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 1.54 2008/05/09 07:26:51 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38 #include <vm/vm_extern.h>
  39 #include <sys/buf.h>
  40 #include <sys/buf2.h>
  41
  42 static int hammer_unload_inode(struct hammer_inode *ip);
  43 static void hammer_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_rec.ino_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 = hammer_get_vnode_type(
 131                                             ip->ino_rec.base.base.obj_type);
 132
 133                         switch(ip->ino_rec.base.base.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_rec.ino_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_RECORD | 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_rec = cursor.record->inode;
 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_RDIRTY |
 326                     HAMMER_INODE_ITIMES;
 327
 328         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 329         RB_INIT(&ip->rec_tree);
 330         TAILQ_INIT(&ip->bio_list);
 331         TAILQ_INIT(&ip->bio_alt_list);
 332         TAILQ_INIT(&ip->target_list);
 333
 334         ip->ino_rec.ino_atime = trans->time;
 335         ip->ino_rec.ino_mtime = trans->time;
 336         ip->ino_rec.ino_size = 0;
 337         ip->ino_rec.ino_nlinks = 0;
 338         /* XXX */
 339         ip->ino_rec.base.base.btype = HAMMER_BTREE_TYPE_RECORD;
 340         ip->ino_rec.base.base.obj_id = ip->obj_id;
 341         ip->ino_rec.base.base.key = 0;
 342         ip->ino_rec.base.base.create_tid = 0;
 343         ip->ino_rec.base.base.delete_tid = 0;
 344         ip->ino_rec.base.base.rec_type = HAMMER_RECTYPE_INODE;
 345         ip->ino_rec.base.base.obj_type = hammer_get_obj_type(vap->va_type);
 346
 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_rec.base.base.obj_id : 0;
 351
 352         switch(ip->ino_rec.base.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_RECORD | 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);
 470                 record->type = HAMMER_MEM_RECORD_GENERAL;
 471                 record->flush_state = HAMMER_FST_FLUSH;
 472                 record->rec.inode = ip->sync_ino_rec;
 473                 record->rec.inode.base.base.create_tid = trans->tid;
 474                 record->rec.inode.base.data_len = sizeof(ip->sync_ino_data);
 475                 record->rec.base.signature = HAMMER_RECORD_SIGNATURE_GOOD;
 476                 record->data = (void *)&ip->sync_ino_data;
 477                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
 478                 for (;;) {
 479                         error = hammer_ip_sync_record_cursor(cursor, record);
 480                         if (hammer_debug_inode)
 481                                 kprintf("GENREC %p rec %08x %d\n",
 482                                         ip, record->flags, error);
 483                         if (error != EDEADLK)
 484                                 break;
 485                         hammer_done_cursor(cursor);
 486                         error = hammer_init_cursor(trans, cursor,
 487                                                    &ip->cache[0], ip);
 488                         if (hammer_debug_inode)
 489                                 kprintf("GENREC reinit %d\n", error);
 490                         if (error)
 491                                 break;
 492                 }
 493                 if (error) {
 494                         kprintf("error %d\n", error);
 495                         Debugger("hammer_update_inode3");
 496                 }
 497
 498                 /*
 499                  * The record isn't managed by the inode's record tree,
 500                  * destroy it whether we succeed or fail.
 501                  */
 502                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 503                 record->flags |= HAMMER_RECF_DELETED_FE;
 504                 record->flush_state = HAMMER_FST_IDLE;
 505                 hammer_rel_mem_record(record);
 506
 507                 /*
 508                  * Finish up.
 509                  */
 510                 if (error == 0) {
 511                         if (hammer_debug_inode)
 512                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
 513                         ip->sync_flags &= ~(HAMMER_INODE_RDIRTY |
 514                                             HAMMER_INODE_DDIRTY |
 515                                             HAMMER_INODE_ITIMES);
 516                         ip->flags &= ~HAMMER_INODE_DELONDISK;
 517
 518                         /*
 519                          * Root volume count of inodes
 520                          */
 521                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
 522                                 hammer_modify_volume_field(trans,
 523                                                            trans->rootvol,
 524                                                            vol0_stat_inodes);
 525                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
 526                                 hammer_modify_volume_done(trans->rootvol);
 527                                 ip->flags |= HAMMER_INODE_ONDISK;
 528                                 if (hammer_debug_inode)
 529                                         kprintf("NOWONDISK %p\n", ip);
 530                         }
 531                 }
 532         }
 533
 534         /*
 535          * If the inode has been destroyed, clean out any left-over flags
 536          * that may have been set by the frontend.
 537          */
 538         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
 539                 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY |
 540                                     HAMMER_INODE_DDIRTY |
 541                                     HAMMER_INODE_ITIMES);
 542         }
 543         return(error);
 544 }
 545
 546 /*
 547  * Update only the itimes fields.  This is done no-historically.  The
 548  * record is updated in-place on the disk.
 549  */
 550 static int
 551 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
 552 {
 553         hammer_transaction_t trans = cursor->trans;
 554         struct hammer_inode_record *rec;
 555         int error;
 556
 557 retry:
 558         error = 0;
 559         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
 560             HAMMER_INODE_ONDISK) {
 561                 hammer_normalize_cursor(cursor);
 562                 cursor->key_beg.obj_id = ip->obj_id;
 563                 cursor->key_beg.key = 0;
 564                 cursor->key_beg.create_tid = 0;
 565                 cursor->key_beg.delete_tid = 0;
 566                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
 567                 cursor->key_beg.obj_type = 0;
 568                 cursor->asof = ip->obj_asof;
 569                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 570                 cursor->flags |= HAMMER_CURSOR_GET_RECORD | HAMMER_CURSOR_ASOF;
 571                 cursor->flags |= HAMMER_CURSOR_BACKEND;
 572
 573                 error = hammer_btree_lookup(cursor);
 574                 if (error) {
 575                         kprintf("error %d\n", error);
 576                         Debugger("hammer_update_itimes1");
 577                 }
 578                 if (error == 0) {
 579                         /*
 580                          * Do not generate UNDO records for atime/mtime
 581                          * updates.
 582                          */
 583                         rec = &cursor->record->inode;
 584                         hammer_modify_record_noundo(trans,
 585                                                     cursor->record_buffer,
 586                                                     cursor->record);
 587                         rec->ino_atime = ip->sync_ino_rec.ino_atime;
 588                         rec->ino_mtime = ip->sync_ino_rec.ino_mtime;
 589                         hammer_modify_record_done(cursor->record_buffer,
 590                                                   cursor->record);
 591                         ip->sync_flags &= ~HAMMER_INODE_ITIMES;
 592                         /* XXX recalculate crc */
 593                         hammer_cache_node(cursor->node, &ip->cache[0]);
 594                 }
 595                 if (error == EDEADLK) {
 596                         hammer_done_cursor(cursor);
 597                         error = hammer_init_cursor(trans, cursor,
 598                                                    &ip->cache[0], ip);
 599                         if (error == 0)
 600                                 goto retry;
 601                 }
 602         }
 603         return(error);
 604 }
 605
 606 /*
 607  * Release a reference on an inode, flush as requested.
 608  *
 609  * On the last reference we queue the inode to the flusher for its final
 610  * disposition.
 611  */
 612 void
 613 hammer_rel_inode(struct hammer_inode *ip, int flush)
 614 {
 615         hammer_mount_t hmp = ip->hmp;
 616
 617         /*
 618          * Handle disposition when dropping the last ref.
 619          */
 620         for (;;) {
 621                 if (ip->lock.refs == 1) {
 622                         /*
 623                          * Determine whether on-disk action is needed for
 624                          * the inode's final disposition.
 625                          */
 626                         KKASSERT(ip->vp == NULL);
 627                         hammer_inode_unloadable_check(ip, 0);
 628                         if (ip->flags & HAMMER_INODE_MODMASK) {
 629                                 hammer_flush_inode(ip, 0);
 630                         } else if (ip->lock.refs == 1) {
 631                                 hammer_unload_inode(ip);
 632                                 break;
 633                         }
 634                 } else {
 635                         if (flush)
 636                                 hammer_flush_inode(ip, 0);
 637
 638                         /*
 639                          * The inode still has multiple refs, try to drop
 640                          * one ref.
 641                          */
 642                         KKASSERT(ip->lock.refs >= 1);
 643                         if (ip->lock.refs > 1) {
 644                                 hammer_unref(&ip->lock);
 645                                 break;
 646                         }
 647                 }
 648         }
 649
 650         /*
 651          * XXX bad hack until I add code to track inodes in SETUP.  We
 652          * can queue a lot of inodes to the syncer but if we don't wake
 653          * it up the undo sets will be too large or too many unflushed
 654          * records will build up and blow our malloc limit.
 655          */
 656         if (++hmp->reclaim_count > 256) {
 657                 hmp->reclaim_count = 0;
 658                 hammer_flusher_async(hmp);
 659         }
 660 }
 661
 662 /*
 663  * Unload and destroy the specified inode.  Must be called with one remaining
 664  * reference.  The reference is disposed of.
 665  *
 666  * This can only be called in the context of the flusher.
 667  */
 668 static int
 669 hammer_unload_inode(struct hammer_inode *ip)
 670 {
 671         KASSERT(ip->lock.refs == 1,
 672                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
 673         KKASSERT(ip->vp == NULL);
 674         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
 675         KKASSERT(ip->cursor_ip_refs == 0);
 676         KKASSERT(ip->lock.lockcount == 0);
 677         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
 678
 679         KKASSERT(RB_EMPTY(&ip->rec_tree));
 680         KKASSERT(TAILQ_EMPTY(&ip->target_list));
 681         KKASSERT(TAILQ_EMPTY(&ip->bio_list));
 682         KKASSERT(TAILQ_EMPTY(&ip->bio_alt_list));
 683
 684         RB_REMOVE(hammer_ino_rb_tree, &ip->hmp->rb_inos_root, ip);
 685
 686         hammer_uncache_node(&ip->cache[0]);
 687         hammer_uncache_node(&ip->cache[1]);
 688         if (ip->objid_cache)
 689                 hammer_clear_objid(ip);
 690         --hammer_count_inodes;
 691         kfree(ip, M_HAMMER);
 692
 693         return(0);
 694 }
 695
 696 /*
 697  * A transaction has modified an inode, requiring updates as specified by
 698  * the passed flags.
 699  *
 700  * HAMMER_INODE_RDIRTY: Inode record has been updated
 701  * HAMMER_INODE_DDIRTY: Inode data has been updated
 702  * HAMMER_INODE_XDIRTY: Dirty in-memory records
 703  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
 704  * HAMMER_INODE_DELETED: Inode record/data must be deleted
 705  * HAMMER_INODE_ITIMES: mtime/atime has been updated
 706  */
 707 void
 708 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
 709 {
 710         KKASSERT ((ip->flags & HAMMER_INODE_RO) == 0 ||
 711                   (flags & (HAMMER_INODE_RDIRTY|HAMMER_INODE_DDIRTY|
 712                    HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|
 713                    HAMMER_INODE_DELETED|HAMMER_INODE_ITIMES)) == 0);
 714
 715         ip->flags |= flags;
 716 }
 717
 718 /*
 719  * Request that an inode be flushed.  This whole mess cannot block and may
 720  * recurse.  Once requested HAMMER will attempt to actively flush it until
 721  * the flush can be done.
 722  *
 723  * The inode may already be flushing, or may be in a setup state.  We can
 724  * place the inode in a flushing state if it is currently idle and flag it
 725  * to reflush if it is currently flushing.
 726  */
 727 void
 728 hammer_flush_inode(hammer_inode_t ip, int flags)
 729 {
 730         hammer_record_t depend;
 731         int r, good;
 732
 733         /*
 734          * Trivial 'nothing to flush' case.  If the inode is ina SETUP
 735          * state we have to put it back into an IDLE state so we can
 736          * drop the extra ref.
 737          */
 738         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
 739                 if (ip->flush_state == HAMMER_FST_SETUP) {
 740                         ip->flush_state = HAMMER_FST_IDLE;
 741                         hammer_rel_inode(ip, 0);
 742                 }
 743                 return;
 744         }
 745
 746         /*
 747          * Our flush action will depend on the current state.
 748          */
 749         switch(ip->flush_state) {
 750         case HAMMER_FST_IDLE:
 751                 /*
 752                  * We have no dependancies and can flush immediately.  Some
 753                  * our children may not be flushable so we have to re-test
 754                  * with that additional knowledge.
 755                  */
 756                 hammer_flush_inode_core(ip, flags);
 757                 break;
 758         case HAMMER_FST_SETUP:
 759                 /*
 760                  * Recurse upwards through dependancies via target_list
 761                  * and start their flusher actions going if possible.
 762                  *
 763                  * 'good' is our connectivity.  -1 means we have none and
 764                  * can't flush, 0 means there weren't any dependancies, and
 765                  * 1 means we have good connectivity.
 766                  */
 767                 good = 0;
 768                 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 769                         r = hammer_setup_parent_inodes(depend);
 770                         if (r < 0 && good == 0)
 771                                 good = -1;
 772                         if (r > 0)
 773                                 good = 1;
 774                 }
 775
 776                 /*
 777                  * We can continue if good >= 0.  Determine how many records
 778                  * under our inode can be flushed (and mark them).
 779                  */
 780                 if (good >= 0) {
 781                         hammer_flush_inode_core(ip, flags);
 782                 } else {
 783                         ip->flags |= HAMMER_INODE_REFLUSH;
 784                         if (flags & HAMMER_FLUSH_SIGNAL) {
 785                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 786                                 hammer_flusher_async(ip->hmp);
 787                         }
 788                 }
 789                 break;
 790         default:
 791                 /*
 792                  * We are already flushing, flag the inode to reflush
 793                  * if needed after it completes its current flush.
 794                  */
 795                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
 796                         ip->flags |= HAMMER_INODE_REFLUSH;
 797                 if (flags & HAMMER_FLUSH_SIGNAL) {
 798                         ip->flags |= HAMMER_INODE_RESIGNAL;
 799                         hammer_flusher_async(ip->hmp);
 800                 }
 801                 break;
 802         }
 803 }
 804
 805 /*
 806  * We are asked to recurse upwards and convert the record from SETUP
 807  * to FLUSH if possible.  record->ip is a parent of the caller's inode,
 808  * and record->target_ip is the caller's inode.
 809  *
 810  * Return 1 if the record gives us connectivity
 811  *
 812  * Return 0 if the record is not relevant
 813  *
 814  * Return -1 if we can't resolve the dependancy and there is no connectivity.
 815  */
 816 static int
 817 hammer_setup_parent_inodes(hammer_record_t record)
 818 {
 819         hammer_mount_t hmp = record->ip->hmp;
 820         hammer_record_t depend;
 821         hammer_inode_t ip;
 822         int r, good;
 823
 824         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
 825         ip = record->ip;
 826
 827         /*
 828          * If the record is already flushing, is it in our flush group?
 829          *
 830          * If it is in our flush group but it is a general record or a
 831          * delete-on-disk, it does not improve our connectivity (return 0),
 832          * and if the target inode is not trying to destroy itself we can't
 833          * allow the operation yet anyway (the second return -1).
 834          */
 835         if (record->flush_state == HAMMER_FST_FLUSH) {
 836                 if (record->flush_group != hmp->flusher_next) {
 837                         ip->flags |= HAMMER_INODE_REFLUSH;
 838                         return(-1);
 839                 }
 840                 if (record->type == HAMMER_MEM_RECORD_ADD)
 841                         return(1);
 842                 /* GENERAL or DEL */
 843                 return(0);
 844         }
 845
 846         /*
 847          * It must be a setup record.  Try to resolve the setup dependancies
 848          * by recursing upwards so we can place ip on the flush list.
 849          */
 850         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
 851
 852         good = 0;
 853         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
 854                 r = hammer_setup_parent_inodes(depend);
 855                 if (r < 0 && good == 0)
 856                         good = -1;
 857                 if (r > 0)
 858                         good = 1;
 859         }
 860
 861         /*
 862          * We can't flush ip because it has no connectivity (XXX also check
 863          * nlinks for pre-existing connectivity!).  Flag it so any resolution
 864          * recurses back down.
 865          */
 866         if (good < 0) {
 867                 ip->flags |= HAMMER_INODE_REFLUSH;
 868                 return(good);
 869         }
 870
 871         /*
 872          * We are go, place the parent inode in a flushing state so we can
 873          * place its record in a flushing state.  Note that the parent
 874          * may already be flushing.  The record must be in the same flush
 875          * group as the parent.
 876          */
 877         if (ip->flush_state != HAMMER_FST_FLUSH)
 878                 hammer_flush_inode_core(ip, HAMMER_FLUSH_RECURSION);
 879         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
 880         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
 881
 882 #if 0
 883         if (record->type == HAMMER_MEM_RECORD_DEL &&
 884             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
 885                 /*
 886                  * Regardless of flushing state we cannot sync this path if the
 887                  * record represents a delete-on-disk but the target inode
 888                  * is not ready to sync its own deletion.
 889                  *
 890                  * XXX need to count effective nlinks to determine whether
 891                  * the flush is ok, otherwise removing a hardlink will
 892                  * just leave the DEL record to rot.
 893                  */
 894                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
 895                 return(-1);
 896         } else
 897 #endif
 898         if (ip->flush_group == ip->hmp->flusher_next) {
 899                 /*
 900                  * This is the record we wanted to synchronize.
 901                  */
 902                 record->flush_state = HAMMER_FST_FLUSH;
 903                 record->flush_group = ip->flush_group;
 904                 hammer_ref(&record->lock);
 905                 if (record->type == HAMMER_MEM_RECORD_ADD)
 906                         return(1);
 907
 908                 /*
 909                  * A general or delete-on-disk record does not contribute
 910                  * to our visibility.  We can still flush it, however.
 911                  */
 912                 return(0);
 913         } else {
 914                 /*
 915                  * We couldn't resolve the dependancies, request that the
 916                  * inode be flushed when the dependancies can be resolved.
 917                  */
 918                 ip->flags |= HAMMER_INODE_REFLUSH;
 919                 return(-1);
 920         }
 921 }
 922
 923 /*
 924  * This is the core routine placing an inode into the FST_FLUSH state.
 925  */
 926 static void
 927 hammer_flush_inode_core(hammer_inode_t ip, int flags)
 928 {
 929         int go_count;
 930
 931         /*
 932          * Set flush state and prevent the flusher from cycling into
 933          * the next flush group.  Do not place the ip on the list yet.
 934          * Inodes not in the idle state get an extra reference.
 935          */
 936         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
 937         if (ip->flush_state == HAMMER_FST_IDLE)
 938                 hammer_ref(&ip->lock);
 939         ip->flush_state = HAMMER_FST_FLUSH;
 940         ip->flush_group = ip->hmp->flusher_next;
 941         ++ip->hmp->flusher_lock;
 942
 943         /*
 944          * We need to be able to vfsync/truncate from the backend.
 945          */
 946         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
 947         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
 948                 ip->flags |= HAMMER_INODE_VHELD;
 949                 vref(ip->vp);
 950         }
 951
 952         /*
 953          * Figure out how many in-memory records we can actually flush
 954          * (not including inode meta-data, buffers, etc).
 955          */
 956         if (flags & HAMMER_FLUSH_RECURSION) {
 957                 go_count = 1;
 958         } else {
 959                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
 960                                    hammer_setup_child_callback, NULL);
 961         }
 962
 963         /*
 964          * This is a more involved test that includes go_count.  If we
 965          * can't flush, flag the inode and return.  If go_count is 0 we
 966          * were are unable to flush any records in our rec_tree and
 967          * must ignore the XDIRTY flag.
 968          */
 969         if (go_count == 0) {
 970                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
 971                         ip->flags |= HAMMER_INODE_REFLUSH;
 972                         ip->flush_state = HAMMER_FST_SETUP;
 973                         if (ip->flags & HAMMER_INODE_VHELD) {
 974                                 ip->flags &= ~HAMMER_INODE_VHELD;
 975                                 vrele(ip->vp);
 976                         }
 977                         if (flags & HAMMER_FLUSH_SIGNAL) {
 978                                 ip->flags |= HAMMER_INODE_RESIGNAL;
 979                                 hammer_flusher_async(ip->hmp);
 980                         }
 981                         if (--ip->hmp->flusher_lock == 0)
 982                                 wakeup(&ip->hmp->flusher_lock);
 983                         return;
 984                 }
 985         }
 986
 987         /*
 988          * Snapshot the state of the inode for the backend flusher.
 989          *
 990          * The truncation must be retained in the frontend until after
 991          * we've actually performed the record deletion.
 992          *
 993          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
 994          * and stays in ip->flags.  Once set, it stays set until the
 995          * inode is destroyed.
 996          */
 997         ip->sync_flags = (ip->flags & HAMMER_INODE_MODMASK);
 998         ip->sync_trunc_off = ip->trunc_off;
 999         ip->sync_ino_rec = ip->ino_rec;
1000         ip->sync_ino_data = ip->ino_data;
1001         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1002
1003         /*
1004          * The flusher list inherits our inode and reference.
1005          */
1006         TAILQ_INSERT_TAIL(&ip->hmp->flush_list, ip, flush_entry);
1007         if (--ip->hmp->flusher_lock == 0)
1008                 wakeup(&ip->hmp->flusher_lock);
1009
1010         if (flags & HAMMER_FLUSH_SIGNAL)
1011                 hammer_flusher_async(ip->hmp);
1012 }
1013
1014 /*
1015  * Callback for scan of ip->rec_tree.  Try to include each record in our
1016  * flush.  ip->flush_group has been set but the inode has not yet been
1017  * moved into a flushing state.
1018  *
1019  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1020  * both inodes.
1021  *
1022  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1023  * the caller from shortcutting the flush.
1024  */
1025 static int
1026 hammer_setup_child_callback(hammer_record_t rec, void *data)
1027 {
1028         hammer_inode_t target_ip;
1029         hammer_inode_t ip;
1030         int r;
1031
1032         /*
1033          * If the record has been deleted by the backend (it's being held
1034          * by the frontend in a race), just ignore it.
1035          */
1036         if (rec->flags & HAMMER_RECF_DELETED_BE)
1037                 return(0);
1038
1039         /*
1040          * If the record is in an idle state it has no dependancies and
1041          * can be flushed.
1042          */
1043         ip = rec->ip;
1044         r = 0;
1045
1046         switch(rec->flush_state) {
1047         case HAMMER_FST_IDLE:
1048                 /*
1049                  * Record has no setup dependancy, we can flush it.
1050                  */
1051                 KKASSERT(rec->target_ip == NULL);
1052                 rec->flush_state = HAMMER_FST_FLUSH;
1053                 rec->flush_group = ip->flush_group;
1054                 hammer_ref(&rec->lock);
1055                 r = 1;
1056                 break;
1057         case HAMMER_FST_SETUP:
1058                 /*
1059                  * Record has a setup dependancy.  Try to include the
1060                  * target ip in the flush.
1061                  *
1062                  * We have to be careful here, if we do not do the right
1063                  * thing we can lose track of dirty inodes and the system
1064                  * will lockup trying to allocate buffers.
1065                  */
1066                 target_ip = rec->target_ip;
1067                 KKASSERT(target_ip != NULL);
1068                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1069                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1070                         /*
1071                          * If the target IP is already flushing in our group
1072                          * we are golden, otherwise make sure the target
1073                          * reflushes.
1074                          */
1075                         if (target_ip->flush_group == ip->flush_group) {
1076                                 rec->flush_state = HAMMER_FST_FLUSH;
1077                                 rec->flush_group = ip->flush_group;
1078                                 hammer_ref(&rec->lock);
1079                                 r = 1;
1080                         } else {
1081                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
1082                         }
1083                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1084                         /*
1085                          * If the target IP is not flushing we can force
1086                          * it to flush, even if it is unable to write out
1087                          * any of its own records we have at least one in
1088                          * hand that we CAN deal with.
1089                          */
1090                         rec->flush_state = HAMMER_FST_FLUSH;
1091                         rec->flush_group = ip->flush_group;
1092                         hammer_ref(&rec->lock);
1093                         hammer_flush_inode_core(target_ip,
1094                                                 HAMMER_FLUSH_RECURSION);
1095                         r = 1;
1096                 } else {
1097                         /*
1098                          * General or delete-on-disk record.
1099                          *
1100                          * XXX this needs help.  If a delete-on-disk we could
1101                          * disconnect the target.  If the target has its own
1102                          * dependancies they really need to be flushed.
1103                          *
1104                          * XXX
1105                          */
1106                         rec->flush_state = HAMMER_FST_FLUSH;
1107                         rec->flush_group = ip->flush_group;
1108                         hammer_ref(&rec->lock);
1109                         hammer_flush_inode_core(target_ip,
1110                                                 HAMMER_FLUSH_RECURSION);
1111                         r = 1;
1112                 }
1113                 break;
1114         case HAMMER_FST_FLUSH:
1115                 /*
1116                  * Record already associated with a flush group.  It had
1117                  * better be ours.
1118                  */
1119                 KKASSERT(rec->flush_group == ip->flush_group);
1120                 r = 1;
1121                 break;
1122         }
1123         return(r);
1124 }
1125
1126 /*
1127  * Wait for a previously queued flush to complete
1128  */
1129 void
1130 hammer_wait_inode(hammer_inode_t ip)
1131 {
1132         while (ip->flush_state != HAMMER_FST_IDLE) {
1133                 ip->flags |= HAMMER_INODE_FLUSHW;
1134                 tsleep(&ip->flags, 0, "hmrwin", 0);
1135         }
1136 }
1137
1138 /*
1139  * Called by the backend code when a flush has been completed.
1140  * The inode has already been removed from the flush list.
1141  *
1142  * A pipelined flush can occur, in which case we must re-enter the
1143  * inode on the list and re-copy its fields.
1144  */
1145 void
1146 hammer_flush_inode_done(hammer_inode_t ip)
1147 {
1148         struct bio *bio;
1149         int dorel = 0;
1150
1151         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1152
1153         /*
1154          * Allow BIOs to queue to the inode's primary bioq again.
1155          */
1156         ip->flags &= ~HAMMER_INODE_WRITE_ALT;
1157
1158         /*
1159          * Merge left-over flags back into the frontend and fix the state.
1160          */
1161         ip->flags |= ip->sync_flags;
1162
1163         /*
1164          * The backend may have adjusted nlinks, so if the adjusted nlinks
1165          * does not match the fronttend set the frontend's RDIRTY flag again.
1166          */
1167         if (ip->ino_rec.ino_nlinks != ip->sync_ino_rec.ino_nlinks)
1168                 ip->flags |= HAMMER_INODE_RDIRTY;
1169
1170         /*
1171          * Reflush any BIOs that wound up in the alt list.  Our inode will
1172          * also wind up at the end of the flusher's list.
1173          */
1174         while ((bio = TAILQ_FIRST(&ip->bio_alt_list)) != NULL) {
1175                 TAILQ_REMOVE(&ip->bio_alt_list, bio, bio_act);
1176                 TAILQ_INSERT_TAIL(&ip->bio_list, bio, bio_act);
1177         }
1178         /*
1179          * Fix up the dirty buffer status.
1180          */
1181         if (TAILQ_FIRST(&ip->bio_list) ||
1182             (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree))) {
1183                 ip->flags |= HAMMER_INODE_BUFS;
1184         }
1185
1186         /*
1187          * Re-set the XDIRTY flag if some of the inode's in-memory records
1188          * could not be flushed.
1189          */
1190         if (RB_ROOT(&ip->rec_tree))
1191                 ip->flags |= HAMMER_INODE_XDIRTY;
1192
1193         /*
1194          * Do not lose track of inodes which no longer have vnode
1195          * assocations, otherwise they may never get flushed again.
1196          */
1197         if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1198                 ip->flags |= HAMMER_INODE_REFLUSH;
1199
1200         /*
1201          * Adjust flush_state.  The target state (idle or setup) shouldn't
1202          * be terribly important since we will reflush if we really need
1203          * to do anything. XXX
1204          */
1205         if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
1206                 ip->flush_state = HAMMER_FST_IDLE;
1207                 dorel = 1;
1208         } else {
1209                 ip->flush_state = HAMMER_FST_SETUP;
1210         }
1211
1212         /*
1213          * Clean up the vnode ref
1214          */
1215         if (ip->flags & HAMMER_INODE_VHELD) {
1216                 ip->flags &= ~HAMMER_INODE_VHELD;
1217                 vrele(ip->vp);
1218         }
1219
1220         /*
1221          * If the frontend made more changes and requested another flush,
1222          * then try to get it running.
1223          */
1224         if (ip->flags & HAMMER_INODE_REFLUSH) {
1225                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1226                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1227                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1228                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1229                 } else {
1230                         hammer_flush_inode(ip, 0);
1231                 }
1232         }
1233
1234         /*
1235          * Finally, if the frontend is waiting for a flush to complete,
1236          * wake it up.
1237          */
1238         if (ip->flush_state != HAMMER_FST_FLUSH) {
1239                 if (ip->flags & HAMMER_INODE_FLUSHW) {
1240                         ip->flags &= ~HAMMER_INODE_FLUSHW;
1241                         wakeup(&ip->flags);
1242                 }
1243         }
1244         if (dorel)
1245                 hammer_rel_inode(ip, 0);
1246 }
1247
1248 /*
1249  * Called from hammer_sync_inode() to synchronize in-memory records
1250  * to the media.
1251  */
1252 static int
1253 hammer_sync_record_callback(hammer_record_t record, void *data)
1254 {
1255         hammer_cursor_t cursor = data;
1256         hammer_transaction_t trans = cursor->trans;
1257         int error;
1258
1259         /*
1260          * Skip records that do not belong to the current flush.
1261          */
1262         if (record->flush_state != HAMMER_FST_FLUSH)
1263                 return(0);
1264         KKASSERT((record->flags & HAMMER_RECF_DELETED_BE) == 0);
1265 #if 1
1266         if (record->flush_group != record->ip->flush_group) {
1267                 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1268                 Debugger("blah2");
1269                 return(0);
1270         }
1271 #endif
1272         KKASSERT(record->flush_group == record->ip->flush_group);
1273
1274         /*
1275          * Interlock the record using the BE flag.  Once BE is set the
1276          * frontend cannot change the state of FE.
1277          *
1278          * NOTE: If FE is set prior to us setting BE we still sync the
1279          * record out, but the flush completion code converts it to
1280          * a delete-on-disk record instead of destroying it.
1281          */
1282         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1283         record->flags |= HAMMER_RECF_INTERLOCK_BE;
1284
1285         /*
1286          * If DELETED_FE is set we may have already sent dependant pieces
1287          * to the disk and we must flush the record as if it hadn't been
1288          * deleted.  This creates a bit of a mess because we have to
1289          * have ip_sync_record convert the record to MEM_RECORD_DEL before
1290          * it inserts the B-Tree record.  Otherwise the media sync might
1291          * be visible to the frontend.
1292          */
1293         if (record->flags & HAMMER_RECF_DELETED_FE) {
1294                 if (record->type == HAMMER_MEM_RECORD_ADD) {
1295                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
1296                 } else {
1297                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
1298                         return(0);
1299                 }
1300         }
1301
1302         /*
1303          * Assign the create_tid for new records.  Deletions already
1304          * have the record's entire key properly set up.
1305          */
1306         if (record->type != HAMMER_MEM_RECORD_DEL)
1307                 record->rec.inode.base.base.create_tid = trans->tid;
1308         for (;;) {
1309                 error = hammer_ip_sync_record_cursor(cursor, record);
1310                 if (error != EDEADLK)
1311                         break;
1312                 hammer_done_cursor(cursor);
1313                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
1314                                            record->ip);
1315                 if (error)
1316                         break;
1317         }
1318         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1319
1320         if (error) {
1321                 error = -error;
1322                 if (error != -ENOSPC) {
1323                         kprintf("hammer_sync_record_callback: sync failed rec "
1324                                 "%p, error %d\n", record, error);
1325                         Debugger("sync failed rec");
1326                 }
1327         }
1328         hammer_flush_record_done(record, error);
1329         return(error);
1330 }
1331
1332 /*
1333  * XXX error handling
1334  */
1335 int
1336 hammer_sync_inode(hammer_inode_t ip)
1337 {
1338         struct hammer_transaction trans;
1339         struct hammer_cursor cursor;
1340         struct bio *bio;
1341         hammer_record_t depend;
1342         hammer_record_t next;
1343         int error, tmp_error;
1344         u_int64_t nlinks;
1345
1346         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
1347                 return(0);
1348
1349         hammer_start_transaction_fls(&trans, ip->hmp);
1350         error = hammer_init_cursor(&trans, &cursor, &ip->cache[0], ip);
1351         if (error)
1352                 goto done;
1353
1354         /*
1355          * Any directory records referencing this inode which are not in
1356          * our current flush group must adjust our nlink count for the
1357          * purposes of synchronization to disk.
1358          *
1359          * Records which are in our flush group can be unlinked from our
1360          * inode now, potentially allowing the inode to be physically
1361          * deleted.
1362          */
1363         nlinks = ip->ino_rec.ino_nlinks;
1364         next = TAILQ_FIRST(&ip->target_list);
1365         while ((depend = next) != NULL) {
1366                 next = TAILQ_NEXT(depend, target_entry);
1367                 if (depend->flush_state == HAMMER_FST_FLUSH &&
1368                     depend->flush_group == ip->hmp->flusher_act) {
1369                         /*
1370                          * If this is an ADD that was deleted by the frontend
1371                          * the frontend nlinks count will have already been
1372                          * decremented, but the backend is going to sync its
1373                          * directory entry and must account for it.  The
1374                          * record will be converted to a delete-on-disk when
1375                          * it gets synced.
1376                          *
1377                          * If the ADD was not deleted by the frontend we
1378                          * can remove the dependancy from our target_list.
1379                          */
1380                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
1381                                 ++nlinks;
1382                         } else {
1383                                 TAILQ_REMOVE(&ip->target_list, depend,
1384                                              target_entry);
1385                                 depend->target_ip = NULL;
1386                         }
1387                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
1388                         /*
1389                          * Not part of our flush group
1390                          */
1391                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
1392                         switch(depend->type) {
1393                         case HAMMER_MEM_RECORD_ADD:
1394                                 --nlinks;
1395                                 break;
1396                         case HAMMER_MEM_RECORD_DEL:
1397                                 ++nlinks;
1398                                 break;
1399                         default:
1400                                 break;
1401                         }
1402                 }
1403         }
1404
1405         /*
1406          * Set dirty if we had to modify the link count.
1407          */
1408         if (ip->sync_ino_rec.ino_nlinks != nlinks) {
1409                 KKASSERT((int64_t)nlinks >= 0);
1410                 ip->sync_ino_rec.ino_nlinks = nlinks;
1411                 ip->sync_flags |= HAMMER_INODE_RDIRTY;
1412         }
1413
1414         /*
1415          * Queue up as many dirty buffers as we can then set a flag to
1416          * cause any further BIOs to go to the alternative queue.
1417          */
1418         if (ip->flags & HAMMER_INODE_VHELD)
1419                 error = vfsync(ip->vp, MNT_NOWAIT, 1, NULL, NULL);
1420         ip->flags |= HAMMER_INODE_WRITE_ALT;
1421
1422         /*
1423          * The buffer cache may contain dirty buffers beyond the inode
1424          * state we copied from the frontend to the backend.  Because
1425          * we are syncing our buffer cache on the backend, resync
1426          * the truncation point and the file size so we don't wipe out
1427          * any data.
1428          *
1429          * Syncing the buffer cache on the frontend has serious problems
1430          * because it prevents us from passively queueing dirty inodes
1431          * to the backend (the BIO's could stall indefinitely).
1432          */
1433         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1434                 ip->sync_trunc_off = ip->trunc_off;
1435                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1436         }
1437         if (ip->sync_ino_rec.ino_size != ip->ino_rec.ino_size) {
1438                 ip->sync_ino_rec.ino_size = ip->ino_rec.ino_size;
1439                 ip->sync_flags |= HAMMER_INODE_RDIRTY;
1440         }
1441
1442         /*
1443          * If there is a trunction queued destroy any data past the (aligned)
1444          * truncation point.  Userland will have dealt with the buffer
1445          * containing the truncation point for us.
1446          *
1447          * We don't flush pending frontend data buffers until after we've
1448          * dealth with the truncation.
1449          *
1450          * Don't bother if the inode is or has been deleted.
1451          */
1452         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
1453                 /*
1454                  * Interlock trunc_off.  The VOP front-end may continue to
1455                  * make adjustments to it while we are blocked.
1456                  */
1457                 off_t trunc_off;
1458                 off_t aligned_trunc_off;
1459
1460                 trunc_off = ip->sync_trunc_off;
1461                 aligned_trunc_off = (trunc_off + HAMMER_BUFMASK) &
1462                                     ~HAMMER_BUFMASK64;
1463
1464                 /*
1465                  * Delete any whole blocks on-media.  The front-end has
1466                  * already cleaned out any partial block and made it
1467                  * pending.  The front-end may have updated trunc_off
1468                  * while we were blocked so do not just unconditionally
1469                  * set it to the maximum offset.
1470                  */
1471                 error = hammer_ip_delete_range(&cursor, ip,
1472                                                 aligned_trunc_off,
1473                                                 0x7FFFFFFFFFFFFFFFLL);
1474                 if (error)
1475                         Debugger("hammer_ip_delete_range errored");
1476                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1477                 if (ip->trunc_off >= trunc_off) {
1478                         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1479                         ip->flags &= ~HAMMER_INODE_TRUNCATED;
1480                 }
1481         } else {
1482                 error = 0;
1483         }
1484
1485         /*
1486          * Now sync related records.  These will typically be directory
1487          * entries or delete-on-disk records.
1488          *
1489          * Not all records will be flushed, but clear XDIRTY anyway.  We
1490          * will set it again in the frontend hammer_flush_inode_done()
1491          * if records remain.
1492          */
1493         if (error == 0) {
1494                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1495                                     hammer_sync_record_callback, &cursor);
1496                 if (tmp_error < 0)
1497                         tmp_error = -error;
1498                 if (tmp_error)
1499                         error = tmp_error;
1500                 if (error == 0)
1501                         ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
1502         }
1503
1504         /*
1505          * If we are deleting the inode the frontend had better not have
1506          * any active references on elements making up the inode.
1507          */
1508         if (error == 0 && ip->sync_ino_rec.ino_nlinks == 0 &&
1509                 RB_EMPTY(&ip->rec_tree)  &&
1510             (ip->sync_flags & HAMMER_INODE_DELETING) &&
1511             (ip->flags & HAMMER_INODE_DELETED) == 0) {
1512                 int count1 = 0;
1513
1514                 hkprintf("Y");
1515                 ip->flags |= HAMMER_INODE_DELETED;
1516                 error = hammer_ip_delete_range_all(&cursor, ip, &count1);
1517                 if (error == 0) {
1518                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
1519                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1520                         KKASSERT(RB_EMPTY(&ip->rec_tree));
1521
1522                         /*
1523                          * Set delete_tid in both the frontend and backend
1524                          * copy of the inode record.  The DELETED flag handles
1525                          * this, do not set RDIRTY.
1526                          */
1527                         ip->ino_rec.base.base.delete_tid = trans.tid;
1528                         ip->sync_ino_rec.base.base.delete_tid = trans.tid;
1529
1530                         /*
1531                          * Adjust the inode count in the volume header
1532                          */
1533                         hammer_modify_volume_field(&trans, trans.rootvol,
1534                                                    vol0_stat_inodes);
1535                         --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1536                         hammer_modify_volume_done(trans.rootvol);
1537                 } else {
1538                         ip->flags &= ~HAMMER_INODE_DELETED;
1539                         Debugger("hammer_ip_delete_range_all errored");
1540                 }
1541         }
1542
1543         /*
1544          * Flush any queued BIOs.  These will just biodone() the IO's if
1545          * the inode has been deleted.
1546          */
1547         while ((bio = TAILQ_FIRST(&ip->bio_list)) != NULL) {
1548                 TAILQ_REMOVE(&ip->bio_list, bio, bio_act);
1549                 tmp_error = hammer_dowrite(&cursor, ip, bio);
1550                 if (tmp_error)
1551                         error = tmp_error;
1552         }
1553         ip->sync_flags &= ~HAMMER_INODE_BUFS;
1554
1555         if (error)
1556                 Debugger("RB_SCAN errored");
1557
1558         /*
1559          * Now update the inode's on-disk inode-data and/or on-disk record.
1560          * DELETED and ONDISK are managed only in ip->flags.
1561          */
1562         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
1563         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
1564                 /*
1565                  * If deleted and on-disk, don't set any additional flags.
1566                  * the delete flag takes care of things.
1567                  *
1568                  * Clear flags which may have been set by the frontend.
1569                  */
1570                 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY|HAMMER_INODE_DDIRTY|
1571                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1572                                     HAMMER_INODE_DELETING);
1573                 break;
1574         case HAMMER_INODE_DELETED:
1575                 /*
1576                  * Take care of the case where a deleted inode was never
1577                  * flushed to the disk in the first place.
1578                  *
1579                  * Clear flags which may have been set by the frontend.
1580                  */
1581                 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY|HAMMER_INODE_DDIRTY|
1582                                     HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1583                                     HAMMER_INODE_DELETING);
1584                 while (RB_ROOT(&ip->rec_tree)) {
1585                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
1586                         hammer_ref(&record->lock);
1587                         KKASSERT(record->lock.refs == 1);
1588                         record->flags |= HAMMER_RECF_DELETED_FE;
1589                         record->flags |= HAMMER_RECF_DELETED_BE;
1590                         hammer_rel_mem_record(record);
1591                 }
1592                 break;
1593         case HAMMER_INODE_ONDISK:
1594                 /*
1595                  * If already on-disk, do not set any additional flags.
1596                  */
1597                 break;
1598         default:
1599                 /*
1600                  * If not on-disk and not deleted, set both dirty flags
1601                  * to force an initial record to be written.  Also set
1602                  * the create_tid for the inode.
1603                  *
1604                  * Set create_tid in both the frontend and backend
1605                  * copy of the inode record.
1606                  */
1607                 ip->ino_rec.base.base.create_tid = trans.tid;
1608                 ip->sync_ino_rec.base.base.create_tid = trans.tid;
1609                 ip->sync_flags |= HAMMER_INODE_RDIRTY | HAMMER_INODE_DDIRTY;
1610                 break;
1611         }
1612
1613         /*
1614          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
1615          * is already on-disk the old record is marked as deleted.
1616          *
1617          * If DELETED is set hammer_update_inode() will delete the existing
1618          * record without writing out a new one.
1619          *
1620          * If *ONLY* the ITIMES flag is set we can update the record in-place.
1621          */
1622         if (ip->flags & HAMMER_INODE_DELETED) {
1623                 error = hammer_update_inode(&cursor, ip);
1624         } else
1625         if ((ip->sync_flags & (HAMMER_INODE_RDIRTY | HAMMER_INODE_DDIRTY |
1626                                HAMMER_INODE_ITIMES)) == HAMMER_INODE_ITIMES) {
1627                 error = hammer_update_itimes(&cursor, ip);
1628         } else
1629         if (ip->sync_flags & (HAMMER_INODE_RDIRTY | HAMMER_INODE_DDIRTY |
1630                               HAMMER_INODE_ITIMES)) {
1631                 error = hammer_update_inode(&cursor, ip);
1632         }
1633         if (error)
1634                 Debugger("hammer_update_itimes/inode errored");
1635 done:
1636         /*
1637          * Save the TID we used to sync the inode with to make sure we
1638          * do not improperly reuse it.
1639          */
1640         hammer_done_cursor(&cursor);
1641         hammer_done_transaction(&trans);
1642         return(error);
1643 }
1644
1645 /*
1646  * This routine is called when the OS is no longer actively referencing
1647  * the inode (but might still be keeping it cached), or when releasing
1648  * the last reference to an inode.
1649  *
1650  * At this point if the inode's nlinks count is zero we want to destroy
1651  * it, which may mean destroying it on-media too.
1652  */
1653 void
1654 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
1655 {
1656         struct vnode *vp;
1657
1658         /*
1659          * Set the DELETING flag when the link count drops to 0 and the
1660          * OS no longer has any opens on the inode.
1661          *
1662          * The backend will clear DELETING (a mod flag) and set DELETED
1663          * (a state flag) when it is actually able to perform the
1664          * operation.
1665          */
1666         if (ip->ino_rec.ino_nlinks == 0 &&
1667             (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
1668                 ip->flags |= HAMMER_INODE_DELETING;
1669                 ip->flags |= HAMMER_INODE_TRUNCATED;
1670                 ip->trunc_off = 0;
1671                 vp = NULL;
1672                 if (getvp) {
1673                         if (hammer_get_vnode(ip, &vp) != 0)
1674                                 return;
1675                 }
1676                 if (ip->vp) {
1677                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
1678                         vnode_pager_setsize(ip->vp, 0);
1679                 }
1680                 if (getvp) {
1681                         vput(vp);
1682                 }
1683         }
1684 }
1685
1686 /*
1687  * Re-test an inode when a dependancy had gone away to see if we
1688  * can chain flush it.
1689  */
1690 void
1691 hammer_test_inode(hammer_inode_t ip)
1692 {
1693         if (ip->flags & HAMMER_INODE_REFLUSH) {
1694                 ip->flags &= ~HAMMER_INODE_REFLUSH;
1695                 hammer_ref(&ip->lock);
1696                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1697                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
1698                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1699                 } else {
1700                         hammer_flush_inode(ip, 0);
1701                 }
1702                 hammer_rel_inode(ip, 0);
1703         }
1704 }
1705