2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
34 * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.50 2008/05/04 09:06:45 dillon Exp $
38 #include <vm/vm_extern.h>
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);
48 * The kernel is not actively referencing this vnode but is still holding
51 * This is called from the frontend.
54 hammer_vop_inactive(struct vop_inactive_args *ap)
56 struct hammer_inode *ip = VTOI(ap->a_vp);
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.
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)
80 * Release the vnode association. This is typically (but not always)
81 * the last reference on the inode.
83 * Once the association is lost we are on our own with regards to
87 hammer_vop_reclaim(struct vop_reclaim_args *ap)
89 struct hammer_inode *ip;
94 if ((ip = vp->v_data) != NULL) {
97 hammer_rel_inode(ip, 1);
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
107 * Called from the frontend.
110 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
116 if ((vp = ip->vp) == NULL) {
117 error = getnewvnode(VT_HAMMER, ip->hmp->mp, vpp, 0, 0);
120 hammer_lock_ex(&ip->lock);
121 if (ip->vp != NULL) {
122 hammer_unlock(&ip->lock);
127 hammer_ref(&ip->lock);
130 vp->v_type = hammer_get_vnode_type(
131 ip->ino_rec.base.base.obj_type);
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);
140 case HAMMER_OBJTYPE_FIFO:
141 vp->v_ops = &ip->hmp->mp->mnt_vn_fifo_ops;
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().
153 if (ip->obj_id == HAMMER_OBJID_ROOT &&
154 ip->obj_asof == ip->hmp->asof) {
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);
168 * loop if the vget fails (aka races), or if the vp
169 * no longer matches ip->vp.
171 if (vget(vp, LK_EXCLUSIVE) == 0) {
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
186 * The flags argument is only applied for newly created inodes, and only
187 * certain flags are inherited.
189 * Called from the frontend.
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)
195 hammer_mount_t hmp = trans->hmp;
196 struct hammer_inode_info iinfo;
197 struct hammer_cursor cursor;
198 struct hammer_inode *ip;
201 * Determine if we already have an inode cached. If we do then
204 iinfo.obj_id = obj_id;
205 iinfo.obj_asof = asof;
207 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
209 hammer_ref(&ip->lock);
214 ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
215 ++hammer_count_inodes;
217 ip->obj_asof = iinfo.obj_asof;
219 ip->flags = flags & HAMMER_INODE_RO;
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);
229 * Locate the on-disk inode.
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 |
243 *errorp = hammer_btree_lookup(&cursor);
244 if (*errorp == EDEADLK) {
245 hammer_done_cursor(&cursor);
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.
256 ip->ino_rec = cursor.record->inode;
257 ip->ino_data = cursor.data->inode;
258 hammer_cache_node(cursor.node, &ip->cache[0]);
260 hammer_cache_node(cursor.node, cache);
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.
268 * The cursor's locked node interlocks against others creating and
269 * destroying ip while we were blocked.
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;
279 hammer_done_cursor(&cursor);
282 ip->flags |= HAMMER_INODE_ONDISK;
284 kprintf("hammer_get_inode: failed ip %p cursor %p error %d\n",
285 ip, &cursor, *errorp);
287 --hammer_count_inodes;
291 hammer_done_cursor(&cursor);
296 * Create a new filesystem object, returning the inode in *ipp. The
297 * returned inode will be referenced.
299 * The inode is created in-memory.
302 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
303 struct ucred *cred, hammer_inode_t dip,
304 struct hammer_inode **ipp)
311 ip = kmalloc(sizeof(*ip), M_HAMMER, M_WAITOK|M_ZERO);
312 ++hammer_count_inodes;
313 ip->obj_id = hammer_alloc_objid(trans, dip);
314 KKASSERT(ip->obj_id != 0);
315 ip->obj_asof = hmp->asof;
317 ip->flush_state = HAMMER_FST_IDLE;
318 ip->flags = HAMMER_INODE_DDIRTY | HAMMER_INODE_RDIRTY |
321 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
322 RB_INIT(&ip->rec_tree);
323 TAILQ_INIT(&ip->bio_list);
324 TAILQ_INIT(&ip->bio_alt_list);
325 TAILQ_INIT(&ip->target_list);
327 ip->ino_rec.ino_atime = trans->time;
328 ip->ino_rec.ino_mtime = trans->time;
329 ip->ino_rec.ino_size = 0;
330 ip->ino_rec.ino_nlinks = 0;
332 ip->ino_rec.base.base.btype = HAMMER_BTREE_TYPE_RECORD;
333 ip->ino_rec.base.base.obj_id = ip->obj_id;
334 ip->ino_rec.base.base.key = 0;
335 ip->ino_rec.base.base.create_tid = 0;
336 ip->ino_rec.base.base.delete_tid = 0;
337 ip->ino_rec.base.base.rec_type = HAMMER_RECTYPE_INODE;
338 ip->ino_rec.base.base.obj_type = hammer_get_obj_type(vap->va_type);
340 ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
341 ip->ino_data.mode = vap->va_mode;
342 ip->ino_data.ctime = trans->time;
343 ip->ino_data.parent_obj_id = (dip) ? dip->ino_rec.base.base.obj_id : 0;
345 switch(ip->ino_rec.base.base.obj_type) {
346 case HAMMER_OBJTYPE_CDEV:
347 case HAMMER_OBJTYPE_BDEV:
348 ip->ino_data.rmajor = vap->va_rmajor;
349 ip->ino_data.rminor = vap->va_rminor;
356 * Calculate default uid/gid and overwrite with information from
359 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
360 ip->ino_data.gid = dip->ino_data.gid;
361 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode, xuid, cred,
363 ip->ino_data.mode = vap->va_mode;
365 if (vap->va_vaflags & VA_UID_UUID_VALID)
366 ip->ino_data.uid = vap->va_uid_uuid;
367 else if (vap->va_uid != (uid_t)VNOVAL)
368 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
369 if (vap->va_vaflags & VA_GID_UUID_VALID)
370 ip->ino_data.gid = vap->va_gid_uuid;
371 else if (vap->va_gid != (gid_t)VNOVAL)
372 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
374 hammer_ref(&ip->lock);
375 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
376 hammer_unref(&ip->lock);
377 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
384 * Called by hammer_sync_inode().
387 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
389 hammer_transaction_t trans = cursor->trans;
390 hammer_record_t record;
397 * If the inode has a presence on-disk then locate it and mark
398 * it deleted, setting DELONDISK.
400 * The record may or may not be physically deleted, depending on
401 * the retention policy.
403 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
404 HAMMER_INODE_ONDISK) {
405 hammer_normalize_cursor(cursor);
406 cursor->key_beg.obj_id = ip->obj_id;
407 cursor->key_beg.key = 0;
408 cursor->key_beg.create_tid = 0;
409 cursor->key_beg.delete_tid = 0;
410 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
411 cursor->key_beg.obj_type = 0;
412 cursor->asof = ip->obj_asof;
413 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
414 cursor->flags |= HAMMER_CURSOR_GET_RECORD | HAMMER_CURSOR_ASOF;
415 cursor->flags |= HAMMER_CURSOR_BACKEND;
417 error = hammer_btree_lookup(cursor);
418 if (hammer_debug_inode)
419 kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
421 kprintf("error %d\n", error);
422 Debugger("hammer_update_inode");
426 error = hammer_ip_delete_record(cursor, trans->tid);
427 if (hammer_debug_inode)
428 kprintf(" error %d\n", error);
429 if (error && error != EDEADLK) {
430 kprintf("error %d\n", error);
431 Debugger("hammer_update_inode2");
434 ip->flags |= HAMMER_INODE_DELONDISK;
437 hammer_cache_node(cursor->node, &ip->cache[0]);
439 if (error == EDEADLK) {
440 hammer_done_cursor(cursor);
441 error = hammer_init_cursor(trans, cursor,
443 if (hammer_debug_inode)
444 kprintf("IPDED %p %d\n", ip, error);
451 * Ok, write out the initial record or a new record (after deleting
452 * the old one), unless the DELETED flag is set. This routine will
453 * clear DELONDISK if it writes out a record.
455 * Update our inode statistics if this is the first application of
458 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
460 * Generate a record and write it to the media
462 record = hammer_alloc_mem_record(ip);
463 record->type = HAMMER_MEM_RECORD_GENERAL;
464 record->flush_state = HAMMER_FST_FLUSH;
465 record->rec.inode = ip->sync_ino_rec;
466 record->rec.inode.base.base.create_tid = trans->tid;
467 record->rec.inode.base.data_len = sizeof(ip->sync_ino_data);
468 record->data = (void *)&ip->sync_ino_data;
469 record->flags |= HAMMER_RECF_INTERLOCK_BE;
471 error = hammer_ip_sync_record_cursor(cursor, record);
472 if (hammer_debug_inode)
473 kprintf("GENREC %p rec %08x %d\n",
474 ip, record->flags, error);
475 if (error != EDEADLK)
477 hammer_done_cursor(cursor);
478 error = hammer_init_cursor(trans, cursor,
480 if (hammer_debug_inode)
481 kprintf("GENREC reinit %d\n", error);
486 kprintf("error %d\n", error);
487 Debugger("hammer_update_inode3");
491 * The record isn't managed by the inode's record tree,
492 * destroy it whether we succeed or fail.
494 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
495 record->flags |= HAMMER_RECF_DELETED_FE;
496 record->flush_state = HAMMER_FST_IDLE;
497 hammer_rel_mem_record(record);
503 if (hammer_debug_inode)
504 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
505 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY |
506 HAMMER_INODE_DDIRTY |
507 HAMMER_INODE_ITIMES);
508 ip->flags &= ~HAMMER_INODE_DELONDISK;
511 * Root volume count of inodes
513 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
514 hammer_modify_volume_field(trans,
517 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
518 hammer_modify_volume_done(trans->rootvol);
519 ip->flags |= HAMMER_INODE_ONDISK;
520 if (hammer_debug_inode)
521 kprintf("NOWONDISK %p\n", ip);
527 * If the inode has been destroyed, clean out any left-over flags
528 * that may have been set by the frontend.
530 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
531 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY |
532 HAMMER_INODE_DDIRTY |
533 HAMMER_INODE_ITIMES);
539 * Update only the itimes fields. This is done no-historically. The
540 * record is updated in-place on the disk.
543 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
545 hammer_transaction_t trans = cursor->trans;
546 struct hammer_inode_record *rec;
551 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
552 HAMMER_INODE_ONDISK) {
553 hammer_normalize_cursor(cursor);
554 cursor->key_beg.obj_id = ip->obj_id;
555 cursor->key_beg.key = 0;
556 cursor->key_beg.create_tid = 0;
557 cursor->key_beg.delete_tid = 0;
558 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
559 cursor->key_beg.obj_type = 0;
560 cursor->asof = ip->obj_asof;
561 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
562 cursor->flags |= HAMMER_CURSOR_GET_RECORD | HAMMER_CURSOR_ASOF;
563 cursor->flags |= HAMMER_CURSOR_BACKEND;
565 error = hammer_btree_lookup(cursor);
567 kprintf("error %d\n", error);
568 Debugger("hammer_update_itimes1");
572 * Do not generate UNDO records for atime/mtime
575 rec = &cursor->record->inode;
576 hammer_modify_buffer(trans, cursor->record_buffer,
578 rec->ino_atime = ip->sync_ino_rec.ino_atime;
579 rec->ino_mtime = ip->sync_ino_rec.ino_mtime;
580 hammer_modify_buffer_done(cursor->record_buffer);
581 ip->sync_flags &= ~HAMMER_INODE_ITIMES;
582 /* XXX recalculate crc */
583 hammer_cache_node(cursor->node, &ip->cache[0]);
585 if (error == EDEADLK) {
586 hammer_done_cursor(cursor);
587 error = hammer_init_cursor(trans, cursor,
597 * Release a reference on an inode, flush as requested.
599 * On the last reference we queue the inode to the flusher for its final
603 hammer_rel_inode(struct hammer_inode *ip, int flush)
605 hammer_mount_t hmp = ip->hmp;
608 * Handle disposition when dropping the last ref.
611 if (ip->lock.refs == 1) {
613 * Determine whether on-disk action is needed for
614 * the inode's final disposition.
616 KKASSERT(ip->vp == NULL);
617 hammer_inode_unloadable_check(ip, 0);
618 if (ip->flags & HAMMER_INODE_MODMASK) {
619 hammer_flush_inode(ip, 0);
620 } else if (ip->lock.refs == 1) {
621 hammer_unload_inode(ip);
626 hammer_flush_inode(ip, 0);
629 * The inode still has multiple refs, try to drop
632 KKASSERT(ip->lock.refs >= 1);
633 if (ip->lock.refs > 1) {
634 hammer_unref(&ip->lock);
641 * XXX bad hack until I add code to track inodes in SETUP. We
642 * can queue a lot of inodes to the syncer but if we don't wake
643 * it up the undo sets will be too large or too many unflushed
644 * records will build up and blow our malloc limit.
646 if (++hmp->reclaim_count > 256) {
647 hmp->reclaim_count = 0;
648 hammer_flusher_async(hmp);
653 * Unload and destroy the specified inode. Must be called with one remaining
654 * reference. The reference is disposed of.
656 * This can only be called in the context of the flusher.
659 hammer_unload_inode(struct hammer_inode *ip)
661 KASSERT(ip->lock.refs == 1,
662 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
663 KKASSERT(ip->vp == NULL);
664 KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
665 KKASSERT(ip->cursor_ip_refs == 0);
666 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
668 KKASSERT(RB_EMPTY(&ip->rec_tree));
669 KKASSERT(TAILQ_EMPTY(&ip->target_list));
670 KKASSERT(TAILQ_EMPTY(&ip->bio_list));
671 KKASSERT(TAILQ_EMPTY(&ip->bio_alt_list));
673 RB_REMOVE(hammer_ino_rb_tree, &ip->hmp->rb_inos_root, ip);
675 hammer_uncache_node(&ip->cache[0]);
676 hammer_uncache_node(&ip->cache[1]);
678 hammer_clear_objid(ip);
679 --hammer_count_inodes;
686 * A transaction has modified an inode, requiring updates as specified by
689 * HAMMER_INODE_RDIRTY: Inode record has been updated
690 * HAMMER_INODE_DDIRTY: Inode data has been updated
691 * HAMMER_INODE_XDIRTY: Dirty in-memory records
692 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
693 * HAMMER_INODE_DELETED: Inode record/data must be deleted
694 * HAMMER_INODE_ITIMES: mtime/atime has been updated
697 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
699 KKASSERT ((ip->flags & HAMMER_INODE_RO) == 0 ||
700 (flags & (HAMMER_INODE_RDIRTY|HAMMER_INODE_DDIRTY|
701 HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|
702 HAMMER_INODE_DELETED|HAMMER_INODE_ITIMES)) == 0);
708 * Request that an inode be flushed. This whole mess cannot block and may
709 * recurse. Once requested HAMMER will attempt to actively flush it until
710 * the flush can be done.
712 * The inode may already be flushing, or may be in a setup state. We can
713 * place the inode in a flushing state if it is currently idle and flag it
714 * to reflush if it is currently flushing.
717 hammer_flush_inode(hammer_inode_t ip, int flags)
719 hammer_record_t depend;
723 * Trivial 'nothing to flush' case. If the inode is ina SETUP
724 * state we have to put it back into an IDLE state so we can
725 * drop the extra ref.
727 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
728 if (ip->flush_state == HAMMER_FST_SETUP) {
729 ip->flush_state = HAMMER_FST_IDLE;
730 hammer_rel_inode(ip, 0);
736 * Our flush action will depend on the current state.
738 switch(ip->flush_state) {
739 case HAMMER_FST_IDLE:
741 * We have no dependancies and can flush immediately. Some
742 * our children may not be flushable so we have to re-test
743 * with that additional knowledge.
745 hammer_flush_inode_core(ip, flags);
747 case HAMMER_FST_SETUP:
749 * Recurse upwards through dependancies via target_list
750 * and start their flusher actions going if possible.
752 * 'good' is our connectivity. -1 means we have none and
753 * can't flush, 0 means there weren't any dependancies, and
754 * 1 means we have good connectivity.
757 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
758 r = hammer_setup_parent_inodes(depend);
759 if (r < 0 && good == 0)
766 * We can continue if good >= 0. Determine how many records
767 * under our inode can be flushed (and mark them).
770 hammer_flush_inode_core(ip, flags);
772 ip->flags |= HAMMER_INODE_REFLUSH;
773 if (flags & HAMMER_FLUSH_SIGNAL) {
774 ip->flags |= HAMMER_INODE_RESIGNAL;
775 hammer_flusher_async(ip->hmp);
781 * We are already flushing, flag the inode to reflush
782 * if needed after it completes its current flush.
784 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
785 ip->flags |= HAMMER_INODE_REFLUSH;
786 if (flags & HAMMER_FLUSH_SIGNAL) {
787 ip->flags |= HAMMER_INODE_RESIGNAL;
788 hammer_flusher_async(ip->hmp);
795 * We are asked to recurse upwards and convert the record from SETUP
796 * to FLUSH if possible. record->ip is a parent of the caller's inode,
797 * and record->target_ip is the caller's inode.
799 * Return 1 if the record gives us connectivity
801 * Return 0 if the record is not relevant
803 * Return -1 if we can't resolve the dependancy and there is no connectivity.
806 hammer_setup_parent_inodes(hammer_record_t record)
808 hammer_mount_t hmp = record->ip->hmp;
809 hammer_record_t depend;
813 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
817 * If the record is already flushing, is it in our flush group?
819 * If it is in our flush group but it is a general record or a
820 * delete-on-disk, it does not improve our connectivity (return 0),
821 * and if the target inode is not trying to destroy itself we can't
822 * allow the operation yet anyway (the second return -1).
824 if (record->flush_state == HAMMER_FST_FLUSH) {
825 if (record->flush_group != hmp->flusher_next) {
826 ip->flags |= HAMMER_INODE_REFLUSH;
829 if (record->type == HAMMER_MEM_RECORD_ADD)
836 * It must be a setup record. Try to resolve the setup dependancies
837 * by recursing upwards so we can place ip on the flush list.
839 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
842 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
843 r = hammer_setup_parent_inodes(depend);
844 if (r < 0 && good == 0)
851 * We can't flush ip because it has no connectivity (XXX also check
852 * nlinks for pre-existing connectivity!). Flag it so any resolution
853 * recurses back down.
856 ip->flags |= HAMMER_INODE_REFLUSH;
861 * We are go, place the parent inode in a flushing state so we can
862 * place its record in a flushing state. Note that the parent
863 * may already be flushing. The record must be in the same flush
864 * group as the parent.
866 if (ip->flush_state != HAMMER_FST_FLUSH)
867 hammer_flush_inode_core(ip, HAMMER_FLUSH_RECURSION);
868 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
869 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
872 if (record->type == HAMMER_MEM_RECORD_DEL &&
873 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
875 * Regardless of flushing state we cannot sync this path if the
876 * record represents a delete-on-disk but the target inode
877 * is not ready to sync its own deletion.
879 * XXX need to count effective nlinks to determine whether
880 * the flush is ok, otherwise removing a hardlink will
881 * just leave the DEL record to rot.
883 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
887 if (ip->flush_group == ip->hmp->flusher_next) {
889 * This is the record we wanted to synchronize.
891 record->flush_state = HAMMER_FST_FLUSH;
892 record->flush_group = ip->flush_group;
893 hammer_ref(&record->lock);
894 if (record->type == HAMMER_MEM_RECORD_ADD)
898 * A general or delete-on-disk record does not contribute
899 * to our visibility. We can still flush it, however.
904 * We couldn't resolve the dependancies, request that the
905 * inode be flushed when the dependancies can be resolved.
907 ip->flags |= HAMMER_INODE_REFLUSH;
913 * This is the core routine placing an inode into the FST_FLUSH state.
916 hammer_flush_inode_core(hammer_inode_t ip, int flags)
921 * Set flush state and prevent the flusher from cycling into
922 * the next flush group. Do not place the ip on the list yet.
923 * Inodes not in the idle state get an extra reference.
925 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
926 if (ip->flush_state == HAMMER_FST_IDLE)
927 hammer_ref(&ip->lock);
928 ip->flush_state = HAMMER_FST_FLUSH;
929 ip->flush_group = ip->hmp->flusher_next;
930 ++ip->hmp->flusher_lock;
933 * We need to be able to vfsync/truncate from the backend.
935 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
936 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
937 ip->flags |= HAMMER_INODE_VHELD;
942 * Figure out how many in-memory records we can actually flush
943 * (not including inode meta-data, buffers, etc).
945 if (flags & HAMMER_FLUSH_RECURSION) {
948 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
949 hammer_setup_child_callback, NULL);
953 * This is a more involved test that includes go_count. If we
954 * can't flush, flag the inode and return. If go_count is 0 we
955 * were are unable to flush any records in our rec_tree and
956 * must ignore the XDIRTY flag.
959 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
960 ip->flags |= HAMMER_INODE_REFLUSH;
961 ip->flush_state = HAMMER_FST_SETUP;
962 if (ip->flags & HAMMER_INODE_VHELD) {
963 ip->flags &= ~HAMMER_INODE_VHELD;
966 if (flags & HAMMER_FLUSH_SIGNAL) {
967 ip->flags |= HAMMER_INODE_RESIGNAL;
968 hammer_flusher_async(ip->hmp);
970 if (--ip->hmp->flusher_lock == 0)
971 wakeup(&ip->hmp->flusher_lock);
977 * Snapshot the state of the inode for the backend flusher.
979 * The truncation must be retained in the frontend until after
980 * we've actually performed the record deletion.
982 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
983 * and stays in ip->flags. Once set, it stays set until the
984 * inode is destroyed.
986 ip->sync_flags = (ip->flags & HAMMER_INODE_MODMASK);
987 ip->sync_trunc_off = ip->trunc_off;
988 ip->sync_ino_rec = ip->ino_rec;
989 ip->sync_ino_data = ip->ino_data;
990 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
993 * The flusher list inherits our inode and reference.
995 TAILQ_INSERT_TAIL(&ip->hmp->flush_list, ip, flush_entry);
996 if (--ip->hmp->flusher_lock == 0)
997 wakeup(&ip->hmp->flusher_lock);
999 if (flags & HAMMER_FLUSH_SIGNAL)
1000 hammer_flusher_async(ip->hmp);
1004 * Callback for scan of ip->rec_tree. Try to include each record in our
1005 * flush. ip->flush_group has been set but the inode has not yet been
1006 * moved into a flushing state.
1008 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1011 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1012 * the caller from shortcutting the flush.
1015 hammer_setup_child_callback(hammer_record_t rec, void *data)
1017 hammer_inode_t target_ip;
1022 * If the record has been deleted by the backend (it's being held
1023 * by the frontend in a race), just ignore it.
1025 if (rec->flags & HAMMER_RECF_DELETED_BE)
1029 * If the record is in an idle state it has no dependancies and
1035 switch(rec->flush_state) {
1036 case HAMMER_FST_IDLE:
1038 * Record has no setup dependancy, we can flush it.
1040 KKASSERT(rec->target_ip == NULL);
1041 rec->flush_state = HAMMER_FST_FLUSH;
1042 rec->flush_group = ip->flush_group;
1043 hammer_ref(&rec->lock);
1046 case HAMMER_FST_SETUP:
1048 * Record has a setup dependancy. Try to include the
1049 * target ip in the flush.
1051 * We have to be careful here, if we do not do the right
1052 * thing we can lose track of dirty inodes and the system
1053 * will lockup trying to allocate buffers.
1055 target_ip = rec->target_ip;
1056 KKASSERT(target_ip != NULL);
1057 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1058 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1060 * If the target IP is already flushing in our group
1061 * we are golden, otherwise make sure the target
1064 if (target_ip->flush_group == ip->flush_group) {
1065 rec->flush_state = HAMMER_FST_FLUSH;
1066 rec->flush_group = ip->flush_group;
1067 hammer_ref(&rec->lock);
1070 target_ip->flags |= HAMMER_INODE_REFLUSH;
1072 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1074 * If the target IP is not flushing we can force
1075 * it to flush, even if it is unable to write out
1076 * any of its own records we have at least one in
1077 * hand that we CAN deal with.
1079 rec->flush_state = HAMMER_FST_FLUSH;
1080 rec->flush_group = ip->flush_group;
1081 hammer_ref(&rec->lock);
1082 hammer_flush_inode_core(target_ip,
1083 HAMMER_FLUSH_RECURSION);
1087 * General or delete-on-disk record.
1089 * XXX this needs help. If a delete-on-disk we could
1090 * disconnect the target. If the target has its own
1091 * dependancies they really need to be flushed.
1095 rec->flush_state = HAMMER_FST_FLUSH;
1096 rec->flush_group = ip->flush_group;
1097 hammer_ref(&rec->lock);
1098 hammer_flush_inode_core(target_ip,
1099 HAMMER_FLUSH_RECURSION);
1103 case HAMMER_FST_FLUSH:
1105 * Record already associated with a flush group. It had
1108 KKASSERT(rec->flush_group == ip->flush_group);
1116 * Wait for a previously queued flush to complete
1119 hammer_wait_inode(hammer_inode_t ip)
1121 while (ip->flush_state != HAMMER_FST_IDLE) {
1122 ip->flags |= HAMMER_INODE_FLUSHW;
1123 tsleep(&ip->flags, 0, "hmrwin", 0);
1128 * Called by the backend code when a flush has been completed.
1129 * The inode has already been removed from the flush list.
1131 * A pipelined flush can occur, in which case we must re-enter the
1132 * inode on the list and re-copy its fields.
1135 hammer_flush_inode_done(hammer_inode_t ip)
1140 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1143 * Allow BIOs to queue to the inode's primary bioq again.
1145 ip->flags &= ~HAMMER_INODE_WRITE_ALT;
1148 * Merge left-over flags back into the frontend and fix the state.
1150 ip->flags |= ip->sync_flags;
1153 * The backend may have adjusted nlinks, so if the adjusted nlinks
1154 * does not match the fronttend set the frontend's RDIRTY flag again.
1156 if (ip->ino_rec.ino_nlinks != ip->sync_ino_rec.ino_nlinks)
1157 ip->flags |= HAMMER_INODE_RDIRTY;
1160 * Reflush any BIOs that wound up in the alt list. Our inode will
1161 * also wind up at the end of the flusher's list.
1163 while ((bio = TAILQ_FIRST(&ip->bio_alt_list)) != NULL) {
1164 TAILQ_REMOVE(&ip->bio_alt_list, bio, bio_act);
1165 TAILQ_INSERT_TAIL(&ip->bio_list, bio, bio_act);
1168 * Fix up the dirty buffer status.
1170 if (TAILQ_FIRST(&ip->bio_list) ||
1171 (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree))) {
1172 ip->flags |= HAMMER_INODE_BUFS;
1176 * Re-set the XDIRTY flag if some of the inode's in-memory records
1177 * could not be flushed.
1179 if (RB_ROOT(&ip->rec_tree))
1180 ip->flags |= HAMMER_INODE_XDIRTY;
1183 * Do not lose track of inodes which no longer have vnode
1184 * assocations, otherwise they may never get flushed again.
1186 if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1187 ip->flags |= HAMMER_INODE_REFLUSH;
1190 * Adjust flush_state. The target state (idle or setup) shouldn't
1191 * be terribly important since we will reflush if we really need
1192 * to do anything. XXX
1194 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
1195 ip->flush_state = HAMMER_FST_IDLE;
1198 ip->flush_state = HAMMER_FST_SETUP;
1202 * Clean up the vnode ref
1204 if (ip->flags & HAMMER_INODE_VHELD) {
1205 ip->flags &= ~HAMMER_INODE_VHELD;
1210 * If the frontend made more changes and requested another flush,
1211 * then try to get it running.
1213 if (ip->flags & HAMMER_INODE_REFLUSH) {
1214 ip->flags &= ~HAMMER_INODE_REFLUSH;
1215 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1216 ip->flags &= ~HAMMER_INODE_RESIGNAL;
1217 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1219 hammer_flush_inode(ip, 0);
1224 * Finally, if the frontend is waiting for a flush to complete,
1227 if (ip->flush_state != HAMMER_FST_FLUSH) {
1228 if (ip->flags & HAMMER_INODE_FLUSHW) {
1229 ip->flags &= ~HAMMER_INODE_FLUSHW;
1234 hammer_rel_inode(ip, 0);
1238 * Called from hammer_sync_inode() to synchronize in-memory records
1242 hammer_sync_record_callback(hammer_record_t record, void *data)
1244 hammer_cursor_t cursor = data;
1245 hammer_transaction_t trans = cursor->trans;
1249 * Skip records that do not belong to the current flush.
1251 if (record->flush_state != HAMMER_FST_FLUSH)
1253 KKASSERT((record->flags & HAMMER_RECF_DELETED_BE) == 0);
1255 if (record->flush_group != record->ip->flush_group) {
1256 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1261 KKASSERT(record->flush_group == record->ip->flush_group);
1264 * Interlock the record using the BE flag. Once BE is set the
1265 * frontend cannot change the state of FE.
1267 * NOTE: If FE is set prior to us setting BE we still sync the
1268 * record out, but the flush completion code converts it to
1269 * a delete-on-disk record instead of destroying it.
1271 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1272 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1275 * If DELETED_FE is set we may have already sent dependant pieces
1276 * to the disk and we must flush the record as if it hadn't been
1277 * deleted. This creates a bit of a mess because we have to
1278 * have ip_sync_record convert the record to MEM_RECORD_DEL before
1279 * it inserts the B-Tree record. Otherwise the media sync might
1280 * be visible to the frontend.
1282 if (record->flags & HAMMER_RECF_DELETED_FE) {
1283 if (record->type == HAMMER_MEM_RECORD_ADD) {
1284 record->flags |= HAMMER_RECF_CONVERT_DELETE;
1286 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
1292 * Assign the create_tid for new records. Deletions already
1293 * have the record's entire key properly set up.
1295 if (record->type != HAMMER_MEM_RECORD_DEL)
1296 record->rec.inode.base.base.create_tid = trans->tid;
1298 error = hammer_ip_sync_record_cursor(cursor, record);
1299 if (error != EDEADLK)
1301 hammer_done_cursor(cursor);
1302 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
1307 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1311 if (error != -ENOSPC) {
1312 kprintf("hammer_sync_record_callback: sync failed rec "
1313 "%p, error %d\n", record, error);
1314 Debugger("sync failed rec");
1317 hammer_flush_record_done(record, error);
1322 * XXX error handling
1325 hammer_sync_inode(hammer_inode_t ip)
1327 struct hammer_transaction trans;
1328 struct hammer_cursor cursor;
1330 hammer_record_t depend;
1331 hammer_record_t next;
1332 int error, tmp_error;
1335 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
1338 hammer_start_transaction_fls(&trans, ip->hmp);
1339 error = hammer_init_cursor(&trans, &cursor, &ip->cache[0], ip);
1344 * Any directory records referencing this inode which are not in
1345 * our current flush group must adjust our nlink count for the
1346 * purposes of synchronization to disk.
1348 * Records which are in our flush group can be unlinked from our
1349 * inode now, allowing the inode to be physically deleted.
1351 nlinks = ip->ino_rec.ino_nlinks;
1352 next = TAILQ_FIRST(&ip->target_list);
1353 while ((depend = next) != NULL) {
1354 next = TAILQ_NEXT(depend, target_entry);
1355 if (depend->flush_state == HAMMER_FST_FLUSH &&
1356 depend->flush_group == ip->hmp->flusher_act) {
1357 TAILQ_REMOVE(&ip->target_list, depend, target_entry);
1358 depend->target_ip = NULL;
1359 /* no need to signal target_ip, it is us */
1360 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
1361 switch(depend->type) {
1362 case HAMMER_MEM_RECORD_ADD:
1365 case HAMMER_MEM_RECORD_DEL:
1375 * Set dirty if we had to modify the link count.
1377 if (ip->sync_ino_rec.ino_nlinks != nlinks) {
1378 KKASSERT((int64_t)nlinks >= 0);
1379 ip->sync_ino_rec.ino_nlinks = nlinks;
1380 ip->sync_flags |= HAMMER_INODE_RDIRTY;
1384 * Queue up as many dirty buffers as we can then set a flag to
1385 * cause any further BIOs to go to the alternative queue.
1387 if (ip->flags & HAMMER_INODE_VHELD)
1388 error = vfsync(ip->vp, MNT_NOWAIT, 1, NULL, NULL);
1389 ip->flags |= HAMMER_INODE_WRITE_ALT;
1392 * The buffer cache may contain dirty buffers beyond the inode
1393 * state we copied from the frontend to the backend. Because
1394 * we are syncing our buffer cache on the backend, resync
1395 * the truncation point and the file size so we don't wipe out
1398 * Syncing the buffer cache on the frontend has serious problems
1399 * because it prevents us from passively queueing dirty inodes
1400 * to the backend (the BIO's could stall indefinitely).
1402 if (ip->flags & HAMMER_INODE_TRUNCATED) {
1403 ip->sync_trunc_off = ip->trunc_off;
1404 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1406 if (ip->sync_ino_rec.ino_size != ip->ino_rec.ino_size) {
1407 ip->sync_ino_rec.ino_size = ip->ino_rec.ino_size;
1408 ip->sync_flags |= HAMMER_INODE_RDIRTY;
1412 * If there is a trunction queued destroy any data past the (aligned)
1413 * truncation point. Userland will have dealt with the buffer
1414 * containing the truncation point for us.
1416 * We don't flush pending frontend data buffers until after we've
1417 * dealth with the truncation.
1419 * Don't bother if the inode is or has been deleted.
1421 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
1423 * Interlock trunc_off. The VOP front-end may continue to
1424 * make adjustments to it while we are blocked.
1427 off_t aligned_trunc_off;
1429 trunc_off = ip->sync_trunc_off;
1430 aligned_trunc_off = (trunc_off + HAMMER_BUFMASK) &
1434 * Delete any whole blocks on-media. The front-end has
1435 * already cleaned out any partial block and made it
1436 * pending. The front-end may have updated trunc_off
1437 * while we were blocked so do not just unconditionally
1438 * set it to the maximum offset.
1440 error = hammer_ip_delete_range(&cursor, ip,
1442 0x7FFFFFFFFFFFFFFFLL);
1444 Debugger("hammer_ip_delete_range errored");
1445 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1446 if (ip->trunc_off >= trunc_off) {
1447 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1448 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1455 * Now sync related records. These will typically be directory
1456 * entries or delete-on-disk records.
1458 * Not all records will be flushed, but clear XDIRTY anyway. We
1459 * will set it again in the frontend hammer_flush_inode_done()
1460 * if records remain.
1463 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1464 hammer_sync_record_callback, &cursor);
1470 ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
1474 * If we are deleting the inode the frontend had better not have
1475 * any active references on elements making up the inode.
1477 if (error == 0 && ip->sync_ino_rec.ino_nlinks == 0 &&
1478 RB_EMPTY(&ip->rec_tree) &&
1479 (ip->sync_flags & HAMMER_INODE_DELETING) &&
1480 (ip->flags & HAMMER_INODE_DELETED) == 0) {
1484 ip->flags |= HAMMER_INODE_DELETED;
1485 error = hammer_ip_delete_range_all(&cursor, ip, &count1);
1487 ip->sync_flags &= ~HAMMER_INODE_DELETING;
1488 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1489 KKASSERT(RB_EMPTY(&ip->rec_tree));
1492 * Set delete_tid in both the frontend and backend
1493 * copy of the inode record. The DELETED flag handles
1494 * this, do not set RDIRTY.
1496 ip->ino_rec.base.base.delete_tid = trans.tid;
1497 ip->sync_ino_rec.base.base.delete_tid = trans.tid;
1500 * Adjust the inode count in the volume header
1502 hammer_modify_volume_field(&trans, trans.rootvol,
1504 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1505 hammer_modify_volume_done(trans.rootvol);
1507 ip->flags &= ~HAMMER_INODE_DELETED;
1508 Debugger("hammer_ip_delete_range_all errored");
1513 * Flush any queued BIOs. These will just biodone() the IO's if
1514 * the inode has been deleted.
1516 while ((bio = TAILQ_FIRST(&ip->bio_list)) != NULL) {
1517 TAILQ_REMOVE(&ip->bio_list, bio, bio_act);
1518 tmp_error = hammer_dowrite(&cursor, ip, bio);
1522 ip->sync_flags &= ~HAMMER_INODE_BUFS;
1525 Debugger("RB_SCAN errored");
1528 * Now update the inode's on-disk inode-data and/or on-disk record.
1529 * DELETED and ONDISK are managed only in ip->flags.
1531 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
1532 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
1534 * If deleted and on-disk, don't set any additional flags.
1535 * the delete flag takes care of things.
1537 * Clear flags which may have been set by the frontend.
1539 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY|HAMMER_INODE_DDIRTY|
1540 HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1541 HAMMER_INODE_DELETING);
1543 case HAMMER_INODE_DELETED:
1545 * Take care of the case where a deleted inode was never
1546 * flushed to the disk in the first place.
1548 * Clear flags which may have been set by the frontend.
1550 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY|HAMMER_INODE_DDIRTY|
1551 HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1552 HAMMER_INODE_DELETING);
1553 while (RB_ROOT(&ip->rec_tree)) {
1554 hammer_record_t record = RB_ROOT(&ip->rec_tree);
1555 hammer_ref(&record->lock);
1556 KKASSERT(record->lock.refs == 1);
1557 record->flags |= HAMMER_RECF_DELETED_FE;
1558 record->flags |= HAMMER_RECF_DELETED_BE;
1559 hammer_rel_mem_record(record);
1562 case HAMMER_INODE_ONDISK:
1564 * If already on-disk, do not set any additional flags.
1569 * If not on-disk and not deleted, set both dirty flags
1570 * to force an initial record to be written. Also set
1571 * the create_tid for the inode.
1573 * Set create_tid in both the frontend and backend
1574 * copy of the inode record.
1576 ip->ino_rec.base.base.create_tid = trans.tid;
1577 ip->sync_ino_rec.base.base.create_tid = trans.tid;
1578 ip->sync_flags |= HAMMER_INODE_RDIRTY | HAMMER_INODE_DDIRTY;
1583 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
1584 * is already on-disk the old record is marked as deleted.
1586 * If DELETED is set hammer_update_inode() will delete the existing
1587 * record without writing out a new one.
1589 * If *ONLY* the ITIMES flag is set we can update the record in-place.
1591 if (ip->flags & HAMMER_INODE_DELETED) {
1592 error = hammer_update_inode(&cursor, ip);
1594 if ((ip->sync_flags & (HAMMER_INODE_RDIRTY | HAMMER_INODE_DDIRTY |
1595 HAMMER_INODE_ITIMES)) == HAMMER_INODE_ITIMES) {
1596 error = hammer_update_itimes(&cursor, ip);
1598 if (ip->sync_flags & (HAMMER_INODE_RDIRTY | HAMMER_INODE_DDIRTY |
1599 HAMMER_INODE_ITIMES)) {
1600 error = hammer_update_inode(&cursor, ip);
1603 Debugger("hammer_update_itimes/inode errored");
1606 * Save the TID we used to sync the inode with to make sure we
1607 * do not improperly reuse it.
1609 hammer_done_cursor(&cursor);
1610 hammer_done_transaction(&trans);
1615 * This routine is called when the OS is no longer actively referencing
1616 * the inode (but might still be keeping it cached), or when releasing
1617 * the last reference to an inode.
1619 * At this point if the inode's nlinks count is zero we want to destroy
1620 * it, which may mean destroying it on-media too.
1623 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
1628 * If the inode is on-media and the link count is 0 we MUST delete
1629 * it on-media. DELETING is a mod flag, DELETED is a state flag.
1631 if (ip->ino_rec.ino_nlinks == 0 &&
1632 (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
1633 ip->flags |= HAMMER_INODE_DELETING;
1634 ip->flags |= HAMMER_INODE_TRUNCATED;
1638 if (hammer_get_vnode(ip, &vp) != 0)
1642 vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
1643 vnode_pager_setsize(ip->vp, 0);
1652 * Re-test an inode when a dependancy had gone away to see if we
1653 * can chain flush it.
1656 hammer_test_inode(hammer_inode_t ip)
1658 if (ip->flags & HAMMER_INODE_REFLUSH) {
1659 ip->flags &= ~HAMMER_INODE_REFLUSH;
1660 hammer_ref(&ip->lock);
1661 if (ip->flags & HAMMER_INODE_RESIGNAL) {
1662 ip->flags &= ~HAMMER_INODE_RESIGNAL;
1663 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
1665 hammer_flush_inode(ip, 0);
1667 hammer_rel_inode(ip, 0);