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 1.114 2008/09/24 00:53:51 dillon Exp $
38 #include <vm/vm_extern.h>
40 static int hammer_unload_inode(struct hammer_inode *ip);
41 static void hammer_free_inode(hammer_inode_t ip);
42 static void hammer_flush_inode_core(hammer_inode_t ip,
43 hammer_flush_group_t flg, int flags);
44 static int hammer_setup_child_callback(hammer_record_t rec, void *data);
46 static int hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
48 static int hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
49 hammer_flush_group_t flg);
50 static int hammer_setup_parent_inodes_helper(hammer_record_t record,
51 int depth, hammer_flush_group_t flg);
52 static void hammer_inode_wakereclaims(hammer_inode_t ip);
55 extern struct hammer_inode *HammerTruncIp;
59 * RB-Tree support for inode structures
62 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
64 if (ip1->obj_localization < ip2->obj_localization)
66 if (ip1->obj_localization > ip2->obj_localization)
68 if (ip1->obj_id < ip2->obj_id)
70 if (ip1->obj_id > ip2->obj_id)
72 if (ip1->obj_asof < ip2->obj_asof)
74 if (ip1->obj_asof > ip2->obj_asof)
80 * RB-Tree support for inode structures / special LOOKUP_INFO
83 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
85 if (info->obj_localization < ip->obj_localization)
87 if (info->obj_localization > ip->obj_localization)
89 if (info->obj_id < ip->obj_id)
91 if (info->obj_id > ip->obj_id)
93 if (info->obj_asof < ip->obj_asof)
95 if (info->obj_asof > ip->obj_asof)
101 * Used by hammer_scan_inode_snapshots() to locate all of an object's
102 * snapshots. Note that the asof field is not tested, which we can get
103 * away with because it is the lowest-priority field.
106 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
108 hammer_inode_info_t info = data;
110 if (ip->obj_localization > info->obj_localization)
112 if (ip->obj_localization < info->obj_localization)
114 if (ip->obj_id > info->obj_id)
116 if (ip->obj_id < info->obj_id)
122 * Used by hammer_unload_pseudofs() to locate all inodes associated with
126 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
128 u_int32_t localization = *(u_int32_t *)data;
129 if (ip->obj_localization > localization)
131 if (ip->obj_localization < localization)
137 * RB-Tree support for pseudofs structures
140 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
142 if (p1->localization < p2->localization)
144 if (p1->localization > p2->localization)
150 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
151 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
152 hammer_inode_info_cmp, hammer_inode_info_t);
153 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
154 hammer_pfs_rb_compare, u_int32_t, localization);
157 * The kernel is not actively referencing this vnode but is still holding
160 * This is called from the frontend.
165 hammer_vop_inactive(struct vop_inactive_args *ap)
167 struct hammer_inode *ip = VTOI(ap->a_vp);
178 * If the inode no longer has visibility in the filesystem try to
179 * recycle it immediately, even if the inode is dirty. Recycling
180 * it quickly allows the system to reclaim buffer cache and VM
181 * resources which can matter a lot in a heavily loaded system.
183 * This can deadlock in vfsync() if we aren't careful.
185 * Do not queue the inode to the flusher if we still have visibility,
186 * otherwise namespace calls such as chmod will unnecessarily generate
187 * multiple inode updates.
189 if (ip->ino_data.nlinks == 0) {
191 hammer_inode_unloadable_check(ip, 0);
192 if (ip->flags & HAMMER_INODE_MODMASK)
193 hammer_flush_inode(ip, 0);
201 * Release the vnode association. This is typically (but not always)
202 * the last reference on the inode.
204 * Once the association is lost we are on our own with regards to
205 * flushing the inode.
208 hammer_vop_reclaim(struct vop_reclaim_args *ap)
210 struct hammer_inode *ip;
216 if ((ip = vp->v_data) != NULL) {
221 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
222 ++hammer_count_reclaiming;
223 ++hmp->inode_reclaims;
224 ip->flags |= HAMMER_INODE_RECLAIM;
226 hammer_rel_inode(ip, 1);
232 * Return a locked vnode for the specified inode. The inode must be
233 * referenced but NOT LOCKED on entry and will remain referenced on
236 * Called from the frontend.
239 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
249 if ((vp = ip->vp) == NULL) {
250 error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
253 hammer_lock_ex(&ip->lock);
254 if (ip->vp != NULL) {
255 hammer_unlock(&ip->lock);
261 hammer_ref(&ip->lock);
265 obj_type = ip->ino_data.obj_type;
266 vp->v_type = hammer_get_vnode_type(obj_type);
268 hammer_inode_wakereclaims(ip);
270 switch(ip->ino_data.obj_type) {
271 case HAMMER_OBJTYPE_CDEV:
272 case HAMMER_OBJTYPE_BDEV:
273 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
274 addaliasu(vp, ip->ino_data.rmajor,
275 ip->ino_data.rminor);
277 case HAMMER_OBJTYPE_FIFO:
278 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
280 case HAMMER_OBJTYPE_REGFILE:
287 * Only mark as the root vnode if the ip is not
288 * historical, otherwise the VFS cache will get
289 * confused. The other half of the special handling
290 * is in hammer_vop_nlookupdotdot().
292 * Pseudo-filesystem roots can be accessed via
293 * non-root filesystem paths and setting VROOT may
294 * confuse the namecache. Set VPFSROOT instead.
296 if (ip->obj_id == HAMMER_OBJID_ROOT &&
297 ip->obj_asof == hmp->asof) {
298 if (ip->obj_localization == 0)
299 vsetflags(vp, VROOT);
301 vsetflags(vp, VPFSROOT);
304 vp->v_data = (void *)ip;
305 /* vnode locked by getnewvnode() */
306 /* make related vnode dirty if inode dirty? */
307 hammer_unlock(&ip->lock);
308 if (vp->v_type == VREG)
309 vinitvmio(vp, ip->ino_data.size);
314 * loop if the vget fails (aka races), or if the vp
315 * no longer matches ip->vp.
317 if (vget(vp, LK_EXCLUSIVE) == 0) {
328 * Locate all copies of the inode for obj_id compatible with the specified
329 * asof, reference, and issue the related call-back. This routine is used
330 * for direct-io invalidation and does not create any new inodes.
333 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
334 int (*callback)(hammer_inode_t ip, void *data),
337 hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
338 hammer_inode_info_cmp_all_history,
343 * Acquire a HAMMER inode. The returned inode is not locked. These functions
344 * do not attach or detach the related vnode (use hammer_get_vnode() for
347 * The flags argument is only applied for newly created inodes, and only
348 * certain flags are inherited.
350 * Called from the frontend.
352 struct hammer_inode *
353 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
354 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
355 int flags, int *errorp)
357 hammer_mount_t hmp = trans->hmp;
358 struct hammer_node_cache *cachep;
359 struct hammer_inode_info iinfo;
360 struct hammer_cursor cursor;
361 struct hammer_inode *ip;
365 * Determine if we already have an inode cached. If we do then
368 * If we find an inode with no vnode we have to mark the
369 * transaction such that hammer_inode_waitreclaims() is
370 * called later on to avoid building up an infinite number
371 * of inodes. Otherwise we can continue to * add new inodes
372 * faster then they can be disposed of, even with the tsleep
375 * If we find a dummy inode we return a failure so dounlink
376 * (which does another lookup) doesn't try to mess with the
377 * link count. hammer_vop_nresolve() uses hammer_get_dummy_inode()
378 * to ref dummy inodes.
380 iinfo.obj_id = obj_id;
381 iinfo.obj_asof = asof;
382 iinfo.obj_localization = localization;
384 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
386 if (ip->flags & HAMMER_INODE_DUMMY) {
390 hammer_ref(&ip->lock);
396 * Allocate a new inode structure and deal with races later.
398 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
399 ++hammer_count_inodes;
402 ip->obj_asof = iinfo.obj_asof;
403 ip->obj_localization = localization;
405 ip->flags = flags & HAMMER_INODE_RO;
406 ip->cache[0].ip = ip;
407 ip->cache[1].ip = ip;
408 ip->cache[2].ip = ip;
409 ip->cache[3].ip = ip;
410 ip->redo_count = SIZE_T_MAX;
412 ip->flags |= HAMMER_INODE_RO;
413 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
414 0x7FFFFFFFFFFFFFFFLL;
415 RB_INIT(&ip->rec_tree);
416 TAILQ_INIT(&ip->target_list);
417 hammer_ref(&ip->lock);
420 * Locate the on-disk inode. If this is a PFS root we always
421 * access the current version of the root inode and (if it is not
422 * a master) always access information under it with a snapshot
425 * We cache recent inode lookups in this directory in dip->cache[2].
426 * If we can't find it we assume the inode we are looking for is
427 * close to the directory inode.
432 if (dip->cache[2].node)
433 cachep = &dip->cache[2];
435 cachep = &dip->cache[0];
437 hammer_init_cursor(trans, &cursor, cachep, NULL);
438 cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
439 cursor.key_beg.obj_id = ip->obj_id;
440 cursor.key_beg.key = 0;
441 cursor.key_beg.create_tid = 0;
442 cursor.key_beg.delete_tid = 0;
443 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
444 cursor.key_beg.obj_type = 0;
446 cursor.asof = iinfo.obj_asof;
447 cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
450 *errorp = hammer_btree_lookup(&cursor);
451 if (*errorp == EDEADLK) {
452 hammer_done_cursor(&cursor);
457 * On success the B-Tree lookup will hold the appropriate
458 * buffer cache buffers and provide a pointer to the requested
459 * information. Copy the information to the in-memory inode
460 * and cache the B-Tree node to improve future operations.
463 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
464 ip->ino_data = cursor.data->inode;
467 * cache[0] tries to cache the location of the object inode.
468 * The assumption is that it is near the directory inode.
470 * cache[1] tries to cache the location of the object data.
471 * We might have something in the governing directory from
472 * scan optimizations (see the strategy code in
475 * We update dip->cache[2], if possible, with the location
476 * of the object inode for future directory shortcuts.
478 hammer_cache_node(&ip->cache[0], cursor.node);
480 if (dip->cache[3].node) {
481 hammer_cache_node(&ip->cache[1],
484 hammer_cache_node(&dip->cache[2], cursor.node);
488 * The file should not contain any data past the file size
489 * stored in the inode. Setting save_trunc_off to the
490 * file size instead of max reduces B-Tree lookup overheads
491 * on append by allowing the flusher to avoid checking for
494 ip->save_trunc_off = ip->ino_data.size;
497 * Locate and assign the pseudofs management structure to
500 if (dip && dip->obj_localization == ip->obj_localization) {
501 ip->pfsm = dip->pfsm;
502 hammer_ref(&ip->pfsm->lock);
504 ip->pfsm = hammer_load_pseudofs(trans,
505 ip->obj_localization,
507 *errorp = 0; /* ignore ENOENT */
512 * The inode is placed on the red-black tree and will be synced to
513 * the media when flushed or by the filesystem sync. If this races
514 * another instantiation/lookup the insertion will fail.
517 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
518 hammer_free_inode(ip);
519 hammer_done_cursor(&cursor);
522 ip->flags |= HAMMER_INODE_ONDISK;
524 if (ip->flags & HAMMER_INODE_RSV_INODES) {
525 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
529 hammer_free_inode(ip);
532 hammer_done_cursor(&cursor);
533 trans->flags |= HAMMER_TRANSF_NEWINODE;
538 * Get a dummy inode to placemark a broken directory entry.
540 struct hammer_inode *
541 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
542 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
543 int flags, int *errorp)
545 hammer_mount_t hmp = trans->hmp;
546 struct hammer_inode_info iinfo;
547 struct hammer_inode *ip;
550 * Determine if we already have an inode cached. If we do then
553 * If we find an inode with no vnode we have to mark the
554 * transaction such that hammer_inode_waitreclaims() is
555 * called later on to avoid building up an infinite number
556 * of inodes. Otherwise we can continue to * add new inodes
557 * faster then they can be disposed of, even with the tsleep
560 * If we find a non-fake inode we return an error. Only fake
561 * inodes can be returned by this routine.
563 iinfo.obj_id = obj_id;
564 iinfo.obj_asof = asof;
565 iinfo.obj_localization = localization;
568 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
570 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
574 hammer_ref(&ip->lock);
579 * Allocate a new inode structure and deal with races later.
581 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
582 ++hammer_count_inodes;
585 ip->obj_asof = iinfo.obj_asof;
586 ip->obj_localization = localization;
588 ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
589 ip->cache[0].ip = ip;
590 ip->cache[1].ip = ip;
591 ip->cache[2].ip = ip;
592 ip->cache[3].ip = ip;
593 ip->redo_count = SIZE_T_MAX;
594 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
595 0x7FFFFFFFFFFFFFFFLL;
596 RB_INIT(&ip->rec_tree);
597 TAILQ_INIT(&ip->target_list);
598 hammer_ref(&ip->lock);
601 * Populate the dummy inode. Leave everything zero'd out.
603 * (ip->ino_leaf and ip->ino_data)
605 * Make the dummy inode a FIFO object which most copy programs
606 * will properly ignore.
608 ip->save_trunc_off = ip->ino_data.size;
609 ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
612 * Locate and assign the pseudofs management structure to
615 if (dip && dip->obj_localization == ip->obj_localization) {
616 ip->pfsm = dip->pfsm;
617 hammer_ref(&ip->pfsm->lock);
619 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
621 *errorp = 0; /* ignore ENOENT */
625 * The inode is placed on the red-black tree and will be synced to
626 * the media when flushed or by the filesystem sync. If this races
627 * another instantiation/lookup the insertion will fail.
629 * NOTE: Do not set HAMMER_INODE_ONDISK. The inode is a fake.
632 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
633 hammer_free_inode(ip);
637 if (ip->flags & HAMMER_INODE_RSV_INODES) {
638 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
641 hammer_free_inode(ip);
644 trans->flags |= HAMMER_TRANSF_NEWINODE;
649 * Return a referenced inode only if it is in our inode cache.
651 * Dummy inodes do not count.
653 struct hammer_inode *
654 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
655 hammer_tid_t asof, u_int32_t localization)
657 hammer_mount_t hmp = trans->hmp;
658 struct hammer_inode_info iinfo;
659 struct hammer_inode *ip;
661 iinfo.obj_id = obj_id;
662 iinfo.obj_asof = asof;
663 iinfo.obj_localization = localization;
665 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
667 if (ip->flags & HAMMER_INODE_DUMMY)
670 hammer_ref(&ip->lock);
676 * Create a new filesystem object, returning the inode in *ipp. The
677 * returned inode will be referenced. The inode is created in-memory.
679 * If pfsm is non-NULL the caller wishes to create the root inode for
683 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
685 hammer_inode_t dip, const char *name, int namelen,
686 hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
697 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
698 ++hammer_count_inodes;
700 trans->flags |= HAMMER_TRANSF_NEWINODE;
703 KKASSERT(pfsm->localization != 0);
704 ip->obj_id = HAMMER_OBJID_ROOT;
705 ip->obj_localization = pfsm->localization;
707 KKASSERT(dip != NULL);
708 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
709 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
710 ip->obj_localization = dip->obj_localization;
713 KKASSERT(ip->obj_id != 0);
714 ip->obj_asof = hmp->asof;
716 ip->flush_state = HAMMER_FST_IDLE;
717 ip->flags = HAMMER_INODE_DDIRTY |
718 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
719 ip->cache[0].ip = ip;
720 ip->cache[1].ip = ip;
721 ip->cache[2].ip = ip;
722 ip->cache[3].ip = ip;
723 ip->redo_count = SIZE_T_MAX;
725 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
726 /* ip->save_trunc_off = 0; (already zero) */
727 RB_INIT(&ip->rec_tree);
728 TAILQ_INIT(&ip->target_list);
730 ip->ino_data.atime = trans->time;
731 ip->ino_data.mtime = trans->time;
732 ip->ino_data.size = 0;
733 ip->ino_data.nlinks = 0;
736 * A nohistory designator on the parent directory is inherited by
737 * the child. We will do this even for pseudo-fs creation... the
738 * sysad can turn it off.
741 ip->ino_data.uflags = dip->ino_data.uflags &
742 (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
745 ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
746 ip->ino_leaf.base.localization = ip->obj_localization +
747 HAMMER_LOCALIZE_INODE;
748 ip->ino_leaf.base.obj_id = ip->obj_id;
749 ip->ino_leaf.base.key = 0;
750 ip->ino_leaf.base.create_tid = 0;
751 ip->ino_leaf.base.delete_tid = 0;
752 ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
753 ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
755 ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
756 ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
757 ip->ino_data.mode = vap->va_mode;
758 ip->ino_data.ctime = trans->time;
761 * If we are running version 2 or greater directory entries are
762 * inode-localized instead of data-localized.
764 if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
765 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
766 ip->ino_data.cap_flags |=
767 HAMMER_INODE_CAP_DIR_LOCAL_INO;
772 * Setup the ".." pointer. This only needs to be done for directories
773 * but we do it for all objects as a recovery aid.
776 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
779 * The parent_obj_localization field only applies to pseudo-fs roots.
780 * XXX this is no longer applicable, PFSs are no longer directly
781 * tied into the parent's directory structure.
783 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
784 ip->obj_id == HAMMER_OBJID_ROOT) {
785 ip->ino_data.ext.obj.parent_obj_localization =
786 dip->obj_localization;
790 switch(ip->ino_leaf.base.obj_type) {
791 case HAMMER_OBJTYPE_CDEV:
792 case HAMMER_OBJTYPE_BDEV:
793 ip->ino_data.rmajor = vap->va_rmajor;
794 ip->ino_data.rminor = vap->va_rminor;
801 * Calculate default uid/gid and overwrite with information from
805 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
806 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
807 xuid, cred, &vap->va_mode);
811 ip->ino_data.mode = vap->va_mode;
813 if (vap->va_vaflags & VA_UID_UUID_VALID)
814 ip->ino_data.uid = vap->va_uid_uuid;
815 else if (vap->va_uid != (uid_t)VNOVAL)
816 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
818 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
820 if (vap->va_vaflags & VA_GID_UUID_VALID)
821 ip->ino_data.gid = vap->va_gid_uuid;
822 else if (vap->va_gid != (gid_t)VNOVAL)
823 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
825 ip->ino_data.gid = dip->ino_data.gid;
827 hammer_ref(&ip->lock);
831 hammer_ref(&pfsm->lock);
833 } else if (dip->obj_localization == ip->obj_localization) {
834 ip->pfsm = dip->pfsm;
835 hammer_ref(&ip->pfsm->lock);
838 ip->pfsm = hammer_load_pseudofs(trans,
839 ip->obj_localization,
841 error = 0; /* ignore ENOENT */
845 hammer_free_inode(ip);
847 } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
848 panic("hammer_create_inode: duplicate obj_id %llx",
849 (long long)ip->obj_id);
851 hammer_free_inode(ip);
858 * Final cleanup / freeing of an inode structure
861 hammer_free_inode(hammer_inode_t ip)
863 struct hammer_mount *hmp;
866 KKASSERT(ip->lock.refs == 1);
867 hammer_uncache_node(&ip->cache[0]);
868 hammer_uncache_node(&ip->cache[1]);
869 hammer_uncache_node(&ip->cache[2]);
870 hammer_uncache_node(&ip->cache[3]);
871 hammer_inode_wakereclaims(ip);
873 hammer_clear_objid(ip);
874 --hammer_count_inodes;
877 hammer_rel_pseudofs(hmp, ip->pfsm);
880 kfree(ip, hmp->m_inodes);
885 * Retrieve pseudo-fs data. NULL will never be returned.
887 * If an error occurs *errorp will be set and a default template is returned,
888 * otherwise *errorp is set to 0. Typically when an error occurs it will
891 hammer_pseudofs_inmem_t
892 hammer_load_pseudofs(hammer_transaction_t trans,
893 u_int32_t localization, int *errorp)
895 hammer_mount_t hmp = trans->hmp;
897 hammer_pseudofs_inmem_t pfsm;
898 struct hammer_cursor cursor;
902 pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
904 hammer_ref(&pfsm->lock);
910 * PFS records are stored in the root inode (not the PFS root inode,
911 * but the real root). Avoid an infinite recursion if loading
912 * the PFS for the real root.
915 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
917 HAMMER_DEF_LOCALIZATION, 0, errorp);
922 pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
923 pfsm->localization = localization;
924 pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
925 pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
927 hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
928 cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
929 HAMMER_LOCALIZE_MISC;
930 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
931 cursor.key_beg.create_tid = 0;
932 cursor.key_beg.delete_tid = 0;
933 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
934 cursor.key_beg.obj_type = 0;
935 cursor.key_beg.key = localization;
936 cursor.asof = HAMMER_MAX_TID;
937 cursor.flags |= HAMMER_CURSOR_ASOF;
940 *errorp = hammer_ip_lookup(&cursor);
942 *errorp = hammer_btree_lookup(&cursor);
944 *errorp = hammer_ip_resolve_data(&cursor);
946 if (cursor.data->pfsd.mirror_flags &
947 HAMMER_PFSD_DELETED) {
950 bytes = cursor.leaf->data_len;
951 if (bytes > sizeof(pfsm->pfsd))
952 bytes = sizeof(pfsm->pfsd);
953 bcopy(cursor.data, &pfsm->pfsd, bytes);
957 hammer_done_cursor(&cursor);
959 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
960 hammer_ref(&pfsm->lock);
962 hammer_rel_inode(ip, 0);
963 if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
964 kfree(pfsm, hmp->m_misc);
971 * Store pseudo-fs data. The backend will automatically delete any prior
972 * on-disk pseudo-fs data but we have to delete in-memory versions.
975 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
977 struct hammer_cursor cursor;
978 hammer_record_t record;
982 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
983 HAMMER_DEF_LOCALIZATION, 0, &error);
985 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
986 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
987 cursor.key_beg.localization = ip->obj_localization +
988 HAMMER_LOCALIZE_MISC;
989 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
990 cursor.key_beg.create_tid = 0;
991 cursor.key_beg.delete_tid = 0;
992 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
993 cursor.key_beg.obj_type = 0;
994 cursor.key_beg.key = pfsm->localization;
995 cursor.asof = HAMMER_MAX_TID;
996 cursor.flags |= HAMMER_CURSOR_ASOF;
999 * Replace any in-memory version of the record.
1001 error = hammer_ip_lookup(&cursor);
1002 if (error == 0 && hammer_cursor_inmem(&cursor)) {
1003 record = cursor.iprec;
1004 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1005 KKASSERT(cursor.deadlk_rec == NULL);
1006 hammer_ref(&record->lock);
1007 cursor.deadlk_rec = record;
1010 record->flags |= HAMMER_RECF_DELETED_FE;
1016 * Allocate replacement general record. The backend flush will
1017 * delete any on-disk version of the record.
1019 if (error == 0 || error == ENOENT) {
1020 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1021 record->type = HAMMER_MEM_RECORD_GENERAL;
1023 record->leaf.base.localization = ip->obj_localization +
1024 HAMMER_LOCALIZE_MISC;
1025 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1026 record->leaf.base.key = pfsm->localization;
1027 record->leaf.data_len = sizeof(pfsm->pfsd);
1028 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1029 error = hammer_ip_add_record(trans, record);
1031 hammer_done_cursor(&cursor);
1032 if (error == EDEADLK)
1034 hammer_rel_inode(ip, 0);
1039 * Create a root directory for a PFS if one does not alredy exist.
1041 * The PFS root stands alone so we must also bump the nlinks count
1042 * to prevent it from being destroyed on release.
1045 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1046 hammer_pseudofs_inmem_t pfsm)
1052 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1053 pfsm->localization, 0, &error);
1058 error = hammer_create_inode(trans, &vap, cred,
1062 ++ip->ino_data.nlinks;
1063 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1067 hammer_rel_inode(ip, 0);
1072 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1073 * if we are unable to disassociate all the inodes.
1077 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1081 hammer_ref(&ip->lock);
1082 if (ip->lock.refs == 2 && ip->vp)
1083 vclean_unlocked(ip->vp);
1084 if (ip->lock.refs == 1 && ip->vp == NULL)
1087 res = -1; /* stop, someone is using the inode */
1088 hammer_rel_inode(ip, 0);
1093 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1098 for (try = res = 0; try < 4; ++try) {
1099 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1100 hammer_inode_pfs_cmp,
1101 hammer_unload_pseudofs_callback,
1103 if (res == 0 && try > 1)
1105 hammer_flusher_sync(trans->hmp);
1114 * Release a reference on a PFS
1117 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1119 hammer_unref(&pfsm->lock);
1120 if (pfsm->lock.refs == 0) {
1121 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1122 kfree(pfsm, hmp->m_misc);
1127 * Called by hammer_sync_inode().
1130 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1132 hammer_transaction_t trans = cursor->trans;
1133 hammer_record_t record;
1141 * If the inode has a presence on-disk then locate it and mark
1142 * it deleted, setting DELONDISK.
1144 * The record may or may not be physically deleted, depending on
1145 * the retention policy.
1147 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1148 HAMMER_INODE_ONDISK) {
1149 hammer_normalize_cursor(cursor);
1150 cursor->key_beg.localization = ip->obj_localization +
1151 HAMMER_LOCALIZE_INODE;
1152 cursor->key_beg.obj_id = ip->obj_id;
1153 cursor->key_beg.key = 0;
1154 cursor->key_beg.create_tid = 0;
1155 cursor->key_beg.delete_tid = 0;
1156 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1157 cursor->key_beg.obj_type = 0;
1158 cursor->asof = ip->obj_asof;
1159 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1160 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1161 cursor->flags |= HAMMER_CURSOR_BACKEND;
1163 error = hammer_btree_lookup(cursor);
1164 if (hammer_debug_inode)
1165 kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1168 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1169 if (hammer_debug_inode)
1170 kprintf(" error %d\n", error);
1172 ip->flags |= HAMMER_INODE_DELONDISK;
1175 hammer_cache_node(&ip->cache[0], cursor->node);
1177 if (error == EDEADLK) {
1178 hammer_done_cursor(cursor);
1179 error = hammer_init_cursor(trans, cursor,
1181 if (hammer_debug_inode)
1182 kprintf("IPDED %p %d\n", ip, error);
1189 * Ok, write out the initial record or a new record (after deleting
1190 * the old one), unless the DELETED flag is set. This routine will
1191 * clear DELONDISK if it writes out a record.
1193 * Update our inode statistics if this is the first application of
1194 * the inode on-disk.
1196 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1198 * Generate a record and write it to the media. We clean-up
1199 * the state before releasing so we do not have to set-up
1202 record = hammer_alloc_mem_record(ip, 0);
1203 record->type = HAMMER_MEM_RECORD_INODE;
1204 record->flush_state = HAMMER_FST_FLUSH;
1205 record->leaf = ip->sync_ino_leaf;
1206 record->leaf.base.create_tid = trans->tid;
1207 record->leaf.data_len = sizeof(ip->sync_ino_data);
1208 record->leaf.create_ts = trans->time32;
1209 record->data = (void *)&ip->sync_ino_data;
1210 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1213 * If this flag is set we cannot sync the new file size
1214 * because we haven't finished related truncations. The
1215 * inode will be flushed in another flush group to finish
1218 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1219 ip->sync_ino_data.size != ip->ino_data.size) {
1221 ip->sync_ino_data.size = ip->ino_data.size;
1227 error = hammer_ip_sync_record_cursor(cursor, record);
1228 if (hammer_debug_inode)
1229 kprintf("GENREC %p rec %08x %d\n",
1230 ip, record->flags, error);
1231 if (error != EDEADLK)
1233 hammer_done_cursor(cursor);
1234 error = hammer_init_cursor(trans, cursor,
1236 if (hammer_debug_inode)
1237 kprintf("GENREC reinit %d\n", error);
1243 * Note: The record was never on the inode's record tree
1244 * so just wave our hands importantly and destroy it.
1246 record->flags |= HAMMER_RECF_COMMITTED;
1247 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1248 record->flush_state = HAMMER_FST_IDLE;
1249 ++ip->rec_generation;
1250 hammer_rel_mem_record(record);
1256 if (hammer_debug_inode)
1257 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1258 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1259 HAMMER_INODE_SDIRTY |
1260 HAMMER_INODE_ATIME |
1261 HAMMER_INODE_MTIME);
1262 ip->flags &= ~HAMMER_INODE_DELONDISK;
1264 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1267 * Root volume count of inodes
1269 hammer_sync_lock_sh(trans);
1270 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1271 hammer_modify_volume_field(trans,
1274 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1275 hammer_modify_volume_done(trans->rootvol);
1276 ip->flags |= HAMMER_INODE_ONDISK;
1277 if (hammer_debug_inode)
1278 kprintf("NOWONDISK %p\n", ip);
1280 hammer_sync_unlock(trans);
1285 * If the inode has been destroyed, clean out any left-over flags
1286 * that may have been set by the frontend.
1288 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1289 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1290 HAMMER_INODE_SDIRTY |
1291 HAMMER_INODE_ATIME |
1292 HAMMER_INODE_MTIME);
1298 * Update only the itimes fields.
1300 * ATIME can be updated without generating any UNDO. MTIME is updated
1301 * with UNDO so it is guaranteed to be synchronized properly in case of
1304 * Neither field is included in the B-Tree leaf element's CRC, which is how
1305 * we can get away with updating ATIME the way we do.
1308 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1310 hammer_transaction_t trans = cursor->trans;
1314 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1315 HAMMER_INODE_ONDISK) {
1319 hammer_normalize_cursor(cursor);
1320 cursor->key_beg.localization = ip->obj_localization +
1321 HAMMER_LOCALIZE_INODE;
1322 cursor->key_beg.obj_id = ip->obj_id;
1323 cursor->key_beg.key = 0;
1324 cursor->key_beg.create_tid = 0;
1325 cursor->key_beg.delete_tid = 0;
1326 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1327 cursor->key_beg.obj_type = 0;
1328 cursor->asof = ip->obj_asof;
1329 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1330 cursor->flags |= HAMMER_CURSOR_ASOF;
1331 cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1332 cursor->flags |= HAMMER_CURSOR_GET_DATA;
1333 cursor->flags |= HAMMER_CURSOR_BACKEND;
1335 error = hammer_btree_lookup(cursor);
1337 hammer_cache_node(&ip->cache[0], cursor->node);
1338 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1340 * Updating MTIME requires an UNDO. Just cover
1341 * both atime and mtime.
1343 hammer_sync_lock_sh(trans);
1344 hammer_modify_buffer(trans, cursor->data_buffer,
1345 HAMMER_ITIMES_BASE(&cursor->data->inode),
1346 HAMMER_ITIMES_BYTES);
1347 cursor->data->inode.atime = ip->sync_ino_data.atime;
1348 cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1349 hammer_modify_buffer_done(cursor->data_buffer);
1350 hammer_sync_unlock(trans);
1351 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1353 * Updating atime only can be done in-place with
1356 hammer_sync_lock_sh(trans);
1357 hammer_modify_buffer(trans, cursor->data_buffer,
1359 cursor->data->inode.atime = ip->sync_ino_data.atime;
1360 hammer_modify_buffer_done(cursor->data_buffer);
1361 hammer_sync_unlock(trans);
1363 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1365 if (error == EDEADLK) {
1366 hammer_done_cursor(cursor);
1367 error = hammer_init_cursor(trans, cursor,
1376 * Release a reference on an inode, flush as requested.
1378 * On the last reference we queue the inode to the flusher for its final
1382 hammer_rel_inode(struct hammer_inode *ip, int flush)
1384 /*hammer_mount_t hmp = ip->hmp;*/
1387 * Handle disposition when dropping the last ref.
1390 if (ip->lock.refs == 1) {
1392 * Determine whether on-disk action is needed for
1393 * the inode's final disposition.
1395 KKASSERT(ip->vp == NULL);
1396 hammer_inode_unloadable_check(ip, 0);
1397 if (ip->flags & HAMMER_INODE_MODMASK) {
1398 hammer_flush_inode(ip, 0);
1399 } else if (ip->lock.refs == 1) {
1400 hammer_unload_inode(ip);
1405 hammer_flush_inode(ip, 0);
1408 * The inode still has multiple refs, try to drop
1411 KKASSERT(ip->lock.refs >= 1);
1412 if (ip->lock.refs > 1) {
1413 hammer_unref(&ip->lock);
1421 * Unload and destroy the specified inode. Must be called with one remaining
1422 * reference. The reference is disposed of.
1424 * The inode must be completely clean.
1427 hammer_unload_inode(struct hammer_inode *ip)
1429 hammer_mount_t hmp = ip->hmp;
1431 KASSERT(ip->lock.refs == 1,
1432 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1433 KKASSERT(ip->vp == NULL);
1434 KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1435 KKASSERT(ip->cursor_ip_refs == 0);
1436 KKASSERT(hammer_notlocked(&ip->lock));
1437 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1439 KKASSERT(RB_EMPTY(&ip->rec_tree));
1440 KKASSERT(TAILQ_EMPTY(&ip->target_list));
1442 RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1444 hammer_free_inode(ip);
1449 * Called during unmounting if a critical error occured. The in-memory
1450 * inode and all related structures are destroyed.
1452 * If a critical error did not occur the unmount code calls the standard
1453 * release and asserts that the inode is gone.
1456 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1458 hammer_record_t rec;
1461 * Get rid of the inodes in-memory records, regardless of their
1462 * state, and clear the mod-mask.
1464 while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1465 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1466 rec->target_ip = NULL;
1467 if (rec->flush_state == HAMMER_FST_SETUP)
1468 rec->flush_state = HAMMER_FST_IDLE;
1470 while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1471 if (rec->flush_state == HAMMER_FST_FLUSH)
1472 --rec->flush_group->refs;
1474 hammer_ref(&rec->lock);
1475 KKASSERT(rec->lock.refs == 1);
1476 rec->flush_state = HAMMER_FST_IDLE;
1477 rec->flush_group = NULL;
1478 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1479 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1480 ++ip->rec_generation;
1481 hammer_rel_mem_record(rec);
1483 ip->flags &= ~HAMMER_INODE_MODMASK;
1484 ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1485 KKASSERT(ip->vp == NULL);
1488 * Remove the inode from any flush group, force it idle. FLUSH
1489 * and SETUP states have an inode ref.
1491 switch(ip->flush_state) {
1492 case HAMMER_FST_FLUSH:
1493 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1494 --ip->flush_group->refs;
1495 ip->flush_group = NULL;
1497 case HAMMER_FST_SETUP:
1498 hammer_unref(&ip->lock);
1499 ip->flush_state = HAMMER_FST_IDLE;
1501 case HAMMER_FST_IDLE:
1506 * There shouldn't be any associated vnode. The unload needs at
1507 * least one ref, if we do have a vp steal its ip ref.
1510 kprintf("hammer_destroy_inode_callback: Unexpected "
1511 "vnode association ip %p vp %p\n", ip, ip->vp);
1512 ip->vp->v_data = NULL;
1515 hammer_ref(&ip->lock);
1517 hammer_unload_inode(ip);
1522 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1523 * the read-only flag for cached inodes.
1525 * This routine is called from a RB_SCAN().
1528 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1530 hammer_mount_t hmp = ip->hmp;
1532 if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1533 ip->flags |= HAMMER_INODE_RO;
1535 ip->flags &= ~HAMMER_INODE_RO;
1540 * A transaction has modified an inode, requiring updates as specified by
1543 * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1544 * and not including size changes due to write-append
1545 * (but other size changes are included).
1546 * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1548 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1549 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1550 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1551 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1554 hammer_modify_inode(hammer_inode_t ip, int flags)
1557 * ronly of 0 or 2 does not trigger assertion.
1558 * 2 is a special error state
1560 KKASSERT(ip->hmp->ronly != 1 ||
1561 (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1562 HAMMER_INODE_SDIRTY |
1563 HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1564 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1565 if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1566 ip->flags |= HAMMER_INODE_RSV_INODES;
1567 ++ip->hmp->rsv_inodes;
1574 * Request that an inode be flushed. This whole mess cannot block and may
1575 * recurse (if not synchronous). Once requested HAMMER will attempt to
1576 * actively flush the inode until the flush can be done.
1578 * The inode may already be flushing, or may be in a setup state. We can
1579 * place the inode in a flushing state if it is currently idle and flag it
1580 * to reflush if it is currently flushing.
1582 * Upon return if the inode could not be flushed due to a setup
1583 * dependancy, then it will be automatically flushed when the dependancy
1587 hammer_flush_inode(hammer_inode_t ip, int flags)
1590 hammer_flush_group_t flg;
1594 * next_flush_group is the first flush group we can place the inode
1595 * in. It may be NULL. If it becomes full we append a new flush
1596 * group and make that the next_flush_group.
1599 while ((flg = hmp->next_flush_group) != NULL) {
1600 KKASSERT(flg->running == 0);
1601 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1603 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1604 hammer_flusher_async(ip->hmp, flg);
1607 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1608 hmp->next_flush_group = flg;
1609 RB_INIT(&flg->flush_tree);
1610 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1614 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1615 * state we have to put it back into an IDLE state so we can
1616 * drop the extra ref.
1618 * If we have a parent dependancy we must still fall through
1621 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1622 if (ip->flush_state == HAMMER_FST_SETUP &&
1623 TAILQ_EMPTY(&ip->target_list)) {
1624 ip->flush_state = HAMMER_FST_IDLE;
1625 hammer_rel_inode(ip, 0);
1627 if (ip->flush_state == HAMMER_FST_IDLE)
1632 * Our flush action will depend on the current state.
1634 switch(ip->flush_state) {
1635 case HAMMER_FST_IDLE:
1637 * We have no dependancies and can flush immediately. Some
1638 * our children may not be flushable so we have to re-test
1639 * with that additional knowledge.
1641 hammer_flush_inode_core(ip, flg, flags);
1643 case HAMMER_FST_SETUP:
1645 * Recurse upwards through dependancies via target_list
1646 * and start their flusher actions going if possible.
1648 * 'good' is our connectivity. -1 means we have none and
1649 * can't flush, 0 means there weren't any dependancies, and
1650 * 1 means we have good connectivity.
1652 good = hammer_setup_parent_inodes(ip, 0, flg);
1656 * We can continue if good >= 0. Determine how
1657 * many records under our inode can be flushed (and
1660 hammer_flush_inode_core(ip, flg, flags);
1663 * Parent has no connectivity, tell it to flush
1664 * us as soon as it does.
1666 * The REFLUSH flag is also needed to trigger
1667 * dependancy wakeups.
1669 ip->flags |= HAMMER_INODE_CONN_DOWN |
1670 HAMMER_INODE_REFLUSH;
1671 if (flags & HAMMER_FLUSH_SIGNAL) {
1672 ip->flags |= HAMMER_INODE_RESIGNAL;
1673 hammer_flusher_async(ip->hmp, flg);
1677 case HAMMER_FST_FLUSH:
1679 * We are already flushing, flag the inode to reflush
1680 * if needed after it completes its current flush.
1682 * The REFLUSH flag is also needed to trigger
1683 * dependancy wakeups.
1685 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1686 ip->flags |= HAMMER_INODE_REFLUSH;
1687 if (flags & HAMMER_FLUSH_SIGNAL) {
1688 ip->flags |= HAMMER_INODE_RESIGNAL;
1689 hammer_flusher_async(ip->hmp, flg);
1696 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1697 * ip which reference our ip.
1699 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1700 * so for now do not ref/deref the structures. Note that if we use the
1701 * ref/rel code later, the rel CAN block.
1704 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1705 hammer_flush_group_t flg)
1707 hammer_record_t depend;
1712 * If we hit our recursion limit and we have parent dependencies
1713 * We cannot continue. Returning < 0 will cause us to be flagged
1714 * for reflush. Returning -2 cuts off additional dependency checks
1715 * because they are likely to also hit the depth limit.
1717 * We cannot return < 0 if there are no dependencies or there might
1718 * not be anything to wakeup (ip).
1720 if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1721 kprintf("HAMMER Warning: depth limit reached on "
1722 "setup recursion, inode %p %016llx\n",
1723 ip, (long long)ip->obj_id);
1731 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1732 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1733 KKASSERT(depend->target_ip == ip);
1734 if (r < 0 && good == 0)
1740 * If we failed due to the recursion depth limit then stop
1750 * This helper function takes a record representing the dependancy between
1751 * the parent inode and child inode.
1753 * record->ip = parent inode
1754 * record->target_ip = child inode
1756 * We are asked to recurse upwards and convert the record from SETUP
1757 * to FLUSH if possible.
1759 * Return 1 if the record gives us connectivity
1761 * Return 0 if the record is not relevant
1763 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1766 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1767 hammer_flush_group_t flg)
1773 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1778 * If the record is already flushing, is it in our flush group?
1780 * If it is in our flush group but it is a general record or a
1781 * delete-on-disk, it does not improve our connectivity (return 0),
1782 * and if the target inode is not trying to destroy itself we can't
1783 * allow the operation yet anyway (the second return -1).
1785 if (record->flush_state == HAMMER_FST_FLUSH) {
1787 * If not in our flush group ask the parent to reflush
1788 * us as soon as possible.
1790 if (record->flush_group != flg) {
1791 pip->flags |= HAMMER_INODE_REFLUSH;
1792 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1797 * If in our flush group everything is already set up,
1798 * just return whether the record will improve our
1799 * visibility or not.
1801 if (record->type == HAMMER_MEM_RECORD_ADD)
1807 * It must be a setup record. Try to resolve the setup dependancies
1808 * by recursing upwards so we can place ip on the flush list.
1810 * Limit ourselves to 20 levels of recursion to avoid blowing out
1811 * the kernel stack. If we hit the recursion limit we can't flush
1812 * until the parent flushes. The parent will flush independantly
1813 * on its own and ultimately a deep recursion will be resolved.
1815 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1817 good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1820 * If good < 0 the parent has no connectivity and we cannot safely
1821 * flush the directory entry, which also means we can't flush our
1822 * ip. Flag us for downward recursion once the parent's
1823 * connectivity is resolved. Flag the parent for [re]flush or it
1824 * may not check for downward recursions.
1827 pip->flags |= HAMMER_INODE_REFLUSH;
1828 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1833 * We are go, place the parent inode in a flushing state so we can
1834 * place its record in a flushing state. Note that the parent
1835 * may already be flushing. The record must be in the same flush
1836 * group as the parent.
1838 if (pip->flush_state != HAMMER_FST_FLUSH)
1839 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1840 KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1841 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1844 if (record->type == HAMMER_MEM_RECORD_DEL &&
1845 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1847 * Regardless of flushing state we cannot sync this path if the
1848 * record represents a delete-on-disk but the target inode
1849 * is not ready to sync its own deletion.
1851 * XXX need to count effective nlinks to determine whether
1852 * the flush is ok, otherwise removing a hardlink will
1853 * just leave the DEL record to rot.
1855 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1859 if (pip->flush_group == flg) {
1861 * Because we have not calculated nlinks yet we can just
1862 * set records to the flush state if the parent is in
1863 * the same flush group as we are.
1865 record->flush_state = HAMMER_FST_FLUSH;
1866 record->flush_group = flg;
1867 ++record->flush_group->refs;
1868 hammer_ref(&record->lock);
1871 * A general directory-add contributes to our visibility.
1873 * Otherwise it is probably a directory-delete or
1874 * delete-on-disk record and does not contribute to our
1875 * visbility (but we can still flush it).
1877 if (record->type == HAMMER_MEM_RECORD_ADD)
1882 * If the parent is not in our flush group we cannot
1883 * flush this record yet, there is no visibility.
1884 * We tell the parent to reflush and mark ourselves
1885 * so the parent knows it should flush us too.
1887 pip->flags |= HAMMER_INODE_REFLUSH;
1888 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1894 * This is the core routine placing an inode into the FST_FLUSH state.
1897 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1902 * Set flush state and prevent the flusher from cycling into
1903 * the next flush group. Do not place the ip on the list yet.
1904 * Inodes not in the idle state get an extra reference.
1906 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1907 if (ip->flush_state == HAMMER_FST_IDLE)
1908 hammer_ref(&ip->lock);
1909 ip->flush_state = HAMMER_FST_FLUSH;
1910 ip->flush_group = flg;
1911 ++ip->hmp->flusher.group_lock;
1912 ++ip->hmp->count_iqueued;
1913 ++hammer_count_iqueued;
1917 * If the flush group reaches the autoflush limit we want to signal
1918 * the flusher. This is particularly important for remove()s.
1920 if (flg->total_count == hammer_autoflush)
1921 flags |= HAMMER_FLUSH_SIGNAL;
1924 * We need to be able to vfsync/truncate from the backend.
1926 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1927 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1928 ip->flags |= HAMMER_INODE_VHELD;
1933 * Figure out how many in-memory records we can actually flush
1934 * (not including inode meta-data, buffers, etc).
1936 KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1937 if (flags & HAMMER_FLUSH_RECURSION) {
1939 * If this is a upwards recursion we do not want to
1940 * recurse down again!
1944 } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1946 * No new records are added if we must complete a flush
1947 * from a previous cycle, but we do have to move the records
1948 * from the previous cycle to the current one.
1951 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1952 hammer_syncgrp_child_callback, NULL);
1958 * Normal flush, scan records and bring them into the flush.
1959 * Directory adds and deletes are usually skipped (they are
1960 * grouped with the related inode rather then with the
1963 * go_count can be negative, which means the scan aborted
1964 * due to the flush group being over-full and we should
1965 * flush what we have.
1967 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1968 hammer_setup_child_callback, NULL);
1972 * This is a more involved test that includes go_count. If we
1973 * can't flush, flag the inode and return. If go_count is 0 we
1974 * were are unable to flush any records in our rec_tree and
1975 * must ignore the XDIRTY flag.
1977 if (go_count == 0) {
1978 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1979 --ip->hmp->count_iqueued;
1980 --hammer_count_iqueued;
1983 ip->flush_state = HAMMER_FST_SETUP;
1984 ip->flush_group = NULL;
1985 if (ip->flags & HAMMER_INODE_VHELD) {
1986 ip->flags &= ~HAMMER_INODE_VHELD;
1991 * REFLUSH is needed to trigger dependancy wakeups
1992 * when an inode is in SETUP.
1994 ip->flags |= HAMMER_INODE_REFLUSH;
1995 if (flags & HAMMER_FLUSH_SIGNAL) {
1996 ip->flags |= HAMMER_INODE_RESIGNAL;
1997 hammer_flusher_async(ip->hmp, flg);
1999 if (--ip->hmp->flusher.group_lock == 0)
2000 wakeup(&ip->hmp->flusher.group_lock);
2006 * Snapshot the state of the inode for the backend flusher.
2008 * We continue to retain save_trunc_off even when all truncations
2009 * have been resolved as an optimization to determine if we can
2010 * skip the B-Tree lookup for overwrite deletions.
2012 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2013 * and stays in ip->flags. Once set, it stays set until the
2014 * inode is destroyed.
2016 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2017 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2018 ip->sync_trunc_off = ip->trunc_off;
2019 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2020 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2021 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2024 * The save_trunc_off used to cache whether the B-Tree
2025 * holds any records past that point is not used until
2026 * after the truncation has succeeded, so we can safely
2029 if (ip->save_trunc_off > ip->sync_trunc_off)
2030 ip->save_trunc_off = ip->sync_trunc_off;
2032 ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2033 ~HAMMER_INODE_TRUNCATED);
2034 ip->sync_ino_leaf = ip->ino_leaf;
2035 ip->sync_ino_data = ip->ino_data;
2036 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2037 #ifdef DEBUG_TRUNCATE
2038 if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2039 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2043 * The flusher list inherits our inode and reference.
2045 KKASSERT(flg->running == 0);
2046 RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2047 if (--ip->hmp->flusher.group_lock == 0)
2048 wakeup(&ip->hmp->flusher.group_lock);
2050 if (flags & HAMMER_FLUSH_SIGNAL) {
2051 hammer_flusher_async(ip->hmp, flg);
2056 * Callback for scan of ip->rec_tree. Try to include each record in our
2057 * flush. ip->flush_group has been set but the inode has not yet been
2058 * moved into a flushing state.
2060 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2063 * We return 1 for any record placed or found in FST_FLUSH, which prevents
2064 * the caller from shortcutting the flush.
2067 hammer_setup_child_callback(hammer_record_t rec, void *data)
2069 hammer_flush_group_t flg;
2070 hammer_inode_t target_ip;
2075 * Records deleted or committed by the backend are ignored.
2076 * Note that the flush detects deleted frontend records at
2077 * multiple points to deal with races. This is just the first
2078 * line of defense. The only time HAMMER_RECF_DELETED_FE cannot
2079 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2080 * messes up link-count calculations.
2082 * NOTE: Don't get confused between record deletion and, say,
2083 * directory entry deletion. The deletion of a directory entry
2084 * which is on-media has nothing to do with the record deletion
2087 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2088 HAMMER_RECF_COMMITTED)) {
2089 if (rec->flush_state == HAMMER_FST_FLUSH) {
2090 KKASSERT(rec->flush_group == rec->ip->flush_group);
2099 * If the record is in an idle state it has no dependancies and
2103 flg = ip->flush_group;
2106 switch(rec->flush_state) {
2107 case HAMMER_FST_IDLE:
2109 * The record has no setup dependancy, we can flush it.
2111 KKASSERT(rec->target_ip == NULL);
2112 rec->flush_state = HAMMER_FST_FLUSH;
2113 rec->flush_group = flg;
2115 hammer_ref(&rec->lock);
2118 case HAMMER_FST_SETUP:
2120 * The record has a setup dependancy. These are typically
2121 * directory entry adds and deletes. Such entries will be
2122 * flushed when their inodes are flushed so we do not
2123 * usually have to add them to the flush here. However,
2124 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2125 * it is asking us to flush this record (and it).
2127 target_ip = rec->target_ip;
2128 KKASSERT(target_ip != NULL);
2129 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2132 * If the target IP is already flushing in our group
2133 * we could associate the record, but target_ip has
2134 * already synced ino_data to sync_ino_data and we
2135 * would also have to adjust nlinks. Plus there are
2136 * ordering issues for adds and deletes.
2138 * Reflush downward if this is an ADD, and upward if
2141 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2142 if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2143 ip->flags |= HAMMER_INODE_REFLUSH;
2145 target_ip->flags |= HAMMER_INODE_REFLUSH;
2150 * Target IP is not yet flushing. This can get complex
2151 * because we have to be careful about the recursion.
2153 * Directories create an issue for us in that if a flush
2154 * of a directory is requested the expectation is to flush
2155 * any pending directory entries, but this will cause the
2156 * related inodes to recursively flush as well. We can't
2157 * really defer the operation so just get as many as we
2161 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2162 (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2164 * We aren't reclaiming and the target ip was not
2165 * previously prevented from flushing due to this
2166 * record dependancy. Do not flush this record.
2171 if (flg->total_count + flg->refs >
2172 ip->hmp->undo_rec_limit) {
2174 * Our flush group is over-full and we risk blowing
2175 * out the UNDO FIFO. Stop the scan, flush what we
2176 * have, then reflush the directory.
2178 * The directory may be forced through multiple
2179 * flush groups before it can be completely
2182 ip->flags |= HAMMER_INODE_RESIGNAL |
2183 HAMMER_INODE_REFLUSH;
2185 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2187 * If the target IP is not flushing we can force
2188 * it to flush, even if it is unable to write out
2189 * any of its own records we have at least one in
2190 * hand that we CAN deal with.
2192 rec->flush_state = HAMMER_FST_FLUSH;
2193 rec->flush_group = flg;
2195 hammer_ref(&rec->lock);
2196 hammer_flush_inode_core(target_ip, flg,
2197 HAMMER_FLUSH_RECURSION);
2201 * General or delete-on-disk record.
2203 * XXX this needs help. If a delete-on-disk we could
2204 * disconnect the target. If the target has its own
2205 * dependancies they really need to be flushed.
2209 rec->flush_state = HAMMER_FST_FLUSH;
2210 rec->flush_group = flg;
2212 hammer_ref(&rec->lock);
2213 hammer_flush_inode_core(target_ip, flg,
2214 HAMMER_FLUSH_RECURSION);
2218 case HAMMER_FST_FLUSH:
2220 * The flush_group should already match.
2222 KKASSERT(rec->flush_group == flg);
2231 * This version just moves records already in a flush state to the new
2232 * flush group and that is it.
2235 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2237 hammer_inode_t ip = rec->ip;
2239 switch(rec->flush_state) {
2240 case HAMMER_FST_FLUSH:
2241 KKASSERT(rec->flush_group == ip->flush_group);
2251 * Wait for a previously queued flush to complete.
2253 * If a critical error occured we don't try to wait.
2256 hammer_wait_inode(hammer_inode_t ip)
2258 hammer_flush_group_t flg;
2261 if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2262 while (ip->flush_state != HAMMER_FST_IDLE &&
2263 (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2264 if (ip->flush_state == HAMMER_FST_SETUP)
2265 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2266 if (ip->flush_state != HAMMER_FST_IDLE) {
2267 ip->flags |= HAMMER_INODE_FLUSHW;
2268 tsleep(&ip->flags, 0, "hmrwin", 0);
2275 * Called by the backend code when a flush has been completed.
2276 * The inode has already been removed from the flush list.
2278 * A pipelined flush can occur, in which case we must re-enter the
2279 * inode on the list and re-copy its fields.
2282 hammer_flush_inode_done(hammer_inode_t ip, int error)
2287 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2292 * Auto-reflush if the backend could not completely flush
2293 * the inode. This fixes a case where a deferred buffer flush
2294 * could cause fsync to return early.
2296 if (ip->sync_flags & HAMMER_INODE_MODMASK)
2297 ip->flags |= HAMMER_INODE_REFLUSH;
2300 * Merge left-over flags back into the frontend and fix the state.
2301 * Incomplete truncations are retained by the backend.
2304 ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2305 ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2308 * The backend may have adjusted nlinks, so if the adjusted nlinks
2309 * does not match the fronttend set the frontend's RDIRTY flag again.
2311 if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2312 ip->flags |= HAMMER_INODE_DDIRTY;
2315 * Fix up the dirty buffer status.
2317 if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2318 ip->flags |= HAMMER_INODE_BUFS;
2322 * Re-set the XDIRTY flag if some of the inode's in-memory records
2323 * could not be flushed.
2325 KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2326 (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2327 (!RB_EMPTY(&ip->rec_tree) &&
2328 (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2331 * Do not lose track of inodes which no longer have vnode
2332 * assocations, otherwise they may never get flushed again.
2334 * The reflush flag can be set superfluously, causing extra pain
2335 * for no reason. If the inode is no longer modified it no longer
2336 * needs to be flushed.
2338 if (ip->flags & HAMMER_INODE_MODMASK) {
2340 ip->flags |= HAMMER_INODE_REFLUSH;
2342 ip->flags &= ~HAMMER_INODE_REFLUSH;
2346 * Adjust the flush state.
2348 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2350 * We were unable to flush out all our records, leave the
2351 * inode in a flush state and in the current flush group.
2352 * The flush group will be re-run.
2354 * This occurs if the UNDO block gets too full or there is
2355 * too much dirty meta-data and allows the flusher to
2356 * finalize the UNDO block and then re-flush.
2358 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2362 * Remove from the flush_group
2364 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2365 ip->flush_group = NULL;
2368 * Clean up the vnode ref and tracking counts.
2370 if (ip->flags & HAMMER_INODE_VHELD) {
2371 ip->flags &= ~HAMMER_INODE_VHELD;
2374 --hmp->count_iqueued;
2375 --hammer_count_iqueued;
2378 * And adjust the state.
2380 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2381 ip->flush_state = HAMMER_FST_IDLE;
2384 ip->flush_state = HAMMER_FST_SETUP;
2389 * If the frontend is waiting for a flush to complete,
2392 if (ip->flags & HAMMER_INODE_FLUSHW) {
2393 ip->flags &= ~HAMMER_INODE_FLUSHW;
2398 * If the frontend made more changes and requested another
2399 * flush, then try to get it running.
2401 * Reflushes are aborted when the inode is errored out.
2403 if (ip->flags & HAMMER_INODE_REFLUSH) {
2404 ip->flags &= ~HAMMER_INODE_REFLUSH;
2405 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2406 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2407 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2409 hammer_flush_inode(ip, 0);
2415 * If we have no parent dependancies we can clear CONN_DOWN
2417 if (TAILQ_EMPTY(&ip->target_list))
2418 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2421 * If the inode is now clean drop the space reservation.
2423 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2424 (ip->flags & HAMMER_INODE_RSV_INODES)) {
2425 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2430 hammer_rel_inode(ip, 0);
2434 * Called from hammer_sync_inode() to synchronize in-memory records
2438 hammer_sync_record_callback(hammer_record_t record, void *data)
2440 hammer_cursor_t cursor = data;
2441 hammer_transaction_t trans = cursor->trans;
2442 hammer_mount_t hmp = trans->hmp;
2446 * Skip records that do not belong to the current flush.
2448 ++hammer_stats_record_iterations;
2449 if (record->flush_state != HAMMER_FST_FLUSH)
2453 if (record->flush_group != record->ip->flush_group) {
2454 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2455 if (hammer_debug_critical)
2460 KKASSERT(record->flush_group == record->ip->flush_group);
2463 * Interlock the record using the BE flag. Once BE is set the
2464 * frontend cannot change the state of FE.
2466 * NOTE: If FE is set prior to us setting BE we still sync the
2467 * record out, but the flush completion code converts it to
2468 * a delete-on-disk record instead of destroying it.
2470 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2471 record->flags |= HAMMER_RECF_INTERLOCK_BE;
2474 * The backend has already disposed of the record.
2476 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2482 * If the whole inode is being deleting all on-disk records will
2483 * be deleted very soon, we can't sync any new records to disk
2484 * because they will be deleted in the same transaction they were
2485 * created in (delete_tid == create_tid), which will assert.
2487 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2488 * that we currently panic on.
2490 if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2491 switch(record->type) {
2492 case HAMMER_MEM_RECORD_DATA:
2494 * We don't have to do anything, if the record was
2495 * committed the space will have been accounted for
2499 case HAMMER_MEM_RECORD_GENERAL:
2501 * Set deleted-by-backend flag. Do not set the
2502 * backend committed flag, because we are throwing
2505 record->flags |= HAMMER_RECF_DELETED_BE;
2506 ++record->ip->rec_generation;
2509 case HAMMER_MEM_RECORD_ADD:
2510 panic("hammer_sync_record_callback: illegal add "
2511 "during inode deletion record %p", record);
2512 break; /* NOT REACHED */
2513 case HAMMER_MEM_RECORD_INODE:
2514 panic("hammer_sync_record_callback: attempt to "
2515 "sync inode record %p?", record);
2516 break; /* NOT REACHED */
2517 case HAMMER_MEM_RECORD_DEL:
2519 * Follow through and issue the on-disk deletion
2526 * If DELETED_FE is set special handling is needed for directory
2527 * entries. Dependant pieces related to the directory entry may
2528 * have already been synced to disk. If this occurs we have to
2529 * sync the directory entry and then change the in-memory record
2530 * from an ADD to a DELETE to cover the fact that it's been
2531 * deleted by the frontend.
2533 * A directory delete covering record (MEM_RECORD_DEL) can never
2534 * be deleted by the frontend.
2536 * Any other record type (aka DATA) can be deleted by the frontend.
2537 * XXX At the moment the flusher must skip it because there may
2538 * be another data record in the flush group for the same block,
2539 * meaning that some frontend data changes can leak into the backend's
2540 * synchronization point.
2542 if (record->flags & HAMMER_RECF_DELETED_FE) {
2543 if (record->type == HAMMER_MEM_RECORD_ADD) {
2545 * Convert a front-end deleted directory-add to
2546 * a directory-delete entry later.
2548 record->flags |= HAMMER_RECF_CONVERT_DELETE;
2551 * Dispose of the record (race case). Mark as
2552 * deleted by backend (and not committed).
2554 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2555 record->flags |= HAMMER_RECF_DELETED_BE;
2556 ++record->ip->rec_generation;
2563 * Assign the create_tid for new records. Deletions already
2564 * have the record's entire key properly set up.
2566 if (record->type != HAMMER_MEM_RECORD_DEL) {
2567 record->leaf.base.create_tid = trans->tid;
2568 record->leaf.create_ts = trans->time32;
2571 error = hammer_ip_sync_record_cursor(cursor, record);
2572 if (error != EDEADLK)
2574 hammer_done_cursor(cursor);
2575 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2580 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2585 hammer_flush_record_done(record, error);
2588 * Do partial finalization if we have built up too many dirty
2589 * buffers. Otherwise a buffer cache deadlock can occur when
2590 * doing things like creating tens of thousands of tiny files.
2592 * We must release our cursor lock to avoid a 3-way deadlock
2593 * due to the exclusive sync lock the finalizer must get.
2595 * WARNING: See warnings in hammer_unlock_cursor() function.
2597 if (hammer_flusher_meta_limit(hmp)) {
2598 hammer_unlock_cursor(cursor);
2599 hammer_flusher_finalize(trans, 0);
2600 hammer_lock_cursor(cursor);
2607 * Backend function called by the flusher to sync an inode to media.
2610 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2612 struct hammer_cursor cursor;
2613 hammer_node_t tmp_node;
2614 hammer_record_t depend;
2615 hammer_record_t next;
2616 int error, tmp_error;
2619 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2622 error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2627 * Any directory records referencing this inode which are not in
2628 * our current flush group must adjust our nlink count for the
2629 * purposes of synchronization to disk.
2631 * Records which are in our flush group can be unlinked from our
2632 * inode now, potentially allowing the inode to be physically
2635 * This cannot block.
2637 nlinks = ip->ino_data.nlinks;
2638 next = TAILQ_FIRST(&ip->target_list);
2639 while ((depend = next) != NULL) {
2640 next = TAILQ_NEXT(depend, target_entry);
2641 if (depend->flush_state == HAMMER_FST_FLUSH &&
2642 depend->flush_group == ip->flush_group) {
2644 * If this is an ADD that was deleted by the frontend
2645 * the frontend nlinks count will have already been
2646 * decremented, but the backend is going to sync its
2647 * directory entry and must account for it. The
2648 * record will be converted to a delete-on-disk when
2651 * If the ADD was not deleted by the frontend we
2652 * can remove the dependancy from our target_list.
2654 if (depend->flags & HAMMER_RECF_DELETED_FE) {
2657 TAILQ_REMOVE(&ip->target_list, depend,
2659 depend->target_ip = NULL;
2661 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2663 * Not part of our flush group and not deleted by
2664 * the front-end, adjust the link count synced to
2665 * the media (undo what the frontend did when it
2666 * queued the record).
2668 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2669 switch(depend->type) {
2670 case HAMMER_MEM_RECORD_ADD:
2673 case HAMMER_MEM_RECORD_DEL:
2683 * Set dirty if we had to modify the link count.
2685 if (ip->sync_ino_data.nlinks != nlinks) {
2686 KKASSERT((int64_t)nlinks >= 0);
2687 ip->sync_ino_data.nlinks = nlinks;
2688 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2692 * If there is a trunction queued destroy any data past the (aligned)
2693 * truncation point. Userland will have dealt with the buffer
2694 * containing the truncation point for us.
2696 * We don't flush pending frontend data buffers until after we've
2697 * dealt with the truncation.
2699 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2701 * Interlock trunc_off. The VOP front-end may continue to
2702 * make adjustments to it while we are blocked.
2705 off_t aligned_trunc_off;
2708 trunc_off = ip->sync_trunc_off;
2709 blkmask = hammer_blocksize(trunc_off) - 1;
2710 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2713 * Delete any whole blocks on-media. The front-end has
2714 * already cleaned out any partial block and made it
2715 * pending. The front-end may have updated trunc_off
2716 * while we were blocked so we only use sync_trunc_off.
2718 * This operation can blow out the buffer cache, EWOULDBLOCK
2719 * means we were unable to complete the deletion. The
2720 * deletion will update sync_trunc_off in that case.
2722 error = hammer_ip_delete_range(&cursor, ip,
2724 0x7FFFFFFFFFFFFFFFLL, 2);
2725 if (error == EWOULDBLOCK) {
2726 ip->flags |= HAMMER_INODE_WOULDBLOCK;
2728 goto defer_buffer_flush;
2735 * Clear the truncation flag on the backend after we have
2736 * complete the deletions. Backend data is now good again
2737 * (including new records we are about to sync, below).
2739 * Leave sync_trunc_off intact. As we write additional
2740 * records the backend will update sync_trunc_off. This
2741 * tells the backend whether it can skip the overwrite
2742 * test. This should work properly even when the backend
2743 * writes full blocks where the truncation point straddles
2744 * the block because the comparison is against the base
2745 * offset of the record.
2747 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2748 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2754 * Now sync related records. These will typically be directory
2755 * entries, records tracking direct-writes, or delete-on-disk records.
2758 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2759 hammer_sync_record_callback, &cursor);
2765 hammer_cache_node(&ip->cache[1], cursor.node);
2768 * Re-seek for inode update, assuming our cache hasn't been ripped
2769 * out from under us.
2772 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2774 hammer_cursor_downgrade(&cursor);
2775 hammer_lock_sh(&tmp_node->lock);
2776 if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2777 hammer_cursor_seek(&cursor, tmp_node, 0);
2778 hammer_unlock(&tmp_node->lock);
2779 hammer_rel_node(tmp_node);
2785 * If we are deleting the inode the frontend had better not have
2786 * any active references on elements making up the inode.
2788 * The call to hammer_ip_delete_clean() cleans up auxillary records
2789 * but not DB or DATA records. Those must have already been deleted
2790 * by the normal truncation mechanic.
2792 if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2793 RB_EMPTY(&ip->rec_tree) &&
2794 (ip->sync_flags & HAMMER_INODE_DELETING) &&
2795 (ip->flags & HAMMER_INODE_DELETED) == 0) {
2798 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2800 ip->flags |= HAMMER_INODE_DELETED;
2801 ip->sync_flags &= ~HAMMER_INODE_DELETING;
2802 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2803 KKASSERT(RB_EMPTY(&ip->rec_tree));
2806 * Set delete_tid in both the frontend and backend
2807 * copy of the inode record. The DELETED flag handles
2808 * this, do not set RDIRTY.
2810 ip->ino_leaf.base.delete_tid = trans->tid;
2811 ip->sync_ino_leaf.base.delete_tid = trans->tid;
2812 ip->ino_leaf.delete_ts = trans->time32;
2813 ip->sync_ino_leaf.delete_ts = trans->time32;
2817 * Adjust the inode count in the volume header
2819 hammer_sync_lock_sh(trans);
2820 if (ip->flags & HAMMER_INODE_ONDISK) {
2821 hammer_modify_volume_field(trans,
2824 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2825 hammer_modify_volume_done(trans->rootvol);
2827 hammer_sync_unlock(trans);
2833 ip->sync_flags &= ~HAMMER_INODE_BUFS;
2837 * Now update the inode's on-disk inode-data and/or on-disk record.
2838 * DELETED and ONDISK are managed only in ip->flags.
2840 * In the case of a defered buffer flush we still update the on-disk
2841 * inode to satisfy visibility requirements if there happen to be
2842 * directory dependancies.
2844 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2845 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2847 * If deleted and on-disk, don't set any additional flags.
2848 * the delete flag takes care of things.
2850 * Clear flags which may have been set by the frontend.
2852 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2853 HAMMER_INODE_SDIRTY |
2854 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2855 HAMMER_INODE_DELETING);
2857 case HAMMER_INODE_DELETED:
2859 * Take care of the case where a deleted inode was never
2860 * flushed to the disk in the first place.
2862 * Clear flags which may have been set by the frontend.
2864 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2865 HAMMER_INODE_SDIRTY |
2866 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2867 HAMMER_INODE_DELETING);
2868 while (RB_ROOT(&ip->rec_tree)) {
2869 hammer_record_t record = RB_ROOT(&ip->rec_tree);
2870 hammer_ref(&record->lock);
2871 KKASSERT(record->lock.refs == 1);
2872 record->flags |= HAMMER_RECF_DELETED_BE;
2873 ++record->ip->rec_generation;
2874 hammer_rel_mem_record(record);
2877 case HAMMER_INODE_ONDISK:
2879 * If already on-disk, do not set any additional flags.
2884 * If not on-disk and not deleted, set DDIRTY to force
2885 * an initial record to be written.
2887 * Also set the create_tid in both the frontend and backend
2888 * copy of the inode record.
2890 ip->ino_leaf.base.create_tid = trans->tid;
2891 ip->ino_leaf.create_ts = trans->time32;
2892 ip->sync_ino_leaf.base.create_tid = trans->tid;
2893 ip->sync_ino_leaf.create_ts = trans->time32;
2894 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2899 * If RDIRTY, DDIRTY, or SDIRTY is set, write out a new record.
2900 * If the inode is already on-disk the old record is marked as
2903 * If DELETED is set hammer_update_inode() will delete the existing
2904 * record without writing out a new one.
2906 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2908 if (ip->flags & HAMMER_INODE_DELETED) {
2909 error = hammer_update_inode(&cursor, ip);
2911 if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
2912 (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2913 error = hammer_update_itimes(&cursor, ip);
2915 if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
2916 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2917 error = hammer_update_inode(&cursor, ip);
2921 hammer_critical_error(ip->hmp, ip, error,
2922 "while syncing inode");
2924 hammer_done_cursor(&cursor);
2929 * This routine is called when the OS is no longer actively referencing
2930 * the inode (but might still be keeping it cached), or when releasing
2931 * the last reference to an inode.
2933 * At this point if the inode's nlinks count is zero we want to destroy
2934 * it, which may mean destroying it on-media too.
2937 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2942 * Set the DELETING flag when the link count drops to 0 and the
2943 * OS no longer has any opens on the inode.
2945 * The backend will clear DELETING (a mod flag) and set DELETED
2946 * (a state flag) when it is actually able to perform the
2949 * Don't reflag the deletion if the flusher is currently syncing
2950 * one that was already flagged. A previously set DELETING flag
2951 * may bounce around flags and sync_flags until the operation is
2954 if (ip->ino_data.nlinks == 0 &&
2955 ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2956 ip->flags |= HAMMER_INODE_DELETING;
2957 ip->flags |= HAMMER_INODE_TRUNCATED;
2961 if (hammer_get_vnode(ip, &vp) != 0)
2969 vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2970 vnode_pager_setsize(ip->vp, 0);
2979 * After potentially resolving a dependancy the inode is tested
2980 * to determine whether it needs to be reflushed.
2983 hammer_test_inode(hammer_inode_t ip)
2985 if (ip->flags & HAMMER_INODE_REFLUSH) {
2986 ip->flags &= ~HAMMER_INODE_REFLUSH;
2987 hammer_ref(&ip->lock);
2988 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2989 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2990 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2992 hammer_flush_inode(ip, 0);
2994 hammer_rel_inode(ip, 0);
2999 * Clear the RECLAIM flag on an inode. This occurs when the inode is
3000 * reassociated with a vp or just before it gets freed.
3002 * Pipeline wakeups to threads blocked due to an excessive number of
3003 * detached inodes. This typically occurs when atime updates accumulate
3004 * while scanning a directory tree.
3007 hammer_inode_wakereclaims(hammer_inode_t ip)
3009 struct hammer_reclaim *reclaim;
3010 hammer_mount_t hmp = ip->hmp;
3012 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3015 --hammer_count_reclaiming;
3016 --hmp->inode_reclaims;
3017 ip->flags &= ~HAMMER_INODE_RECLAIM;
3019 while ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3020 if (reclaim->count > 0 && --reclaim->count == 0) {
3021 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3024 if (hmp->inode_reclaims > hammer_limit_reclaim / 2)
3030 * Setup our reclaim pipeline. We only let so many detached (and dirty)
3031 * inodes build up before we start blocking. This routine is called
3032 * if a new inode is created or an inode is loaded from media.
3034 * When we block we don't care *which* inode has finished reclaiming,
3035 * as lone as one does.
3038 hammer_inode_waitreclaims(hammer_mount_t hmp)
3040 struct hammer_reclaim reclaim;
3042 if (hmp->inode_reclaims < hammer_limit_reclaim)
3045 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3046 tsleep(&reclaim, 0, "hmrrcm", hz);
3047 if (reclaim.count > 0)
3048 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3054 * XXX not used, doesn't work very well due to the large batching nature
3057 * A larger then normal backlog of inodes is sitting in the flusher,
3058 * enforce a general slowdown to let it catch up. This routine is only
3059 * called on completion of a non-flusher-related transaction which
3060 * performed B-Tree node I/O.
3062 * It is possible for the flusher to stall in a continuous load.
3063 * blogbench -i1000 -o seems to do a good job generating this sort of load.
3064 * If the flusher is unable to catch up the inode count can bloat until
3065 * we run out of kvm.
3067 * This is a bit of a hack.
3070 hammer_inode_waithard(hammer_mount_t hmp)
3075 if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3076 if (hmp->inode_reclaims < hammer_limit_reclaim / 2 &&
3077 hmp->count_iqueued < hmp->count_inodes / 20) {
3078 hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3082 if (hmp->inode_reclaims < hammer_limit_reclaim ||
3083 hmp->count_iqueued < hmp->count_inodes / 10) {
3086 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3090 * Block for one flush cycle.
3092 hammer_flusher_wait_next(hmp);