIGNAL)
1014                 hammer_flusher_async(ip->hmp);
1015 }
1016
1017 /*
1018  * Callback for scan of ip->rec_tree.  Try to include each record in our
1019  * flush.  ip->flush_group has been set but the inode has not yet been
1020  * moved into a flushing state.
1021  *
1022  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1023  * both inodes.
1024  *
1025  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1026  * the caller from shortcutting the flush.
1027  */
1028 static int
1029 hammer_setup_child_callback(hammer_record_t rec, void *data)
1030 {
1031         hammer_inode_t target_ip;
1032         hammer_inode_t ip;
1033         int r;
1034
1035         /*
1036          * If the record has been deleted by the backend (it's being held
1037          * by the frontend in a race), just ignore it.
1038          */
1039         if (rec->flags & HAMMER_RECF_DELETED_BE)
1040                 return(0);
1041
1042         /*
1043          * If the record is in an idle state it has no dependancies and
1044          * can be flushed.
1045          */
1046         ip = rec->ip;
1047         r = 0;
1048
1049         switch(rec->flush_state) {
1050         case HAMMER_FST_IDLE:
1051                 /*
1052                  * Record has no setup dependancy, we can flush it.
1053                  */
1054                 KKASSERT(rec->target_ip == NULL);
1055                 rec->flush_state = HAMMER_FST_FLUSH;
1056                 rec->flush_group = ip->flush_group;
1057                 hammer_ref(&rec->lock);
1058                 r = 1;
1059                 break;
1060         case HAMMER_FST_SETUP:
1061                 /*
1062                  * Record has a setup dependancy.  Try to include the
1063                  * target ip in the flush.
1064                  *
1065                  * We have to be careful here, if we do not do the right
1066                  * thing we can lose track of dirty inodes and the system
1067                  * will lockup trying to allocate buffers.
1068                  */
1069                 target_ip = rec->target_ip;
1070                 KKASSERT(target_ip != NULL);
1071                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1072                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1073                         /*
1074                          * If the target IP is already flushing in our group
1075                          * we are golden, otherwise make sure the target
1076                          * reflushes.
1077                          */
1078                         if (target_ip->flush_group == ip->flush_group) {
1079                                 rec->flush_state = HAMMER_FST_FLUSH;
1080                                 rec->flush_group = ip->flush_group;
1081                                 hammer_ref(&rec->lock);
1082                                 r = 1;
1083                         } else {
1084                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1085                         }
1086                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1087                         /*
1088                          * If the target IP is not flushing we can force
1089                          * it to flush, even if it is unable to write out
1090                          * any of its own records we have at least one in
1091                          * hand that we CAN deal with.
1092                          */
1093                         rec->flush_state = HAMMER_FST_FLUSH;
1094                         rec->flush_group = ip->flush_group;
1095                         hammer_ref(&rec->lock);
1096                         hammer_flush_inode_core(target_ip,
1097                                                 HAMMER_FLUSH_RECURSION);
1098                         r = 1;
1099                 } else {
1100                         /*
1101                          * General or delete-on-disk record.
1102                          *
1103                          * XXX this needs help.  If a delete-on-disk we could
1104                          * disconnect the target.  If the target has its own
1105                          * dependancies they really need to be flushed.
1106                          *
1107                          * XXX
1108                          */
1109                         rec->flush_state = HAMMER_FST_FLUSH;
1110                         rec->flush_group = ip->flush_group;
1111                         hammer_ref(&rec->lock);
1112                         hammer_flush_inode_core(target_ip,
1113                                                 HAMMER_FLUSH_RECURSION);
1114                         r = 1;
1115                 }
1116                 break;
1117         case HAMMER_FST_FLUSH:
1118                 /*
1119                  * Record already associated with a flush group.  It had
1120                  * better be ours.
1121                  */
1122                 KKASSERT(rec->flush_group == ip->flush_group);
1123                 r = 1;
1124                 break;
1125         }
1126         return(r);
1127 }
1128
1129 /*
1130  * Wait for a previously queued flush to complete
1131  */
1132 void
1133 hammer_wait_inode(hammer_inode_t ip)
1134 {
1135         while (ip->flush_state != HAMMER_FST_IDLE) {
1136                 ip->flags |= HAMMER_INODE_FLUSHW;
1137                 tsleep(&ip->flags, 0, "hmrwin", 0);
1138         }
1139 }
1140
1141 /*
1142  * Called by the backend code when a flush has been completed.
1143  * The inode has already been removed from the flush list.
1144  *
1145  * A pipelined flush can occur, in which case we must re-enter the
1146  * inode on the list and re-copy its fields.
1147  */
1148 void
1149 hammer_flush_inode_done(hammer_inode_t ip)
1150 {
1151         struct bio *bio;
1152         int dorel = 0;
1153
1154         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1155
1156         /*
1157          * Allow BIOs to queue to the inode's primary bioq again.
1158          */
1159         ip->flags &= ~HAMMER_INODE_WRITE_ALT;
1160
1161         /*
1162          * Merge left-over flags back into the frontend and fix the state.
1163          */
1164         ip->flags |= ip->sync_flags;
1165
1166         /*
1167          * The backend may have adjusted nlinks, so if the adjusted nlinks
1168          * does not match the fronttend set the frontend's RDIRTY flag again.
1169          */
1170         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
1171                 ip->flags |= HAMMER_INODE_DDIRTY;
1172
1173         /*
1174          * Reflush any BIOs that wound up in the alt list.  Our inode will
1175          * also wind up at the end of the flusher's list.
1176          */
1177         while ((bio = TAILQ_FIRST(&ip->bio_alt_list)) != NULL) {
1178                 TAILQ_REMOVE(&ip->bio_alt_list, bio, bio_act);
1179                 TAILQ_INSERT_TAIL(&ip->bio_list, bio, bio_act);
1180         }
1181         /*
1182          * Fix up the dirty buffer status.
1183          */
1184         if (TAILQ_FIRST(&ip->bio_list) ||
1185             (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree))) {
1186                 ip->flags |= HAMMER_INODE_BUFS;
1187         }
1188
1189         /*
1190          * Re-set the XDIRTY flag if some of the inode's in-memory records
1191          * could not be flushed.
1192          */
1193         if (RB_ROOT(&ip->rec_tree))
1194                 ip->flags |= HAMMER_INODE_XDIRTY;
1195
1196         /*
1197          * Do not lose track of inodes which no longer have vnode
1198          * assocations, otherwise they may never get flushed again.
1199          */
1200         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1201                 ip->flags |= HAMMER_INODE_REFLUSH;
1202
1203         /*
1204          * Adjust flush_state.  The target state (idle or setup) shouldn't
1205          * be terribly important since we will reflush if we really need
1206          * to do anything. XXX
1207          */
1208         if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
1209                 ip->flush_state = HAMMER_FST_IDLE;
1210                 dorel = 1;
1211         } else {
1212                 ip->flush_state = HAMMER_FST_SETUP;
1213         }
1214
1215         /*
1216          * Clean up the vnode ref
1217          */
1218         if (ip->flags & HAMMER_INODE_VHELD) {
1219                 ip->flags &= ~HAMMER_INODE_VHELD;
1220                 vrele(ip->vp);
1221         }
1222
1223         /*
1224          * If the frontend made more changes and requested another flush,
1225          * then try to get it running.
1226          */
1227         if (ip->flags & HAMMER_INODE_REFLUSH) {
1228                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1229                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1230                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1231                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1232                 } else {
1233                         hammer_flush_inode(ip, 0);
1234                 }
1235         }
1236
1237         /*
1238          * Finally, if the frontend is waiting for a flush to complete,
1239          * wake it up.
1240          */
1241         if (ip->flush_state != HAMMER_FST_FLUSH) {
1242                 if (ip->flags & HAMMER_INODE_FLUSHW) {
1243                         ip->flags &= ~HAMMER_INODE_FLUSHW;
1244                         wakeup(&ip->flags);
1245                 }
1246         }
1247         if (dorel)
1248                 hammer_rel_inode(ip, 0);
1249 }
1250
1251 /*
1252  * Called from hammer_sync_inode() to synchronize in-memory records
1253  * to the media.
1254  */
1255 static int
1256 hammer_sync_record_callback(hammer_record_t record, void *data)
1257 {
1258         hammer_cursor_t cursor = data;
1259         hammer_transaction_t trans = cursor->trans;
1260         int error;
1261
1262         /*
1263          * Skip records that do not belong to the current flush.
1264          */
1265         if (record->flush_state != HAMMER_FST_FLUSH)
1266                 return(0);
1267         KKASSERT((record->flags & HAMMER_RECF_DELETED_BE) == 0);
1268 #if 1
1269         if (record->flush_group != record->ip->flush_group) {
1270                 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1271                 Debugger("blah2");
1272                 return(0);
1273         }
1274 #endif
1275         KKASSERT(record->flush_group == record->ip->flush_group);
1276
1277         /*
1278          * Interlock the record using the BE flag.  Once BE is set the
1279          * frontend cannot change the state of FE.
1280          *
1281          * NOTE: If FE is set prior to us setting BE we still sync the
1282          * record out, but the flush completion code converts it to
1283          * a delete-on-disk record instead of destroying it.
1284          */
1285         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1286         record->flags |= HAMMER_RECF_INTERLOCK_BE;
1287
1288         /*
1289          * If the whole inode is being deleting all on-disk records will
1290          * be deleted very soon, we can't sync any new records to disk
1291          * because they will be deleted in the same transaction they were
1292          * created in (delete_tid == create_tid), which will assert.
1293          *
1294          * XXX There may be a case with RECORD_ADD with DELETED_FE set
1295          * that we currently panic on.
1296          */
1297         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
1298                 switch(record->type) {
1299                 case HAMMER_MEM_RECORD_GENERAL:
1300                         record->flags |= HAMMER_RECF_DELETED_FE;
1301                         record->flags |= HAMMER_RECF_DELETED_BE;
1302                         error = 0;
1303                         goto done;
1304                 case HAMMER_MEM_RECORD_ADD:
1305                         panic("hammer_sync_record_callback: illegal add "
1306                               "during inode deletion record %p", record);
1307                         break; /* NOT REACHED */
1308                 case HAMMER_MEM_RECORD_INODE:
1309                         panic("hammer_sync_record_callback: attempt to "
1310                               "sync inode record %p?", record);
1311                         break; /* NOT REACHED */
1312                 case HAMMER_MEM_RECORD_DEL:
1313                         /*
1314                          * Follow through and issue the on-disk deletion
1315                          */
1316                         break;
1317                 }
1318         }
1319
1320         /*
1321          * If DELETED_FE is set we may have already sent dependant pieces
1322          * to the disk and we must flush the record as if it hadn't been
1323          * deleted.  This creates a bit of a mess because we have to
1324          * have ip_sync_record convert the record to MEM_RECORD_DEL before
1325          * it inserts the B-Tree record.  Otherwise the media sync might
1326          * be visible to the frontend.
1327          */
1328         if (record->flags & HAMMER_RECF_DELETED_FE) {
1329                 if (record->type == HAMMER_MEM_RECORD_ADD) {
1330                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
1331                 } else {
1332                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
1333                         return(0);
1334                 }
1335         }
1336
1337         /*
1338          * Assign the create_tid for new records.  Deletions already
1339          * have the record's entire key properly set up.
1340          */
1341         if (record->type != HAMMER_MEM_RECORD_DEL)
1342                 record->leaf.base.create_tid = trans->tid;
1343         for (;;) {
1344                 error = hammer_ip_sync_record_cursor(cursor, record);
1345                 if (error != EDEADLK)
1346                         break;
1347                 hammer_done_cursor(cursor);
1348                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
1349                                            record->ip);
1350                 if (error)
1351                         break;
1352         }
1353         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1354
1355         if (error) {
1356                 error = -error;
1357                 if (error != -ENOSPC) {
1358                         kprintf("hammer_sync_record_callback: sync failed rec "
1359                                 "%p, error %d\n", record, error);
1360                         Debugger("sync failed rec");
1361                 }
1362         }
1363 done:
1364         hammer_flush_record_done(record, error);
1365         return(error);
1366 }
1367
1368 /*
1369  * XXX error handling
1370  */
1371 int
1372 hammer_sync_inode(hammer_inode_t ip)
1373 {
1374         struct hammer_transaction trans;
1375         struct hammer_cursor cursor;
1376         struct bio *bio;
1377         hammer_record_t depend;
1378         hammer_record_t next;
1379         int error, tmp_error;
1380         u_int64_t nlinks;
1381
1382         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
1383                 return(0);
1384
1385         hammer_start_transaction_fls(&trans, ip->hmp);
1386         error = hammer_init_cursor(&trans, &cursor, &ip->cache[0], ip);
1387         if (error)
1388                 goto done;
1389
1390         /*
1391          * Any directory records referencing this inode which are not in
1392          * our current flush group must adjust our nlink count for the
1393          * purposes of synchronization to disk.
1394          *
1395          * Records which are in our flush group can be unlinked from our
1396          * inode now, potentially allowing the inode to be physically
1397          * deleted.
1398          */
1399         nlinks = ip->ino_data.nlinks;
1400         next = TAILQ_FIRST(&ip->target_list);
1401         while ((depend = next) != NULL) {
1402                 next = TAILQ_NEXT(depend, target_entry);
1403                 if (depend->flush_state == HAMMER_FST_FLUSH &&
1404                     depend->flush_group == ip->hmp->flusher_act) {
1405                         /*
1406                          * If this is an ADD that was deleted by the frontend
1407                          * the frontend nlinks count will have already been
1408                          * decremented, but the backend is going to sync its
1409                          * directory entry and must account for it.  The
1410                          * record will be converted to a delete-on-disk when
1411                          * it gets synced.
1412                          *
1413                          * If the ADD was not deleted by the frontend we
1414                          * can remove the dependancy from our target_list.
1415                          */
1416                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
1417                                 ++nlinks;
1418                         } else {
1419                                 TAILQ_REMOVE(&ip->target_list, depend,
1420                                              target_entry);
1421                                 depend->target_ip = NULL;
1422                         }
1423                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
1424                         /*
1425                          * Not part of our flush group
1426                          */
1427                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
1428                         switch(depend->type) {
1429                         case HAMMER_MEM_RECORD_ADD:
1430                                 --nlinks;
1431                                 break;
1432                         case HAMMER_MEM_RECORD_DEL:
1433                                 ++nlinks;
1434                                 break;
1435                         default:
1436                                 break;
1437                         }
1438                 }
1439         }
1440
1441         /*
1442          * Set dirty if we had to modify the link count.
1443          */
1444         if (ip->sync_ino_data.nlinks != nlinks) {
1445                 KKASSERT((int64_t)nlinks >= 0);
1446                 ip->sync_ino_data.nlinks = nlinks;
1447                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1448         }
1449
1450         /*
1451          * Queue up as many dirty buffers as we can then set a flag to
1452          * cause any further BIOs to go to the alternative queue.
1453          */
1454         if (ip->flags & HAMMER_INODE_VHELD)
1455                 error = vfsync(ip->vp, MNT_NOWAIT, 1, NULL, NULL);
1456         ip->flags |= HAMMER_INODE_WRITE_ALT;
1457
1458         /*
1459          * The buffer cache may contain dirty buffers beyond the inode
1460          * state we copied from the frontend to the backend.  Because
1461          * we are syncing our buffer cache on the backend, resync
1462          * the truncation point and the file size so we don't wipe out
1463          * any data.
1464          *
1465          * Syncing the buffer cache on the frontend has serious problems
1466          * because it prevents us from passively queueing dirty inodes
1467          * to the backend (the BIO's could stall indefinitely).
1468          */
1469         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1470                 ip->sync_trunc_off = ip->trunc_off;
1471                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1472         }
1473         if (ip->sync_ino_data.size != ip->ino_data.size) {
1474                 ip->sync_ino_data.size = ip->ino_data.size;
1475                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1476         }
1477
1478         /*
1479          * If there is a trunction queued destroy any data past the (aligned)
1480          * truncation point.  Userland will have dealt with the buffer
1481          * containing the truncation point for us.
1482          *
1483          * We don't flush pending frontend data buffers until after we've
1484          * dealth with the truncation.
1485          *
1486          * Don't bother if the inode is or has been deleted.
1487          */
1488         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
1489                 /*
1490                  * Interlock trunc_off.  The VOP front-end may continue to
1491                  * make adjustments to it while we are blocked.
1492                  */
1493                 off_t trunc_off;
1494                 off_t aligned_trunc_off;
1495
1496                 trunc_off = ip->sync_trunc_off;
1497                 aligned_trunc_off = (trunc_off + HAMMER_BUFMASK) &
1498                                     ~HAMMER_BUFMASK64;
1499
1500                 /*
1501                  * Delete any whole blocks on-media.  The front-end has
1502                  * already cleaned out any partial block and made it
1503                  * pending.  The front-end may have updated trunc_off
1504                  * while we were blocked so do not just unconditionally
1505                  * set it to the maximum offset.
1506                  */
1507                 error = hammer_ip_delete_range(&cursor, ip,
1508                                                 aligned_trunc_off,
1509                                                 0x7FFFFFFFFFFFFFFFLL);
1510                 if (error)
1511                         Debugger("hammer_ip_delete_range errored");
1512                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1513                 if (ip->trunc_off >= trunc_off) {
1514                         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1515                         ip->flags &= ~HAMMER_INODE_TRUNCATED;
1516                 }
1517         } else {
1518                 error = 0;
1519         }
1520
1521         /*
1522          * Now sync related records.  These will typically be directory
1523          * entries or delete-on-disk records.
1524          *
1525          * Not all records will be flushed, but clear XDIRTY anyway.  We
1526          * will set it again in the frontend hammer_flush_inode_done()
1527          * if records remain.
1528          */
1529         if (error == 0) {
1530                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1531                                     hammer_sync_record_callback, &cursor);
1532                 if (tmp_error < 0)
1533                         tmp_error = -error;
1534                 if (tmp_error)
1535                         error = tmp_error;
1536                 if (RB_EMPTY(&ip->rec_tree))
1537                         ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
1538         }
1539
1540         /*
1541          * If we are deleting the inode the frontend had better not have
1542          * any active references on elements making up the inode.
1543          */
1544         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
1545                 RB_EMPTY(&ip->rec_tree)  &&
1546             (ip->sync_flags & HAMMER_INODE_DELETING) &&
1547             (ip->flags & HAMMER_INODE_DELETED) == 0) {
1548                 int count1 = 0;
1549
1550                 hkprintf("Y");
1551                 ip->flags |= HAMMER_INODE_DELETED;
1552                 error = hammer_ip_delete_range_all(&cursor, ip, &count1);
1553                 if (error == 0) {
1554                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
1555                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1556                         KKASSERT(RB_EMPTY(&ip->rec_tree));
1557
1558                         /*
1559                          * Set delete_tid in both the frontend and backend
1560                          * copy of the inode record.  The DELETED flag handles
1561                          * this, do not set RDIRTY.
1562                          */
1563                         ip->ino_leaf.base.delete_tid = trans.tid;
1564                         ip->sync_ino_leaf.base.delete_tid = trans.tid;
1565
1566                         /*
1567                          * Adjust the inode count in the volume header
1568                          */
1569                         if (ip->flags & HAMMER_INODE_ONDISK) {
1570                                 hammer_modify_volume_field(&trans,
1571                                                            trans.rootvol,
1572                                                            vol0_stat_inodes);
1573                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1574                                 hammer_modify_volume_done(trans.rootvol);
1575                         }
1576                 } else {
1577                         ip->flags &= ~HAMMER_INODE_DELETED;
1578                         Debugger("hammer_ip_delete_range_all errored");
1579                 }
1580         }
1581
1582         /*
1583          * Flush any queued BIOs.  These will just biodone() the IO's if
1584          * the inode has been deleted.
1585          */
1586         while ((bio = TAILQ_FIRST(&ip->bio_list)) != NULL) {
1587                 TAILQ_REMOVE(&ip->bio_list, bio, bio_act);
1588                 tmp_error = hammer_dowrite(&cursor, ip, bio);
1589                 if (tmp_error)
1590                         error = tmp_error;
1591         }
1592         ip->sync_flags &= ~HAMMER_INODE_BUFS;
1593
1594         if (error)
1595                 Debugger("RB_SCAN errored");
1596
1597         /*
1598          * Now update the inode's on-disk inode-data and/or on-disk record.
1599          * DELETED and ONDISK are managed only in ip->flags.
1600          */
1601         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
1602         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
1603                 /*
1604                  * If deleted and on-disk, don't set any additional flags.
1605                  * the delete flag takes care of things.
1606                  *
1607                  * Clear flags which may have been set by the frontend.
1608                  */
1609                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY|
1610                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1611                                     HAMMER_INODE_DELETING);
1612                 break;
1613         case HAMMER_INODE_DELETED:
1614                 /*
1615                  * Take care of the case where a deleted inode was never
1616                  * flushed to the disk in the first place.
1617                  *
1618                  * Clear flags which may have been set by the frontend.
1619                  */
1620                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY|
1621                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1622                                     HAMMER_INODE_DELETING);
1623                 while (RB_ROOT(&ip->rec_tree)) {
1624                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
1625                         hammer_ref(&record->lock);
1626                         KKASSERT(record->lock.refs == 1);
1627                         record->flags |= HAMMER_RECF_DELETED_FE;
1628                         record->flags |= HAMMER_RECF_DELETED_BE;
1629                         hammer_rel_mem_record(record);
1630                 }
1631                 break;
1632         case HAMMER_INODE_ONDISK:
1633                 /*
1634                  * If already on-disk, do not set any additional flags.
1635                  */
1636                 break;
1637         default:
1638                 /*
1639                  * If not on-disk and not deleted, set both dirty flags
1640                  * to force an initial record to be written.  Also set
1641                  * the create_tid for the inode.
1642                  *
1643                  * Set create_tid in both the frontend and backend
1644                  * copy of the inode record.
1645                  */
1646                 ip->ino_leaf.base.create_tid = trans.tid;
1647                 ip->sync_ino_leaf.base.create_tid = trans.tid;
1648                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1649                 break;
1650         }
1651
1652         /*
1653          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
1654          * is already on-disk the old record is marked as deleted.
1655          *
1656          * If DELETED is set hammer_update_inode() will delete the existing
1657          * record without writing out a new one.
1658          *
1659          * If *ONLY* the ITIMES flag is set we can update the record in-place.
1660          */
1661         if (ip->flags & HAMMER_INODE_DELETED) {
1662                 error = hammer_update_inode(&cursor, ip);
1663         } else
1664         if ((ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES)) ==
1665             HAMMER_INODE_ITIMES) {
1666                 error = hammer_update_itimes(&cursor, ip);
1667         } else
1668         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES)) {
1669                 error = hammer_update_inode(&cursor, ip);
1670         }
1671         if (error)
1672                 Debugger("hammer_update_itimes/inode errored");
1673 done:
1674         /*
1675          * Save the TID we used to sync the inode with to make sure we
1676          * do not improperly reuse it.
1677          */
1678         hammer_done_cursor(&cursor);
1679         hammer_done_transaction(&trans);
1680         return(error);
1681 }
1682
1683 /*
1684  * This routine is called when the OS is no longer actively referencing
1685  * the inode (but might still be keeping it cached), or when releasing
1686  * the last reference to an inode.
1687  *
1688  * At this point if the inode's nlinks count is zero we want to destroy
1689  * it, which may mean destroying it on-media too.
1690  */
1691 void
1692 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
1693 {
1694         struct vnode *vp;
1695         struct bio *bio;
1696
1697         /*
1698          * Set the DELETING flag when the link count drops to 0 and the
1699          * OS no longer has any opens on the inode.
1700          *
1701          * The backend will clear DELETING (a mod flag) and set DELETED
1702          * (a state flag) when it is actually able to perform the
1703          * operation.
1704          */
1705         if (ip->ino_data.nlinks == 0 &&
1706             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
1707                 ip->flags |= HAMMER_INODE_DELETING;
1708                 ip->flags |= HAMMER_INODE_TRUNCATED;
1709                 ip->trunc_off = 0;
1710                 vp = NULL;
1711                 if (getvp) {
1712                         if (hammer_get_vnode(ip, &vp) != 0)
1713                                 return;
1714                 }
1715
1716                 /*
1717                  * biodone any buffers with pending IO.  These buffers are
1718                  * holding a BUF_KERNPROC() exclusive lock and our
1719                  * vtruncbuf() call will deadlock if any remain.
1720                  *
1721                  * (interlocked against hammer_vop_strategy_write via
1722                  *  HAMMER_INODE_DELETING|HAMMER_INODE_DELETED).
1723                  */
1724                 while ((bio = TAILQ_FIRST(&ip->bio_list)) != NULL) {
1725                         TAILQ_REMOVE(&ip->bio_list, bio, bio_act);
1726                         bio->bio_buf->b_resid = 0;
1727                         biodone(bio);
1728                 }
1729                 while ((bio = TAILQ_FIRST(&ip->bio_alt_list)) != NULL) {
1730                         TAILQ_REMOVE(&ip->bio_alt_list, bio, bio_act);
1731                         bio->bio_buf->b_resid = 0;
1732                         biodone(bio);
1733                 }
1734
1735                 /*
1736                  * Final cleanup
1737                  */
1738                 if (ip->vp) {
1739                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
1740                         vnode_pager_setsize(ip->vp, 0);
1741                 }
1742                 if (getvp) {
1743                         vput(vp);
1744                 }
1745         }
1746 }
1747
1748 /*
1749  * Re-test an inode when a dependancy had gone away to see if we
1750  * can chain flush it.
1751  */
1752 void
1753 hammer_test_inode(hammer_inode_t ip)
1754 {
1755         if (ip->flags & HAMMER_INODE_REFLUSH) {
1756                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1757                 hammer_ref(&ip->lock);
1758                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1759                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1760                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1761                 } else {
1762                         hammer_flush_inode(ip, 0);
1763                 }
1764                 hammer_rel_inode(ip, 0);
1765         }
1766 }
1767