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
37 static int hammer_unload_inode(struct hammer_inode *ip);
38 static void hammer_free_inode(hammer_inode_t ip);
39 static void hammer_flush_inode_core(hammer_inode_t ip,
40 hammer_flush_group_t flg, int flags);
41 static int hammer_setup_child_callback(hammer_record_t rec, void *data);
43 static int hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
45 static int hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
46 hammer_flush_group_t flg);
47 static int hammer_setup_parent_inodes_helper(hammer_record_t record,
48 int depth, hammer_flush_group_t flg);
49 static void hammer_inode_wakereclaims(hammer_inode_t ip);
50 static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
54 extern struct hammer_inode *HammerTruncIp;
57 struct krate hammer_gen_krate = { 1 };
60 * RB-Tree support for inode structures
63 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
65 if (ip1->obj_localization < ip2->obj_localization)
67 if (ip1->obj_localization > ip2->obj_localization)
69 if (ip1->obj_id < ip2->obj_id)
71 if (ip1->obj_id > ip2->obj_id)
73 if (ip1->obj_asof < ip2->obj_asof)
75 if (ip1->obj_asof > ip2->obj_asof)
81 hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
83 if (ip1->redo_fifo_start < ip2->redo_fifo_start)
85 if (ip1->redo_fifo_start > ip2->redo_fifo_start)
91 * RB-Tree support for inode structures / special LOOKUP_INFO
94 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
96 if (info->obj_localization < ip->obj_localization)
98 if (info->obj_localization > ip->obj_localization)
100 if (info->obj_id < ip->obj_id)
102 if (info->obj_id > ip->obj_id)
104 if (info->obj_asof < ip->obj_asof)
106 if (info->obj_asof > ip->obj_asof)
112 * Used by hammer_scan_inode_snapshots() to locate all of an object's
113 * snapshots. Note that the asof field is not tested, which we can get
114 * away with because it is the lowest-priority field.
117 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
119 hammer_inode_info_t info = data;
121 if (ip->obj_localization > info->obj_localization)
123 if (ip->obj_localization < info->obj_localization)
125 if (ip->obj_id > info->obj_id)
127 if (ip->obj_id < info->obj_id)
133 * Used by hammer_unload_pseudofs() to locate all inodes associated with
137 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
139 u_int32_t localization = *(u_int32_t *)data;
140 if (ip->obj_localization > localization)
142 if (ip->obj_localization < localization)
148 * RB-Tree support for pseudofs structures
151 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
153 if (p1->localization < p2->localization)
155 if (p1->localization > p2->localization)
161 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
162 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
163 hammer_inode_info_cmp, hammer_inode_info_t);
164 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
165 hammer_pfs_rb_compare, u_int32_t, localization);
168 * The kernel is not actively referencing this vnode but is still holding
171 * This is called from the frontend.
176 hammer_vop_inactive(struct vop_inactive_args *ap)
178 struct hammer_inode *ip = VTOI(ap->a_vp);
190 * If the inode no longer has visibility in the filesystem try to
191 * recycle it immediately, even if the inode is dirty. Recycling
192 * it quickly allows the system to reclaim buffer cache and VM
193 * resources which can matter a lot in a heavily loaded system.
195 * This can deadlock in vfsync() if we aren't careful.
197 * Do not queue the inode to the flusher if we still have visibility,
198 * otherwise namespace calls such as chmod will unnecessarily generate
199 * multiple inode updates.
201 if (ip->ino_data.nlinks == 0) {
203 lwkt_gettoken(&hmp->fs_token);
204 hammer_inode_unloadable_check(ip, 0);
205 if (ip->flags & HAMMER_INODE_MODMASK)
206 hammer_flush_inode(ip, 0);
207 lwkt_reltoken(&hmp->fs_token);
214 * Release the vnode association. This is typically (but not always)
215 * the last reference on the inode.
217 * Once the association is lost we are on our own with regards to
218 * flushing the inode.
220 * We must interlock ip->vp so hammer_get_vnode() can avoid races.
223 hammer_vop_reclaim(struct vop_reclaim_args *ap)
225 struct hammer_inode *ip;
231 if ((ip = vp->v_data) != NULL) {
233 lwkt_gettoken(&hmp->fs_token);
234 hammer_lock_ex(&ip->lock);
238 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
239 ++hammer_count_reclaims;
240 ++hmp->count_reclaims;
241 ip->flags |= HAMMER_INODE_RECLAIM;
243 hammer_unlock(&ip->lock);
245 hammer_rel_inode(ip, 1);
246 lwkt_reltoken(&hmp->fs_token);
252 * Inform the kernel that the inode is dirty. This will be checked
255 * Theoretically in order to reclaim a vnode the hammer_vop_reclaim()
256 * must be called which will interlock against our inode lock, so
257 * if VRECLAIMED is not set vp->v_mount (as used by vsetisdirty())
258 * should be stable without having to acquire any new locks.
261 hammer_inode_dirty(struct hammer_inode *ip)
265 if ((ip->flags & HAMMER_INODE_MODMASK) &&
266 (vp = ip->vp) != NULL &&
267 (vp->v_flag & (VRECLAIMED | VISDIRTY)) == 0) {
273 * Return a locked vnode for the specified inode. The inode must be
274 * referenced but NOT LOCKED on entry and will remain referenced on
277 * Called from the frontend.
280 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
290 if ((vp = ip->vp) == NULL) {
291 error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
294 hammer_lock_ex(&ip->lock);
295 if (ip->vp != NULL) {
296 hammer_unlock(&ip->lock);
302 hammer_ref(&ip->lock);
306 obj_type = ip->ino_data.obj_type;
307 vp->v_type = hammer_get_vnode_type(obj_type);
309 hammer_inode_wakereclaims(ip);
311 switch(ip->ino_data.obj_type) {
312 case HAMMER_OBJTYPE_CDEV:
313 case HAMMER_OBJTYPE_BDEV:
314 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
315 addaliasu(vp, ip->ino_data.rmajor,
316 ip->ino_data.rminor);
318 case HAMMER_OBJTYPE_FIFO:
319 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
321 case HAMMER_OBJTYPE_REGFILE:
328 * Only mark as the root vnode if the ip is not
329 * historical, otherwise the VFS cache will get
330 * confused. The other half of the special handling
331 * is in hammer_vop_nlookupdotdot().
333 * Pseudo-filesystem roots can be accessed via
334 * non-root filesystem paths and setting VROOT may
335 * confuse the namecache. Set VPFSROOT instead.
337 if (ip->obj_id == HAMMER_OBJID_ROOT) {
338 if (ip->obj_asof == hmp->asof) {
339 if (ip->obj_localization == 0)
340 vsetflags(vp, VROOT);
342 vsetflags(vp, VPFSROOT);
344 vsetflags(vp, VPFSROOT);
348 vp->v_data = (void *)ip;
349 /* vnode locked by getnewvnode() */
350 /* make related vnode dirty if inode dirty? */
351 hammer_unlock(&ip->lock);
352 if (vp->v_type == VREG) {
353 vinitvmio(vp, ip->ino_data.size,
354 hammer_blocksize(ip->ino_data.size),
355 hammer_blockoff(ip->ino_data.size));
361 * Interlock vnode clearing. This does not prevent the
362 * vnode from going into a reclaimed state but it does
363 * prevent it from being destroyed or reused so the vget()
364 * will properly fail.
366 hammer_lock_ex(&ip->lock);
367 if ((vp = ip->vp) == NULL) {
368 hammer_unlock(&ip->lock);
372 hammer_unlock(&ip->lock);
375 * loop if the vget fails (aka races), or if the vp
376 * no longer matches ip->vp.
378 if (vget(vp, LK_EXCLUSIVE) == 0) {
392 * Locate all copies of the inode for obj_id compatible with the specified
393 * asof, reference, and issue the related call-back. This routine is used
394 * for direct-io invalidation and does not create any new inodes.
397 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
398 int (*callback)(hammer_inode_t ip, void *data),
401 hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
402 hammer_inode_info_cmp_all_history,
407 * Acquire a HAMMER inode. The returned inode is not locked. These functions
408 * do not attach or detach the related vnode (use hammer_get_vnode() for
411 * The flags argument is only applied for newly created inodes, and only
412 * certain flags are inherited.
414 * Called from the frontend.
416 struct hammer_inode *
417 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
418 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
419 int flags, int *errorp)
421 hammer_mount_t hmp = trans->hmp;
422 struct hammer_node_cache *cachep;
423 struct hammer_inode_info iinfo;
424 struct hammer_cursor cursor;
425 struct hammer_inode *ip;
429 * Determine if we already have an inode cached. If we do then
432 * If we find an inode with no vnode we have to mark the
433 * transaction such that hammer_inode_waitreclaims() is
434 * called later on to avoid building up an infinite number
435 * of inodes. Otherwise we can continue to * add new inodes
436 * faster then they can be disposed of, even with the tsleep
439 * If we find a dummy inode we return a failure so dounlink
440 * (which does another lookup) doesn't try to mess with the
441 * link count. hammer_vop_nresolve() uses hammer_get_dummy_inode()
442 * to ref dummy inodes.
444 iinfo.obj_id = obj_id;
445 iinfo.obj_asof = asof;
446 iinfo.obj_localization = localization;
448 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
450 if (ip->flags & HAMMER_INODE_DUMMY) {
454 hammer_ref(&ip->lock);
460 * Allocate a new inode structure and deal with races later.
462 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
463 ++hammer_count_inodes;
466 ip->obj_asof = iinfo.obj_asof;
467 ip->obj_localization = localization;
469 ip->flags = flags & HAMMER_INODE_RO;
470 ip->cache[0].ip = ip;
471 ip->cache[1].ip = ip;
472 ip->cache[2].ip = ip;
473 ip->cache[3].ip = ip;
475 ip->flags |= HAMMER_INODE_RO;
476 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
477 0x7FFFFFFFFFFFFFFFLL;
478 RB_INIT(&ip->rec_tree);
479 TAILQ_INIT(&ip->target_list);
480 hammer_ref(&ip->lock);
483 * Locate the on-disk inode. If this is a PFS root we always
484 * access the current version of the root inode and (if it is not
485 * a master) always access information under it with a snapshot
488 * We cache recent inode lookups in this directory in dip->cache[2].
489 * If we can't find it we assume the inode we are looking for is
490 * close to the directory inode.
495 if (dip->cache[2].node)
496 cachep = &dip->cache[2];
498 cachep = &dip->cache[0];
500 hammer_init_cursor(trans, &cursor, cachep, NULL);
501 cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
502 cursor.key_beg.obj_id = ip->obj_id;
503 cursor.key_beg.key = 0;
504 cursor.key_beg.create_tid = 0;
505 cursor.key_beg.delete_tid = 0;
506 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
507 cursor.key_beg.obj_type = 0;
509 cursor.asof = iinfo.obj_asof;
510 cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
513 *errorp = hammer_btree_lookup(&cursor);
514 if (*errorp == EDEADLK) {
515 hammer_done_cursor(&cursor);
520 * On success the B-Tree lookup will hold the appropriate
521 * buffer cache buffers and provide a pointer to the requested
522 * information. Copy the information to the in-memory inode
523 * and cache the B-Tree node to improve future operations.
526 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
527 ip->ino_data = cursor.data->inode;
530 * cache[0] tries to cache the location of the object inode.
531 * The assumption is that it is near the directory inode.
533 * cache[1] tries to cache the location of the object data.
534 * We might have something in the governing directory from
535 * scan optimizations (see the strategy code in
538 * We update dip->cache[2], if possible, with the location
539 * of the object inode for future directory shortcuts.
541 hammer_cache_node(&ip->cache[0], cursor.node);
543 if (dip->cache[3].node) {
544 hammer_cache_node(&ip->cache[1],
547 hammer_cache_node(&dip->cache[2], cursor.node);
551 * The file should not contain any data past the file size
552 * stored in the inode. Setting save_trunc_off to the
553 * file size instead of max reduces B-Tree lookup overheads
554 * on append by allowing the flusher to avoid checking for
557 ip->save_trunc_off = ip->ino_data.size;
560 * Locate and assign the pseudofs management structure to
563 if (dip && dip->obj_localization == ip->obj_localization) {
564 ip->pfsm = dip->pfsm;
565 hammer_ref(&ip->pfsm->lock);
567 ip->pfsm = hammer_load_pseudofs(trans,
568 ip->obj_localization,
570 *errorp = 0; /* ignore ENOENT */
575 * The inode is placed on the red-black tree and will be synced to
576 * the media when flushed or by the filesystem sync. If this races
577 * another instantiation/lookup the insertion will fail.
580 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
581 hammer_free_inode(ip);
582 hammer_done_cursor(&cursor);
585 ip->flags |= HAMMER_INODE_ONDISK;
587 if (ip->flags & HAMMER_INODE_RSV_INODES) {
588 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
592 hammer_free_inode(ip);
595 hammer_done_cursor(&cursor);
598 * NEWINODE is only set if the inode becomes dirty later,
599 * setting it here just leads to unnecessary stalls.
601 * trans->flags |= HAMMER_TRANSF_NEWINODE;
607 * Get a dummy inode to placemark a broken directory entry.
609 struct hammer_inode *
610 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
611 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
612 int flags, int *errorp)
614 hammer_mount_t hmp = trans->hmp;
615 struct hammer_inode_info iinfo;
616 struct hammer_inode *ip;
619 * Determine if we already have an inode cached. If we do then
622 * If we find an inode with no vnode we have to mark the
623 * transaction such that hammer_inode_waitreclaims() is
624 * called later on to avoid building up an infinite number
625 * of inodes. Otherwise we can continue to * add new inodes
626 * faster then they can be disposed of, even with the tsleep
629 * If we find a non-fake inode we return an error. Only fake
630 * inodes can be returned by this routine.
632 iinfo.obj_id = obj_id;
633 iinfo.obj_asof = asof;
634 iinfo.obj_localization = localization;
637 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
639 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
643 hammer_ref(&ip->lock);
648 * Allocate a new inode structure and deal with races later.
650 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
651 ++hammer_count_inodes;
654 ip->obj_asof = iinfo.obj_asof;
655 ip->obj_localization = localization;
657 ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
658 ip->cache[0].ip = ip;
659 ip->cache[1].ip = ip;
660 ip->cache[2].ip = ip;
661 ip->cache[3].ip = ip;
662 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
663 0x7FFFFFFFFFFFFFFFLL;
664 RB_INIT(&ip->rec_tree);
665 TAILQ_INIT(&ip->target_list);
666 hammer_ref(&ip->lock);
669 * Populate the dummy inode. Leave everything zero'd out.
671 * (ip->ino_leaf and ip->ino_data)
673 * Make the dummy inode a FIFO object which most copy programs
674 * will properly ignore.
676 ip->save_trunc_off = ip->ino_data.size;
677 ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
680 * Locate and assign the pseudofs management structure to
683 if (dip && dip->obj_localization == ip->obj_localization) {
684 ip->pfsm = dip->pfsm;
685 hammer_ref(&ip->pfsm->lock);
687 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
689 *errorp = 0; /* ignore ENOENT */
693 * The inode is placed on the red-black tree and will be synced to
694 * the media when flushed or by the filesystem sync. If this races
695 * another instantiation/lookup the insertion will fail.
697 * NOTE: Do not set HAMMER_INODE_ONDISK. The inode is a fake.
700 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
701 hammer_free_inode(ip);
705 if (ip->flags & HAMMER_INODE_RSV_INODES) {
706 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
709 hammer_free_inode(ip);
712 trans->flags |= HAMMER_TRANSF_NEWINODE;
717 * Return a referenced inode only if it is in our inode cache.
719 * Dummy inodes do not count.
721 struct hammer_inode *
722 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
723 hammer_tid_t asof, u_int32_t localization)
725 hammer_mount_t hmp = trans->hmp;
726 struct hammer_inode_info iinfo;
727 struct hammer_inode *ip;
729 iinfo.obj_id = obj_id;
730 iinfo.obj_asof = asof;
731 iinfo.obj_localization = localization;
733 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
735 if (ip->flags & HAMMER_INODE_DUMMY)
738 hammer_ref(&ip->lock);
744 * Create a new filesystem object, returning the inode in *ipp. The
745 * returned inode will be referenced. The inode is created in-memory.
747 * If pfsm is non-NULL the caller wishes to create the root inode for
751 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
753 hammer_inode_t dip, const char *name, int namelen,
754 hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
766 * Disallow the creation of new inodes in directories which
767 * have been deleted. In HAMMER, this will cause a record
768 * syncing assertion later on in the flush code.
770 if (dip && dip->ino_data.nlinks == 0) {
778 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
779 ++hammer_count_inodes;
781 trans->flags |= HAMMER_TRANSF_NEWINODE;
784 KKASSERT(pfsm->localization != 0);
785 ip->obj_id = HAMMER_OBJID_ROOT;
786 ip->obj_localization = pfsm->localization;
788 KKASSERT(dip != NULL);
789 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
790 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
791 ip->obj_localization = dip->obj_localization;
794 KKASSERT(ip->obj_id != 0);
795 ip->obj_asof = hmp->asof;
797 ip->flush_state = HAMMER_FST_IDLE;
798 ip->flags = HAMMER_INODE_DDIRTY |
799 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
800 ip->cache[0].ip = ip;
801 ip->cache[1].ip = ip;
802 ip->cache[2].ip = ip;
803 ip->cache[3].ip = ip;
805 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
806 /* ip->save_trunc_off = 0; (already zero) */
807 RB_INIT(&ip->rec_tree);
808 TAILQ_INIT(&ip->target_list);
810 ip->ino_data.atime = trans->time;
811 ip->ino_data.mtime = trans->time;
812 ip->ino_data.size = 0;
813 ip->ino_data.nlinks = 0;
816 * A nohistory designator on the parent directory is inherited by
817 * the child. We will do this even for pseudo-fs creation... the
818 * sysad can turn it off.
821 ip->ino_data.uflags = dip->ino_data.uflags &
822 (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
825 ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
826 ip->ino_leaf.base.localization = ip->obj_localization +
827 HAMMER_LOCALIZE_INODE;
828 ip->ino_leaf.base.obj_id = ip->obj_id;
829 ip->ino_leaf.base.key = 0;
830 ip->ino_leaf.base.create_tid = 0;
831 ip->ino_leaf.base.delete_tid = 0;
832 ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
833 ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
835 ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
836 ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
837 ip->ino_data.mode = vap->va_mode;
838 ip->ino_data.ctime = trans->time;
841 * If we are running version 2 or greater directory entries are
842 * inode-localized instead of data-localized.
844 if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
845 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
846 ip->ino_data.cap_flags |=
847 HAMMER_INODE_CAP_DIR_LOCAL_INO;
850 if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
851 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
852 ip->ino_data.cap_flags |=
853 HAMMER_INODE_CAP_DIRHASH_ALG1;
858 * Setup the ".." pointer. This only needs to be done for directories
859 * but we do it for all objects as a recovery aid if dip exists.
860 * The inode is probably a PFS root if dip is NULL.
863 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
866 * The parent_obj_localization field only applies to pseudo-fs roots.
867 * XXX this is no longer applicable, PFSs are no longer directly
868 * tied into the parent's directory structure.
870 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
871 ip->obj_id == HAMMER_OBJID_ROOT) {
872 ip->ino_data.ext.obj.parent_obj_localization =
873 dip->obj_localization;
877 switch(ip->ino_leaf.base.obj_type) {
878 case HAMMER_OBJTYPE_CDEV:
879 case HAMMER_OBJTYPE_BDEV:
880 ip->ino_data.rmajor = vap->va_rmajor;
881 ip->ino_data.rminor = vap->va_rminor;
888 * Calculate default uid/gid and overwrite with information from
892 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
893 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
894 xuid, cred, &vap->va_mode);
898 ip->ino_data.mode = vap->va_mode;
900 if (vap->va_vaflags & VA_UID_UUID_VALID)
901 ip->ino_data.uid = vap->va_uid_uuid;
902 else if (vap->va_uid != (uid_t)VNOVAL)
903 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
905 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
907 if (vap->va_vaflags & VA_GID_UUID_VALID)
908 ip->ino_data.gid = vap->va_gid_uuid;
909 else if (vap->va_gid != (gid_t)VNOVAL)
910 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
912 ip->ino_data.gid = dip->ino_data.gid;
914 hammer_ref(&ip->lock);
918 hammer_ref(&pfsm->lock);
920 } else if (dip->obj_localization == ip->obj_localization) {
921 ip->pfsm = dip->pfsm;
922 hammer_ref(&ip->pfsm->lock);
925 ip->pfsm = hammer_load_pseudofs(trans,
926 ip->obj_localization,
928 error = 0; /* ignore ENOENT */
932 hammer_free_inode(ip);
934 } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
935 panic("hammer_create_inode: duplicate obj_id %llx",
936 (long long)ip->obj_id);
938 hammer_free_inode(ip);
945 * Final cleanup / freeing of an inode structure
948 hammer_free_inode(hammer_inode_t ip)
950 struct hammer_mount *hmp;
953 KKASSERT(hammer_oneref(&ip->lock));
954 hammer_uncache_node(&ip->cache[0]);
955 hammer_uncache_node(&ip->cache[1]);
956 hammer_uncache_node(&ip->cache[2]);
957 hammer_uncache_node(&ip->cache[3]);
958 hammer_inode_wakereclaims(ip);
960 hammer_clear_objid(ip);
961 --hammer_count_inodes;
964 hammer_rel_pseudofs(hmp, ip->pfsm);
967 kfree(ip, hmp->m_inodes);
972 * Retrieve pseudo-fs data. NULL will never be returned.
974 * If an error occurs *errorp will be set and a default template is returned,
975 * otherwise *errorp is set to 0. Typically when an error occurs it will
978 hammer_pseudofs_inmem_t
979 hammer_load_pseudofs(hammer_transaction_t trans,
980 u_int32_t localization, int *errorp)
982 hammer_mount_t hmp = trans->hmp;
984 hammer_pseudofs_inmem_t pfsm;
985 struct hammer_cursor cursor;
989 pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
991 hammer_ref(&pfsm->lock);
997 * PFS records are associated with the root inode (not the PFS root
998 * inode, but the real root). Avoid an infinite recursion if loading
999 * the PFS for the real root.
1002 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
1004 HAMMER_DEF_LOCALIZATION, 0, errorp);
1009 pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
1010 pfsm->localization = localization;
1011 pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
1012 pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
1014 hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
1015 cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
1016 HAMMER_LOCALIZE_MISC;
1017 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1018 cursor.key_beg.create_tid = 0;
1019 cursor.key_beg.delete_tid = 0;
1020 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1021 cursor.key_beg.obj_type = 0;
1022 cursor.key_beg.key = localization;
1023 cursor.asof = HAMMER_MAX_TID;
1024 cursor.flags |= HAMMER_CURSOR_ASOF;
1027 *errorp = hammer_ip_lookup(&cursor);
1029 *errorp = hammer_btree_lookup(&cursor);
1031 *errorp = hammer_ip_resolve_data(&cursor);
1033 if (cursor.data->pfsd.mirror_flags &
1034 HAMMER_PFSD_DELETED) {
1037 bytes = cursor.leaf->data_len;
1038 if (bytes > sizeof(pfsm->pfsd))
1039 bytes = sizeof(pfsm->pfsd);
1040 bcopy(cursor.data, &pfsm->pfsd, bytes);
1044 hammer_done_cursor(&cursor);
1046 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1047 hammer_ref(&pfsm->lock);
1049 hammer_rel_inode(ip, 0);
1050 if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
1051 kfree(pfsm, hmp->m_misc);
1058 * Store pseudo-fs data. The backend will automatically delete any prior
1059 * on-disk pseudo-fs data but we have to delete in-memory versions.
1062 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
1064 struct hammer_cursor cursor;
1065 hammer_record_t record;
1070 * PFS records are associated with the root inode (not the PFS root
1071 * inode, but the real root).
1073 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1074 HAMMER_DEF_LOCALIZATION, 0, &error);
1076 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1077 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1078 cursor.key_beg.localization = ip->obj_localization +
1079 HAMMER_LOCALIZE_MISC;
1080 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1081 cursor.key_beg.create_tid = 0;
1082 cursor.key_beg.delete_tid = 0;
1083 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1084 cursor.key_beg.obj_type = 0;
1085 cursor.key_beg.key = pfsm->localization;
1086 cursor.asof = HAMMER_MAX_TID;
1087 cursor.flags |= HAMMER_CURSOR_ASOF;
1090 * Replace any in-memory version of the record.
1092 error = hammer_ip_lookup(&cursor);
1093 if (error == 0 && hammer_cursor_inmem(&cursor)) {
1094 record = cursor.iprec;
1095 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1096 KKASSERT(cursor.deadlk_rec == NULL);
1097 hammer_ref(&record->lock);
1098 cursor.deadlk_rec = record;
1101 record->flags |= HAMMER_RECF_DELETED_FE;
1107 * Allocate replacement general record. The backend flush will
1108 * delete any on-disk version of the record.
1110 if (error == 0 || error == ENOENT) {
1111 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1112 record->type = HAMMER_MEM_RECORD_GENERAL;
1114 record->leaf.base.localization = ip->obj_localization +
1115 HAMMER_LOCALIZE_MISC;
1116 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1117 record->leaf.base.key = pfsm->localization;
1118 record->leaf.data_len = sizeof(pfsm->pfsd);
1119 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1120 error = hammer_ip_add_record(trans, record);
1122 hammer_done_cursor(&cursor);
1123 if (error == EDEADLK)
1125 hammer_rel_inode(ip, 0);
1130 * Create a root directory for a PFS if one does not alredy exist.
1132 * The PFS root stands alone so we must also bump the nlinks count
1133 * to prevent it from being destroyed on release.
1136 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1137 hammer_pseudofs_inmem_t pfsm)
1143 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1144 pfsm->localization, 0, &error);
1149 error = hammer_create_inode(trans, &vap, cred,
1153 ++ip->ino_data.nlinks;
1154 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1158 hammer_rel_inode(ip, 0);
1163 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1164 * if we are unable to disassociate all the inodes.
1168 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1172 hammer_ref(&ip->lock);
1173 if (ip->vp && (ip->vp->v_flag & VPFSROOT)) {
1175 * The hammer pfs-upgrade directive itself might have the
1176 * root of the pfs open. Just allow it.
1181 * Don't allow any subdirectories or files to be open.
1183 if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1184 vclean_unlocked(ip->vp);
1185 if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1188 res = -1; /* stop, someone is using the inode */
1190 hammer_rel_inode(ip, 0);
1195 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1200 for (try = res = 0; try < 4; ++try) {
1201 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1202 hammer_inode_pfs_cmp,
1203 hammer_unload_pseudofs_callback,
1205 if (res == 0 && try > 1)
1207 hammer_flusher_sync(trans->hmp);
1216 * Release a reference on a PFS
1219 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1221 hammer_rel(&pfsm->lock);
1222 if (hammer_norefs(&pfsm->lock)) {
1223 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1224 kfree(pfsm, hmp->m_misc);
1229 * Called by hammer_sync_inode().
1232 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1234 hammer_transaction_t trans = cursor->trans;
1235 hammer_record_t record;
1243 * If the inode has a presence on-disk then locate it and mark
1244 * it deleted, setting DELONDISK.
1246 * The record may or may not be physically deleted, depending on
1247 * the retention policy.
1249 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1250 HAMMER_INODE_ONDISK) {
1251 hammer_normalize_cursor(cursor);
1252 cursor->key_beg.localization = ip->obj_localization +
1253 HAMMER_LOCALIZE_INODE;
1254 cursor->key_beg.obj_id = ip->obj_id;
1255 cursor->key_beg.key = 0;
1256 cursor->key_beg.create_tid = 0;
1257 cursor->key_beg.delete_tid = 0;
1258 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1259 cursor->key_beg.obj_type = 0;
1260 cursor->asof = ip->obj_asof;
1261 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1262 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1263 cursor->flags |= HAMMER_CURSOR_BACKEND;
1265 error = hammer_btree_lookup(cursor);
1266 if (hammer_debug_inode)
1267 kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1270 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1271 if (hammer_debug_inode)
1272 kprintf(" error %d\n", error);
1274 ip->flags |= HAMMER_INODE_DELONDISK;
1277 hammer_cache_node(&ip->cache[0], cursor->node);
1279 if (error == EDEADLK) {
1280 hammer_done_cursor(cursor);
1281 error = hammer_init_cursor(trans, cursor,
1283 if (hammer_debug_inode)
1284 kprintf("IPDED %p %d\n", ip, error);
1291 * Ok, write out the initial record or a new record (after deleting
1292 * the old one), unless the DELETED flag is set. This routine will
1293 * clear DELONDISK if it writes out a record.
1295 * Update our inode statistics if this is the first application of
1296 * the inode on-disk.
1298 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1300 * Generate a record and write it to the media. We clean-up
1301 * the state before releasing so we do not have to set-up
1304 record = hammer_alloc_mem_record(ip, 0);
1305 record->type = HAMMER_MEM_RECORD_INODE;
1306 record->flush_state = HAMMER_FST_FLUSH;
1307 record->leaf = ip->sync_ino_leaf;
1308 record->leaf.base.create_tid = trans->tid;
1309 record->leaf.data_len = sizeof(ip->sync_ino_data);
1310 record->leaf.create_ts = trans->time32;
1311 record->data = (void *)&ip->sync_ino_data;
1312 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1315 * If this flag is set we cannot sync the new file size
1316 * because we haven't finished related truncations. The
1317 * inode will be flushed in another flush group to finish
1320 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1321 ip->sync_ino_data.size != ip->ino_data.size) {
1323 ip->sync_ino_data.size = ip->ino_data.size;
1329 error = hammer_ip_sync_record_cursor(cursor, record);
1330 if (hammer_debug_inode)
1331 kprintf("GENREC %p rec %08x %d\n",
1332 ip, record->flags, error);
1333 if (error != EDEADLK)
1335 hammer_done_cursor(cursor);
1336 error = hammer_init_cursor(trans, cursor,
1338 if (hammer_debug_inode)
1339 kprintf("GENREC reinit %d\n", error);
1345 * Note: The record was never on the inode's record tree
1346 * so just wave our hands importantly and destroy it.
1348 record->flags |= HAMMER_RECF_COMMITTED;
1349 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1350 record->flush_state = HAMMER_FST_IDLE;
1351 ++ip->rec_generation;
1352 hammer_rel_mem_record(record);
1358 if (hammer_debug_inode)
1359 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1360 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1361 HAMMER_INODE_SDIRTY |
1362 HAMMER_INODE_ATIME |
1363 HAMMER_INODE_MTIME);
1364 ip->flags &= ~HAMMER_INODE_DELONDISK;
1366 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1369 * Root volume count of inodes
1371 hammer_sync_lock_sh(trans);
1372 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1373 hammer_modify_volume_field(trans,
1376 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1377 hammer_modify_volume_done(trans->rootvol);
1378 ip->flags |= HAMMER_INODE_ONDISK;
1379 if (hammer_debug_inode)
1380 kprintf("NOWONDISK %p\n", ip);
1382 hammer_sync_unlock(trans);
1387 * If the inode has been destroyed, clean out any left-over flags
1388 * that may have been set by the frontend.
1390 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1391 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1392 HAMMER_INODE_SDIRTY |
1393 HAMMER_INODE_ATIME |
1394 HAMMER_INODE_MTIME);
1400 * Update only the itimes fields.
1402 * ATIME can be updated without generating any UNDO. MTIME is updated
1403 * with UNDO so it is guaranteed to be synchronized properly in case of
1406 * Neither field is included in the B-Tree leaf element's CRC, which is how
1407 * we can get away with updating ATIME the way we do.
1410 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1412 hammer_transaction_t trans = cursor->trans;
1416 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1417 HAMMER_INODE_ONDISK) {
1421 hammer_normalize_cursor(cursor);
1422 cursor->key_beg.localization = ip->obj_localization +
1423 HAMMER_LOCALIZE_INODE;
1424 cursor->key_beg.obj_id = ip->obj_id;
1425 cursor->key_beg.key = 0;
1426 cursor->key_beg.create_tid = 0;
1427 cursor->key_beg.delete_tid = 0;
1428 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1429 cursor->key_beg.obj_type = 0;
1430 cursor->asof = ip->obj_asof;
1431 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1432 cursor->flags |= HAMMER_CURSOR_ASOF;
1433 cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1434 cursor->flags |= HAMMER_CURSOR_GET_DATA;
1435 cursor->flags |= HAMMER_CURSOR_BACKEND;
1437 error = hammer_btree_lookup(cursor);
1439 hammer_cache_node(&ip->cache[0], cursor->node);
1440 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1442 * Updating MTIME requires an UNDO. Just cover
1443 * both atime and mtime.
1445 hammer_sync_lock_sh(trans);
1446 hammer_modify_buffer(trans, cursor->data_buffer,
1447 HAMMER_ITIMES_BASE(&cursor->data->inode),
1448 HAMMER_ITIMES_BYTES);
1449 cursor->data->inode.atime = ip->sync_ino_data.atime;
1450 cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1451 hammer_modify_buffer_done(cursor->data_buffer);
1452 hammer_sync_unlock(trans);
1453 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1455 * Updating atime only can be done in-place with
1458 hammer_sync_lock_sh(trans);
1459 hammer_modify_buffer(trans, cursor->data_buffer,
1461 cursor->data->inode.atime = ip->sync_ino_data.atime;
1462 hammer_modify_buffer_done(cursor->data_buffer);
1463 hammer_sync_unlock(trans);
1465 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1467 if (error == EDEADLK) {
1468 hammer_done_cursor(cursor);
1469 error = hammer_init_cursor(trans, cursor, &ip->cache[0], ip);
1477 * Release a reference on an inode, flush as requested.
1479 * On the last reference we queue the inode to the flusher for its final
1483 hammer_rel_inode(struct hammer_inode *ip, int flush)
1486 * Handle disposition when dropping the last ref.
1489 if (hammer_oneref(&ip->lock)) {
1491 * Determine whether on-disk action is needed for
1492 * the inode's final disposition.
1494 KKASSERT(ip->vp == NULL);
1495 hammer_inode_unloadable_check(ip, 0);
1496 if (ip->flags & HAMMER_INODE_MODMASK) {
1497 hammer_flush_inode(ip, 0);
1498 } else if (hammer_oneref(&ip->lock)) {
1499 hammer_unload_inode(ip);
1504 hammer_flush_inode(ip, 0);
1507 * The inode still has multiple refs, try to drop
1510 KKASSERT(hammer_isactive(&ip->lock) >= 1);
1511 if (hammer_isactive(&ip->lock) > 1) {
1512 hammer_rel(&ip->lock);
1520 * Unload and destroy the specified inode. Must be called with one remaining
1521 * reference. The reference is disposed of.
1523 * The inode must be completely clean.
1526 hammer_unload_inode(struct hammer_inode *ip)
1528 hammer_mount_t hmp = ip->hmp;
1530 KASSERT(hammer_oneref(&ip->lock),
1531 ("hammer_unload_inode: %d refs", hammer_isactive(&ip->lock)));
1532 KKASSERT(ip->vp == NULL);
1533 KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1534 KKASSERT(ip->cursor_ip_refs == 0);
1535 KKASSERT(hammer_notlocked(&ip->lock));
1536 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1538 KKASSERT(RB_EMPTY(&ip->rec_tree));
1539 KKASSERT(TAILQ_EMPTY(&ip->target_list));
1541 if (ip->flags & HAMMER_INODE_RDIRTY) {
1542 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1543 ip->flags &= ~HAMMER_INODE_RDIRTY;
1545 RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1547 hammer_free_inode(ip);
1552 * Called during unmounting if a critical error occured. The in-memory
1553 * inode and all related structures are destroyed.
1555 * If a critical error did not occur the unmount code calls the standard
1556 * release and asserts that the inode is gone.
1559 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1561 hammer_record_t rec;
1564 * Get rid of the inodes in-memory records, regardless of their
1565 * state, and clear the mod-mask.
1567 while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1568 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1569 rec->target_ip = NULL;
1570 if (rec->flush_state == HAMMER_FST_SETUP)
1571 rec->flush_state = HAMMER_FST_IDLE;
1573 while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1574 if (rec->flush_state == HAMMER_FST_FLUSH)
1575 --rec->flush_group->refs;
1577 hammer_ref(&rec->lock);
1578 KKASSERT(hammer_oneref(&rec->lock));
1579 rec->flush_state = HAMMER_FST_IDLE;
1580 rec->flush_group = NULL;
1581 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1582 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1583 ++ip->rec_generation;
1584 hammer_rel_mem_record(rec);
1586 ip->flags &= ~HAMMER_INODE_MODMASK;
1587 ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1588 KKASSERT(ip->vp == NULL);
1591 * Remove the inode from any flush group, force it idle. FLUSH
1592 * and SETUP states have an inode ref.
1594 switch(ip->flush_state) {
1595 case HAMMER_FST_FLUSH:
1596 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1597 --ip->flush_group->refs;
1598 ip->flush_group = NULL;
1600 case HAMMER_FST_SETUP:
1601 hammer_rel(&ip->lock);
1602 ip->flush_state = HAMMER_FST_IDLE;
1604 case HAMMER_FST_IDLE:
1609 * There shouldn't be any associated vnode. The unload needs at
1610 * least one ref, if we do have a vp steal its ip ref.
1613 kprintf("hammer_destroy_inode_callback: Unexpected "
1614 "vnode association ip %p vp %p\n", ip, ip->vp);
1615 ip->vp->v_data = NULL;
1618 hammer_ref(&ip->lock);
1620 hammer_unload_inode(ip);
1625 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1626 * the read-only flag for cached inodes.
1628 * This routine is called from a RB_SCAN().
1631 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1633 hammer_mount_t hmp = ip->hmp;
1635 if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1636 ip->flags |= HAMMER_INODE_RO;
1638 ip->flags &= ~HAMMER_INODE_RO;
1643 * A transaction has modified an inode, requiring updates as specified by
1646 * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1647 * and not including size changes due to write-append
1648 * (but other size changes are included).
1649 * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1651 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1652 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1653 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1654 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1657 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1660 * ronly of 0 or 2 does not trigger assertion.
1661 * 2 is a special error state
1663 KKASSERT(ip->hmp->ronly != 1 ||
1664 (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1665 HAMMER_INODE_SDIRTY |
1666 HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1667 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1668 if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1669 ip->flags |= HAMMER_INODE_RSV_INODES;
1670 ++ip->hmp->rsv_inodes;
1674 * Set the NEWINODE flag in the transaction if the inode
1675 * transitions to a dirty state. This is used to track
1676 * the load on the inode cache.
1679 (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1680 (flags & HAMMER_INODE_MODMASK)) {
1681 trans->flags |= HAMMER_TRANSF_NEWINODE;
1683 if (flags & HAMMER_INODE_MODMASK)
1684 hammer_inode_dirty(ip);
1689 * Attempt to quickly update the atime for a hammer inode. Return 0 on
1690 * success, -1 on failure.
1692 * We attempt to update the atime with only the ip lock and not the
1693 * whole filesystem lock in order to improve concurrency. We can only
1694 * do this safely if the ATIME flag is already pending on the inode.
1696 * This function is called via a vnops path (ip pointer is stable) without
1700 hammer_update_atime_quick(hammer_inode_t ip)
1705 if ((ip->flags & HAMMER_INODE_RO) ||
1706 (ip->hmp->mp->mnt_flag & MNT_NOATIME)) {
1708 * Silently indicate success on read-only mount/snap
1711 } else if (ip->flags & HAMMER_INODE_ATIME) {
1713 * Double check with inode lock held against backend. This
1714 * is only safe if all we need to do is update
1718 hammer_lock_ex(&ip->lock);
1719 if (ip->flags & HAMMER_INODE_ATIME) {
1720 ip->ino_data.atime =
1721 (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec;
1724 hammer_unlock(&ip->lock);
1730 * Request that an inode be flushed. This whole mess cannot block and may
1731 * recurse (if not synchronous). Once requested HAMMER will attempt to
1732 * actively flush the inode until the flush can be done.
1734 * The inode may already be flushing, or may be in a setup state. We can
1735 * place the inode in a flushing state if it is currently idle and flag it
1736 * to reflush if it is currently flushing.
1738 * Upon return if the inode could not be flushed due to a setup
1739 * dependancy, then it will be automatically flushed when the dependancy
1743 hammer_flush_inode(hammer_inode_t ip, int flags)
1746 hammer_flush_group_t flg;
1750 * fill_flush_group is the first flush group we may be able to
1751 * continue filling, it may be open or closed but it will always
1752 * be past the currently flushing (running) flg.
1754 * next_flush_group is the next open flush group.
1757 while ((flg = hmp->fill_flush_group) != NULL) {
1758 KKASSERT(flg->running == 0);
1759 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1760 flg->total_count <= hammer_autoflush) {
1763 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1764 hammer_flusher_async(ip->hmp, flg);
1767 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1768 flg->seq = hmp->flusher.next++;
1769 if (hmp->next_flush_group == NULL)
1770 hmp->next_flush_group = flg;
1771 if (hmp->fill_flush_group == NULL)
1772 hmp->fill_flush_group = flg;
1773 RB_INIT(&flg->flush_tree);
1774 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1778 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1779 * state we have to put it back into an IDLE state so we can
1780 * drop the extra ref.
1782 * If we have a parent dependancy we must still fall through
1785 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1786 if (ip->flush_state == HAMMER_FST_SETUP &&
1787 TAILQ_EMPTY(&ip->target_list)) {
1788 ip->flush_state = HAMMER_FST_IDLE;
1789 hammer_rel_inode(ip, 0);
1791 if (ip->flush_state == HAMMER_FST_IDLE)
1796 * Our flush action will depend on the current state.
1798 switch(ip->flush_state) {
1799 case HAMMER_FST_IDLE:
1801 * We have no dependancies and can flush immediately. Some
1802 * our children may not be flushable so we have to re-test
1803 * with that additional knowledge.
1805 hammer_flush_inode_core(ip, flg, flags);
1807 case HAMMER_FST_SETUP:
1809 * Recurse upwards through dependancies via target_list
1810 * and start their flusher actions going if possible.
1812 * 'good' is our connectivity. -1 means we have none and
1813 * can't flush, 0 means there weren't any dependancies, and
1814 * 1 means we have good connectivity.
1816 good = hammer_setup_parent_inodes(ip, 0, flg);
1820 * We can continue if good >= 0. Determine how
1821 * many records under our inode can be flushed (and
1824 hammer_flush_inode_core(ip, flg, flags);
1827 * Parent has no connectivity, tell it to flush
1828 * us as soon as it does.
1830 * The REFLUSH flag is also needed to trigger
1831 * dependancy wakeups.
1833 ip->flags |= HAMMER_INODE_CONN_DOWN |
1834 HAMMER_INODE_REFLUSH;
1835 if (flags & HAMMER_FLUSH_SIGNAL) {
1836 ip->flags |= HAMMER_INODE_RESIGNAL;
1837 hammer_flusher_async(ip->hmp, flg);
1841 case HAMMER_FST_FLUSH:
1843 * We are already flushing, flag the inode to reflush
1844 * if needed after it completes its current flush.
1846 * The REFLUSH flag is also needed to trigger
1847 * dependancy wakeups.
1849 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1850 ip->flags |= HAMMER_INODE_REFLUSH;
1851 if (flags & HAMMER_FLUSH_SIGNAL) {
1852 ip->flags |= HAMMER_INODE_RESIGNAL;
1853 hammer_flusher_async(ip->hmp, flg);
1860 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1861 * ip which reference our ip.
1863 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1864 * so for now do not ref/deref the structures. Note that if we use the
1865 * ref/rel code later, the rel CAN block.
1868 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1869 hammer_flush_group_t flg)
1871 hammer_record_t depend;
1876 * If we hit our recursion limit and we have parent dependencies
1877 * We cannot continue. Returning < 0 will cause us to be flagged
1878 * for reflush. Returning -2 cuts off additional dependency checks
1879 * because they are likely to also hit the depth limit.
1881 * We cannot return < 0 if there are no dependencies or there might
1882 * not be anything to wakeup (ip).
1884 if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1885 if (hammer_debug_general & 0x10000)
1886 krateprintf(&hammer_gen_krate,
1887 "HAMMER Warning: depth limit reached on "
1888 "setup recursion, inode %p %016llx\n",
1889 ip, (long long)ip->obj_id);
1897 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1898 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1899 KKASSERT(depend->target_ip == ip);
1900 if (r < 0 && good == 0)
1906 * If we failed due to the recursion depth limit then stop
1916 * This helper function takes a record representing the dependancy between
1917 * the parent inode and child inode.
1919 * record = record in question (*rec in below)
1920 * record->ip = parent inode (*pip in below)
1921 * record->target_ip = child inode (*ip in below)
1923 * *pip--------------\
1926 * \ip /\\\\\ rbtree of recs from parent inode's view
1930 * \------*rec------target_ip------>*ip
1931 * ...target_entry<----...----->target_list<---...
1932 * list of recs from inode's view
1934 * We are asked to recurse upwards and convert the record from SETUP
1935 * to FLUSH if possible.
1937 * Return 1 if the record gives us connectivity
1939 * Return 0 if the record is not relevant
1941 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1944 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1945 hammer_flush_group_t flg)
1950 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1954 * If the record is already flushing, is it in our flush group?
1956 * If it is in our flush group but it is a general record or a
1957 * delete-on-disk, it does not improve our connectivity (return 0),
1958 * and if the target inode is not trying to destroy itself we can't
1959 * allow the operation yet anyway (the second return -1).
1961 if (record->flush_state == HAMMER_FST_FLUSH) {
1963 * If not in our flush group ask the parent to reflush
1964 * us as soon as possible.
1966 if (record->flush_group != flg) {
1967 pip->flags |= HAMMER_INODE_REFLUSH;
1968 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1973 * If in our flush group everything is already set up,
1974 * just return whether the record will improve our
1975 * visibility or not.
1977 if (record->type == HAMMER_MEM_RECORD_ADD)
1983 * It must be a setup record. Try to resolve the setup dependancies
1984 * by recursing upwards so we can place ip on the flush list.
1986 * Limit ourselves to 20 levels of recursion to avoid blowing out
1987 * the kernel stack. If we hit the recursion limit we can't flush
1988 * until the parent flushes. The parent will flush independantly
1989 * on its own and ultimately a deep recursion will be resolved.
1991 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1993 good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1996 * If good < 0 the parent has no connectivity and we cannot safely
1997 * flush the directory entry, which also means we can't flush our
1998 * ip. Flag us for downward recursion once the parent's
1999 * connectivity is resolved. Flag the parent for [re]flush or it
2000 * may not check for downward recursions.
2003 pip->flags |= HAMMER_INODE_REFLUSH;
2004 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2009 * We are go, place the parent inode in a flushing state so we can
2010 * place its record in a flushing state. Note that the parent
2011 * may already be flushing. The record must be in the same flush
2012 * group as the parent.
2014 if (pip->flush_state != HAMMER_FST_FLUSH)
2015 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
2016 KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
2019 * It is possible for a rename to create a loop in the recursion
2020 * and revisit a record. This will result in the record being
2021 * placed in a flush state unexpectedly. This check deals with
2024 if (record->flush_state == HAMMER_FST_FLUSH) {
2025 if (record->type == HAMMER_MEM_RECORD_ADD)
2030 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
2033 if (record->type == HAMMER_MEM_RECORD_DEL &&
2034 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
2036 * Regardless of flushing state we cannot sync this path if the
2037 * record represents a delete-on-disk but the target inode
2038 * is not ready to sync its own deletion.
2040 * XXX need to count effective nlinks to determine whether
2041 * the flush is ok, otherwise removing a hardlink will
2042 * just leave the DEL record to rot.
2044 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
2048 if (pip->flush_group == flg) {
2050 * Because we have not calculated nlinks yet we can just
2051 * set records to the flush state if the parent is in
2052 * the same flush group as we are.
2054 record->flush_state = HAMMER_FST_FLUSH;
2055 record->flush_group = flg;
2056 ++record->flush_group->refs;
2057 hammer_ref(&record->lock);
2060 * A general directory-add contributes to our visibility.
2062 * Otherwise it is probably a directory-delete or
2063 * delete-on-disk record and does not contribute to our
2064 * visbility (but we can still flush it).
2066 if (record->type == HAMMER_MEM_RECORD_ADD)
2071 * If the parent is not in our flush group we cannot
2072 * flush this record yet, there is no visibility.
2073 * We tell the parent to reflush and mark ourselves
2074 * so the parent knows it should flush us too.
2076 pip->flags |= HAMMER_INODE_REFLUSH;
2077 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2083 * This is the core routine placing an inode into the FST_FLUSH state.
2086 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
2088 hammer_mount_t hmp = ip->hmp;
2092 * Set flush state and prevent the flusher from cycling into
2093 * the next flush group. Do not place the ip on the list yet.
2094 * Inodes not in the idle state get an extra reference.
2096 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
2097 if (ip->flush_state == HAMMER_FST_IDLE)
2098 hammer_ref(&ip->lock);
2099 ip->flush_state = HAMMER_FST_FLUSH;
2100 ip->flush_group = flg;
2101 ++hmp->flusher.group_lock;
2102 ++hmp->count_iqueued;
2103 ++hammer_count_iqueued;
2105 hammer_redo_fifo_start_flush(ip);
2109 * We need to be able to vfsync/truncate from the backend.
2111 * XXX Any truncation from the backend will acquire the vnode
2114 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2115 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2116 ip->flags |= HAMMER_INODE_VHELD;
2122 * Figure out how many in-memory records we can actually flush
2123 * (not including inode meta-data, buffers, etc).
2125 KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2126 if (flags & HAMMER_FLUSH_RECURSION) {
2128 * If this is a upwards recursion we do not want to
2129 * recurse down again!
2133 } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2135 * No new records are added if we must complete a flush
2136 * from a previous cycle, but we do have to move the records
2137 * from the previous cycle to the current one.
2140 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2141 hammer_syncgrp_child_callback, NULL);
2147 * Normal flush, scan records and bring them into the flush.
2148 * Directory adds and deletes are usually skipped (they are
2149 * grouped with the related inode rather then with the
2152 * go_count can be negative, which means the scan aborted
2153 * due to the flush group being over-full and we should
2154 * flush what we have.
2156 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2157 hammer_setup_child_callback, NULL);
2161 * This is a more involved test that includes go_count. If we
2162 * can't flush, flag the inode and return. If go_count is 0 we
2163 * were are unable to flush any records in our rec_tree and
2164 * must ignore the XDIRTY flag.
2166 if (go_count == 0) {
2167 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2168 --hmp->count_iqueued;
2169 --hammer_count_iqueued;
2172 ip->flush_state = HAMMER_FST_SETUP;
2173 ip->flush_group = NULL;
2174 if (flags & HAMMER_FLUSH_SIGNAL) {
2175 ip->flags |= HAMMER_INODE_REFLUSH |
2176 HAMMER_INODE_RESIGNAL;
2178 ip->flags |= HAMMER_INODE_REFLUSH;
2181 if (ip->flags & HAMMER_INODE_VHELD) {
2182 ip->flags &= ~HAMMER_INODE_VHELD;
2188 * REFLUSH is needed to trigger dependancy wakeups
2189 * when an inode is in SETUP.
2191 ip->flags |= HAMMER_INODE_REFLUSH;
2192 if (--hmp->flusher.group_lock == 0)
2193 wakeup(&hmp->flusher.group_lock);
2199 * Snapshot the state of the inode for the backend flusher.
2201 * We continue to retain save_trunc_off even when all truncations
2202 * have been resolved as an optimization to determine if we can
2203 * skip the B-Tree lookup for overwrite deletions.
2205 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2206 * and stays in ip->flags. Once set, it stays set until the
2207 * inode is destroyed.
2209 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2210 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2211 ip->sync_trunc_off = ip->trunc_off;
2212 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2213 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2214 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2217 * The save_trunc_off used to cache whether the B-Tree
2218 * holds any records past that point is not used until
2219 * after the truncation has succeeded, so we can safely
2222 if (ip->save_trunc_off > ip->sync_trunc_off)
2223 ip->save_trunc_off = ip->sync_trunc_off;
2225 ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2226 ~HAMMER_INODE_TRUNCATED);
2227 ip->sync_ino_leaf = ip->ino_leaf;
2228 ip->sync_ino_data = ip->ino_data;
2229 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2230 #ifdef DEBUG_TRUNCATE
2231 if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2232 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2236 * The flusher list inherits our inode and reference.
2238 KKASSERT(flg->running == 0);
2239 RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2240 if (--hmp->flusher.group_lock == 0)
2241 wakeup(&hmp->flusher.group_lock);
2244 * Auto-flush the group if it grows too large. Make sure the
2245 * inode reclaim wait pipeline continues to work.
2247 if (flg->total_count >= hammer_autoflush ||
2248 flg->total_count >= hammer_limit_reclaims / 4) {
2249 if (hmp->fill_flush_group == flg)
2250 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2251 hammer_flusher_async(hmp, flg);
2256 * Callback for scan of ip->rec_tree. Try to include each record in our
2257 * flush. ip->flush_group has been set but the inode has not yet been
2258 * moved into a flushing state.
2260 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2263 * We return 1 for any record placed or found in FST_FLUSH, which prevents
2264 * the caller from shortcutting the flush.
2267 hammer_setup_child_callback(hammer_record_t rec, void *data)
2269 hammer_flush_group_t flg;
2270 hammer_inode_t target_ip;
2275 * Records deleted or committed by the backend are ignored.
2276 * Note that the flush detects deleted frontend records at
2277 * multiple points to deal with races. This is just the first
2278 * line of defense. The only time HAMMER_RECF_DELETED_FE cannot
2279 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2280 * messes up link-count calculations.
2282 * NOTE: Don't get confused between record deletion and, say,
2283 * directory entry deletion. The deletion of a directory entry
2284 * which is on-media has nothing to do with the record deletion
2287 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2288 HAMMER_RECF_COMMITTED)) {
2289 if (rec->flush_state == HAMMER_FST_FLUSH) {
2290 KKASSERT(rec->flush_group == rec->ip->flush_group);
2299 * If the record is in an idle state it has no dependancies and
2303 flg = ip->flush_group;
2306 switch(rec->flush_state) {
2307 case HAMMER_FST_IDLE:
2309 * The record has no setup dependancy, we can flush it.
2311 KKASSERT(rec->target_ip == NULL);
2312 rec->flush_state = HAMMER_FST_FLUSH;
2313 rec->flush_group = flg;
2315 hammer_ref(&rec->lock);
2318 case HAMMER_FST_SETUP:
2320 * The record has a setup dependancy. These are typically
2321 * directory entry adds and deletes. Such entries will be
2322 * flushed when their inodes are flushed so we do not
2323 * usually have to add them to the flush here. However,
2324 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2325 * it is asking us to flush this record (and it).
2327 target_ip = rec->target_ip;
2328 KKASSERT(target_ip != NULL);
2329 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2332 * If the target IP is already flushing in our group
2333 * we could associate the record, but target_ip has
2334 * already synced ino_data to sync_ino_data and we
2335 * would also have to adjust nlinks. Plus there are
2336 * ordering issues for adds and deletes.
2338 * Reflush downward if this is an ADD, and upward if
2341 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2342 if (rec->type == HAMMER_MEM_RECORD_ADD)
2343 ip->flags |= HAMMER_INODE_REFLUSH;
2345 target_ip->flags |= HAMMER_INODE_REFLUSH;
2350 * Target IP is not yet flushing. This can get complex
2351 * because we have to be careful about the recursion.
2353 * Directories create an issue for us in that if a flush
2354 * of a directory is requested the expectation is to flush
2355 * any pending directory entries, but this will cause the
2356 * related inodes to recursively flush as well. We can't
2357 * really defer the operation so just get as many as we
2361 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2362 (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2364 * We aren't reclaiming and the target ip was not
2365 * previously prevented from flushing due to this
2366 * record dependancy. Do not flush this record.
2371 if (flg->total_count + flg->refs >
2372 ip->hmp->undo_rec_limit) {
2374 * Our flush group is over-full and we risk blowing
2375 * out the UNDO FIFO. Stop the scan, flush what we
2376 * have, then reflush the directory.
2378 * The directory may be forced through multiple
2379 * flush groups before it can be completely
2382 ip->flags |= HAMMER_INODE_RESIGNAL |
2383 HAMMER_INODE_REFLUSH;
2385 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2387 * If the target IP is not flushing we can force
2388 * it to flush, even if it is unable to write out
2389 * any of its own records we have at least one in
2390 * hand that we CAN deal with.
2392 rec->flush_state = HAMMER_FST_FLUSH;
2393 rec->flush_group = flg;
2395 hammer_ref(&rec->lock);
2396 hammer_flush_inode_core(target_ip, flg,
2397 HAMMER_FLUSH_RECURSION);
2401 * General or delete-on-disk record.
2403 * XXX this needs help. If a delete-on-disk we could
2404 * disconnect the target. If the target has its own
2405 * dependancies they really need to be flushed.
2409 rec->flush_state = HAMMER_FST_FLUSH;
2410 rec->flush_group = flg;
2412 hammer_ref(&rec->lock);
2413 hammer_flush_inode_core(target_ip, flg,
2414 HAMMER_FLUSH_RECURSION);
2418 case HAMMER_FST_FLUSH:
2420 * The record could be part of a previous flush group if the
2421 * inode is a directory (the record being a directory entry).
2422 * Once the flush group was closed a hammer_test_inode()
2423 * function can cause a new flush group to be setup, placing
2424 * the directory inode itself in a new flush group.
2426 * When associated with a previous flush group we count it
2427 * as if it were in our current flush group, since it will
2428 * effectively be flushed by the time we flush our current
2432 rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2433 rec->flush_group == flg);
2442 * This version just moves records already in a flush state to the new
2443 * flush group and that is it.
2446 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2448 hammer_inode_t ip = rec->ip;
2450 switch(rec->flush_state) {
2451 case HAMMER_FST_FLUSH:
2452 KKASSERT(rec->flush_group == ip->flush_group);
2462 * Wait for a previously queued flush to complete.
2464 * If a critical error occured we don't try to wait.
2467 hammer_wait_inode(hammer_inode_t ip)
2470 * The inode can be in a SETUP state in which case RESIGNAL
2471 * should be set. If RESIGNAL is not set then the previous
2472 * flush completed and a later operation placed the inode
2473 * in a passive setup state again, so we're done.
2475 * The inode can be in a FLUSH state in which case we
2476 * can just wait for completion.
2478 while (ip->flush_state == HAMMER_FST_FLUSH ||
2479 (ip->flush_state == HAMMER_FST_SETUP &&
2480 (ip->flags & HAMMER_INODE_RESIGNAL))) {
2482 * Don't try to flush on a critical error
2484 if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
2488 * If the inode was already being flushed its flg
2489 * may not have been queued to the backend. We
2490 * have to make sure it gets queued or we can wind
2491 * up blocked or deadlocked (particularly if we are
2492 * the vnlru thread).
2494 if (ip->flush_state == HAMMER_FST_FLUSH) {
2495 KKASSERT(ip->flush_group);
2496 if (ip->flush_group->closed == 0) {
2497 if (hammer_debug_inode) {
2498 kprintf("HAMMER: debug: forcing "
2499 "async flush ip %016jx\n",
2500 (intmax_t)ip->obj_id);
2502 hammer_flusher_async(ip->hmp, ip->flush_group);
2503 continue; /* retest */
2508 * In a flush state with the flg queued to the backend
2509 * or in a setup state with RESIGNAL set, we can safely
2512 ip->flags |= HAMMER_INODE_FLUSHW;
2513 tsleep(&ip->flags, 0, "hmrwin", 0);
2518 * The inode may have been in a passive setup state,
2519 * call flush to make sure we get signaled.
2521 if (ip->flush_state == HAMMER_FST_SETUP)
2522 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2528 * Called by the backend code when a flush has been completed.
2529 * The inode has already been removed from the flush list.
2531 * A pipelined flush can occur, in which case we must re-enter the
2532 * inode on the list and re-copy its fields.
2535 hammer_flush_inode_done(hammer_inode_t ip, int error)
2540 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2545 * Auto-reflush if the backend could not completely flush
2546 * the inode. This fixes a case where a deferred buffer flush
2547 * could cause fsync to return early.
2549 if (ip->sync_flags & HAMMER_INODE_MODMASK)
2550 ip->flags |= HAMMER_INODE_REFLUSH;
2553 * Merge left-over flags back into the frontend and fix the state.
2554 * Incomplete truncations are retained by the backend.
2557 ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2558 ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2561 * The backend may have adjusted nlinks, so if the adjusted nlinks
2562 * does not match the fronttend set the frontend's DDIRTY flag again.
2564 if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2565 ip->flags |= HAMMER_INODE_DDIRTY;
2568 * Fix up the dirty buffer status.
2570 if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2571 ip->flags |= HAMMER_INODE_BUFS;
2573 hammer_redo_fifo_end_flush(ip);
2576 * Re-set the XDIRTY flag if some of the inode's in-memory records
2577 * could not be flushed.
2579 KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2580 (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2581 (!RB_EMPTY(&ip->rec_tree) &&
2582 (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2585 * Do not lose track of inodes which no longer have vnode
2586 * assocations, otherwise they may never get flushed again.
2588 * The reflush flag can be set superfluously, causing extra pain
2589 * for no reason. If the inode is no longer modified it no longer
2590 * needs to be flushed.
2592 if (ip->flags & HAMMER_INODE_MODMASK) {
2594 ip->flags |= HAMMER_INODE_REFLUSH;
2596 ip->flags &= ~HAMMER_INODE_REFLUSH;
2600 * The fs token is held but the inode lock is not held. Because this
2601 * is a backend flush it is possible that the vnode has no references
2602 * and cause a reclaim race inside vsetisdirty() if/when it blocks.
2604 * Therefore, we must lock the inode around this particular dirtying
2605 * operation. We don't have to around other dirtying operations
2606 * where the vnode is implicitly or explicitly held.
2608 if (ip->flags & HAMMER_INODE_MODMASK) {
2609 hammer_lock_ex(&ip->lock);
2610 hammer_inode_dirty(ip);
2611 hammer_unlock(&ip->lock);
2615 * Adjust the flush state.
2617 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2619 * We were unable to flush out all our records, leave the
2620 * inode in a flush state and in the current flush group.
2621 * The flush group will be re-run.
2623 * This occurs if the UNDO block gets too full or there is
2624 * too much dirty meta-data and allows the flusher to
2625 * finalize the UNDO block and then re-flush.
2627 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2631 * Remove from the flush_group
2633 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2634 ip->flush_group = NULL;
2638 * Clean up the vnode ref and tracking counts.
2640 if (ip->flags & HAMMER_INODE_VHELD) {
2641 ip->flags &= ~HAMMER_INODE_VHELD;
2645 --hmp->count_iqueued;
2646 --hammer_count_iqueued;
2649 * And adjust the state.
2651 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2652 ip->flush_state = HAMMER_FST_IDLE;
2655 ip->flush_state = HAMMER_FST_SETUP;
2660 * If the frontend is waiting for a flush to complete,
2663 if (ip->flags & HAMMER_INODE_FLUSHW) {
2664 ip->flags &= ~HAMMER_INODE_FLUSHW;
2669 * If the frontend made more changes and requested another
2670 * flush, then try to get it running.
2672 * Reflushes are aborted when the inode is errored out.
2674 if (ip->flags & HAMMER_INODE_REFLUSH) {
2675 ip->flags &= ~HAMMER_INODE_REFLUSH;
2676 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2677 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2678 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2680 hammer_flush_inode(ip, 0);
2686 * If we have no parent dependancies we can clear CONN_DOWN
2688 if (TAILQ_EMPTY(&ip->target_list))
2689 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2692 * If the inode is now clean drop the space reservation.
2694 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2695 (ip->flags & HAMMER_INODE_RSV_INODES)) {
2696 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2700 ip->flags &= ~HAMMER_INODE_SLAVEFLUSH;
2703 hammer_rel_inode(ip, 0);
2707 * Called from hammer_sync_inode() to synchronize in-memory records
2711 hammer_sync_record_callback(hammer_record_t record, void *data)
2713 hammer_cursor_t cursor = data;
2714 hammer_transaction_t trans = cursor->trans;
2715 hammer_mount_t hmp = trans->hmp;
2719 * Skip records that do not belong to the current flush.
2721 ++hammer_stats_record_iterations;
2722 if (record->flush_state != HAMMER_FST_FLUSH)
2726 if (record->flush_group != record->ip->flush_group) {
2727 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2728 if (hammer_debug_critical)
2733 KKASSERT(record->flush_group == record->ip->flush_group);
2736 * Interlock the record using the BE flag. Once BE is set the
2737 * frontend cannot change the state of FE.
2739 * NOTE: If FE is set prior to us setting BE we still sync the
2740 * record out, but the flush completion code converts it to
2741 * a delete-on-disk record instead of destroying it.
2743 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2744 record->flags |= HAMMER_RECF_INTERLOCK_BE;
2747 * The backend has already disposed of the record.
2749 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2755 * If the whole inode is being deleted and all on-disk records will
2756 * be deleted very soon, we can't sync any new records to disk
2757 * because they will be deleted in the same transaction they were
2758 * created in (delete_tid == create_tid), which will assert.
2760 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2761 * that we currently panic on.
2763 if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2764 switch(record->type) {
2765 case HAMMER_MEM_RECORD_DATA:
2767 * We don't have to do anything, if the record was
2768 * committed the space will have been accounted for
2772 case HAMMER_MEM_RECORD_GENERAL:
2774 * Set deleted-by-backend flag. Do not set the
2775 * backend committed flag, because we are throwing
2778 record->flags |= HAMMER_RECF_DELETED_BE;
2779 ++record->ip->rec_generation;
2782 case HAMMER_MEM_RECORD_ADD:
2783 panic("hammer_sync_record_callback: illegal add "
2784 "during inode deletion record %p", record);
2785 break; /* NOT REACHED */
2786 case HAMMER_MEM_RECORD_INODE:
2787 panic("hammer_sync_record_callback: attempt to "
2788 "sync inode record %p?", record);
2789 break; /* NOT REACHED */
2790 case HAMMER_MEM_RECORD_DEL:
2792 * Follow through and issue the on-disk deletion
2799 * If DELETED_FE is set special handling is needed for directory
2800 * entries. Dependant pieces related to the directory entry may
2801 * have already been synced to disk. If this occurs we have to
2802 * sync the directory entry and then change the in-memory record
2803 * from an ADD to a DELETE to cover the fact that it's been
2804 * deleted by the frontend.
2806 * A directory delete covering record (MEM_RECORD_DEL) can never
2807 * be deleted by the frontend.
2809 * Any other record type (aka DATA) can be deleted by the frontend.
2810 * XXX At the moment the flusher must skip it because there may
2811 * be another data record in the flush group for the same block,
2812 * meaning that some frontend data changes can leak into the backend's
2813 * synchronization point.
2815 if (record->flags & HAMMER_RECF_DELETED_FE) {
2816 if (record->type == HAMMER_MEM_RECORD_ADD) {
2818 * Convert a front-end deleted directory-add to
2819 * a directory-delete entry later.
2821 record->flags |= HAMMER_RECF_CONVERT_DELETE;
2824 * Dispose of the record (race case). Mark as
2825 * deleted by backend (and not committed).
2827 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2828 record->flags |= HAMMER_RECF_DELETED_BE;
2829 ++record->ip->rec_generation;
2836 * Assign the create_tid for new records. Deletions already
2837 * have the record's entire key properly set up.
2839 if (record->type != HAMMER_MEM_RECORD_DEL) {
2840 record->leaf.base.create_tid = trans->tid;
2841 record->leaf.create_ts = trans->time32;
2845 * This actually moves the record to the on-media B-Tree. We
2846 * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2847 * indicating that the related REDO_WRITE(s) have been committed.
2849 * During recovery any REDO_TERM's within the nominal recovery span
2850 * are ignored since the related meta-data is being undone, causing
2851 * any matching REDO_WRITEs to execute. The REDO_TERMs outside
2852 * the nominal recovery span will match against REDO_WRITEs and
2853 * prevent them from being executed (because the meta-data has
2854 * already been synchronized).
2856 if (record->flags & HAMMER_RECF_REDO) {
2857 KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2858 hammer_generate_redo(trans, record->ip,
2859 record->leaf.base.key -
2860 record->leaf.data_len,
2861 HAMMER_REDO_TERM_WRITE,
2863 record->leaf.data_len);
2867 error = hammer_ip_sync_record_cursor(cursor, record);
2868 if (error != EDEADLK)
2870 hammer_done_cursor(cursor);
2871 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2876 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2881 hammer_flush_record_done(record, error);
2884 * Do partial finalization if we have built up too many dirty
2885 * buffers. Otherwise a buffer cache deadlock can occur when
2886 * doing things like creating tens of thousands of tiny files.
2888 * We must release our cursor lock to avoid a 3-way deadlock
2889 * due to the exclusive sync lock the finalizer must get.
2891 * WARNING: See warnings in hammer_unlock_cursor() function.
2893 if (hammer_flusher_meta_limit(hmp) ||
2894 vm_page_count_severe()) {
2895 hammer_unlock_cursor(cursor);
2896 hammer_flusher_finalize(trans, 0);
2897 hammer_lock_cursor(cursor);
2903 * Backend function called by the flusher to sync an inode to media.
2906 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2908 struct hammer_cursor cursor;
2909 hammer_node_t tmp_node;
2910 hammer_record_t depend;
2911 hammer_record_t next;
2912 int error, tmp_error;
2915 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2918 error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2923 * Any directory records referencing this inode which are not in
2924 * our current flush group must adjust our nlink count for the
2925 * purposes of synchronizating to disk.
2927 * Records which are in our flush group can be unlinked from our
2928 * inode now, potentially allowing the inode to be physically
2931 * This cannot block.
2933 nlinks = ip->ino_data.nlinks;
2934 next = TAILQ_FIRST(&ip->target_list);
2935 while ((depend = next) != NULL) {
2936 next = TAILQ_NEXT(depend, target_entry);
2937 if (depend->flush_state == HAMMER_FST_FLUSH &&
2938 depend->flush_group == ip->flush_group) {
2940 * If this is an ADD that was deleted by the frontend
2941 * the frontend nlinks count will have already been
2942 * decremented, but the backend is going to sync its
2943 * directory entry and must account for it. The
2944 * record will be converted to a delete-on-disk when
2947 * If the ADD was not deleted by the frontend we
2948 * can remove the dependancy from our target_list.
2950 if (depend->flags & HAMMER_RECF_DELETED_FE) {
2953 TAILQ_REMOVE(&ip->target_list, depend,
2955 depend->target_ip = NULL;
2957 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2959 * Not part of our flush group and not deleted by
2960 * the front-end, adjust the link count synced to
2961 * the media (undo what the frontend did when it
2962 * queued the record).
2964 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2965 switch(depend->type) {
2966 case HAMMER_MEM_RECORD_ADD:
2969 case HAMMER_MEM_RECORD_DEL:
2979 * Set dirty if we had to modify the link count.
2981 if (ip->sync_ino_data.nlinks != nlinks) {
2982 KKASSERT((int64_t)nlinks >= 0);
2983 ip->sync_ino_data.nlinks = nlinks;
2984 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2988 * If there is a trunction queued destroy any data past the (aligned)
2989 * truncation point. Userland will have dealt with the buffer
2990 * containing the truncation point for us.
2992 * We don't flush pending frontend data buffers until after we've
2993 * dealt with the truncation.
2995 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2997 * Interlock trunc_off. The VOP front-end may continue to
2998 * make adjustments to it while we are blocked.
3001 off_t aligned_trunc_off;
3004 trunc_off = ip->sync_trunc_off;
3005 blkmask = hammer_blocksize(trunc_off) - 1;
3006 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
3009 * Delete any whole blocks on-media. The front-end has
3010 * already cleaned out any partial block and made it
3011 * pending. The front-end may have updated trunc_off
3012 * while we were blocked so we only use sync_trunc_off.
3014 * This operation can blow out the buffer cache, EWOULDBLOCK
3015 * means we were unable to complete the deletion. The
3016 * deletion will update sync_trunc_off in that case.
3018 error = hammer_ip_delete_range(&cursor, ip,
3020 0x7FFFFFFFFFFFFFFFLL, 2);
3021 if (error == EWOULDBLOCK) {
3022 ip->flags |= HAMMER_INODE_WOULDBLOCK;
3024 goto defer_buffer_flush;
3031 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
3033 * XXX we do this even if we did not previously generate
3034 * a REDO_TRUNC record. This operation may enclosed the
3035 * range for multiple prior truncation entries in the REDO
3038 if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
3039 (ip->flags & HAMMER_INODE_RDIRTY)) {
3040 hammer_generate_redo(trans, ip, aligned_trunc_off,
3041 HAMMER_REDO_TERM_TRUNC,
3046 * Clear the truncation flag on the backend after we have
3047 * completed the deletions. Backend data is now good again
3048 * (including new records we are about to sync, below).
3050 * Leave sync_trunc_off intact. As we write additional
3051 * records the backend will update sync_trunc_off. This
3052 * tells the backend whether it can skip the overwrite
3053 * test. This should work properly even when the backend
3054 * writes full blocks where the truncation point straddles
3055 * the block because the comparison is against the base
3056 * offset of the record.
3058 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3059 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
3065 * Now sync related records. These will typically be directory
3066 * entries, records tracking direct-writes, or delete-on-disk records.
3069 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
3070 hammer_sync_record_callback, &cursor);
3076 hammer_cache_node(&ip->cache[1], cursor.node);
3079 * Re-seek for inode update, assuming our cache hasn't been ripped
3080 * out from under us.
3083 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
3085 hammer_cursor_downgrade(&cursor);
3086 hammer_lock_sh(&tmp_node->lock);
3087 if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
3088 hammer_cursor_seek(&cursor, tmp_node, 0);
3089 hammer_unlock(&tmp_node->lock);
3090 hammer_rel_node(tmp_node);
3096 * If we are deleting the inode the frontend had better not have
3097 * any active references on elements making up the inode.
3099 * The call to hammer_ip_delete_clean() cleans up auxillary records
3100 * but not DB or DATA records. Those must have already been deleted
3101 * by the normal truncation mechanic.
3103 if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
3104 RB_EMPTY(&ip->rec_tree) &&
3105 (ip->sync_flags & HAMMER_INODE_DELETING) &&
3106 (ip->flags & HAMMER_INODE_DELETED) == 0) {
3109 error = hammer_ip_delete_clean(&cursor, ip, &count1);
3111 ip->flags |= HAMMER_INODE_DELETED;
3112 ip->sync_flags &= ~HAMMER_INODE_DELETING;
3113 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3114 KKASSERT(RB_EMPTY(&ip->rec_tree));
3117 * Set delete_tid in both the frontend and backend
3118 * copy of the inode record. The DELETED flag handles
3119 * this, do not set DDIRTY.
3121 ip->ino_leaf.base.delete_tid = trans->tid;
3122 ip->sync_ino_leaf.base.delete_tid = trans->tid;
3123 ip->ino_leaf.delete_ts = trans->time32;
3124 ip->sync_ino_leaf.delete_ts = trans->time32;
3128 * Adjust the inode count in the volume header
3130 hammer_sync_lock_sh(trans);
3131 if (ip->flags & HAMMER_INODE_ONDISK) {
3132 hammer_modify_volume_field(trans,
3135 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
3136 hammer_modify_volume_done(trans->rootvol);
3138 hammer_sync_unlock(trans);
3144 ip->sync_flags &= ~HAMMER_INODE_BUFS;
3148 * Now update the inode's on-disk inode-data and/or on-disk record.
3149 * DELETED and ONDISK are managed only in ip->flags.
3151 * In the case of a defered buffer flush we still update the on-disk
3152 * inode to satisfy visibility requirements if there happen to be
3153 * directory dependancies.
3155 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
3156 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3158 * If deleted and on-disk, don't set any additional flags.
3159 * the delete flag takes care of things.
3161 * Clear flags which may have been set by the frontend.
3163 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3164 HAMMER_INODE_SDIRTY |
3165 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3166 HAMMER_INODE_DELETING);
3168 case HAMMER_INODE_DELETED:
3170 * Take care of the case where a deleted inode was never
3171 * flushed to the disk in the first place.
3173 * Clear flags which may have been set by the frontend.
3175 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3176 HAMMER_INODE_SDIRTY |
3177 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3178 HAMMER_INODE_DELETING);
3179 while (RB_ROOT(&ip->rec_tree)) {
3180 hammer_record_t record = RB_ROOT(&ip->rec_tree);
3181 hammer_ref(&record->lock);
3182 KKASSERT(hammer_oneref(&record->lock));
3183 record->flags |= HAMMER_RECF_DELETED_BE;
3184 ++record->ip->rec_generation;
3185 hammer_rel_mem_record(record);
3188 case HAMMER_INODE_ONDISK:
3190 * If already on-disk, do not set any additional flags.
3195 * If not on-disk and not deleted, set DDIRTY to force
3196 * an initial record to be written.
3198 * Also set the create_tid in both the frontend and backend
3199 * copy of the inode record.
3201 ip->ino_leaf.base.create_tid = trans->tid;
3202 ip->ino_leaf.create_ts = trans->time32;
3203 ip->sync_ino_leaf.base.create_tid = trans->tid;
3204 ip->sync_ino_leaf.create_ts = trans->time32;
3205 ip->sync_flags |= HAMMER_INODE_DDIRTY;
3210 * If DDIRTY or SDIRTY is set, write out a new record.
3211 * If the inode is already on-disk the old record is marked as
3214 * If DELETED is set hammer_update_inode() will delete the existing
3215 * record without writing out a new one.
3217 * If *ONLY* the ITIMES flag is set we can update the record in-place.
3219 if (ip->flags & HAMMER_INODE_DELETED) {
3220 error = hammer_update_inode(&cursor, ip);
3222 if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3223 (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3224 error = hammer_update_itimes(&cursor, ip);
3226 if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3227 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3228 error = hammer_update_inode(&cursor, ip);
3231 if (ip->flags & HAMMER_INODE_MODMASK)
3232 hammer_inode_dirty(ip);
3234 hammer_critical_error(ip->hmp, ip, error,
3235 "while syncing inode");
3237 hammer_done_cursor(&cursor);
3242 * This routine is called when the OS is no longer actively referencing
3243 * the inode (but might still be keeping it cached), or when releasing
3244 * the last reference to an inode.
3246 * At this point if the inode's nlinks count is zero we want to destroy
3247 * it, which may mean destroying it on-media too.
3250 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3255 * Set the DELETING flag when the link count drops to 0 and the
3256 * OS no longer has any opens on the inode.
3258 * The backend will clear DELETING (a mod flag) and set DELETED
3259 * (a state flag) when it is actually able to perform the
3262 * Don't reflag the deletion if the flusher is currently syncing
3263 * one that was already flagged. A previously set DELETING flag
3264 * may bounce around flags and sync_flags until the operation is
3267 * Do not attempt to modify a snapshot inode (one set to read-only).
3269 if (ip->ino_data.nlinks == 0 &&
3270 ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3271 ip->flags |= HAMMER_INODE_DELETING;
3272 ip->flags |= HAMMER_INODE_TRUNCATED;
3276 if (hammer_get_vnode(ip, &vp) != 0)
3284 nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0, 0);
3285 if (ip->flags & HAMMER_INODE_MODMASK)
3286 hammer_inode_dirty(ip);
3293 * After potentially resolving a dependancy the inode is tested
3294 * to determine whether it needs to be reflushed.
3297 hammer_test_inode(hammer_inode_t ip)
3299 if (ip->flags & HAMMER_INODE_REFLUSH) {
3300 ip->flags &= ~HAMMER_INODE_REFLUSH;
3301 hammer_ref(&ip->lock);
3302 if (ip->flags & HAMMER_INODE_RESIGNAL) {
3303 ip->flags &= ~HAMMER_INODE_RESIGNAL;
3304 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3306 hammer_flush_inode(ip, 0);
3308 hammer_rel_inode(ip, 0);
3313 * Clear the RECLAIM flag on an inode. This occurs when the inode is
3314 * reassociated with a vp or just before it gets freed.
3316 * Pipeline wakeups to threads blocked due to an excessive number of
3317 * detached inodes. This typically occurs when atime updates accumulate
3318 * while scanning a directory tree.
3321 hammer_inode_wakereclaims(hammer_inode_t ip)
3323 struct hammer_reclaim *reclaim;
3324 hammer_mount_t hmp = ip->hmp;
3326 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3329 --hammer_count_reclaims;
3330 --hmp->count_reclaims;
3331 ip->flags &= ~HAMMER_INODE_RECLAIM;
3333 if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3334 KKASSERT(reclaim->count > 0);
3335 if (--reclaim->count == 0) {
3336 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3343 * Setup our reclaim pipeline. We only let so many detached (and dirty)
3344 * inodes build up before we start blocking. This routine is called
3345 * if a new inode is created or an inode is loaded from media.
3347 * When we block we don't care *which* inode has finished reclaiming,
3348 * as long as one does.
3350 * The reclaim pipeline is primarily governed by the auto-flush which is
3351 * 1/4 hammer_limit_reclaims. We don't want to block if the count is
3352 * less than 1/2 hammer_limit_reclaims. From 1/2 to full count is
3353 * dynamically governed.
3356 hammer_inode_waitreclaims(hammer_transaction_t trans)
3358 hammer_mount_t hmp = trans->hmp;
3359 struct hammer_reclaim reclaim;
3363 * Track inode load, delay if the number of reclaiming inodes is
3364 * between 2/4 and 4/4 hammer_limit_reclaims, depending.
3366 if (curthread->td_proc) {
3367 struct hammer_inostats *stats;
3369 stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3372 if (stats->count > hammer_limit_reclaims / 2)
3373 stats->count = hammer_limit_reclaims / 2;
3374 lower_limit = hammer_limit_reclaims - stats->count;
3375 if (hammer_debug_general & 0x10000) {
3376 kprintf("pid %5d limit %d\n",
3377 (int)curthread->td_proc->p_pid, lower_limit);
3380 lower_limit = hammer_limit_reclaims * 3 / 4;
3382 if (hmp->count_reclaims >= lower_limit) {
3384 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3385 tsleep(&reclaim, 0, "hmrrcm", hz);
3386 if (reclaim.count > 0)
3387 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3392 * Keep track of reclaim statistics on a per-pid basis using a loose
3393 * 4-way set associative hash table. Collisions inherit the count of
3394 * the previous entry.
3396 * NOTE: We want to be careful here to limit the chain size. If the chain
3397 * size is too large a pid will spread its stats out over too many
3398 * entries under certain types of heavy filesystem activity and
3399 * wind up not delaying long enough.
3402 struct hammer_inostats *
3403 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3405 struct hammer_inostats *stats;
3408 static volatile int iterator; /* we don't care about MP races */
3411 * Chain up to 4 times to find our entry.
3413 for (chain = 0; chain < 4; ++chain) {
3414 stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3415 if (stats->pid == pid)
3420 * Replace one of the four chaining entries with our new entry.
3423 stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3424 HAMMER_INOSTATS_HMASK];
3431 if (stats->count && stats->ltick != ticks) {
3432 delta = ticks - stats->ltick;
3433 stats->ltick = ticks;
3434 if (delta <= 0 || delta > hz * 60)
3437 stats->count = stats->count * hz / (hz + delta);
3439 if (hammer_debug_general & 0x10000)
3440 kprintf("pid %5d stats %d\n", (int)pid, stats->count);
3447 * XXX not used, doesn't work very well due to the large batching nature
3450 * A larger then normal backlog of inodes is sitting in the flusher,
3451 * enforce a general slowdown to let it catch up. This routine is only
3452 * called on completion of a non-flusher-related transaction which
3453 * performed B-Tree node I/O.
3455 * It is possible for the flusher to stall in a continuous load.
3456 * blogbench -i1000 -o seems to do a good job generating this sort of load.
3457 * If the flusher is unable to catch up the inode count can bloat until
3458 * we run out of kvm.
3460 * This is a bit of a hack.
3463 hammer_inode_waithard(hammer_mount_t hmp)
3468 if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3469 if (hmp->count_reclaims < hammer_limit_reclaims / 2 &&
3470 hmp->count_iqueued < hmp->count_inodes / 20) {
3471 hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3475 if (hmp->count_reclaims < hammer_limit_reclaims ||
3476 hmp->count_iqueued < hmp->count_inodes / 10) {
3479 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3483 * Block for one flush cycle.
3485 hammer_flusher_wait_next(hmp);