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;
411 ip->flags |= HAMMER_INODE_RO;
412 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
413 0x7FFFFFFFFFFFFFFFLL;
414 RB_INIT(&ip->rec_tree);
415 TAILQ_INIT(&ip->target_list);
416 hammer_ref(&ip->lock);
419 * Locate the on-disk inode. If this is a PFS root we always
420 * access the current version of the root inode and (if it is not
421 * a master) always access information under it with a snapshot
424 * We cache recent inode lookups in this directory in dip->cache[2].
425 * If we can't find it we assume the inode we are looking for is
426 * close to the directory inode.
431 if (dip->cache[2].node)
432 cachep = &dip->cache[2];
434 cachep = &dip->cache[0];
436 hammer_init_cursor(trans, &cursor, cachep, NULL);
437 cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
438 cursor.key_beg.obj_id = ip->obj_id;
439 cursor.key_beg.key = 0;
440 cursor.key_beg.create_tid = 0;
441 cursor.key_beg.delete_tid = 0;
442 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
443 cursor.key_beg.obj_type = 0;
445 cursor.asof = iinfo.obj_asof;
446 cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
449 *errorp = hammer_btree_lookup(&cursor);
450 if (*errorp == EDEADLK) {
451 hammer_done_cursor(&cursor);
456 * On success the B-Tree lookup will hold the appropriate
457 * buffer cache buffers and provide a pointer to the requested
458 * information. Copy the information to the in-memory inode
459 * and cache the B-Tree node to improve future operations.
462 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
463 ip->ino_data = cursor.data->inode;
466 * cache[0] tries to cache the location of the object inode.
467 * The assumption is that it is near the directory inode.
469 * cache[1] tries to cache the location of the object data.
470 * We might have something in the governing directory from
471 * scan optimizations (see the strategy code in
474 * We update dip->cache[2], if possible, with the location
475 * of the object inode for future directory shortcuts.
477 hammer_cache_node(&ip->cache[0], cursor.node);
479 if (dip->cache[3].node) {
480 hammer_cache_node(&ip->cache[1],
483 hammer_cache_node(&dip->cache[2], cursor.node);
487 * The file should not contain any data past the file size
488 * stored in the inode. Setting save_trunc_off to the
489 * file size instead of max reduces B-Tree lookup overheads
490 * on append by allowing the flusher to avoid checking for
493 ip->save_trunc_off = ip->ino_data.size;
496 * Locate and assign the pseudofs management structure to
499 if (dip && dip->obj_localization == ip->obj_localization) {
500 ip->pfsm = dip->pfsm;
501 hammer_ref(&ip->pfsm->lock);
503 ip->pfsm = hammer_load_pseudofs(trans,
504 ip->obj_localization,
506 *errorp = 0; /* ignore ENOENT */
511 * The inode is placed on the red-black tree and will be synced to
512 * the media when flushed or by the filesystem sync. If this races
513 * another instantiation/lookup the insertion will fail.
516 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
517 hammer_free_inode(ip);
518 hammer_done_cursor(&cursor);
521 ip->flags |= HAMMER_INODE_ONDISK;
523 if (ip->flags & HAMMER_INODE_RSV_INODES) {
524 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
528 hammer_free_inode(ip);
531 hammer_done_cursor(&cursor);
532 trans->flags |= HAMMER_TRANSF_NEWINODE;
537 * Get a dummy inode to placemark a broken directory entry.
539 struct hammer_inode *
540 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
541 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
542 int flags, int *errorp)
544 hammer_mount_t hmp = trans->hmp;
545 struct hammer_inode_info iinfo;
546 struct hammer_inode *ip;
549 * Determine if we already have an inode cached. If we do then
552 * If we find an inode with no vnode we have to mark the
553 * transaction such that hammer_inode_waitreclaims() is
554 * called later on to avoid building up an infinite number
555 * of inodes. Otherwise we can continue to * add new inodes
556 * faster then they can be disposed of, even with the tsleep
559 * If we find a non-fake inode we return an error. Only fake
560 * inodes can be returned by this routine.
562 iinfo.obj_id = obj_id;
563 iinfo.obj_asof = asof;
564 iinfo.obj_localization = localization;
567 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
569 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
573 hammer_ref(&ip->lock);
578 * Allocate a new inode structure and deal with races later.
580 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
581 ++hammer_count_inodes;
584 ip->obj_asof = iinfo.obj_asof;
585 ip->obj_localization = localization;
587 ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
588 ip->cache[0].ip = ip;
589 ip->cache[1].ip = ip;
590 ip->cache[2].ip = ip;
591 ip->cache[3].ip = ip;
592 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
593 0x7FFFFFFFFFFFFFFFLL;
594 RB_INIT(&ip->rec_tree);
595 TAILQ_INIT(&ip->target_list);
596 hammer_ref(&ip->lock);
599 * Populate the dummy inode. Leave everything zero'd out.
601 * (ip->ino_leaf and ip->ino_data)
603 * Make the dummy inode a FIFO object which most copy programs
604 * will properly ignore.
606 ip->save_trunc_off = ip->ino_data.size;
607 ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
610 * Locate and assign the pseudofs management structure to
613 if (dip && dip->obj_localization == ip->obj_localization) {
614 ip->pfsm = dip->pfsm;
615 hammer_ref(&ip->pfsm->lock);
617 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
619 *errorp = 0; /* ignore ENOENT */
623 * The inode is placed on the red-black tree and will be synced to
624 * the media when flushed or by the filesystem sync. If this races
625 * another instantiation/lookup the insertion will fail.
627 * NOTE: Do not set HAMMER_INODE_ONDISK. The inode is a fake.
630 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
631 hammer_free_inode(ip);
635 if (ip->flags & HAMMER_INODE_RSV_INODES) {
636 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
639 hammer_free_inode(ip);
642 trans->flags |= HAMMER_TRANSF_NEWINODE;
647 * Return a referenced inode only if it is in our inode cache.
649 * Dummy inodes do not count.
651 struct hammer_inode *
652 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
653 hammer_tid_t asof, u_int32_t localization)
655 hammer_mount_t hmp = trans->hmp;
656 struct hammer_inode_info iinfo;
657 struct hammer_inode *ip;
659 iinfo.obj_id = obj_id;
660 iinfo.obj_asof = asof;
661 iinfo.obj_localization = localization;
663 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
665 if (ip->flags & HAMMER_INODE_DUMMY)
668 hammer_ref(&ip->lock);
674 * Create a new filesystem object, returning the inode in *ipp. The
675 * returned inode will be referenced. The inode is created in-memory.
677 * If pfsm is non-NULL the caller wishes to create the root inode for
681 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
683 hammer_inode_t dip, const char *name, int namelen,
684 hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
695 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
696 ++hammer_count_inodes;
698 trans->flags |= HAMMER_TRANSF_NEWINODE;
701 KKASSERT(pfsm->localization != 0);
702 ip->obj_id = HAMMER_OBJID_ROOT;
703 ip->obj_localization = pfsm->localization;
705 KKASSERT(dip != NULL);
706 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
707 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
708 ip->obj_localization = dip->obj_localization;
711 KKASSERT(ip->obj_id != 0);
712 ip->obj_asof = hmp->asof;
714 ip->flush_state = HAMMER_FST_IDLE;
715 ip->flags = HAMMER_INODE_DDIRTY |
716 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
717 ip->cache[0].ip = ip;
718 ip->cache[1].ip = ip;
719 ip->cache[2].ip = ip;
720 ip->cache[3].ip = ip;
722 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
723 /* ip->save_trunc_off = 0; (already zero) */
724 RB_INIT(&ip->rec_tree);
725 TAILQ_INIT(&ip->target_list);
727 ip->ino_data.atime = trans->time;
728 ip->ino_data.mtime = trans->time;
729 ip->ino_data.size = 0;
730 ip->ino_data.nlinks = 0;
733 * A nohistory designator on the parent directory is inherited by
734 * the child. We will do this even for pseudo-fs creation... the
735 * sysad can turn it off.
738 ip->ino_data.uflags = dip->ino_data.uflags &
739 (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
742 ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
743 ip->ino_leaf.base.localization = ip->obj_localization +
744 HAMMER_LOCALIZE_INODE;
745 ip->ino_leaf.base.obj_id = ip->obj_id;
746 ip->ino_leaf.base.key = 0;
747 ip->ino_leaf.base.create_tid = 0;
748 ip->ino_leaf.base.delete_tid = 0;
749 ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
750 ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
752 ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
753 ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
754 ip->ino_data.mode = vap->va_mode;
755 ip->ino_data.ctime = trans->time;
758 * If we are running version 2 or greater directory entries are
759 * inode-localized instead of data-localized.
761 if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
762 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
763 ip->ino_data.cap_flags |=
764 HAMMER_INODE_CAP_DIR_LOCAL_INO;
769 * Setup the ".." pointer. This only needs to be done for directories
770 * but we do it for all objects as a recovery aid.
773 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
776 * The parent_obj_localization field only applies to pseudo-fs roots.
777 * XXX this is no longer applicable, PFSs are no longer directly
778 * tied into the parent's directory structure.
780 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
781 ip->obj_id == HAMMER_OBJID_ROOT) {
782 ip->ino_data.ext.obj.parent_obj_localization =
783 dip->obj_localization;
787 switch(ip->ino_leaf.base.obj_type) {
788 case HAMMER_OBJTYPE_CDEV:
789 case HAMMER_OBJTYPE_BDEV:
790 ip->ino_data.rmajor = vap->va_rmajor;
791 ip->ino_data.rminor = vap->va_rminor;
798 * Calculate default uid/gid and overwrite with information from
802 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
803 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
804 xuid, cred, &vap->va_mode);
808 ip->ino_data.mode = vap->va_mode;
810 if (vap->va_vaflags & VA_UID_UUID_VALID)
811 ip->ino_data.uid = vap->va_uid_uuid;
812 else if (vap->va_uid != (uid_t)VNOVAL)
813 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
815 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
817 if (vap->va_vaflags & VA_GID_UUID_VALID)
818 ip->ino_data.gid = vap->va_gid_uuid;
819 else if (vap->va_gid != (gid_t)VNOVAL)
820 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
822 ip->ino_data.gid = dip->ino_data.gid;
824 hammer_ref(&ip->lock);
828 hammer_ref(&pfsm->lock);
830 } else if (dip->obj_localization == ip->obj_localization) {
831 ip->pfsm = dip->pfsm;
832 hammer_ref(&ip->pfsm->lock);
835 ip->pfsm = hammer_load_pseudofs(trans,
836 ip->obj_localization,
838 error = 0; /* ignore ENOENT */
842 hammer_free_inode(ip);
844 } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
845 panic("hammer_create_inode: duplicate obj_id %llx",
846 (long long)ip->obj_id);
848 hammer_free_inode(ip);
855 * Final cleanup / freeing of an inode structure
858 hammer_free_inode(hammer_inode_t ip)
860 struct hammer_mount *hmp;
863 KKASSERT(ip->lock.refs == 1);
864 hammer_uncache_node(&ip->cache[0]);
865 hammer_uncache_node(&ip->cache[1]);
866 hammer_uncache_node(&ip->cache[2]);
867 hammer_uncache_node(&ip->cache[3]);
868 hammer_inode_wakereclaims(ip);
870 hammer_clear_objid(ip);
871 --hammer_count_inodes;
874 hammer_rel_pseudofs(hmp, ip->pfsm);
877 kfree(ip, hmp->m_inodes);
882 * Retrieve pseudo-fs data. NULL will never be returned.
884 * If an error occurs *errorp will be set and a default template is returned,
885 * otherwise *errorp is set to 0. Typically when an error occurs it will
888 hammer_pseudofs_inmem_t
889 hammer_load_pseudofs(hammer_transaction_t trans,
890 u_int32_t localization, int *errorp)
892 hammer_mount_t hmp = trans->hmp;
894 hammer_pseudofs_inmem_t pfsm;
895 struct hammer_cursor cursor;
899 pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
901 hammer_ref(&pfsm->lock);
907 * PFS records are stored in the root inode (not the PFS root inode,
908 * but the real root). Avoid an infinite recursion if loading
909 * the PFS for the real root.
912 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
914 HAMMER_DEF_LOCALIZATION, 0, errorp);
919 pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
920 pfsm->localization = localization;
921 pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
922 pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
924 hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
925 cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
926 HAMMER_LOCALIZE_MISC;
927 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
928 cursor.key_beg.create_tid = 0;
929 cursor.key_beg.delete_tid = 0;
930 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
931 cursor.key_beg.obj_type = 0;
932 cursor.key_beg.key = localization;
933 cursor.asof = HAMMER_MAX_TID;
934 cursor.flags |= HAMMER_CURSOR_ASOF;
937 *errorp = hammer_ip_lookup(&cursor);
939 *errorp = hammer_btree_lookup(&cursor);
941 *errorp = hammer_ip_resolve_data(&cursor);
943 if (cursor.data->pfsd.mirror_flags &
944 HAMMER_PFSD_DELETED) {
947 bytes = cursor.leaf->data_len;
948 if (bytes > sizeof(pfsm->pfsd))
949 bytes = sizeof(pfsm->pfsd);
950 bcopy(cursor.data, &pfsm->pfsd, bytes);
954 hammer_done_cursor(&cursor);
956 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
957 hammer_ref(&pfsm->lock);
959 hammer_rel_inode(ip, 0);
960 if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
961 kfree(pfsm, hmp->m_misc);
968 * Store pseudo-fs data. The backend will automatically delete any prior
969 * on-disk pseudo-fs data but we have to delete in-memory versions.
972 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
974 struct hammer_cursor cursor;
975 hammer_record_t record;
979 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
980 HAMMER_DEF_LOCALIZATION, 0, &error);
982 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
983 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
984 cursor.key_beg.localization = ip->obj_localization +
985 HAMMER_LOCALIZE_MISC;
986 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
987 cursor.key_beg.create_tid = 0;
988 cursor.key_beg.delete_tid = 0;
989 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
990 cursor.key_beg.obj_type = 0;
991 cursor.key_beg.key = pfsm->localization;
992 cursor.asof = HAMMER_MAX_TID;
993 cursor.flags |= HAMMER_CURSOR_ASOF;
996 * Replace any in-memory version of the record.
998 error = hammer_ip_lookup(&cursor);
999 if (error == 0 && hammer_cursor_inmem(&cursor)) {
1000 record = cursor.iprec;
1001 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1002 KKASSERT(cursor.deadlk_rec == NULL);
1003 hammer_ref(&record->lock);
1004 cursor.deadlk_rec = record;
1007 record->flags |= HAMMER_RECF_DELETED_FE;
1013 * Allocate replacement general record. The backend flush will
1014 * delete any on-disk version of the record.
1016 if (error == 0 || error == ENOENT) {
1017 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1018 record->type = HAMMER_MEM_RECORD_GENERAL;
1020 record->leaf.base.localization = ip->obj_localization +
1021 HAMMER_LOCALIZE_MISC;
1022 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1023 record->leaf.base.key = pfsm->localization;
1024 record->leaf.data_len = sizeof(pfsm->pfsd);
1025 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1026 error = hammer_ip_add_record(trans, record);
1028 hammer_done_cursor(&cursor);
1029 if (error == EDEADLK)
1031 hammer_rel_inode(ip, 0);
1036 * Create a root directory for a PFS if one does not alredy exist.
1038 * The PFS root stands alone so we must also bump the nlinks count
1039 * to prevent it from being destroyed on release.
1042 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1043 hammer_pseudofs_inmem_t pfsm)
1049 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1050 pfsm->localization, 0, &error);
1055 error = hammer_create_inode(trans, &vap, cred,
1059 ++ip->ino_data.nlinks;
1060 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1064 hammer_rel_inode(ip, 0);
1069 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1070 * if we are unable to disassociate all the inodes.
1074 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1078 hammer_ref(&ip->lock);
1079 if (ip->lock.refs == 2 && ip->vp)
1080 vclean_unlocked(ip->vp);
1081 if (ip->lock.refs == 1 && ip->vp == NULL)
1084 res = -1; /* stop, someone is using the inode */
1085 hammer_rel_inode(ip, 0);
1090 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1095 for (try = res = 0; try < 4; ++try) {
1096 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1097 hammer_inode_pfs_cmp,
1098 hammer_unload_pseudofs_callback,
1100 if (res == 0 && try > 1)
1102 hammer_flusher_sync(trans->hmp);
1111 * Release a reference on a PFS
1114 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1116 hammer_unref(&pfsm->lock);
1117 if (pfsm->lock.refs == 0) {
1118 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1119 kfree(pfsm, hmp->m_misc);
1124 * Called by hammer_sync_inode().
1127 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1129 hammer_transaction_t trans = cursor->trans;
1130 hammer_record_t record;
1138 * If the inode has a presence on-disk then locate it and mark
1139 * it deleted, setting DELONDISK.
1141 * The record may or may not be physically deleted, depending on
1142 * the retention policy.
1144 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1145 HAMMER_INODE_ONDISK) {
1146 hammer_normalize_cursor(cursor);
1147 cursor->key_beg.localization = ip->obj_localization +
1148 HAMMER_LOCALIZE_INODE;
1149 cursor->key_beg.obj_id = ip->obj_id;
1150 cursor->key_beg.key = 0;
1151 cursor->key_beg.create_tid = 0;
1152 cursor->key_beg.delete_tid = 0;
1153 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1154 cursor->key_beg.obj_type = 0;
1155 cursor->asof = ip->obj_asof;
1156 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1157 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1158 cursor->flags |= HAMMER_CURSOR_BACKEND;
1160 error = hammer_btree_lookup(cursor);
1161 if (hammer_debug_inode)
1162 kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1165 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1166 if (hammer_debug_inode)
1167 kprintf(" error %d\n", error);
1169 ip->flags |= HAMMER_INODE_DELONDISK;
1172 hammer_cache_node(&ip->cache[0], cursor->node);
1174 if (error == EDEADLK) {
1175 hammer_done_cursor(cursor);
1176 error = hammer_init_cursor(trans, cursor,
1178 if (hammer_debug_inode)
1179 kprintf("IPDED %p %d\n", ip, error);
1186 * Ok, write out the initial record or a new record (after deleting
1187 * the old one), unless the DELETED flag is set. This routine will
1188 * clear DELONDISK if it writes out a record.
1190 * Update our inode statistics if this is the first application of
1191 * the inode on-disk.
1193 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1195 * Generate a record and write it to the media. We clean-up
1196 * the state before releasing so we do not have to set-up
1199 record = hammer_alloc_mem_record(ip, 0);
1200 record->type = HAMMER_MEM_RECORD_INODE;
1201 record->flush_state = HAMMER_FST_FLUSH;
1202 record->leaf = ip->sync_ino_leaf;
1203 record->leaf.base.create_tid = trans->tid;
1204 record->leaf.data_len = sizeof(ip->sync_ino_data);
1205 record->leaf.create_ts = trans->time32;
1206 record->data = (void *)&ip->sync_ino_data;
1207 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1210 * If this flag is set we cannot sync the new file size
1211 * because we haven't finished related truncations. The
1212 * inode will be flushed in another flush group to finish
1215 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1216 ip->sync_ino_data.size != ip->ino_data.size) {
1218 ip->sync_ino_data.size = ip->ino_data.size;
1224 error = hammer_ip_sync_record_cursor(cursor, record);
1225 if (hammer_debug_inode)
1226 kprintf("GENREC %p rec %08x %d\n",
1227 ip, record->flags, error);
1228 if (error != EDEADLK)
1230 hammer_done_cursor(cursor);
1231 error = hammer_init_cursor(trans, cursor,
1233 if (hammer_debug_inode)
1234 kprintf("GENREC reinit %d\n", error);
1240 * Note: The record was never on the inode's record tree
1241 * so just wave our hands importantly and destroy it.
1243 record->flags |= HAMMER_RECF_COMMITTED;
1244 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1245 record->flush_state = HAMMER_FST_IDLE;
1246 ++ip->rec_generation;
1247 hammer_rel_mem_record(record);
1253 if (hammer_debug_inode)
1254 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1255 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1256 HAMMER_INODE_ATIME |
1257 HAMMER_INODE_MTIME);
1258 ip->flags &= ~HAMMER_INODE_DELONDISK;
1260 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1263 * Root volume count of inodes
1265 hammer_sync_lock_sh(trans);
1266 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1267 hammer_modify_volume_field(trans,
1270 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1271 hammer_modify_volume_done(trans->rootvol);
1272 ip->flags |= HAMMER_INODE_ONDISK;
1273 if (hammer_debug_inode)
1274 kprintf("NOWONDISK %p\n", ip);
1276 hammer_sync_unlock(trans);
1281 * If the inode has been destroyed, clean out any left-over flags
1282 * that may have been set by the frontend.
1284 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1285 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1286 HAMMER_INODE_ATIME |
1287 HAMMER_INODE_MTIME);
1293 * Update only the itimes fields.
1295 * ATIME can be updated without generating any UNDO. MTIME is updated
1296 * with UNDO so it is guaranteed to be synchronized properly in case of
1299 * Neither field is included in the B-Tree leaf element's CRC, which is how
1300 * we can get away with updating ATIME the way we do.
1303 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1305 hammer_transaction_t trans = cursor->trans;
1309 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1310 HAMMER_INODE_ONDISK) {
1314 hammer_normalize_cursor(cursor);
1315 cursor->key_beg.localization = ip->obj_localization +
1316 HAMMER_LOCALIZE_INODE;
1317 cursor->key_beg.obj_id = ip->obj_id;
1318 cursor->key_beg.key = 0;
1319 cursor->key_beg.create_tid = 0;
1320 cursor->key_beg.delete_tid = 0;
1321 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1322 cursor->key_beg.obj_type = 0;
1323 cursor->asof = ip->obj_asof;
1324 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1325 cursor->flags |= HAMMER_CURSOR_ASOF;
1326 cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1327 cursor->flags |= HAMMER_CURSOR_GET_DATA;
1328 cursor->flags |= HAMMER_CURSOR_BACKEND;
1330 error = hammer_btree_lookup(cursor);
1332 hammer_cache_node(&ip->cache[0], cursor->node);
1333 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1335 * Updating MTIME requires an UNDO. Just cover
1336 * both atime and mtime.
1338 hammer_sync_lock_sh(trans);
1339 hammer_modify_buffer(trans, cursor->data_buffer,
1340 HAMMER_ITIMES_BASE(&cursor->data->inode),
1341 HAMMER_ITIMES_BYTES);
1342 cursor->data->inode.atime = ip->sync_ino_data.atime;
1343 cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1344 hammer_modify_buffer_done(cursor->data_buffer);
1345 hammer_sync_unlock(trans);
1346 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1348 * Updating atime only can be done in-place with
1351 hammer_sync_lock_sh(trans);
1352 hammer_modify_buffer(trans, cursor->data_buffer,
1354 cursor->data->inode.atime = ip->sync_ino_data.atime;
1355 hammer_modify_buffer_done(cursor->data_buffer);
1356 hammer_sync_unlock(trans);
1358 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1360 if (error == EDEADLK) {
1361 hammer_done_cursor(cursor);
1362 error = hammer_init_cursor(trans, cursor,
1371 * Release a reference on an inode, flush as requested.
1373 * On the last reference we queue the inode to the flusher for its final
1377 hammer_rel_inode(struct hammer_inode *ip, int flush)
1379 /*hammer_mount_t hmp = ip->hmp;*/
1382 * Handle disposition when dropping the last ref.
1385 if (ip->lock.refs == 1) {
1387 * Determine whether on-disk action is needed for
1388 * the inode's final disposition.
1390 KKASSERT(ip->vp == NULL);
1391 hammer_inode_unloadable_check(ip, 0);
1392 if (ip->flags & HAMMER_INODE_MODMASK) {
1393 hammer_flush_inode(ip, 0);
1394 } else if (ip->lock.refs == 1) {
1395 hammer_unload_inode(ip);
1400 hammer_flush_inode(ip, 0);
1403 * The inode still has multiple refs, try to drop
1406 KKASSERT(ip->lock.refs >= 1);
1407 if (ip->lock.refs > 1) {
1408 hammer_unref(&ip->lock);
1416 * Unload and destroy the specified inode. Must be called with one remaining
1417 * reference. The reference is disposed of.
1419 * The inode must be completely clean.
1422 hammer_unload_inode(struct hammer_inode *ip)
1424 hammer_mount_t hmp = ip->hmp;
1426 KASSERT(ip->lock.refs == 1,
1427 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1428 KKASSERT(ip->vp == NULL);
1429 KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1430 KKASSERT(ip->cursor_ip_refs == 0);
1431 KKASSERT(hammer_notlocked(&ip->lock));
1432 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1434 KKASSERT(RB_EMPTY(&ip->rec_tree));
1435 KKASSERT(TAILQ_EMPTY(&ip->target_list));
1437 RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1439 hammer_free_inode(ip);
1444 * Called during unmounting if a critical error occured. The in-memory
1445 * inode and all related structures are destroyed.
1447 * If a critical error did not occur the unmount code calls the standard
1448 * release and asserts that the inode is gone.
1451 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1453 hammer_record_t rec;
1456 * Get rid of the inodes in-memory records, regardless of their
1457 * state, and clear the mod-mask.
1459 while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1460 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1461 rec->target_ip = NULL;
1462 if (rec->flush_state == HAMMER_FST_SETUP)
1463 rec->flush_state = HAMMER_FST_IDLE;
1465 while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1466 if (rec->flush_state == HAMMER_FST_FLUSH)
1467 --rec->flush_group->refs;
1469 hammer_ref(&rec->lock);
1470 KKASSERT(rec->lock.refs == 1);
1471 rec->flush_state = HAMMER_FST_IDLE;
1472 rec->flush_group = NULL;
1473 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1474 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1475 ++ip->rec_generation;
1476 hammer_rel_mem_record(rec);
1478 ip->flags &= ~HAMMER_INODE_MODMASK;
1479 ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1480 KKASSERT(ip->vp == NULL);
1483 * Remove the inode from any flush group, force it idle. FLUSH
1484 * and SETUP states have an inode ref.
1486 switch(ip->flush_state) {
1487 case HAMMER_FST_FLUSH:
1488 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1489 --ip->flush_group->refs;
1490 ip->flush_group = NULL;
1492 case HAMMER_FST_SETUP:
1493 hammer_unref(&ip->lock);
1494 ip->flush_state = HAMMER_FST_IDLE;
1496 case HAMMER_FST_IDLE:
1501 * There shouldn't be any associated vnode. The unload needs at
1502 * least one ref, if we do have a vp steal its ip ref.
1505 kprintf("hammer_destroy_inode_callback: Unexpected "
1506 "vnode association ip %p vp %p\n", ip, ip->vp);
1507 ip->vp->v_data = NULL;
1510 hammer_ref(&ip->lock);
1512 hammer_unload_inode(ip);
1517 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1518 * the read-only flag for cached inodes.
1520 * This routine is called from a RB_SCAN().
1523 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1525 hammer_mount_t hmp = ip->hmp;
1527 if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1528 ip->flags |= HAMMER_INODE_RO;
1530 ip->flags &= ~HAMMER_INODE_RO;
1535 * A transaction has modified an inode, requiring updates as specified by
1538 * HAMMER_INODE_DDIRTY: Inode data has been updated
1539 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1540 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1541 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1542 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1545 hammer_modify_inode(hammer_inode_t ip, int flags)
1548 * ronly of 0 or 2 does not trigger assertion.
1549 * 2 is a special error state
1551 KKASSERT(ip->hmp->ronly != 1 ||
1552 (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1553 HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1554 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1555 if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1556 ip->flags |= HAMMER_INODE_RSV_INODES;
1557 ++ip->hmp->rsv_inodes;
1564 * Request that an inode be flushed. This whole mess cannot block and may
1565 * recurse (if not synchronous). Once requested HAMMER will attempt to
1566 * actively flush the inode until the flush can be done.
1568 * The inode may already be flushing, or may be in a setup state. We can
1569 * place the inode in a flushing state if it is currently idle and flag it
1570 * to reflush if it is currently flushing.
1572 * Upon return if the inode could not be flushed due to a setup
1573 * dependancy, then it will be automatically flushed when the dependancy
1577 hammer_flush_inode(hammer_inode_t ip, int flags)
1580 hammer_flush_group_t flg;
1584 * next_flush_group is the first flush group we can place the inode
1585 * in. It may be NULL. If it becomes full we append a new flush
1586 * group and make that the next_flush_group.
1589 while ((flg = hmp->next_flush_group) != NULL) {
1590 KKASSERT(flg->running == 0);
1591 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1593 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1594 hammer_flusher_async(ip->hmp, flg);
1597 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1598 hmp->next_flush_group = flg;
1599 RB_INIT(&flg->flush_tree);
1600 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1604 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1605 * state we have to put it back into an IDLE state so we can
1606 * drop the extra ref.
1608 * If we have a parent dependancy we must still fall through
1611 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1612 if (ip->flush_state == HAMMER_FST_SETUP &&
1613 TAILQ_EMPTY(&ip->target_list)) {
1614 ip->flush_state = HAMMER_FST_IDLE;
1615 hammer_rel_inode(ip, 0);
1617 if (ip->flush_state == HAMMER_FST_IDLE)
1622 * Our flush action will depend on the current state.
1624 switch(ip->flush_state) {
1625 case HAMMER_FST_IDLE:
1627 * We have no dependancies and can flush immediately. Some
1628 * our children may not be flushable so we have to re-test
1629 * with that additional knowledge.
1631 hammer_flush_inode_core(ip, flg, flags);
1633 case HAMMER_FST_SETUP:
1635 * Recurse upwards through dependancies via target_list
1636 * and start their flusher actions going if possible.
1638 * 'good' is our connectivity. -1 means we have none and
1639 * can't flush, 0 means there weren't any dependancies, and
1640 * 1 means we have good connectivity.
1642 good = hammer_setup_parent_inodes(ip, 0, flg);
1646 * We can continue if good >= 0. Determine how
1647 * many records under our inode can be flushed (and
1650 hammer_flush_inode_core(ip, flg, flags);
1653 * Parent has no connectivity, tell it to flush
1654 * us as soon as it does.
1656 * The REFLUSH flag is also needed to trigger
1657 * dependancy wakeups.
1659 ip->flags |= HAMMER_INODE_CONN_DOWN |
1660 HAMMER_INODE_REFLUSH;
1661 if (flags & HAMMER_FLUSH_SIGNAL) {
1662 ip->flags |= HAMMER_INODE_RESIGNAL;
1663 hammer_flusher_async(ip->hmp, flg);
1667 case HAMMER_FST_FLUSH:
1669 * We are already flushing, flag the inode to reflush
1670 * if needed after it completes its current flush.
1672 * The REFLUSH flag is also needed to trigger
1673 * dependancy wakeups.
1675 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1676 ip->flags |= HAMMER_INODE_REFLUSH;
1677 if (flags & HAMMER_FLUSH_SIGNAL) {
1678 ip->flags |= HAMMER_INODE_RESIGNAL;
1679 hammer_flusher_async(ip->hmp, flg);
1686 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1687 * ip which reference our ip.
1689 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1690 * so for now do not ref/deref the structures. Note that if we use the
1691 * ref/rel code later, the rel CAN block.
1694 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1695 hammer_flush_group_t flg)
1697 hammer_record_t depend;
1702 * If we hit our recursion limit and we have parent dependencies
1703 * We cannot continue. Returning < 0 will cause us to be flagged
1704 * for reflush. Returning -2 cuts off additional dependency checks
1705 * because they are likely to also hit the depth limit.
1707 * We cannot return < 0 if there are no dependencies or there might
1708 * not be anything to wakeup (ip).
1710 if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1711 kprintf("HAMMER Warning: depth limit reached on "
1712 "setup recursion, inode %p %016llx\n",
1713 ip, (long long)ip->obj_id);
1721 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1722 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1723 KKASSERT(depend->target_ip == ip);
1724 if (r < 0 && good == 0)
1730 * If we failed due to the recursion depth limit then stop
1740 * This helper function takes a record representing the dependancy between
1741 * the parent inode and child inode.
1743 * record->ip = parent inode
1744 * record->target_ip = child inode
1746 * We are asked to recurse upwards and convert the record from SETUP
1747 * to FLUSH if possible.
1749 * Return 1 if the record gives us connectivity
1751 * Return 0 if the record is not relevant
1753 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1756 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1757 hammer_flush_group_t flg)
1763 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1768 * If the record is already flushing, is it in our flush group?
1770 * If it is in our flush group but it is a general record or a
1771 * delete-on-disk, it does not improve our connectivity (return 0),
1772 * and if the target inode is not trying to destroy itself we can't
1773 * allow the operation yet anyway (the second return -1).
1775 if (record->flush_state == HAMMER_FST_FLUSH) {
1777 * If not in our flush group ask the parent to reflush
1778 * us as soon as possible.
1780 if (record->flush_group != flg) {
1781 pip->flags |= HAMMER_INODE_REFLUSH;
1782 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1787 * If in our flush group everything is already set up,
1788 * just return whether the record will improve our
1789 * visibility or not.
1791 if (record->type == HAMMER_MEM_RECORD_ADD)
1797 * It must be a setup record. Try to resolve the setup dependancies
1798 * by recursing upwards so we can place ip on the flush list.
1800 * Limit ourselves to 20 levels of recursion to avoid blowing out
1801 * the kernel stack. If we hit the recursion limit we can't flush
1802 * until the parent flushes. The parent will flush independantly
1803 * on its own and ultimately a deep recursion will be resolved.
1805 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1807 good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1810 * If good < 0 the parent has no connectivity and we cannot safely
1811 * flush the directory entry, which also means we can't flush our
1812 * ip. Flag us for downward recursion once the parent's
1813 * connectivity is resolved. Flag the parent for [re]flush or it
1814 * may not check for downward recursions.
1817 pip->flags |= HAMMER_INODE_REFLUSH;
1818 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1823 * We are go, place the parent inode in a flushing state so we can
1824 * place its record in a flushing state. Note that the parent
1825 * may already be flushing. The record must be in the same flush
1826 * group as the parent.
1828 if (pip->flush_state != HAMMER_FST_FLUSH)
1829 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1830 KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1831 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1834 if (record->type == HAMMER_MEM_RECORD_DEL &&
1835 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1837 * Regardless of flushing state we cannot sync this path if the
1838 * record represents a delete-on-disk but the target inode
1839 * is not ready to sync its own deletion.
1841 * XXX need to count effective nlinks to determine whether
1842 * the flush is ok, otherwise removing a hardlink will
1843 * just leave the DEL record to rot.
1845 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1849 if (pip->flush_group == flg) {
1851 * Because we have not calculated nlinks yet we can just
1852 * set records to the flush state if the parent is in
1853 * the same flush group as we are.
1855 record->flush_state = HAMMER_FST_FLUSH;
1856 record->flush_group = flg;
1857 ++record->flush_group->refs;
1858 hammer_ref(&record->lock);
1861 * A general directory-add contributes to our visibility.
1863 * Otherwise it is probably a directory-delete or
1864 * delete-on-disk record and does not contribute to our
1865 * visbility (but we can still flush it).
1867 if (record->type == HAMMER_MEM_RECORD_ADD)
1872 * If the parent is not in our flush group we cannot
1873 * flush this record yet, there is no visibility.
1874 * We tell the parent to reflush and mark ourselves
1875 * so the parent knows it should flush us too.
1877 pip->flags |= HAMMER_INODE_REFLUSH;
1878 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1884 * This is the core routine placing an inode into the FST_FLUSH state.
1887 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1892 * Set flush state and prevent the flusher from cycling into
1893 * the next flush group. Do not place the ip on the list yet.
1894 * Inodes not in the idle state get an extra reference.
1896 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1897 if (ip->flush_state == HAMMER_FST_IDLE)
1898 hammer_ref(&ip->lock);
1899 ip->flush_state = HAMMER_FST_FLUSH;
1900 ip->flush_group = flg;
1901 ++ip->hmp->flusher.group_lock;
1902 ++ip->hmp->count_iqueued;
1903 ++hammer_count_iqueued;
1907 * If the flush group reaches the autoflush limit we want to signal
1908 * the flusher. This is particularly important for remove()s.
1910 if (flg->total_count == hammer_autoflush)
1911 flags |= HAMMER_FLUSH_SIGNAL;
1914 * We need to be able to vfsync/truncate from the backend.
1916 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1917 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1918 ip->flags |= HAMMER_INODE_VHELD;
1923 * Figure out how many in-memory records we can actually flush
1924 * (not including inode meta-data, buffers, etc).
1926 KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1927 if (flags & HAMMER_FLUSH_RECURSION) {
1929 * If this is a upwards recursion we do not want to
1930 * recurse down again!
1934 } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1936 * No new records are added if we must complete a flush
1937 * from a previous cycle, but we do have to move the records
1938 * from the previous cycle to the current one.
1941 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1942 hammer_syncgrp_child_callback, NULL);
1948 * Normal flush, scan records and bring them into the flush.
1949 * Directory adds and deletes are usually skipped (they are
1950 * grouped with the related inode rather then with the
1953 * go_count can be negative, which means the scan aborted
1954 * due to the flush group being over-full and we should
1955 * flush what we have.
1957 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1958 hammer_setup_child_callback, NULL);
1962 * This is a more involved test that includes go_count. If we
1963 * can't flush, flag the inode and return. If go_count is 0 we
1964 * were are unable to flush any records in our rec_tree and
1965 * must ignore the XDIRTY flag.
1967 if (go_count == 0) {
1968 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1969 --ip->hmp->count_iqueued;
1970 --hammer_count_iqueued;
1973 ip->flush_state = HAMMER_FST_SETUP;
1974 ip->flush_group = NULL;
1975 if (ip->flags & HAMMER_INODE_VHELD) {
1976 ip->flags &= ~HAMMER_INODE_VHELD;
1981 * REFLUSH is needed to trigger dependancy wakeups
1982 * when an inode is in SETUP.
1984 ip->flags |= HAMMER_INODE_REFLUSH;
1985 if (flags & HAMMER_FLUSH_SIGNAL) {
1986 ip->flags |= HAMMER_INODE_RESIGNAL;
1987 hammer_flusher_async(ip->hmp, flg);
1989 if (--ip->hmp->flusher.group_lock == 0)
1990 wakeup(&ip->hmp->flusher.group_lock);
1996 * Snapshot the state of the inode for the backend flusher.
1998 * We continue to retain save_trunc_off even when all truncations
1999 * have been resolved as an optimization to determine if we can
2000 * skip the B-Tree lookup for overwrite deletions.
2002 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2003 * and stays in ip->flags. Once set, it stays set until the
2004 * inode is destroyed.
2006 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2007 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2008 ip->sync_trunc_off = ip->trunc_off;
2009 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2010 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2011 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2014 * The save_trunc_off used to cache whether the B-Tree
2015 * holds any records past that point is not used until
2016 * after the truncation has succeeded, so we can safely
2019 if (ip->save_trunc_off > ip->sync_trunc_off)
2020 ip->save_trunc_off = ip->sync_trunc_off;
2022 ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2023 ~HAMMER_INODE_TRUNCATED);
2024 ip->sync_ino_leaf = ip->ino_leaf;
2025 ip->sync_ino_data = ip->ino_data;
2026 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2027 #ifdef DEBUG_TRUNCATE
2028 if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2029 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2033 * The flusher list inherits our inode and reference.
2035 KKASSERT(flg->running == 0);
2036 RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2037 if (--ip->hmp->flusher.group_lock == 0)
2038 wakeup(&ip->hmp->flusher.group_lock);
2040 if (flags & HAMMER_FLUSH_SIGNAL) {
2041 hammer_flusher_async(ip->hmp, flg);
2046 * Callback for scan of ip->rec_tree. Try to include each record in our
2047 * flush. ip->flush_group has been set but the inode has not yet been
2048 * moved into a flushing state.
2050 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2053 * We return 1 for any record placed or found in FST_FLUSH, which prevents
2054 * the caller from shortcutting the flush.
2057 hammer_setup_child_callback(hammer_record_t rec, void *data)
2059 hammer_flush_group_t flg;
2060 hammer_inode_t target_ip;
2065 * Records deleted or committed by the backend are ignored.
2066 * Note that the flush detects deleted frontend records at
2067 * multiple points to deal with races. This is just the first
2068 * line of defense. The only time HAMMER_RECF_DELETED_FE cannot
2069 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2070 * messes up link-count calculations.
2072 * NOTE: Don't get confused between record deletion and, say,
2073 * directory entry deletion. The deletion of a directory entry
2074 * which is on-media has nothing to do with the record deletion
2077 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2078 HAMMER_RECF_COMMITTED)) {
2079 if (rec->flush_state == HAMMER_FST_FLUSH) {
2080 KKASSERT(rec->flush_group == rec->ip->flush_group);
2089 * If the record is in an idle state it has no dependancies and
2093 flg = ip->flush_group;
2096 switch(rec->flush_state) {
2097 case HAMMER_FST_IDLE:
2099 * The record has no setup dependancy, we can flush it.
2101 KKASSERT(rec->target_ip == NULL);
2102 rec->flush_state = HAMMER_FST_FLUSH;
2103 rec->flush_group = flg;
2105 hammer_ref(&rec->lock);
2108 case HAMMER_FST_SETUP:
2110 * The record has a setup dependancy. These are typically
2111 * directory entry adds and deletes. Such entries will be
2112 * flushed when their inodes are flushed so we do not
2113 * usually have to add them to the flush here. However,
2114 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2115 * it is asking us to flush this record (and it).
2117 target_ip = rec->target_ip;
2118 KKASSERT(target_ip != NULL);
2119 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2122 * If the target IP is already flushing in our group
2123 * we could associate the record, but target_ip has
2124 * already synced ino_data to sync_ino_data and we
2125 * would also have to adjust nlinks. Plus there are
2126 * ordering issues for adds and deletes.
2128 * Reflush downward if this is an ADD, and upward if
2131 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2132 if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2133 ip->flags |= HAMMER_INODE_REFLUSH;
2135 target_ip->flags |= HAMMER_INODE_REFLUSH;
2140 * Target IP is not yet flushing. This can get complex
2141 * because we have to be careful about the recursion.
2143 * Directories create an issue for us in that if a flush
2144 * of a directory is requested the expectation is to flush
2145 * any pending directory entries, but this will cause the
2146 * related inodes to recursively flush as well. We can't
2147 * really defer the operation so just get as many as we
2151 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2152 (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2154 * We aren't reclaiming and the target ip was not
2155 * previously prevented from flushing due to this
2156 * record dependancy. Do not flush this record.
2161 if (flg->total_count + flg->refs >
2162 ip->hmp->undo_rec_limit) {
2164 * Our flush group is over-full and we risk blowing
2165 * out the UNDO FIFO. Stop the scan, flush what we
2166 * have, then reflush the directory.
2168 * The directory may be forced through multiple
2169 * flush groups before it can be completely
2172 ip->flags |= HAMMER_INODE_RESIGNAL |
2173 HAMMER_INODE_REFLUSH;
2175 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2177 * If the target IP is not flushing we can force
2178 * it to flush, even if it is unable to write out
2179 * any of its own records we have at least one in
2180 * hand that we CAN deal with.
2182 rec->flush_state = HAMMER_FST_FLUSH;
2183 rec->flush_group = flg;
2185 hammer_ref(&rec->lock);
2186 hammer_flush_inode_core(target_ip, flg,
2187 HAMMER_FLUSH_RECURSION);
2191 * General or delete-on-disk record.
2193 * XXX this needs help. If a delete-on-disk we could
2194 * disconnect the target. If the target has its own
2195 * dependancies they really need to be flushed.
2199 rec->flush_state = HAMMER_FST_FLUSH;
2200 rec->flush_group = flg;
2202 hammer_ref(&rec->lock);
2203 hammer_flush_inode_core(target_ip, flg,
2204 HAMMER_FLUSH_RECURSION);
2208 case HAMMER_FST_FLUSH:
2210 * The flush_group should already match.
2212 KKASSERT(rec->flush_group == flg);
2221 * This version just moves records already in a flush state to the new
2222 * flush group and that is it.
2225 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2227 hammer_inode_t ip = rec->ip;
2229 switch(rec->flush_state) {
2230 case HAMMER_FST_FLUSH:
2231 KKASSERT(rec->flush_group == ip->flush_group);
2241 * Wait for a previously queued flush to complete.
2243 * If a critical error occured we don't try to wait.
2246 hammer_wait_inode(hammer_inode_t ip)
2248 hammer_flush_group_t flg;
2251 if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2252 while (ip->flush_state != HAMMER_FST_IDLE &&
2253 (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2254 if (ip->flush_state == HAMMER_FST_SETUP)
2255 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2256 if (ip->flush_state != HAMMER_FST_IDLE) {
2257 ip->flags |= HAMMER_INODE_FLUSHW;
2258 tsleep(&ip->flags, 0, "hmrwin", 0);
2265 * Called by the backend code when a flush has been completed.
2266 * The inode has already been removed from the flush list.
2268 * A pipelined flush can occur, in which case we must re-enter the
2269 * inode on the list and re-copy its fields.
2272 hammer_flush_inode_done(hammer_inode_t ip, int error)
2277 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2282 * Auto-reflush if the backend could not completely flush
2283 * the inode. This fixes a case where a deferred buffer flush
2284 * could cause fsync to return early.
2286 if (ip->sync_flags & HAMMER_INODE_MODMASK)
2287 ip->flags |= HAMMER_INODE_REFLUSH;
2290 * Merge left-over flags back into the frontend and fix the state.
2291 * Incomplete truncations are retained by the backend.
2294 ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2295 ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2298 * The backend may have adjusted nlinks, so if the adjusted nlinks
2299 * does not match the fronttend set the frontend's RDIRTY flag again.
2301 if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2302 ip->flags |= HAMMER_INODE_DDIRTY;
2305 * Fix up the dirty buffer status.
2307 if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2308 ip->flags |= HAMMER_INODE_BUFS;
2312 * Re-set the XDIRTY flag if some of the inode's in-memory records
2313 * could not be flushed.
2315 KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2316 (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2317 (!RB_EMPTY(&ip->rec_tree) &&
2318 (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2321 * Do not lose track of inodes which no longer have vnode
2322 * assocations, otherwise they may never get flushed again.
2324 * The reflush flag can be set superfluously, causing extra pain
2325 * for no reason. If the inode is no longer modified it no longer
2326 * needs to be flushed.
2328 if (ip->flags & HAMMER_INODE_MODMASK) {
2330 ip->flags |= HAMMER_INODE_REFLUSH;
2332 ip->flags &= ~HAMMER_INODE_REFLUSH;
2336 * Adjust the flush state.
2338 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2340 * We were unable to flush out all our records, leave the
2341 * inode in a flush state and in the current flush group.
2342 * The flush group will be re-run.
2344 * This occurs if the UNDO block gets too full or there is
2345 * too much dirty meta-data and allows the flusher to
2346 * finalize the UNDO block and then re-flush.
2348 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2352 * Remove from the flush_group
2354 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2355 ip->flush_group = NULL;
2358 * Clean up the vnode ref and tracking counts.
2360 if (ip->flags & HAMMER_INODE_VHELD) {
2361 ip->flags &= ~HAMMER_INODE_VHELD;
2364 --hmp->count_iqueued;
2365 --hammer_count_iqueued;
2368 * And adjust the state.
2370 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2371 ip->flush_state = HAMMER_FST_IDLE;
2374 ip->flush_state = HAMMER_FST_SETUP;
2379 * If the frontend is waiting for a flush to complete,
2382 if (ip->flags & HAMMER_INODE_FLUSHW) {
2383 ip->flags &= ~HAMMER_INODE_FLUSHW;
2388 * If the frontend made more changes and requested another
2389 * flush, then try to get it running.
2391 * Reflushes are aborted when the inode is errored out.
2393 if (ip->flags & HAMMER_INODE_REFLUSH) {
2394 ip->flags &= ~HAMMER_INODE_REFLUSH;
2395 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2396 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2397 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2399 hammer_flush_inode(ip, 0);
2405 * If we have no parent dependancies we can clear CONN_DOWN
2407 if (TAILQ_EMPTY(&ip->target_list))
2408 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2411 * If the inode is now clean drop the space reservation.
2413 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2414 (ip->flags & HAMMER_INODE_RSV_INODES)) {
2415 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2420 hammer_rel_inode(ip, 0);
2424 * Called from hammer_sync_inode() to synchronize in-memory records
2428 hammer_sync_record_callback(hammer_record_t record, void *data)
2430 hammer_cursor_t cursor = data;
2431 hammer_transaction_t trans = cursor->trans;
2432 hammer_mount_t hmp = trans->hmp;
2436 * Skip records that do not belong to the current flush.
2438 ++hammer_stats_record_iterations;
2439 if (record->flush_state != HAMMER_FST_FLUSH)
2443 if (record->flush_group != record->ip->flush_group) {
2444 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2445 if (hammer_debug_critical)
2450 KKASSERT(record->flush_group == record->ip->flush_group);
2453 * Interlock the record using the BE flag. Once BE is set the
2454 * frontend cannot change the state of FE.
2456 * NOTE: If FE is set prior to us setting BE we still sync the
2457 * record out, but the flush completion code converts it to
2458 * a delete-on-disk record instead of destroying it.
2460 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2461 record->flags |= HAMMER_RECF_INTERLOCK_BE;
2464 * The backend has already disposed of the record.
2466 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2472 * If the whole inode is being deleting all on-disk records will
2473 * be deleted very soon, we can't sync any new records to disk
2474 * because they will be deleted in the same transaction they were
2475 * created in (delete_tid == create_tid), which will assert.
2477 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2478 * that we currently panic on.
2480 if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2481 switch(record->type) {
2482 case HAMMER_MEM_RECORD_DATA:
2484 * We don't have to do anything, if the record was
2485 * committed the space will have been accounted for
2489 case HAMMER_MEM_RECORD_GENERAL:
2491 * Set deleted-by-backend flag. Do not set the
2492 * backend committed flag, because we are throwing
2495 record->flags |= HAMMER_RECF_DELETED_BE;
2496 ++record->ip->rec_generation;
2499 case HAMMER_MEM_RECORD_ADD:
2500 panic("hammer_sync_record_callback: illegal add "
2501 "during inode deletion record %p", record);
2502 break; /* NOT REACHED */
2503 case HAMMER_MEM_RECORD_INODE:
2504 panic("hammer_sync_record_callback: attempt to "
2505 "sync inode record %p?", record);
2506 break; /* NOT REACHED */
2507 case HAMMER_MEM_RECORD_DEL:
2509 * Follow through and issue the on-disk deletion
2516 * If DELETED_FE is set special handling is needed for directory
2517 * entries. Dependant pieces related to the directory entry may
2518 * have already been synced to disk. If this occurs we have to
2519 * sync the directory entry and then change the in-memory record
2520 * from an ADD to a DELETE to cover the fact that it's been
2521 * deleted by the frontend.
2523 * A directory delete covering record (MEM_RECORD_DEL) can never
2524 * be deleted by the frontend.
2526 * Any other record type (aka DATA) can be deleted by the frontend.
2527 * XXX At the moment the flusher must skip it because there may
2528 * be another data record in the flush group for the same block,
2529 * meaning that some frontend data changes can leak into the backend's
2530 * synchronization point.
2532 if (record->flags & HAMMER_RECF_DELETED_FE) {
2533 if (record->type == HAMMER_MEM_RECORD_ADD) {
2535 * Convert a front-end deleted directory-add to
2536 * a directory-delete entry later.
2538 record->flags |= HAMMER_RECF_CONVERT_DELETE;
2541 * Dispose of the record (race case). Mark as
2542 * deleted by backend (and not committed).
2544 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2545 record->flags |= HAMMER_RECF_DELETED_BE;
2546 ++record->ip->rec_generation;
2553 * Assign the create_tid for new records. Deletions already
2554 * have the record's entire key properly set up.
2556 if (record->type != HAMMER_MEM_RECORD_DEL) {
2557 record->leaf.base.create_tid = trans->tid;
2558 record->leaf.create_ts = trans->time32;
2561 error = hammer_ip_sync_record_cursor(cursor, record);
2562 if (error != EDEADLK)
2564 hammer_done_cursor(cursor);
2565 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2570 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2575 hammer_flush_record_done(record, error);
2578 * Do partial finalization if we have built up too many dirty
2579 * buffers. Otherwise a buffer cache deadlock can occur when
2580 * doing things like creating tens of thousands of tiny files.
2582 * We must release our cursor lock to avoid a 3-way deadlock
2583 * due to the exclusive sync lock the finalizer must get.
2585 * WARNING: See warnings in hammer_unlock_cursor() function.
2587 if (hammer_flusher_meta_limit(hmp)) {
2588 hammer_unlock_cursor(cursor);
2589 hammer_flusher_finalize(trans, 0);
2590 hammer_lock_cursor(cursor);
2597 * Backend function called by the flusher to sync an inode to media.
2600 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2602 struct hammer_cursor cursor;
2603 hammer_node_t tmp_node;
2604 hammer_record_t depend;
2605 hammer_record_t next;
2606 int error, tmp_error;
2609 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2612 error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2617 * Any directory records referencing this inode which are not in
2618 * our current flush group must adjust our nlink count for the
2619 * purposes of synchronization to disk.
2621 * Records which are in our flush group can be unlinked from our
2622 * inode now, potentially allowing the inode to be physically
2625 * This cannot block.
2627 nlinks = ip->ino_data.nlinks;
2628 next = TAILQ_FIRST(&ip->target_list);
2629 while ((depend = next) != NULL) {
2630 next = TAILQ_NEXT(depend, target_entry);
2631 if (depend->flush_state == HAMMER_FST_FLUSH &&
2632 depend->flush_group == ip->flush_group) {
2634 * If this is an ADD that was deleted by the frontend
2635 * the frontend nlinks count will have already been
2636 * decremented, but the backend is going to sync its
2637 * directory entry and must account for it. The
2638 * record will be converted to a delete-on-disk when
2641 * If the ADD was not deleted by the frontend we
2642 * can remove the dependancy from our target_list.
2644 if (depend->flags & HAMMER_RECF_DELETED_FE) {
2647 TAILQ_REMOVE(&ip->target_list, depend,
2649 depend->target_ip = NULL;
2651 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2653 * Not part of our flush group and not deleted by
2654 * the front-end, adjust the link count synced to
2655 * the media (undo what the frontend did when it
2656 * queued the record).
2658 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2659 switch(depend->type) {
2660 case HAMMER_MEM_RECORD_ADD:
2663 case HAMMER_MEM_RECORD_DEL:
2673 * Set dirty if we had to modify the link count.
2675 if (ip->sync_ino_data.nlinks != nlinks) {
2676 KKASSERT((int64_t)nlinks >= 0);
2677 ip->sync_ino_data.nlinks = nlinks;
2678 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2682 * If there is a trunction queued destroy any data past the (aligned)
2683 * truncation point. Userland will have dealt with the buffer
2684 * containing the truncation point for us.
2686 * We don't flush pending frontend data buffers until after we've
2687 * dealt with the truncation.
2689 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2691 * Interlock trunc_off. The VOP front-end may continue to
2692 * make adjustments to it while we are blocked.
2695 off_t aligned_trunc_off;
2698 trunc_off = ip->sync_trunc_off;
2699 blkmask = hammer_blocksize(trunc_off) - 1;
2700 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2703 * Delete any whole blocks on-media. The front-end has
2704 * already cleaned out any partial block and made it
2705 * pending. The front-end may have updated trunc_off
2706 * while we were blocked so we only use sync_trunc_off.
2708 * This operation can blow out the buffer cache, EWOULDBLOCK
2709 * means we were unable to complete the deletion. The
2710 * deletion will update sync_trunc_off in that case.
2712 error = hammer_ip_delete_range(&cursor, ip,
2714 0x7FFFFFFFFFFFFFFFLL, 2);
2715 if (error == EWOULDBLOCK) {
2716 ip->flags |= HAMMER_INODE_WOULDBLOCK;
2718 goto defer_buffer_flush;
2725 * Clear the truncation flag on the backend after we have
2726 * complete the deletions. Backend data is now good again
2727 * (including new records we are about to sync, below).
2729 * Leave sync_trunc_off intact. As we write additional
2730 * records the backend will update sync_trunc_off. This
2731 * tells the backend whether it can skip the overwrite
2732 * test. This should work properly even when the backend
2733 * writes full blocks where the truncation point straddles
2734 * the block because the comparison is against the base
2735 * offset of the record.
2737 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2738 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2744 * Now sync related records. These will typically be directory
2745 * entries, records tracking direct-writes, or delete-on-disk records.
2748 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2749 hammer_sync_record_callback, &cursor);
2755 hammer_cache_node(&ip->cache[1], cursor.node);
2758 * Re-seek for inode update, assuming our cache hasn't been ripped
2759 * out from under us.
2762 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2764 hammer_cursor_downgrade(&cursor);
2765 hammer_lock_sh(&tmp_node->lock);
2766 if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2767 hammer_cursor_seek(&cursor, tmp_node, 0);
2768 hammer_unlock(&tmp_node->lock);
2769 hammer_rel_node(tmp_node);
2775 * If we are deleting the inode the frontend had better not have
2776 * any active references on elements making up the inode.
2778 * The call to hammer_ip_delete_clean() cleans up auxillary records
2779 * but not DB or DATA records. Those must have already been deleted
2780 * by the normal truncation mechanic.
2782 if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2783 RB_EMPTY(&ip->rec_tree) &&
2784 (ip->sync_flags & HAMMER_INODE_DELETING) &&
2785 (ip->flags & HAMMER_INODE_DELETED) == 0) {
2788 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2790 ip->flags |= HAMMER_INODE_DELETED;
2791 ip->sync_flags &= ~HAMMER_INODE_DELETING;
2792 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2793 KKASSERT(RB_EMPTY(&ip->rec_tree));
2796 * Set delete_tid in both the frontend and backend
2797 * copy of the inode record. The DELETED flag handles
2798 * this, do not set RDIRTY.
2800 ip->ino_leaf.base.delete_tid = trans->tid;
2801 ip->sync_ino_leaf.base.delete_tid = trans->tid;
2802 ip->ino_leaf.delete_ts = trans->time32;
2803 ip->sync_ino_leaf.delete_ts = trans->time32;
2807 * Adjust the inode count in the volume header
2809 hammer_sync_lock_sh(trans);
2810 if (ip->flags & HAMMER_INODE_ONDISK) {
2811 hammer_modify_volume_field(trans,
2814 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2815 hammer_modify_volume_done(trans->rootvol);
2817 hammer_sync_unlock(trans);
2823 ip->sync_flags &= ~HAMMER_INODE_BUFS;
2827 * Now update the inode's on-disk inode-data and/or on-disk record.
2828 * DELETED and ONDISK are managed only in ip->flags.
2830 * In the case of a defered buffer flush we still update the on-disk
2831 * inode to satisfy visibility requirements if there happen to be
2832 * directory dependancies.
2834 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2835 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2837 * If deleted and on-disk, don't set any additional flags.
2838 * the delete flag takes care of things.
2840 * Clear flags which may have been set by the frontend.
2842 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2843 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2844 HAMMER_INODE_DELETING);
2846 case HAMMER_INODE_DELETED:
2848 * Take care of the case where a deleted inode was never
2849 * flushed to the disk in the first place.
2851 * Clear flags which may have been set by the frontend.
2853 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2854 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2855 HAMMER_INODE_DELETING);
2856 while (RB_ROOT(&ip->rec_tree)) {
2857 hammer_record_t record = RB_ROOT(&ip->rec_tree);
2858 hammer_ref(&record->lock);
2859 KKASSERT(record->lock.refs == 1);
2860 record->flags |= HAMMER_RECF_DELETED_BE;
2861 ++record->ip->rec_generation;
2862 hammer_rel_mem_record(record);
2865 case HAMMER_INODE_ONDISK:
2867 * If already on-disk, do not set any additional flags.
2872 * If not on-disk and not deleted, set DDIRTY to force
2873 * an initial record to be written.
2875 * Also set the create_tid in both the frontend and backend
2876 * copy of the inode record.
2878 ip->ino_leaf.base.create_tid = trans->tid;
2879 ip->ino_leaf.create_ts = trans->time32;
2880 ip->sync_ino_leaf.base.create_tid = trans->tid;
2881 ip->sync_ino_leaf.create_ts = trans->time32;
2882 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2887 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2888 * is already on-disk the old record is marked as deleted.
2890 * If DELETED is set hammer_update_inode() will delete the existing
2891 * record without writing out a new one.
2893 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2895 if (ip->flags & HAMMER_INODE_DELETED) {
2896 error = hammer_update_inode(&cursor, ip);
2898 if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2899 (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2900 error = hammer_update_itimes(&cursor, ip);
2902 if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2903 error = hammer_update_inode(&cursor, ip);
2907 hammer_critical_error(ip->hmp, ip, error,
2908 "while syncing inode");
2910 hammer_done_cursor(&cursor);
2915 * This routine is called when the OS is no longer actively referencing
2916 * the inode (but might still be keeping it cached), or when releasing
2917 * the last reference to an inode.
2919 * At this point if the inode's nlinks count is zero we want to destroy
2920 * it, which may mean destroying it on-media too.
2923 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2928 * Set the DELETING flag when the link count drops to 0 and the
2929 * OS no longer has any opens on the inode.
2931 * The backend will clear DELETING (a mod flag) and set DELETED
2932 * (a state flag) when it is actually able to perform the
2935 * Don't reflag the deletion if the flusher is currently syncing
2936 * one that was already flagged. A previously set DELETING flag
2937 * may bounce around flags and sync_flags until the operation is
2940 if (ip->ino_data.nlinks == 0 &&
2941 ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2942 ip->flags |= HAMMER_INODE_DELETING;
2943 ip->flags |= HAMMER_INODE_TRUNCATED;
2947 if (hammer_get_vnode(ip, &vp) != 0)
2955 vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2956 vnode_pager_setsize(ip->vp, 0);
2965 * After potentially resolving a dependancy the inode is tested
2966 * to determine whether it needs to be reflushed.
2969 hammer_test_inode(hammer_inode_t ip)
2971 if (ip->flags & HAMMER_INODE_REFLUSH) {
2972 ip->flags &= ~HAMMER_INODE_REFLUSH;
2973 hammer_ref(&ip->lock);
2974 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2975 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2976 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2978 hammer_flush_inode(ip, 0);
2980 hammer_rel_inode(ip, 0);
2985 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2986 * reassociated with a vp or just before it gets freed.
2988 * Pipeline wakeups to threads blocked due to an excessive number of
2989 * detached inodes. This typically occurs when atime updates accumulate
2990 * while scanning a directory tree.
2993 hammer_inode_wakereclaims(hammer_inode_t ip)
2995 struct hammer_reclaim *reclaim;
2996 hammer_mount_t hmp = ip->hmp;
2998 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3001 --hammer_count_reclaiming;
3002 --hmp->inode_reclaims;
3003 ip->flags &= ~HAMMER_INODE_RECLAIM;
3005 while ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3006 if (reclaim->count > 0 && --reclaim->count == 0) {
3007 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3010 if (hmp->inode_reclaims > hammer_limit_reclaim / 2)
3016 * Setup our reclaim pipeline. We only let so many detached (and dirty)
3017 * inodes build up before we start blocking. This routine is called
3018 * if a new inode is created or an inode is loaded from media.
3020 * When we block we don't care *which* inode has finished reclaiming,
3021 * as lone as one does.
3024 hammer_inode_waitreclaims(hammer_mount_t hmp)
3026 struct hammer_reclaim reclaim;
3028 if (hmp->inode_reclaims < hammer_limit_reclaim)
3031 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3032 tsleep(&reclaim, 0, "hmrrcm", hz);
3033 if (reclaim.count > 0)
3034 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3040 * XXX not used, doesn't work very well due to the large batching nature
3043 * A larger then normal backlog of inodes is sitting in the flusher,
3044 * enforce a general slowdown to let it catch up. This routine is only
3045 * called on completion of a non-flusher-related transaction which
3046 * performed B-Tree node I/O.
3048 * It is possible for the flusher to stall in a continuous load.
3049 * blogbench -i1000 -o seems to do a good job generating this sort of load.
3050 * If the flusher is unable to catch up the inode count can bloat until
3051 * we run out of kvm.
3053 * This is a bit of a hack.
3056 hammer_inode_waithard(hammer_mount_t hmp)
3061 if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3062 if (hmp->inode_reclaims < hammer_limit_reclaim / 2 &&
3063 hmp->count_iqueued < hmp->count_inodes / 20) {
3064 hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3068 if (hmp->inode_reclaims < hammer_limit_reclaim ||
3069 hmp->count_iqueued < hmp->count_inodes / 10) {
3072 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3076 * Block for one flush cycle.
3078 hammer_flusher_wait_next(hmp);
projects / dragonfly.git / blob