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 1.54 2008/05/09 07:26:51 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;
285 * Do not panic on read-only accesses which fail, particularly
286 * historical accesses where the snapshot might not have
287 * complete connectivity.
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);
294 --hammer_count_inodes;
298 hammer_done_cursor(&cursor);
303 * Create a new filesystem object, returning the inode in *ipp. The
304 * returned inode will be referenced.
306 * The inode is created in-memory.
309 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
310 struct ucred *cred, hammer_inode_t dip,
311 struct hammer_inode **ipp)
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;
324 ip->flush_state = HAMMER_FST_IDLE;
325 ip->flags = HAMMER_INODE_DDIRTY | HAMMER_INODE_RDIRTY |
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);
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;
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);
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;
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;
363 * Calculate default uid/gid and overwrite with information from
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,
370 ip->ino_data.mode = vap->va_mode;
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);
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);
391 * Called by hammer_sync_inode().
394 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
396 hammer_transaction_t trans = cursor->trans;
397 hammer_record_t record;
404 * If the inode has a presence on-disk then locate it and mark
405 * it deleted, setting DELONDISK.
407 * The record may or may not be physically deleted, depending on
408 * the retention policy.
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;
424 error = hammer_btree_lookup(cursor);
425 if (hammer_debug_inode)
426 kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
428 kprintf("error %d\n", error);
429 Debugger("hammer_update_inode");
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");
441 ip->flags |= HAMMER_INODE_DELONDISK;
444 hammer_cache_node(cursor->node, &ip->cache[0]);
446 if (error == EDEADLK) {
447 hammer_done_cursor(cursor);
448 error = hammer_init_cursor(trans, cursor,
450 if (hammer_debug_inode)
451 kprintf("IPDED %p %d\n", ip, error);
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.
462 * Update our inode statistics if this is the first application of
465 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
467 * Generate a record and write it to the media
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;
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)
485 hammer_done_cursor(cursor);
486 error = hammer_init_cursor(trans, cursor,
488 if (hammer_debug_inode)
489 kprintf("GENREC reinit %d\n", error);
494 kprintf("error %d\n", error);
495 Debugger("hammer_update_inode3");
499 * The record isn't managed by the inode's record tree,
500 * destroy it whether we succeed or fail.
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);
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;
519 * Root volume count of inodes
521 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
522 hammer_modify_volume_field(trans,
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);
535 * If the inode has been destroyed, clean out any left-over flags
536 * that may have been set by the frontend.
538 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
539 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY |
540 HAMMER_INODE_DDIRTY |
541 HAMMER_INODE_ITIMES);
547 * Update only the itimes fields. This is done no-historically. The
548 * record is updated in-place on the disk.
551 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
553 hammer_transaction_t trans = cursor->trans;
554 struct hammer_inode_record *rec;
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;
573 error = hammer_btree_lookup(cursor);
575 kprintf("error %d\n", error);
576 Debugger("hammer_update_itimes1");
580 * Do not generate UNDO records for atime/mtime
583 rec = &cursor->record->inode;
584 hammer_modify_record_noundo(trans,
585 cursor->record_buffer,
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,
591 ip->sync_flags &= ~HAMMER_INODE_ITIMES;
592 /* XXX recalculate crc */
593 hammer_cache_node(cursor->node, &ip->cache[0]);
595 if (error == EDEADLK) {
596 hammer_done_cursor(cursor);
597 error = hammer_init_cursor(trans, cursor,
607 * Release a reference on an inode, flush as requested.
609 * On the last reference we queue the inode to the flusher for its final
613 hammer_rel_inode(struct hammer_inode *ip, int flush)
615 hammer_mount_t hmp = ip->hmp;
618 * Handle disposition when dropping the last ref.
621 if (ip->lock.refs == 1) {
623 * Determine whether on-disk action is needed for
624 * the inode's final disposition.
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);
636 hammer_flush_inode(ip, 0);
639 * The inode still has multiple refs, try to drop
642 KKASSERT(ip->lock.refs >= 1);
643 if (ip->lock.refs > 1) {
644 hammer_unref(&ip->lock);
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.
656 if (++hmp->reclaim_count > 256) {
657 hmp->reclaim_count = 0;
658 hammer_flusher_async(hmp);
663 * Unload and destroy the specified inode. Must be called with one remaining
664 * reference. The reference is disposed of.
666 * This can only be called in the context of the flusher.
669 hammer_unload_inode(struct hammer_inode *ip)
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);
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));
684 RB_REMOVE(hammer_ino_rb_tree, &ip->hmp->rb_inos_root, ip);
686 hammer_uncache_node(&ip->cache[0]);
687 hammer_uncache_node(&ip->cache[1]);
689 hammer_clear_objid(ip);
690 --hammer_count_inodes;
697 * A transaction has modified an inode, requiring updates as specified by
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
708 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
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);
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.
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.
728 hammer_flush_inode(hammer_inode_t ip, int flags)
730 hammer_record_t depend;
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.
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);
747 * Our flush action will depend on the current state.
749 switch(ip->flush_state) {
750 case HAMMER_FST_IDLE:
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.
756 hammer_flush_inode_core(ip, flags);
758 case HAMMER_FST_SETUP:
760 * Recurse upwards through dependancies via target_list
761 * and start their flusher actions going if possible.
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.
768 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
769 r = hammer_setup_parent_inodes(depend);
770 if (r < 0 && good == 0)
777 * We can continue if good >= 0. Determine how many records
778 * under our inode can be flushed (and mark them).
781 hammer_flush_inode_core(ip, flags);
783 ip->flags |= HAMMER_INODE_REFLUSH;
784 if (flags & HAMMER_FLUSH_SIGNAL) {
785 ip->flags |= HAMMER_INODE_RESIGNAL;
786 hammer_flusher_async(ip->hmp);
792 * We are already flushing, flag the inode to reflush
793 * if needed after it completes its current flush.
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);
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.
810 * Return 1 if the record gives us connectivity
812 * Return 0 if the record is not relevant
814 * Return -1 if we can't resolve the dependancy and there is no connectivity.
817 hammer_setup_parent_inodes(hammer_record_t record)
819 hammer_mount_t hmp = record->ip->hmp;
820 hammer_record_t depend;
824 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
828 * If the record is already flushing, is it in our flush group?
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).
835 if (record->flush_state == HAMMER_FST_FLUSH) {
836 if (record->flush_group != hmp->flusher_next) {
837 ip->flags |= HAMMER_INODE_REFLUSH;
840 if (record->type == HAMMER_MEM_RECORD_ADD)
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.
850 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
853 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
854 r = hammer_setup_parent_inodes(depend);
855 if (r < 0 && good == 0)
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.
867 ip->flags |= HAMMER_INODE_REFLUSH;
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.
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);
883 if (record->type == HAMMER_MEM_RECORD_DEL &&
884 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
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.
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.
894 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
898 if (ip->flush_group == ip->hmp->flusher_next) {
900 * This is the record we wanted to synchronize.
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)
909 * A general or delete-on-disk record does not contribute
910 * to our visibility. We can still flush it, however.
915 * We couldn't resolve the dependancies, request that the
916 * inode be flushed when the dependancies can be resolved.
918 ip->flags |= HAMMER_INODE_REFLUSH;
924 * This is the core routine placing an inode into the FST_FLUSH state.
927 hammer_flush_inode_core(hammer_inode_t ip, int flags)
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.
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;
944 * We need to be able to vfsync/truncate from the backend.
946 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
947 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
948 ip->flags |= HAMMER_INODE_VHELD;
953 * Figure out how many in-memory records we can actually flush
954 * (not including inode meta-data, buffers, etc).
956 if (flags & HAMMER_FLUSH_RECURSION) {
959 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
960 hammer_setup_child_callback, NULL);
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.
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;
977 if (flags & HAMMER_FLUSH_SIGNAL) {
978 ip->flags |= HAMMER_INODE_RESIGNAL;
979 hammer_flusher_async(ip->hmp);
981 if (--ip->hmp->flusher_lock == 0)
982 wakeup(&ip->hmp->flusher_lock);
988 * Snapshot the state of the inode for the backend flusher.
990 * The truncation must be retained in the frontend until after
991 * we've actually performed the record deletion.
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.
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;
1004 * The flusher list inherits our inode and reference.
1006 TAILQ_INSERT_TAIL(&ip->hmp->flush_list, ip, flush_entry);
1007 if (--ip->hmp->flusher_lock == 0)
1008 wakeup(&ip->hmp->flusher_lock);
1010 if (flags & HAMMER_FLUSH_SIGNAL)
1011 hammer_flusher_async(ip->hmp);
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.
1019 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1022 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1023 * the caller from shortcutting the flush.
1026 hammer_setup_child_callback(hammer_record_t rec, void *data)
1028 hammer_inode_t target_ip;
1033 * If the record has been deleted by the backend (it's being held
1034 * by the frontend in a race), just ignore it.
1036 if (rec->flags & HAMMER_RECF_DELETED_BE)
1040 * If the record is in an idle state it has no dependancies and
1046 switch(rec->flush_state) {
1047 case HAMMER_FST_IDLE:
1049 * Record has no setup dependancy, we can flush it.
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);
1057 case HAMMER_FST_SETUP:
1059 * Record has a setup dependancy. Try to include the
1060 * target ip in the flush.
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.
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) {
1071 * If the target IP is already flushing in our group
1072 * we are golden, otherwise make sure the target
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);
1081 target_ip->flags |= HAMMER_INODE_REFLUSH;
1083 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
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.
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);
1098 * General or delete-on-disk record.
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.
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);
1114 case HAMMER_FST_FLUSH:
1116 * Record already associated with a flush group. It had
1119 KKASSERT(rec->flush_group == ip->flush_group);
1127 * Wait for a previously queued flush to complete
1130 hammer_wait_inode(hammer_inode_t ip)
1132 while (ip->flush_state != HAMMER_FST_IDLE) {
1133 ip->flags |= HAMMER_INODE_FLUSHW;
1134 tsleep(&ip->flags, 0, "hmrwin", 0);
1139 * Called by the backend code when a flush has been completed.
1140 * The inode has already been removed from the flush list.
1142 * A pipelined flush can occur, in which case we must re-enter the
1143 * inode on the list and re-copy its fields.
1146 hammer_flush_inode_done(hammer_inode_t ip)
1151 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1154 * Allow BIOs to queue to the inode's primary bioq again.
1156 ip->flags &= ~HAMMER_INODE_WRITE_ALT;
1159 * Merge left-over flags back into the frontend and fix the state.
1161 ip->flags |= ip->sync_flags;
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.
1167 if (ip->ino_rec.ino_nlinks != ip->sync_ino_rec.ino_nlinks)
1168 ip->flags |= HAMMER_INODE_RDIRTY;
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.
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);
1179 * Fix up the dirty buffer status.
1181 if (TAILQ_FIRST(&ip->bio_list) ||
1182 (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree))) {
1183 ip->flags |= HAMMER_INODE_BUFS;
1187 * Re-set the XDIRTY flag if some of the inode's in-memory records
1188 * could not be flushed.
1190 if (RB_ROOT(&ip->rec_tree))
1191 ip->flags |= HAMMER_INODE_XDIRTY;
1194 * Do not lose track of inodes which no longer have vnode
1195 * assocations, otherwise they may never get flushed again.
1197 if ((ip->flags & HAMMER_INODE_MODMASK) && ip->vp == NULL)
1198 ip->flags |= HAMMER_INODE_REFLUSH;
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
1205 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
1206 ip->flush_state = HAMMER_FST_IDLE;
1209 ip->flush_state = HAMMER_FST_SETUP;
1213 * Clean up the vnode ref
1215 if (ip->flags & HAMMER_INODE_VHELD) {
1216 ip->flags &= ~HAMMER_INODE_VHELD;
1221 * If the frontend made more changes and requested another flush,
1222 * then try to get it running.
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);
1230 hammer_flush_inode(ip, 0);
1235 * Finally, if the frontend is waiting for a flush to complete,
1238 if (ip->flush_state != HAMMER_FST_FLUSH) {
1239 if (ip->flags & HAMMER_INODE_FLUSHW) {
1240 ip->flags &= ~HAMMER_INODE_FLUSHW;
1245 hammer_rel_inode(ip, 0);
1249 * Called from hammer_sync_inode() to synchronize in-memory records
1253 hammer_sync_record_callback(hammer_record_t record, void *data)
1255 hammer_cursor_t cursor = data;
1256 hammer_transaction_t trans = cursor->trans;
1260 * Skip records that do not belong to the current flush.
1262 if (record->flush_state != HAMMER_FST_FLUSH)
1264 KKASSERT((record->flags & HAMMER_RECF_DELETED_BE) == 0);
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);
1272 KKASSERT(record->flush_group == record->ip->flush_group);
1275 * Interlock the record using the BE flag. Once BE is set the
1276 * frontend cannot change the state of FE.
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.
1282 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1283 record->flags |= HAMMER_RECF_INTERLOCK_BE;
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.
1293 if (record->flags & HAMMER_RECF_DELETED_FE) {
1294 if (record->type == HAMMER_MEM_RECORD_ADD) {
1295 record->flags |= HAMMER_RECF_CONVERT_DELETE;
1297 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
1303 * Assign the create_tid for new records. Deletions already
1304 * have the record's entire key properly set up.
1306 if (record->type != HAMMER_MEM_RECORD_DEL)
1307 record->rec.inode.base.base.create_tid = trans->tid;
1309 error = hammer_ip_sync_record_cursor(cursor, record);
1310 if (error != EDEADLK)
1312 hammer_done_cursor(cursor);
1313 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
1318 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
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");
1328 hammer_flush_record_done(record, error);
1333 * XXX error handling
1336 hammer_sync_inode(hammer_inode_t ip)
1338 struct hammer_transaction trans;
1339 struct hammer_cursor cursor;
1341 hammer_record_t depend;
1342 hammer_record_t next;
1343 int error, tmp_error;
1346 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
1349 hammer_start_transaction_fls(&trans, ip->hmp);
1350 error = hammer_init_cursor(&trans, &cursor, &ip->cache[0], ip);
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.
1359 * Records which are in our flush group can be unlinked from our
1360 * inode now, potentially allowing the inode to be physically
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) {
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
1377 * If the ADD was not deleted by the frontend we
1378 * can remove the dependancy from our target_list.
1380 if (depend->flags & HAMMER_RECF_DELETED_FE) {
1383 TAILQ_REMOVE(&ip->target_list, depend,
1385 depend->target_ip = NULL;
1387 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
1389 * Not part of our flush group
1391 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
1392 switch(depend->type) {
1393 case HAMMER_MEM_RECORD_ADD:
1396 case HAMMER_MEM_RECORD_DEL:
1406 * Set dirty if we had to modify the link count.
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;
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.
1418 if (ip->flags & HAMMER_INODE_VHELD)
1419 error = vfsync(ip->vp, MNT_NOWAIT, 1, NULL, NULL);
1420 ip->flags |= HAMMER_INODE_WRITE_ALT;
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
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).
1433 if (ip->flags & HAMMER_INODE_TRUNCATED) {
1434 ip->sync_trunc_off = ip->trunc_off;
1435 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
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;
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.
1447 * We don't flush pending frontend data buffers until after we've
1448 * dealth with the truncation.
1450 * Don't bother if the inode is or has been deleted.
1452 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
1454 * Interlock trunc_off. The VOP front-end may continue to
1455 * make adjustments to it while we are blocked.
1458 off_t aligned_trunc_off;
1460 trunc_off = ip->sync_trunc_off;
1461 aligned_trunc_off = (trunc_off + HAMMER_BUFMASK) &
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.
1471 error = hammer_ip_delete_range(&cursor, ip,
1473 0x7FFFFFFFFFFFFFFFLL);
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;
1486 * Now sync related records. These will typically be directory
1487 * entries or delete-on-disk records.
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.
1494 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1495 hammer_sync_record_callback, &cursor);
1501 ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
1505 * If we are deleting the inode the frontend had better not have
1506 * any active references on elements making up the inode.
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) {
1515 ip->flags |= HAMMER_INODE_DELETED;
1516 error = hammer_ip_delete_range_all(&cursor, ip, &count1);
1518 ip->sync_flags &= ~HAMMER_INODE_DELETING;
1519 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
1520 KKASSERT(RB_EMPTY(&ip->rec_tree));
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.
1527 ip->ino_rec.base.base.delete_tid = trans.tid;
1528 ip->sync_ino_rec.base.base.delete_tid = trans.tid;
1531 * Adjust the inode count in the volume header
1533 hammer_modify_volume_field(&trans, trans.rootvol,
1535 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1536 hammer_modify_volume_done(trans.rootvol);
1538 ip->flags &= ~HAMMER_INODE_DELETED;
1539 Debugger("hammer_ip_delete_range_all errored");
1544 * Flush any queued BIOs. These will just biodone() the IO's if
1545 * the inode has been deleted.
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);
1553 ip->sync_flags &= ~HAMMER_INODE_BUFS;
1556 Debugger("RB_SCAN errored");
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.
1562 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
1563 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
1565 * If deleted and on-disk, don't set any additional flags.
1566 * the delete flag takes care of things.
1568 * Clear flags which may have been set by the frontend.
1570 ip->sync_flags &= ~(HAMMER_INODE_RDIRTY|HAMMER_INODE_DDIRTY|
1571 HAMMER_INODE_XDIRTY|HAMMER_INODE_ITIMES|
1572 HAMMER_INODE_DELETING);
1574 case HAMMER_INODE_DELETED:
1576 * Take care of the case where a deleted inode was never
1577 * flushed to the disk in the first place.
1579 * Clear flags which may have been set by the frontend.
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);
1593 case HAMMER_INODE_ONDISK:
1595 * If already on-disk, do not set any additional flags.
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.
1604 * Set create_tid in both the frontend and backend
1605 * copy of the inode record.
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;
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.
1617 * If DELETED is set hammer_update_inode() will delete the existing
1618 * record without writing out a new one.
1620 * If *ONLY* the ITIMES flag is set we can update the record in-place.
1622 if (ip->flags & HAMMER_INODE_DELETED) {
1623 error = hammer_update_inode(&cursor, ip);
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);
1629 if (ip->sync_flags & (HAMMER_INODE_RDIRTY | HAMMER_INODE_DDIRTY |
1630 HAMMER_INODE_ITIMES)) {
1631 error = hammer_update_inode(&cursor, ip);
1634 Debugger("hammer_update_itimes/inode errored");
1637 * Save the TID we used to sync the inode with to make sure we
1638 * do not improperly reuse it.
1640 hammer_done_cursor(&cursor);
1641 hammer_done_transaction(&trans);
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.
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.
1654 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
1659 * Set the DELETING flag when the link count drops to 0 and the
1660 * OS no longer has any opens on the inode.
1662 * The backend will clear DELETING (a mod flag) and set DELETED
1663 * (a state flag) when it is actually able to perform the
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;
1673 if (hammer_get_vnode(ip, &vp) != 0)
1677 vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
1678 vnode_pager_setsize(ip->vp, 0);
1687 * Re-test an inode when a dependancy had gone away to see if we
1688 * can chain flush it.
1691 hammer_test_inode(hammer_inode_t ip)
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);
1700 hammer_flush_inode(ip, 0);
1702 hammer_rel_inode(ip, 0);