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 1.57 2008/05/13 00:15:28 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         hammer_inode_unloadable_check(ip, 0);
  72         if (ip->flags & HAMMER_INODE_MODMASK)
  73                 hammer_flush_inode(ip, 0);
  74         else if (ip->ino_data.nlinks == 0)
  75                 vrecycle(ap->a_vp);
  76         return(0);
  77 }
  78
  79 /*
  80  * Release the vnode association.  This is typically (but not always)
  81  * the last reference on the inode.
  82  *
  83  * Once the association is lost we are on our own with regards to
  84  * flushing the inode.
  85  */
  86 int
  87 hammer_vop_reclaim(struct vop_reclaim_args *ap)
  88 {
  89         struct hammer_inode *ip;
  90         struct vnode *vp;
  91
  92         vp = ap->a_vp;
  93
  94         if ((ip = vp->v_data) != NULL) {
  95                 vp->v_data = NULL;
  96                 ip->vp = NULL;
  97                 hammer_rel_inode(ip, 1);
  98         }
  99         return(0);
 100 }
 101
 102 /*
 103  * Return a locked vnode for the specified inode.  The inode must be
 104  * referenced but NOT LOCKED on entry and will remain referenced on
 105  * return.
 106  *
 107  * Called from the frontend.
 108  */
 109 int
 110 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 111 {
 112         struct vnode *vp;
 113         int error = 0;
 114
 115         for (;;) {
 116                 if ((vp = ip->vp) == NULL) {
 117                         error = getnewvnode(VT_HAMMER, ip->hmp->mp, vpp, 0, 0);
 118                         if (error)
 119                                 break;
 120                         hammer_lock_ex(&ip->lock);
 121                         if (ip->vp != NULL) {
 122                                 hammer_unlock(&ip->lock);
 123                                 vp->v_type = VBAD;
 124                                 vx_put(vp);
 125                                 continue;
 126                         }
 127                         hammer_ref(&ip->lock);
 128                         vp = *vpp;
 129                         ip->vp = vp;
 130                         vp->v_type =
 131                                 hammer_get_vnode_type(ip->ino_data.obj_type);
 132
 133                         switch(ip->ino_data.obj_type) {
 134                         case HAMMER_OBJTYPE_CDEV:
 135                         case HAMMER_OBJTYPE_BDEV:
 136                                 vp->v_ops = &ip->hmp->mp->mnt_vn_spec_ops;
 137                                 addaliasu(vp, ip->ino_data.rmajor,
 138                                           ip->ino_data.rminor);
 139                                 break;
 140                         case HAMMER_OBJTYPE_FIFO:
 141                                 vp->v_ops = &ip->hmp->mp->mnt_vn_fifo_ops;
 142                                 break;
 143                         default:
 144                                 break;
 145                         }
 146
 147                         /*
 148                          * Only mark as the root vnode if the ip is not
 149                          * historical, otherwise the VFS cache will get
 150                          * confused.  The other half of the special handling
 151                          * is in hammer_vop_nlookupdotdot().
 152                          */
 153                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
 154                             ip->obj_asof == ip->hmp->asof) {
 155                                 vp->v_flag |= VROOT;
 156                         }
 157
 158                         vp->v_data = (void *)ip;
 159                         /* vnode locked by getnewvnode() */
 160                         /* make related vnode dirty if inode dirty? */
 161                         hammer_unlock(&ip->lock);
 162                         if (vp->v_type == VREG)
 163                                 vinitvmio(vp, ip->ino_data.size);
 164                         break;
 165                 }
 166
 167                 /*
 168                  * loop if the vget fails (aka races), or if the vp
 169                  * no longer matches ip->vp.
 170                  */
 171                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 172                         if (vp == ip->vp)
 173                                 break;
 174                         vput(vp);
 175                 }
 176         }
 177         *vpp = vp;
 178         return(error);
 179 }
 180
 181 /*
 182  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 183  * do not attach or detach the related vnode (use hammer_get_vnode() for
 184  * that).
 185  *
 186  * The flags argument is only applied for newly created inodes, and only
 187  * certain flags are inherited.
 188  *
 189  * Called from the frontend.
 190  */
 191 struct hammer_inode *
 192 hammer_get_inode(hammer_transaction_t trans, struct hammer_node **cache,
 193                  u_int64_t obj_id, hammer_tid_t asof, int flags, int *errorp)
 194 {
 195         hammer_mount_t hmp = trans->hmp;
 196         struct hammer_inode_info iinfo;
 197         struct hammer_cursor cursor;
 198         struct hammer_inode *ip;
 199
 200         /*
 201          * Determine if we already have an inode cached.  If we do then
 202          * we are golden.
 203          */
 204         iinfo.obj_id = obj_id;
 205         iinfo.obj_asof = asof;
 206 loop:
 207         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 208         if (ip) {
 209                 hammer_ref(&ip->lock);
 210                 *errorp = 0;
 211                 return(ip);
 212         }
 213
 214         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 215         ++hammer_count_inodes;
 216         ip->obj_id = obj_id;
 217         ip->obj_asof = iinfo.obj_asof;
 218         ip->hmp = hmp;
 219         ip->flags = flags & HAMMER_INODE_RO;
 220         if (hmp->ronly)
 221                 ip->flags |= HAMMER_INODE_RO;
 222         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 223         RB_INIT(&ip->rec_tree);
 224         TAILQ_INIT(&ip->bio_list);
 225         TAILQ_INIT(&ip->bio_alt_list);
 226         TAILQ_INIT(&ip->target_list);
 227
 228         /*
 229          * Locate the on-disk inode.
 230          */
 231 retry:
 232         hammer_init_cursor(trans, &cursor, cache, NULL);
 233         cursor.key_beg.obj_id = ip->obj_id;
 234         cursor.key_beg.key = 0;
 235         cursor.key_beg.create_tid = 0;
 236         cursor.key_beg.delete_tid = 0;
 237         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 238         cursor.key_beg.obj_type = 0;
 239         cursor.asof = iinfo.obj_asof;
 240         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 241                        HAMMER_CURSOR_ASOF;
 242
 243         *errorp = hammer_btree_lookup(&cursor);
 244         if (*errorp == EDEADLK) {
 245                 hammer_done_cursor(&cursor);
 246                 goto retry;
 247         }
 248
 249         /*
 250          * On success the B-Tree lookup will hold the appropriate
 251          * buffer cache buffers and provide a pointer to the requested
 252          * information.  Copy the information to the in-memory inode
 253          * and cache the B-Tree node to improve future operations.
 254          */
 255         if (*errorp == 0) {
 256                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 257                 ip->ino_data = cursor.data->inode;
 258                 hammer_cache_node(cursor.node, &ip->cache[0]);
 259                 if (cache)
 260                         hammer_cache_node(cursor.node, cache);
 261         }
 262
 263         /*
 264          * On success load the inode's record and data and insert the
 265          * inode into the B-Tree.  It is possible to race another lookup
 266          * insertion of the same inode so deal with that condition too.
 267          *
 268          * The cursor's locked node interlocks against others creating and
 269          * destroying ip while we were blocked.
 270          */
 271         if (*errorp == 0) {
 272                 hammer_ref(&ip->lock);
 273                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 274                         hammer_uncache_node(&ip->cache[0]);
 275                         hammer_uncache_node(&ip->cache[1]);
 276                         KKASSERT(ip->lock.refs == 1);
 277                         --hammer_count_inodes;
 278                         kfree(ip, M_HAMMER);
 279                         hammer_done_cursor(&cursor);
 280                         goto loop;
 281                 }
 282                 ip->flags |= HAMMER_INODE_ONDISK;
 283         } else {
 284                 /*
 285                  * Do not panic on read-only accesses which fail, particularly
 286                  * historical accesses where the snapshot might not have
 287                  * complete connectivity.
 288                  */
 289                 if ((flags & HAMMER_INODE_RO) == 0) {
 290                         kprintf("hammer_get_inode: failed ip %p obj_id %016llx cursor %p error %d\n",
 291                                 ip, ip->obj_id, &cursor, *errorp);
 292                         Debugger("x");
 293                 }
 294                 --hammer_count_inodes;
 295                 kfree(ip, M_HAMMER);
 296                 ip = NULL;
 297         }
 298         hammer_done_cursor(&cursor);
 299         return (ip);
 300 }
 301
 302 /*
 303  * Create a new filesystem object, returning the inode in *ipp.  The
 304  * returned inode will be referenced.
 305  *
 306  * The inode is created in-memory.
 307  */
 308 int
 309 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 310                     struct ucred *cred, hammer_inode_t dip,
 311                     struct hammer_inode **ipp)
 312 {
 313         hammer_mount_t hmp;
 314         hammer_inode_t ip;
 315         uid_t xuid;
 316
 317         hmp = trans->hmp;
 318         ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
 319         ++hammer_count_inodes;
 320         ip->obj_id = hammer_alloc_objid(trans, dip);
 321         KKASSERT(ip->obj_id != 0);
 322         ip->obj_asof = hmp->asof;
 323         ip->hmp = hmp;
 324         ip->flush_state = HAMMER_FST_IDLE;
 325         ip->flags = HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES;
 326
 327         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 328         RB_INIT(&ip->rec_tree);
 329         TAILQ_INIT(&ip->bio_list);
 330         TAILQ_INIT(&ip->bio_alt_list);
 331         TAILQ_INIT(&ip->target_list);
 332
 333         ip->ino_leaf.atime = trans->time;
 334         ip->ino_data.mtime = trans->time;
 335         ip->ino_data.size = 0;
 336         ip->ino_data.nlinks = 0;
 337         /* XXX */
 338         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 339         ip->ino_leaf.base.obj_id = ip->obj_id;
 340         ip->ino_leaf.base.key = 0;
 341         ip->ino_leaf.base.create_tid = 0;
 342         ip->ino_leaf.base.delete_tid = 0;
 343         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 344         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 345
 346         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 347         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 348         ip->ino_data.mode = vap->va_mode;
 349         ip->ino_data.ctime = trans->time;
 350         ip->ino_data.parent_obj_id = (dip) ? dip->ino_leaf.base.obj_id : 0;
 351
 352         switch(ip->ino_leaf.base.obj_type) {
 353         case HAMMER_OBJTYPE_CDEV:
 354         case HAMMER_OBJTYPE_BDEV:
 355                 ip->ino_data.rmajor = vap->va_rmajor;
 356                 ip->ino_data.rminor = vap->va_rminor;
 357                 break;
 358         default:
 359                 break;
 360         }
 361
 362         /*
 363          * Calculate default uid/gid and overwrite with information from
 364          * the vap.
 365          */
 366         xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 367         ip->ino_data.gid = dip->ino_data.gid;
 368         xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode, xuid, cred,
 369                                      &vap->va_mode);
 370         ip->ino_data.mode = vap->va_mode;
 371
 372         if (vap->va_vaflags & VA_UID_UUID_VALID)
 373                 ip->ino_data.uid = vap->va_uid_uuid;
 374         else if (vap->va_uid != (uid_t)VNOVAL)
 375                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 376         if (vap->va_vaflags & VA_GID_UUID_VALID)
 377                 ip->ino_data.gid = vap->va_gid_uuid;
 378         else if (vap->va_gid != (gid_t)VNOVAL)
 379                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 380
 381         hammer_ref(&ip->lock);
 382         if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 383                 hammer_unref(&ip->lock);
 384                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
 385         }
 386         *ipp = ip;
 387         return(0);
 388 }
 389
 390 /*
 391  * Called by hammer_sync_inode().
 392  */
 393 static int
 394 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
 395 {
 396         hammer_transaction_t trans = cursor->trans;
 397         hammer_record_t record;
 398         int error;
 399
 400 retry:
 401         error = 0;
 402
 403         /*
 404          * If the inode has a presence on-disk then locate it and mark
 405          * it deleted, setting DELONDISK.
 406          *
 407          * The record may or may not be physically deleted, depending on
 408          * the retention policy.
 409          */
 410         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 411             HAMMER_INODE_ONDISK) {
 412                 hammer_normalize_cursor(cursor);
 413                 cursor->key_beg.obj_id = ip->obj_id;
 414                 cursor->key_beg.key = 0;
 415                 cursor->key_beg.create_tid = 0;
 416                 cursor->key_beg.delete_tid = 0;
 417                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 418                 cursor->key_beg.obj_type = 0;
 419                 cursor->asof = ip->obj_asof;
 420                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 421                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 422                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 423
 424                 error = hammer_btree_lookup(cursor);
 425                 if (hammer_debug_inode)
 426                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
 427                 if (error) {
 428                         kprintf("error %d\n", error);
 429                         Debugger("hammer_update_inode");
 430                 }
 431
 432                 if (error == 0) {
 433                         error = hammer_ip_delete_record(cursor, trans->tid);
 434                         if (hammer_debug_inode)
 435                                 kprintf(" error %d\n", error);
 436                         if (error && error != EDEADLK) {
 437                                 kprintf("error %d\n", error);
 438                                 Debugger("hammer_update_inode2");
 439                         }
 440                         if (error == 0) {
 441                                 ip->flags |= HAMMER_INODE_DELONDISK;
 442                         }
 443                         if (cursor->node)
 444                                 hammer_cache_node(cursor->node, &ip->cache[0]);
 445                 }
 446                 if (error == EDEADLK) {
 447                         hammer_done_cursor(cursor);
 448                         error = hammer_init_cursor(trans, cursor,
 449                                                    &ip->cache[0], ip);
 450                         if (hammer_debug_inode)
 451                                 kprintf("IPDED %p %d\n", ip, error);
 452                         if (error == 0)
 453                                 goto retry;
 454                 }
 455         }
 456
 457         /*
 458          * Ok, write out the initial record or a new record (after deleting
 459          * the old one), unless the DELETED flag is set.  This routine will
 460          * clear DELONDISK if it writes out a record.
 461          *
 462          * Update our inode statistics if this is the first application of
 463          * the inode on-disk.
 464          */
 465         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
 466                 /*
 467                  * Generate a record and write it to the media
 468                  */
 469                 record = hammer_alloc_mem_record(ip, 0);
 470                 record->type = HAMMER_MEM_RECORD_INODE;
 471                 record->flush_state = HAMMER_FST_FLUSH;
 472                 record->leaf = ip->sync_ino_leaf;
 473                 record->leaf.base.create_tid = trans->tid;
 474                 record->leaf.data_len = sizeof(ip->sync_ino_data);
 475                 record->data = (void *)&ip->sync_ino_data;
 476                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
 477                 for (;;) {
 478                         error = hammer_ip_sync_record_cursor(cursor, record);
 479                         if (hammer_debug_inode)
 480                                 kprintf("GENREC %p rec %08x %d\n",
 481                                         ip, record->flags, error);
 482                         if (error != EDEADLK)
 483                                 break;
 484                         hammer_done_cursor(cursor);
 485                         error = hammer_init_cursor(trans, cursor,
 486                                                    &ip->cache[0], ip);
 487                         if (hammer_debug_inode)
 488                                 kprintf("GENREC reinit %d\n", error);
 489                         if (error)
 490                                 break;
 491                 }
 492                 if (error) {
 493                         kprintf("error %d\n", error);
 494                         Debugger("hammer_update_inode3");
 495                 }
 496
 497                 /*
 498                  * The record isn't managed by the inode's record tree,
 499                  * destroy it whether we succeed or fail.
 500                  */
 501                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 502                 record->flags |= HAMMER_RECF_DELETED_FE;
 503                 record->flush_state = HAMMER_FST_IDLE;
 504                 hammer_rel_mem_record(record);
 505
 506                 /*
 507                  * Finish up.
 508                  */
 509                 if (error == 0) {
 510                         if (hammer_debug_inode)
 511                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
 512                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
 513                                             HAMMER_INODE_ITIMES);
 514                         ip->flags &= ~HAMMER_INODE_DELONDISK;
 515
 516                         /*
 517                          * Root volume count of inodes
 518                          */
 519                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
 520                                 hammer_modify_volume_field(trans,
 521                                                            trans->rootvol,
 522                                                            vol0_stat_inodes);
 523                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
 524                                 hammer_modify_volume_done(trans->rootvol);
 525                                 ip->flags |= HAMMER_INODE_ONDISK;
 526                                 if (hammer_debug_inode)
 527                                         kprintf("NOWONDISK %p\n", ip);
 528                         }
 529                 }
 530         }
 531
 532         /*
 533          * If the inode has been destroyed, clean out any left-over flags
 534          * that may have been set by the frontend.
 535          */
 536         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
 537                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
 538                                     HAMMER_INODE_ITIMES);
 539         }
 540         return(error);
 541 }
 542
 543 /*
 544  * Update only the itimes fields.  This is done no-historically.  The
 545  * record is updated in-place on the disk.
 546  */
 547 static int
 548 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
 549 {
 550         hammer_transaction_t trans = cursor->trans;
 551         struct hammer_btree_leaf_elm *leaf;
 552         int error;
 553
 554 retry:
 555         error = 0;
 556         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 557             HAMMER_INODE_ONDISK) {
 558                 hammer_normalize_cursor(cursor);
 559                 cursor->key_beg.obj_id = ip->obj_id;
 560                 cursor->key_beg.key = 0;
 561                 cursor->key_beg.create_tid = 0;
 562                 cursor->key_beg.delete_tid = 0;
 563                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 564                 cursor->key_beg.obj_type = 0;
 565                 cursor->asof = ip->obj_asof;
 566                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 567                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
 568                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 569
 570                 error = hammer_btree_lookup(cursor);
 571                 if (error) {
 572                         kprintf("error %d\n", error);
 573                         Debugger("hammer_update_itimes1");
 574                 }
 575                 if (error == 0) {
 576                         /*
 577                          * Do not generate UNDO records for atime updates.
 578                          */
 579                         leaf = cursor->leaf;
 580                         hammer_modify_node(trans, cursor->node,
 581                                            &leaf->atime, sizeof(leaf->atime));
 582                         leaf->atime = ip->sync_ino_leaf.atime;
 583                         hammer_modify_node_done(cursor->node);
 584                         /*rec->ino_mtime = ip->sync_ino_rec.ino_mtime;*/
 585                         ip->sync_flags &= ~HAMMER_INODE_ITIMES;
 586                         /* XXX recalculate crc */
 587                         hammer_cache_node(cursor->node, &ip->cache[0]);
 588                 }
 589                 if (error == EDEADLK) {
 590                         hammer_done_cursor(cursor);
 591                         error = hammer_init_cursor(trans, cursor,
 592                                                    &ip->cache[0], ip);
 593                         if (error == 0)
 594                                 goto retry;
 595                 }
 596         }
 597         return(error);
 598 }
 599
 600 /*
 601  * Release a reference on an inode, flush as requested.
 602  *
 603  * On the last reference we queue the inode to the flusher for its final
 604  * disposition.
 605  */
 606 void
 607 hammer_rel_inode(struct hammer_inode *ip, int flush)
 608 {
 609         hammer_mount_t hmp = ip->hmp;
 610
 611         /*
 612          * Handle disposition when dropping the last ref.
 613          */
 614         for (;;) {
 615                 if (ip->lock.refs == 1) {
 616                         /*
 617                          * Determine whether on-disk action is needed for
 618                          * the inode's final disposition.
 619                          */
 620                         KKASSERT(ip->vp == NULL);
 621                         hammer_inode_unloadable_check(ip, 0);
 622                         if (ip->flags & HAMMER_INODE_MODMASK) {
 623                                 hammer_flush_inode(ip, 0);
 624                         } else if (ip->lock.refs == 1) {
 625                                 hammer_unload_inode(ip);
 626                                 break;
 627                         }
 628                 } else {
 629                         if (flush)
 630                                 hammer_flush_inode(ip, 0);
 631
 632                         /*
 633                          * The inode still has multiple refs, try to drop
 634                          * one ref.
 635                          */
 636                         KKASSERT(ip->lock.refs >= 1);
 637                         if (ip->lock.refs > 1) {
 638                                 hammer_unref(&ip->lock);
 639                                 break;
 640                         }
 641                 }
 642         }
 643
 644         /*
 645          * XXX bad hack until I add code to track inodes in SETUP.  We
 646          * can queue a lot of inodes to the syncer but if we don't wake
 647          * it up the undo sets will be too large or too many unflushed
 648          * records will build up and blow our malloc limit.
 649          */
 650         if (++hmp->reclaim_count > 256) {
 651                 hmp->reclaim_count = 0;
 652                 hammer_flusher_async(hmp);
 653         }
 654 }
 655
 656 /*
 657  * Unload and destroy the specified inode.  Must be called with one remaining
 658  * reference.  The reference is disposed of.
 659  *
 660  * This can only be called in the context of the flusher.
 661  */
 662 static int
 663 hammer_unload_inode(struct hammer_inode *ip)
 664 {
 665         KASSERT(ip->lock.refs == 1,
 666                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
 667         KKASSERT(ip->vp == NULL);
 668         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
 669         KKASSERT(ip->cursor_ip_refs == 0);
 670         KKASSERT(ip->lock.lockcount == 0);
 671         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
 672
 673         KKASSERT(RB_EMPTY(&ip->rec_tree));
 674         KKASSERT(TAILQ_EMPTY(&ip->target_list));
 675         KKASSERT(TAILQ_EMPTY(&ip->bio_list));
 676         KKASSERT(TAILQ_EMPTY(&ip->bio_alt_list));
 677
 678         RB_REMOVE(hammer_ino_rb_tree, &ip->hmp->rb_inos_root, ip);
 679
 680         hammer_uncache_node(&ip->cache[0]);
 681         hammer_uncache_node(&ip->cache[1]);
 682         if (ip->objid_cache)
 683                 hammer_clear_objid(ip);
 684         --hammer_count_inodes;
 685         kfree(ip, M_HAMMER);
 686
 687         return(0);
 688 }
 689
 690 /*
 691  * A transaction has modified an inode, requiring updates as specified by
 692  * the passed flags.
 693  *
 694  * HAMMER_INODE_DDIRTY: Inode data has been updated
 695  * HAMMER_INODE_XDIRTY: Dirty in-memory records
 696  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
 697  * HAMMER_INODE_DELETED: Inode record/data must be deleted
 698  * HAMMER_INODE_ITIMES: mtime/atime has been updated
 699  */
 700 void
 701 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
 702 {
 703         KKASSERT ((ip->flags & HAMMER_INODE_RO) == 0 ||
 704                   (flags & (HAMMER_INODE_DDIRTY |
 705                             HAMMER_INODE_XDIRTY | HAMMER_INODE_BUFS |
 706                             HAMMER_INODE_DELETED | HAMMER_INODE_ITIMES)) == 0);
 707
 708         ip->flags |= flags;
 709 }
 710
 711 /*
 712  * Request that an inode be flushed.  This whole mess cannot block and may
 713  * recurse.  Once requested HAMMER will attempt to actively flush it until
 714  * the flush can be done.
 715  *
 716  * The inode may already be flushing, or may be in a setup state.  We can
 717  * place the inode in a flushing state if it is currently idle and flag it
 718  * to reflush if it is currently flushing.
 719  */
 720 void
 721 hammer_flush_inode(hammer_inode_t ip, int flags)
 722 {
 723         hammer_record_t depend;
 724         int r, good;
 725
 726         /*
 727          * Trivial 'nothing to flush' case.  If the inode is ina SETUP
 728          * state we have to put it back into an IDLE state so we can
 729          * drop the extra ref.
 730          */
 731         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
 732                 if (ip->flush_state == HAMMER_FST_SETUP) {
 733                         ip->flush_state = HAMMER_FST_IDLE;
 734                         hammer_rel_inode(ip, 0);
 735                 }
 736                 return;
 737         }
 738
 739         /*
 740          * Our flush action will depend on the current state.
 741          */
 742         switch(ip->flush_state) {
 743         case HAMMER_FST_IDLE:
 744                 /*
 745                  * We have no dependancies and can flush immediately.  Some
 746                  * our children may not be flushable so we have to re-test
 747                  * with that additional knowledge.
 748                  */
 749                 hammer_flush_inode_core(ip, flags);
 750                 break;
 751         case HAMMER_FST_SETUP:
 752                 /*
 753                  * Recurse upwards through dependancies via target_list
 754                  * and start their flusher actions going if possible.
 755                  *
 756                  * 'good' is our connectivity.  -1 means we have none and
 757                  * can't flush, 0 means there weren't any dependancies, and
 758                  * 1 means we have good connectivity.
 759                  */
 760                 good = 0;
 761                 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 762                         r = hammer_setup_parent_inodes(depend);
 763                         if (r < 0 && good == 0)
 764                                 good = -1;
 765                         if (r > 0)
 766                                 good = 1;
 767                 }
 768
 769                 /*
 770                  * We can continue if good >= 0.  Determine how many records
 771                  * under our inode can be flushed (and mark them).
 772                  */
 773                 if (good >= 0) {
 774                         hammer_flush_inode_core(ip, flags);
 775                 } else {
 776                         ip->flags |= HAMMER_INODE_REFLUSH;
 777                         if (flags & HAMMER_FLUSH_SIGNAL) {
 778                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 779                                 hammer_flusher_async(ip->hmp);
 780                         }
 781                 }
 782                 break;
 783         default:
 784                 /*
 785                  * We are already flushing, flag the inode to reflush
 786                  * if needed after it completes its current flush.
 787                  */
 788                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
 789                         ip->flags |= HAMMER_INODE_REFLUSH;
 790                 if (flags & HAMMER_FLUSH_SIGNAL) {
 791                         ip->flags |= HAMMER_INODE_RESIGNAL;
 792                         hammer_flusher_async(ip->hmp);
 793                 }
 794                 break;
 795         }
 796 }
 797
 798 /*
 799  * We are asked to recurse upwards and convert the record from SETUP
 800  * to FLUSH if possible.  record->ip is a parent of the caller's inode,
 801  * and record->target_ip is the caller's inode.
 802  *
 803  * Return 1 if the record gives us connectivity
 804  *
 805  * Return 0 if the record is not relevant
 806  *
 807  * Return -1 if we can't resolve the dependancy and there is no connectivity.
 808  */
 809 static int
 810 hammer_setup_parent_inodes(hammer_record_t record)
 811 {
 812         hammer_mount_t hmp = record->ip->hmp;
 813         hammer_record_t depend;
 814         hammer_inode_t ip;
 815         int r, good;
 816
 817         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
 818         ip = record->ip;
 819
 820         /*
 821          * If the record is already flushing, is it in our flush group?
 822          *
 823          * If it is in our flush group but it is a general record or a
 824          * delete-on-disk, it does not improve our connectivity (return 0),
 825          * and if the target inode is not trying to destroy itself we can't
 826          * allow the operation yet anyway (the second return -1).
 827          */
 828         if (record->flush_state == HAMMER_FST_FLUSH) {
 829                 if (record->flush_group != hmp->flusher_next) {
 830                         ip->flags |= HAMMER_INODE_REFLUSH;
 831                         return(-1);
 832                 }
 833                 if (record->type == HAMMER_MEM_RECORD_ADD)
 834                         return(1);
 835                 /* GENERAL or DEL */
 836                 return(0);
 837         }
 838
 839         /*
 840          * It must be a setup record.  Try to resolve the setup dependancies
 841          * by recursing upwards so we can place ip on the flush list.
 842          */
 843         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
 844
 845         good = 0;
 846         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 847                 r = hammer_setup_parent_inodes(depend);
 848                 if (r < 0 && good == 0)
 849                         good = -1;
 850                 if (r > 0)
 851                         good = 1;
 852         }
 853
 854         /*
 855          * We can't flush ip because it has no connectivity (XXX also check
 856          * nlinks for pre-existing connectivity!).  Flag it so any resolution
 857          * recurses back down.
 858          */
 859         if (good < 0) {
 860                 ip->flags |= HAMMER_INODE_REFLUSH;
 861                 return(good);
 862         }
 863
 864         /*
 865          * We are go, place the parent inode in a flushing state so we can
 866          * place its record in a flushing state.  Note that the parent
 867          * may already be flushing.  The record must be in the same flush
 868          * group as the parent.
 869          */
 870         if (ip->flush_state != HAMMER_FST_FLUSH)
 871                 hammer_flush_inode_core(ip, HAMMER_FLUSH_RECURSION);
 872         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
 873         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
 874
 875 #if 0
 876         if (record->type == HAMMER_MEM_RECORD_DEL &&
 877             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
 878                 /*
 879                  * Regardless of flushing state we cannot sync this path if the
 880                  * record represents a delete-on-disk but the target inode
 881                  * is not ready to sync its own deletion.
 882                  *
 883                  * XXX need to count effective nlinks to determine whether
 884                  * the flush is ok, otherwise removing a hardlink will
 885                  * just leave the DEL record to rot.
 886                  */
 887                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
 888                 return(-1);
 889         } else
 890 #endif
 891         if (ip->flush_group == ip->hmp->flusher_next) {
 892                 /*
 893                  * This is the record we wanted to synchronize.
 894                  */
 895                 record->flush_state = HAMMER_FST_FLUSH;
 896                 record->flush_group = ip->flush_group;
 897                 hammer_ref(&record->lock);
 898                 if (record->type == HAMMER_MEM_RECORD_ADD)
 899                         return(1);
 900
 901                 /*
 902                  * A general or delete-on-disk record does not contribute
 903                  * to our visibility.  We can still flush it, however.
 904                  */
 905                 return(0);
 906         } else {
 907                 /*
 908                  * We couldn't resolve the dependancies, request that the
 909                  * inode be flushed when the dependancies can be resolved.
 910                  */
 911                 ip->flags |= HAMMER_INODE_REFLUSH;
 912                 return(-1);
 913         }
 914 }
 915
 916 /*
 917  * This is the core routine placing an inode into the FST_FLUSH state.
 918  */
 919 static void
 920 hammer_flush_inode_core(hammer_inode_t ip, int flags)
 921 {
 922         int go_count;
 923
 924         /*
 925          * Set flush state and prevent the flusher from cycling into
 926          * the next flush group.  Do not place the ip on the list yet.
 927          * Inodes not in the idle state get an extra reference.
 928          */
 929         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
 930         if (ip->flush_state == HAMMER_FST_IDLE)
 931                 hammer_ref(&ip->lock);
 932         ip->flush_state = HAMMER_FST_FLUSH;
 933         ip->flush_group = ip->hmp->flusher_next;
 934         ++ip->hmp->flusher_lock;
 935
 936         /*
 937          * We need to be able to vfsync/truncate from the backend.
 938          */
 939         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
 940         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
 941                 ip->flags |= HAMMER_INODE_VHELD;
 942                 vref(ip->vp);
 943         }
 944
 945         /*
 946          * Figure out how many in-memory records we can actually flush
 947          * (not including inode meta-data, buffers, etc).
 948          */
 949         if (flags & HAMMER_FLUSH_RECURSION) {
 950                 go_count = 1;
 951         } else {
 952                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
 953                                    hammer_setup_child_callback, NULL);
 954         }
 955
 956         /*
 957          * This is a more involved test that includes go_count.  If we
 958          * can't flush, flag the inode and return.  If go_count is 0 we
 959          * were are unable to flush any records in our rec_tree and
 960          * must ignore the XDIRTY flag.
 961          */
 962         if (go_count == 0) {
 963                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
 964                         ip->flags |= HAMMER_INODE_REFLUSH;
 965                         ip->flush_state = HAMMER_FST_SETUP;
 966                         if (ip->flags & HAMMER_INODE_VHELD) {
 967                                 ip->flags &= ~HAMMER_INODE_VHELD;
 968                                 vrele(ip->vp);
 969                         }
 970                         if (flags & HAMMER_FLUSH_SIGNAL) {
 971                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 972                                 hammer_flusher_async(ip->hmp);
 973                         }
 974                         if (--ip->hmp->flusher_lock == 0)
 975                                 wakeup(&ip->hmp->flusher_lock);
 976                         return;
 977                 }
 978         }
 979
 980         /*
 981          * Snapshot the state of the inode for the backend flusher.
 982          *
 983          * The truncation must be retained in the frontend until after
 984          * we've actually performed the record deletion.
 985          *
 986          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
 987          * and stays in ip->flags.  Once set, it stays set until the
 988          * inode is destroyed.
 989          */
 990         ip->sync_flags = (ip->flags & HAMMER_INODE_MODMASK);
 991         ip->sync_trunc_off = ip->trunc_off;
 992         ip->sync_ino_leaf = ip->ino_leaf;
 993         ip->sync_ino_data = ip->ino_data;
 994         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
 995
 996         /*
 997          * The flusher list inherits our inode and reference.
 998          */
 999         TAILQ_INSERT_TAIL(&ip->hmp->flush_list, ip, flush_entry);
1000         if (--ip->hmp->flusher_lock == 0)
1001                 wakeup(&ip->hmp->flusher_lock);
1002
1003         if (flags & HAMMER_FLUSH_SIGNAL)
1004                 hammer_flusher_async(ip->hmp);
1005 }
1006
1007 /*
1008  * Callback for scan of ip->rec_tree.  Try to include each record in our
1009  * flush.  ip->flush_group has been set but the inode has not yet been
1010  * moved into a flushing state.
1011  *
1012  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1013  * both inodes.
1014  *
1015  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1016  * the caller from shortcutting the flush.
1017  */
1018 static int
1019 hammer_setup_child_callback(hammer_record_t rec, void *data)
1020 {
1021         hammer_inode_t target_ip;
1022         hammer_inode_t ip;
1023         int r;
1024
1025         /*
1026          * If the record has been deleted by the backend (it's being held
1027          * by the frontend in a race), just ignore it.
1028          */
1029         if (rec->flags & HAMMER_RECF_DELETED_BE)
1030                 return(0);
1031
1032         /*
1033          * If the record is in an idle state it has no dependancies and
1034          * can be flushed.
1035          */
1036         ip = rec->ip;
1037         r = 0;
1038
1039         switch(rec->flush_state) {
1040         case HAMMER_FST_IDLE:
1041                 /*
1042                  * Record has no setup dependancy, we can flush it.
1043                  */
1044                 KKASSERT(rec->target_ip == NULL);
1045                 rec->flush_state = HAMMER_FST_FLUSH;
1046                 rec->flush_group = ip->flush_group;
1047                 hammer_ref(&rec->lock);
1048                 r = 1;
1049                 break;
1050         case HAMMER_FST_SETUP:
1051                 /*
1052                  * Record has a setup dependancy.  Try to include the
1053                  * target ip in the flush.
1054                  *
1055                  * We have to be careful here, if we do not do the right
1056                  * thing we can lose track of dirty inodes and the system
1057                  * will lockup trying to allocate buffers.
1058                  */
1059                 target_ip = rec->target_ip;
1060                 KKASSERT(target_ip != NULL);
1061                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1062                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1063                         /*
1064                          * If the target IP is already flushing in our group
1065                          * we are golden, otherwise make sure the target
1066                          * reflushes.
1067                          */
1068                         if (target_ip->flush_group == ip->flush_group) {
1069                                 rec->flush_state = HAMMER_FST_FLUSH;
1070                                 rec->flush_group = ip->flush_group;
1071                                 hammer_ref(&rec->lock);
1072                                 r = 1;
1073                         } else {
1074                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1075                         }
1076                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1077                         /*
1078                          * If the target IP is not flushing we can force
1079                          * it to flush, even if it is unable to write out
1080                          * any of its own records we have at least one in
1081                          * hand that we CAN deal with.
1082                          */
1083                         rec->flush_state = HAMMER_FST_FLUSH;
1084                         rec->flush_group = ip->flush_group;
1085                         hammer_ref(&rec->lock);
1086                         hammer_flush_inode_core(target_ip,
1087                                                 HAMMER_FLUSH_RECURSION);
1088                         r = 1;
1089                 } else {
1090                         /*
1091                          * General or delete-on-disk record.
1092                          *
1093                          * XXX this needs help.  If a delete-on-disk we could
1094                          * disconnect the target.  If the target has its own
1095                          * dependancies they really need to be flushed.
1096                          *
1097                          * XXX
1098                          */
1099                         rec->flush_state = HAMMER_FST_FLUSH;
1100                         rec->flush_group = ip->flush_group;
1101                         hammer_ref(&rec->lock);
1102                         hammer_flush_inode_core(target_ip,
1103                                                 HAMMER_FLUSH_RECURSION);
1104                         r = 1;
1105                 }
1106                 break;
1107         case HAMMER_FST_FLUSH:
1108                 /*
1109                  * Record already associated with a flush group.  It had
1110                  * better be ours.
1111                  */
1112                 KKASSERT(rec->flush_group == ip->flush_group);
1113                 r = 1;
1114                 break;
1115         }
1116         return(r);
1117 }
1118
1119 /*
1120  * Wait for a previously queued flush to complete
1121  */
1122 void
1123 hammer_wait_inode(hammer_inode_t ip)
1124 {
1125         while (ip->flush_state != HAMMER_FST_IDLE) {
1126                 ip->flags |= HAMMER_INODE_FLUSHW;
1127                 tsleep(&ip->flags, 0, "hmrwin", 0);
1128         }
1129 }
1130
1131 /*
1132  * Called by the backend code when a flush has been completed.
1133  * The inode has already been removed from the flush list.
1134  *
1135  * A pipelined flush can occur, in which case we must re-enter the
1136  * inode on the list and re-copy its fields.
1137  */
1138 void
1139 hammer_flush_inode_done(hammer_inode_t ip)
1140 {
1141         struct bio *bio;
1142         int dorel = 0;
1143
1144         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1145
1146         /*
1147          * Allow BIOs to queue to the inode's primary bioq again.
1148          */
1149         ip->flags &= ~HAMMER_INODE_WRITE_ALT;
1150
1151         /*
1152          * Merge left-over flags back into the frontend and fix the state.
1153          */
1154         ip->flags |= ip->sync_flags;
1155
1156         /*
1157          * The backend may have adjusted nlinks, so if the adjusted nlinks
1158          * does not match the fronttend set the frontend's RDIRTY flag again.
1159          */
1160         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
1161                 ip->flags |= HAMMER_INODE_DDIRTY;
1162
1163         /*
1164          * Reflush any BIOs that wound up in the alt list.  Our inode will
1165          * also wind up at the end of the flusher's list.
1166          */
1167         while ((bio = TAILQ_FIRST(&ip->bio_alt_list)) != NULL) {
1168                 TAILQ_REMOVE(&ip->bio_alt_list, bio, bio_act);
1169                 TAILQ_INSERT_TAIL(&ip->bio_list, bio, bio_act);
1170         }
1171         /*
1172          * Fix up the dirty buffer status.
1173          */
1174         if (TAILQ_FIRST(&ip->bio_list) ||
1175             (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree))) {
1176                 ip->flags |= HAMMER_INODE_BUFS;
1177         }
1178
1179         /*
1180          * Re-set the XDIRTY flag if some of the inode's in-memory records
1181          * could not be flushed.
1182          */
1183         if (RB_ROOT(&ip->rec_tree))
1184                 ip->flags |= HAMMER_INODE_XDIRTY;
1185
1186         /*
1187          * Do not lose track of inodes which no longer have vnode
1188          * assocations, otherwise they may never get flushed again.
1189          */
1190         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1191                 ip->flags |= HAMMER_INODE_REFLUSH;
1192
1193         /*
1194          * Adjust flush_state.  The target state (idle or setup) shouldn't
1195          * be terribly important since we will reflush if we really need
1196          * to do anything. XXX
1197          */
1198         if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
1199                 ip->flush_state = HAMMER_FST_IDLE;
1200                 dorel = 1;
1201         } else {
1202                 ip->flush_state = HAMMER_FST_SETUP;
1203         }
1204
1205         /*
1206          * Clean up the vnode ref
1207          */
1208         if (ip->flags & HAMMER_INODE_VHELD) {
1209                 ip->flags &= ~HAMMER_INODE_VHELD;
1210                 vrele(ip->vp);
1211         }
1212
1213         /*
1214          * If the frontend made more changes and requested another flush,
1215          * then try to get it running.
1216          */
1217         if (ip->flags & HAMMER_INODE_REFLUSH) {
1218                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1219                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1220                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1221                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1222                 } else {
1223                         hammer_flush_inode(ip, 0);
1224                 }
1225         }
1226
1227         /*
1228          * Finally, if the frontend is waiting for a flush to complete,
1229          * wake it up.
1230          */
1231         if (ip->flush_state != HAMMER_FST_FLUSH) {
1232                 if (ip->flags & HAMMER_INODE_FLUSHW) {
1233                         ip->flags &= ~HAMMER_INODE_FLUSHW;
1234                         wakeup(&ip->flags);
1235                 }
1236         }
1237         if (dorel)
1238                 hammer_rel_inode(ip, 0);
1239 }
1240
1241 /*
1242  * Called from hammer_sync_inode() to synchronize in-memory records
1243  * to the media.
1244  */
1245 static int
1246 hammer_sync_record_callback(hammer_record_t record, void *data)
1247 {
1248         hammer_cursor_t cursor = data;
1249         hammer_transaction_t trans = cursor->trans;
1250         int error;
1251
1252         /*
1253          * Skip records that do not belong to the current flush.
1254          */
1255         if (record->flush_state != HAMMER_FST_FLUSH)
1256                 return(0);
1257         KKASSERT((record->flags & HAMMER_RECF_DELETED_BE) == 0);
1258 #if 1
1259         if (record->flush_group != record->ip->flush_group) {
1260                 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1261                 Debugger("blah2");
1262                 return(0);
1263         }
1264 #endif
1265         KKASSERT(record->flush_group == record->ip->flush_group);
1266
1267         /*
1268          * Interlock the record using the BE flag.  Once BE is set the
1269          * frontend cannot change the state of FE.
1270          *
1271          * NOTE: If FE is set prior to us setting BE we still sync the
1272          * record out, but the flush completion code converts it to
1273          * a delete-on-disk record instead of destroying it.
1274          */
1275         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1276         record->flags |= HAMMER_RECF_INTERLOCK_BE;
1277
1278         /*
1279          * If the whole inode is being deleting all on-disk records will
1280          * be deleted very soon, we can't sync any new records to disk
1281          * because they will be deleted in the same transaction they were
1282          * created in (delete_tid == create_tid), which will assert.
1283          *
1284          * XXX There may be a case with RECORD_ADD with DELETED_FE set
1285          * that we currently panic on.
1286          */
1287         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
1288                 switch(record->type) {
1289                 case HAMMER_MEM_RECORD_GENERAL:
1290                         record->flags |= HAMMER_RECF_DELETED_FE;
1291                         record->flags |= HAMMER_RECF_DELETED_BE;
1292                         error = 0;
1293                         goto done;
1294                 case HAMMER_MEM_RECORD_ADD:
1295                         panic("hammer_sync_record_callback: illegal add "
1296                               "during inode deletion record %p", record);
1297                         break; /* NOT REACHED */
1298                 case HAMMER_MEM_RECORD_INODE:
1299                         panic("hammer_sync_record_callback: attempt to "
1300                               "sync inode record %p?", record);
1301                         break; /* NOT REACHED */
1302                 case HAMMER_MEM_RECORD_DEL:
1303                         /*
1304                          * Follow through and issue the on-disk deletion
1305                          */
1306                         break;
1307                 }
1308         }
1309
1310         /*
1311          * If DELETED_FE is set we may have already sent dependant pieces
1312          * to the disk and we must flush the record as if it hadn't been
1313          * deleted.  This creates a bit of a mess because we have to
1314          * have ip_sync_record convert the record to MEM_RECORD_DEL before
1315          * it inserts the B-Tree record.  Otherwise the media sync might
1316          * be visible to the frontend.
1317          */
1318         if (record->flags & HAMMER_RECF_DELETED_FE) {
1319                 if (record->type == HAMMER_MEM_RECORD_ADD) {
1320                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
1321                 } else {
1322                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
1323                         return(0);
1324                 }
1325         }
1326
1327         /*
1328          * Assign the create_tid for new records.  Deletions already
1329          * have the record's entire key properly set up.
1330          */
1331         if (record->type != HAMMER_MEM_RECORD_DEL)
1332                 record->leaf.base.create_tid = trans->tid;
1333         for (;;) {
1334                 error = hammer_ip_sync_record_cursor(cursor, record);
1335                 if (error != EDEADLK)
1336                         break;
1337                 hammer_done_cursor(cursor);
1338                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
1339                                            record->ip);
1340                 if (error)
1341                         break;
1342         }
1343         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1344
1345         if (error) {
1346                 error = -error;
1347                 if (error != -ENOSPC) {
1348                         kprintf("hammer_sync_record_callback: sync failed rec "
1349                                 "%p, error %d\n", record, error);
1350                         Debugger("sync failed rec");
1351                 }
1352         }
1353 done:
1354         hammer_flush_record_done(record, error);
1355         return(error);
1356 }
1357
1358 /*
1359  * XXX error handling
1360  */
1361 int
1362 hammer_sync_inode(hammer_inode_t ip)
1363 {
1364         struct hammer_transaction trans;
1365         struct hammer_cursor cursor;
1366         struct bio *bio;
1367         hammer_record_t depend;
1368         hammer_record_t next;
1369         int error, tmp_error;
1370         u_int64_t nlinks;
1371
1372         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
1373                 return(0);
1374
1375         hammer_start_transaction_fls(&trans, ip->hmp);
1376         error = hammer_init_cursor(&trans, &cursor, &ip->cache[0], ip);
1377         if (error)
1378                 goto done;
1379
1380         /*
1381          * Any directory records referencing this inode which are not in
1382          * our current flush group must adjust our nlink count for the
1383          * purposes of synchronization to disk.
1384          *
1385          * Records which are in our flush group can be unlinked from our
1386          * inode now, potentially allowing the inode to be physically
1387          * deleted.
1388          */
1389         nlinks = ip->ino_data.nlinks;
1390         next = TAILQ_FIRST(&ip->target_list);
1391         while ((depend = next) != NULL) {
1392                 next = TAILQ_NEXT(depend, target_entry);
1393                 if (depend->flush_state == HAMMER_FST_FLUSH &&
1394                     depend->flush_group == ip->hmp->flusher_act) {
1395                         /*
1396                          * If this is an ADD that was deleted by the frontend
1397                          * the frontend nlinks count will have already been
1398                          * decremented, but the backend is going to sync its
1399                          * directory entry and must account for it.  The
1400                          * record will be converted to a delete-on-disk when
1401                          * it gets synced.
1402                          *
1403                          * If the ADD was not deleted by the frontend we
1404                          * can remove the dependancy from our target_list.
1405                          */
1406                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
1407                                 ++nlinks;
1408                         } else {
1409                                 TAILQ_REMOVE(&ip->target_list, depend,
1410                                              target_entry);
1411                                 depend->target_ip = NULL;
1412                         }
1413                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
1414                         /*
1415                          * Not part of our flush group
1416                          */
1417                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
1418                         switch(depend->type) {
1419                         case HAMMER_MEM_RECORD_ADD:
1420                                 --nlinks;
1421                                 break;
1422                         case HAMMER_MEM_RECORD_DEL:
1423                                 ++nlinks;
1424                                 break;
1425                         default:
1426                                 break;
1427                         }
1428                 }
1429         }
1430
1431         /*
1432          * Set dirty if we had to modify the link count.
1433          */
1434         if (ip->sync_ino_data.nlinks != nlinks) {
1435                 KKASSERT((int64_t)nlinks >= 0);
1436                 ip->sync_ino_data.nlinks = nlinks;
1437                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1438         }
1439
1440         /*
1441          * Queue up as many dirty buffers as we can then set a flag to
1442          * cause any further BIOs to go to the alternative queue.
1443          */
1444         if (ip->flags & HAMMER_INODE_VHELD)
1445                 error = vfsync(ip->vp, MNT_NOWAIT, 1, NULL, NULL);
1446         ip->flags |= HAMMER_INODE_WRITE_ALT;
1447
1448         /*
1449          * The buffer cache may contain dirty buffers beyond the inode
1450          * state we copied from the frontend to the backend.  Because
1451          * we are syncing our buffer cache on the backend, resync
1452          * the truncation point and the file size so we don't wipe out
1453          * any data.
1454          *
1455          * Syncing the buffer cache on the frontend has serious problems
1456          * because it prevents us from passively queueing dirty inodes
1457          * to the backend (the BIO's could stall indefinitely).
1458          */
1459         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1460                 ip->sync_trunc_off = ip->trunc_off;
1461                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1462         }
1463         if (ip->sync_ino_data.size != ip->ino_data.size) {
1464                 ip->sync_ino_data.size = ip->ino_data.size;
1465                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1466         }
1467
1468         /*
1469          * If there is a trunction queued destroy any data past the (aligned)
1470          * truncation point.  Userland will have dealt with the buffer
1471          * containing the truncation point for us.
1472          *
1473          * We don't flush pending frontend data buffers until after we've
1474          * dealth with the truncation.
1475          *
1476          * Don't bother if the inode is or has been deleted.
1477          */
1478         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
1479                 /*
1480                  * Interlock trunc_off.  The VOP front-end may continue to
1481                  * make adjustments to it while we are blocked.
1482                  */
1483                 off_t trunc_off;
1484                 off_t aligned_trunc_off;
1485
1486                 trunc_off = ip->sync_trunc_off;
1487                 aligned_trunc_off = (trunc_off + HAMMER_BUFMASK) &
1488                                     ~HAMMER_BUFMASK64;
1489
1490                 /*
1491                  * Delete any whole blocks on-media.  The front-end has
1492                  * already cleaned out any partial block and made it
1493                  * pending.  The front-end may have updated trunc_off
1494                  * while we were blocked so do not just unconditionally
1495                  * set it to the maximum offset.
1496                  */
1497                 error = hammer_ip_delete_range(&cursor, ip,
1498                                                 aligned_trunc_off,
1499                                                 0x7FFFFFFFFFFFFFFFLL);
1500                 if (error)
1501                         Debugger("hammer_ip_delete_range errored");
1502                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1503                 if (ip->trunc_off >= trunc_off) {
1504                         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1505                         ip->flags &= ~HAMMER_INODE_TRUNCATED;
1506                 }
1507         } else {
1508                 error = 0;
1509         }
1510
1511         /*
1512          * Now sync related records.  These will typically be directory
1513          * entries or delete-on-disk records.
1514          *
1515          * Not all records will be flushed, but clear XDIRTY anyway.  We
1516          * will set it again in the frontend hammer_flush_inode_done()
1517          * if records remain.
1518          */
1519         if (error == 0) {
1520                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1521                                     hammer_sync_record_callback, &cursor);
1522                 if (tmp_error < 0)
1523                         tmp_error = -error;
1524                 if (tmp_error)
1525                         error = tmp_error;
1526                 if (RB_EMPTY(&ip->rec_tree))
1527                         ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
1528         }
1529
1530         /*
1531          * If we are deleting the inode the frontend had better not have
1532          * any active references on elements making up the inode.
1533          */
1534         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
1535                 RB_EMPTY(&ip->rec_tree)  &&
1536             (ip->sync_flags & HAMMER_INODE_DELETING) &&
1537             (ip->flags & HAMMER_INODE_DELETED) == 0) {
1538                 int count1 = 0;
1539
1540                 hkprintf("Y");
1541                 ip->flags |= HAMMER_INODE_DELETED;
1542                 error = hammer_ip_delete_range_all(&cursor, ip, &count1);
1543                 if (error == 0) {
1544                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
1545                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1546                         KKASSERT(RB_EMPTY(&ip->rec_tree));
1547
1548                         /*
1549                          * Set delete_tid in both the frontend and backend
1550                          * copy of the inode record.  The DELETED flag handles
1551                          * this, do not set RDIRTY.
1552                          */
1553                         ip->ino_leaf.base.delete_tid = trans.tid;
1554                         ip->sync_ino_leaf.base.delete_tid = trans.tid;
1555
1556                         /*
1557                          * Adjust the inode count in the volume header
1558                          */
1559                         hammer_modify_volume_field(&trans, trans.rootvol,
1560                                                    vol0_stat_inodes);
1561                         --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1562                         hammer_modify_volume_done(trans.rootvol);
1563                 } else {
1564                         ip->flags &= ~HAMMER_INODE_DELETED;
1565                         Debugger("hammer_ip_delete_range_all errored");
1566                 }
1567         }
1568
1569         /*
1570          * Flush any queued BIOs.  These will just biodone() the IO's if
1571          * the inode has been deleted.
1572          */
1573         while ((bio = TAILQ_FIRST(&ip->bio_list)) != NULL) {
1574                 TAILQ_REMOVE(&ip->bio_list, bio, bio_act);
1575                 tmp_error = hammer_dowrite(&cursor, ip, bio);
1576                 if (tmp_error)
1577                         error = tmp_error;
1578         }
1579         ip->sync_flags &= ~HAMMER_INODE_BUFS;
1580
1581         if (error)
1582                 Debugger("RB_SCAN errored");
1583
1584         /*
1585          * Now update the inode's on-disk inode-data and/or on-disk record.
1586          * DELETED and ONDISK are managed only in ip->flags.
1587          */
1588         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
1589         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
1590                 /*
1591                  * If deleted and on-disk, don't set any additional flags.
1592                  * the delete flag takes care of things.
1593                  *
1594                  * Clear flags which may have been set by the frontend.
1595                  */
1596                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY|
1597                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1598                                     HAMMER_INODE_DELETING);
1599                 break;
1600         case HAMMER_INODE_DELETED:
1601                 /*
1602                  * Take care of the case where a deleted inode was never
1603                  * flushed to the disk in the first place.
1604                  *
1605                  * Clear flags which may have been set by the frontend.
1606                  */
1607                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY|
1608                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1609                                     HAMMER_INODE_DELETING);
1610                 while (RB_ROOT(&ip->rec_tree)) {
1611                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
1612                         hammer_ref(&record->lock);
1613                         KKASSERT(record->lock.refs == 1);
1614                         record->flags |= HAMMER_RECF_DELETED_FE;
1615                         record->flags |= HAMMER_RECF_DELETED_BE;
1616                         hammer_rel_mem_record(record);
1617                 }
1618                 break;
1619         case HAMMER_INODE_ONDISK:
1620                 /*
1621                  * If already on-disk, do not set any additional flags.
1622                  */
1623                 break;
1624         default:
1625                 /*
1626                  * If not on-disk and not deleted, set both dirty flags
1627                  * to force an initial record to be written.  Also set
1628                  * the create_tid for the inode.
1629                  *
1630                  * Set create_tid in both the frontend and backend
1631                  * copy of the inode record.
1632                  */
1633                 ip->ino_leaf.base.create_tid = trans.tid;
1634                 ip->sync_ino_leaf.base.create_tid = trans.tid;
1635                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1636                 break;
1637         }
1638
1639         /*
1640          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
1641          * is already on-disk the old record is marked as deleted.
1642          *
1643          * If DELETED is set hammer_update_inode() will delete the existing
1644          * record without writing out a new one.
1645          *
1646          * If *ONLY* the ITIMES flag is set we can update the record in-place.
1647          */
1648         if (ip->flags & HAMMER_INODE_DELETED) {
1649                 error = hammer_update_inode(&cursor, ip);
1650         } else
1651         if ((ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES)) ==
1652             HAMMER_INODE_ITIMES) {
1653                 error = hammer_update_itimes(&cursor, ip);
1654         } else
1655         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ITIMES)) {
1656                 error = hammer_update_inode(&cursor, ip);
1657         }
1658         if (error)
1659                 Debugger("hammer_update_itimes/inode errored");
1660 done:
1661         /*
1662          * Save the TID we used to sync the inode with to make sure we
1663          * do not improperly reuse it.
1664          */
1665         hammer_done_cursor(&cursor);
1666         hammer_done_transaction(&trans);
1667         return(error);
1668 }
1669
1670 /*
1671  * This routine is called when the OS is no longer actively referencing
1672  * the inode (but might still be keeping it cached), or when releasing
1673  * the last reference to an inode.
1674  *
1675  * At this point if the inode's nlinks count is zero we want to destroy
1676  * it, which may mean destroying it on-media too.
1677  */
1678 void
1679 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
1680 {
1681         struct vnode *vp;
1682
1683         /*
1684          * Set the DELETING flag when the link count drops to 0 and the
1685          * OS no longer has any opens on the inode.
1686          *
1687          * The backend will clear DELETING (a mod flag) and set DELETED
1688          * (a state flag) when it is actually able to perform the
1689          * operation.
1690          */
1691         if (ip->ino_data.nlinks == 0 &&
1692             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
1693                 ip->flags |= HAMMER_INODE_DELETING;
1694                 ip->flags |= HAMMER_INODE_TRUNCATED;
1695                 ip->trunc_off = 0;
1696                 vp = NULL;
1697                 if (getvp) {
1698                         if (hammer_get_vnode(ip, &vp) != 0)
1699                                 return;
1700                 }
1701                 if (ip->vp) {
1702                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
1703                         vnode_pager_setsize(ip->vp, 0);
1704                 }
1705                 if (getvp) {
1706                         vput(vp);
1707                 }
1708         }
1709 }
1710
1711 /*
1712  * Re-test an inode when a dependancy had gone away to see if we
1713  * can chain flush it.
1714  */
1715 void
1716 hammer_test_inode(hammer_inode_t ip)
1717 {
1718         if (ip->flags & HAMMER_INODE_REFLUSH) {
1719                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1720                 hammer_ref(&ip->lock);
1721                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1722                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1723                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1724                 } else {
1725                         hammer_flush_inode(ip, 0);
1726                 }
1727                 hammer_rel_inode(ip, 0);
1728         }
1729 }
1730