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
35 #include <vm/vm_page2.h>
39 static int hammer_unload_inode(struct hammer_inode *ip);
40 static void hammer_free_inode(hammer_inode_t ip);
41 static void hammer_flush_inode_core(hammer_inode_t ip,
42 hammer_flush_group_t flg, int flags);
43 static int hammer_setup_child_callback(hammer_record_t rec, void *data);
45 static int hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
47 static int hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
48 hammer_flush_group_t flg);
49 static int hammer_setup_parent_inodes_helper(hammer_record_t record,
50 int depth, hammer_flush_group_t flg);
51 static void hammer_inode_wakereclaims(hammer_inode_t ip);
52 static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
54 static struct hammer_inode *__hammer_find_inode(hammer_transaction_t trans,
55 int64_t obj_id, hammer_tid_t asof,
56 uint32_t localization);
58 struct krate hammer_gen_krate = { 1 };
61 * RB-Tree support for inode structures
64 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
66 if (ip1->obj_localization < ip2->obj_localization)
68 if (ip1->obj_localization > ip2->obj_localization)
70 if (ip1->obj_id < ip2->obj_id)
72 if (ip1->obj_id > ip2->obj_id)
74 if (ip1->obj_asof < ip2->obj_asof)
76 if (ip1->obj_asof > ip2->obj_asof)
82 hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
84 if (ip1->redo_fifo_start < ip2->redo_fifo_start)
86 if (ip1->redo_fifo_start > ip2->redo_fifo_start)
92 * RB-Tree support for inode structures / special LOOKUP_INFO
95 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
97 if (info->obj_localization < ip->obj_localization)
99 if (info->obj_localization > ip->obj_localization)
101 if (info->obj_id < ip->obj_id)
103 if (info->obj_id > ip->obj_id)
105 if (info->obj_asof < ip->obj_asof)
107 if (info->obj_asof > ip->obj_asof)
113 * Used by hammer_scan_inode_snapshots() to locate all of an object's
114 * snapshots. Note that the asof field is not tested, which we can get
115 * away with because it is the lowest-priority field.
118 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
120 hammer_inode_info_t info = data;
122 if (ip->obj_localization > info->obj_localization)
124 if (ip->obj_localization < info->obj_localization)
126 if (ip->obj_id > info->obj_id)
128 if (ip->obj_id < info->obj_id)
134 * Used by hammer_unload_pseudofs() to locate all inodes associated with
138 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
140 uint32_t localization = *(uint32_t *)data;
141 if (ip->obj_localization > localization)
143 if (ip->obj_localization < localization)
149 * RB-Tree support for pseudofs structures
152 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
154 if (p1->localization < p2->localization)
156 if (p1->localization > p2->localization)
162 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
163 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
164 hammer_inode_info_cmp, hammer_inode_info_t);
165 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
166 hammer_pfs_rb_compare, uint32_t, localization);
169 * The kernel is not actively referencing this vnode but is still holding
172 * This is called from the frontend.
177 hammer_vop_inactive(struct vop_inactive_args *ap)
179 struct hammer_inode *ip = VTOI(ap->a_vp);
191 * If the inode no longer has visibility in the filesystem try to
192 * recycle it immediately, even if the inode is dirty. Recycling
193 * it quickly allows the system to reclaim buffer cache and VM
194 * resources which can matter a lot in a heavily loaded system.
196 * This can deadlock in vfsync() if we aren't careful.
198 * Do not queue the inode to the flusher if we still have visibility,
199 * otherwise namespace calls such as chmod will unnecessarily generate
200 * multiple inode updates.
202 if (ip->ino_data.nlinks == 0) {
204 lwkt_gettoken(&hmp->fs_token);
205 hammer_inode_unloadable_check(ip, 0);
206 if (ip->flags & HAMMER_INODE_MODMASK)
207 hammer_flush_inode(ip, 0);
208 lwkt_reltoken(&hmp->fs_token);
215 * Release the vnode association. This is typically (but not always)
216 * the last reference on the inode.
218 * Once the association is lost we are on our own with regards to
219 * flushing the inode.
221 * We must interlock ip->vp so hammer_get_vnode() can avoid races.
224 hammer_vop_reclaim(struct vop_reclaim_args *ap)
226 struct hammer_inode *ip;
232 if ((ip = vp->v_data) != NULL) {
234 lwkt_gettoken(&hmp->fs_token);
235 hammer_lock_ex(&ip->lock);
239 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
240 ++hammer_count_reclaims;
241 ++hmp->count_reclaims;
242 ip->flags |= HAMMER_INODE_RECLAIM;
244 hammer_unlock(&ip->lock);
246 hammer_rel_inode(ip, 1);
247 lwkt_reltoken(&hmp->fs_token);
253 * Inform the kernel that the inode is dirty. This will be checked
256 * Theoretically in order to reclaim a vnode the hammer_vop_reclaim()
257 * must be called which will interlock against our inode lock, so
258 * if VRECLAIMED is not set vp->v_mount (as used by vsetisdirty())
259 * should be stable without having to acquire any new locks.
262 hammer_inode_dirty(struct hammer_inode *ip)
266 if ((ip->flags & HAMMER_INODE_MODMASK) &&
267 (vp = ip->vp) != NULL &&
268 (vp->v_flag & (VRECLAIMED | VISDIRTY)) == 0) {
274 * Return a locked vnode for the specified inode. The inode must be
275 * referenced but NOT LOCKED on entry and will remain referenced on
278 * Called from the frontend.
281 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
291 if ((vp = ip->vp) == NULL) {
292 error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
295 hammer_lock_ex(&ip->lock);
296 if (ip->vp != NULL) {
297 hammer_unlock(&ip->lock);
303 hammer_ref(&ip->lock);
307 obj_type = ip->ino_data.obj_type;
308 vp->v_type = hammer_get_vnode_type(obj_type);
310 hammer_inode_wakereclaims(ip);
312 switch(ip->ino_data.obj_type) {
313 case HAMMER_OBJTYPE_CDEV:
314 case HAMMER_OBJTYPE_BDEV:
315 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
316 addaliasu(vp, ip->ino_data.rmajor,
317 ip->ino_data.rminor);
319 case HAMMER_OBJTYPE_FIFO:
320 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
322 case HAMMER_OBJTYPE_REGFILE:
329 * Only mark as the root vnode if the ip is not
330 * historical, otherwise the VFS cache will get
331 * confused. The other half of the special handling
332 * is in hammer_vop_nlookupdotdot().
334 * Pseudo-filesystem roots can be accessed via
335 * non-root filesystem paths and setting VROOT may
336 * confuse the namecache. Set VPFSROOT instead.
338 if (ip->obj_id == HAMMER_OBJID_ROOT) {
339 if (ip->obj_asof == hmp->asof) {
340 if (ip->obj_localization ==
341 HAMMER_DEF_LOCALIZATION)
342 vsetflags(vp, VROOT);
344 vsetflags(vp, VPFSROOT);
346 vsetflags(vp, VPFSROOT);
350 vp->v_data = (void *)ip;
351 /* vnode locked by getnewvnode() */
352 /* make related vnode dirty if inode dirty? */
353 hammer_unlock(&ip->lock);
354 if (vp->v_type == VREG) {
355 vinitvmio(vp, ip->ino_data.size,
356 hammer_blocksize(ip->ino_data.size),
357 hammer_blockoff(ip->ino_data.size));
363 * Interlock vnode clearing. This does not prevent the
364 * vnode from going into a reclaimed state but it does
365 * prevent it from being destroyed or reused so the vget()
366 * will properly fail.
368 hammer_lock_ex(&ip->lock);
369 if ((vp = ip->vp) == NULL) {
370 hammer_unlock(&ip->lock);
374 hammer_unlock(&ip->lock);
377 * loop if the vget fails (aka races), or if the vp
378 * no longer matches ip->vp.
380 if (vget(vp, LK_EXCLUSIVE) == 0) {
394 * Locate all copies of the inode for obj_id compatible with the specified
395 * asof, reference, and issue the related call-back. This routine is used
396 * for direct-io invalidation and does not create any new inodes.
399 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
400 int (*callback)(hammer_inode_t ip, void *data),
403 hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
404 hammer_inode_info_cmp_all_history,
409 * Acquire a HAMMER inode. The returned inode is not locked. These functions
410 * do not attach or detach the related vnode (use hammer_get_vnode() for
413 * The flags argument is only applied for newly created inodes, and only
414 * certain flags are inherited.
416 * Called from the frontend.
418 struct hammer_inode *
419 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
420 int64_t obj_id, hammer_tid_t asof, uint32_t localization,
421 int flags, int *errorp)
423 hammer_mount_t hmp = trans->hmp;
424 struct hammer_node_cache *cachep;
425 struct hammer_cursor cursor;
426 struct hammer_inode *ip;
430 * Determine if we already have an inode cached. If we do then
433 * If we find an inode with no vnode we have to mark the
434 * transaction such that hammer_inode_waitreclaims() is
435 * called later on to avoid building up an infinite number
436 * of inodes. Otherwise we can continue to * add new inodes
437 * faster then they can be disposed of, even with the tsleep
440 * If we find a dummy inode we return a failure so dounlink
441 * (which does another lookup) doesn't try to mess with the
442 * link count. hammer_vop_nresolve() uses hammer_get_dummy_inode()
443 * to ref dummy inodes.
447 ip = __hammer_find_inode(trans, obj_id, asof, localization);
449 if (ip->flags & HAMMER_INODE_DUMMY) {
453 hammer_ref(&ip->lock);
458 * Allocate a new inode structure and deal with races later.
460 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
461 ++hammer_count_inodes;
465 ip->obj_localization = localization;
467 ip->flags = flags & HAMMER_INODE_RO;
468 ip->cache[0].ip = ip;
469 ip->cache[1].ip = ip;
470 ip->cache[2].ip = ip;
471 ip->cache[3].ip = ip;
473 ip->flags |= HAMMER_INODE_RO;
474 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
475 0x7FFFFFFFFFFFFFFFLL;
476 RB_INIT(&ip->rec_tree);
477 TAILQ_INIT(&ip->target_list);
478 hammer_ref(&ip->lock);
481 * Locate the on-disk inode. If this is a PFS root we always
482 * access the current version of the root inode and (if it is not
483 * a master) always access information under it with a snapshot
486 * We cache recent inode lookups in this directory in dip->cache[2].
487 * If we can't find it we assume the inode we are looking for is
488 * close to the directory inode.
493 if (dip->cache[2].node)
494 cachep = &dip->cache[2];
496 cachep = &dip->cache[0];
498 hammer_init_cursor(trans, &cursor, cachep, NULL);
499 cursor.key_beg.localization = localization | HAMMER_LOCALIZE_INODE;
500 cursor.key_beg.obj_id = ip->obj_id;
501 cursor.key_beg.key = 0;
502 cursor.key_beg.create_tid = 0;
503 cursor.key_beg.delete_tid = 0;
504 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
505 cursor.key_beg.obj_type = 0;
508 cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
511 *errorp = hammer_btree_lookup(&cursor);
512 if (*errorp == EDEADLK) {
513 hammer_done_cursor(&cursor);
518 * On success the B-Tree lookup will hold the appropriate
519 * buffer cache buffers and provide a pointer to the requested
520 * information. Copy the information to the in-memory inode
521 * and cache the B-Tree node to improve future operations.
524 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
525 ip->ino_data = cursor.data->inode;
528 * cache[0] tries to cache the location of the object inode.
529 * The assumption is that it is near the directory inode.
531 * cache[1] tries to cache the location of the object data.
532 * We might have something in the governing directory from
533 * scan optimizations (see the strategy code in
536 * We update dip->cache[2], if possible, with the location
537 * of the object inode for future directory shortcuts.
539 hammer_cache_node(&ip->cache[0], cursor.node);
541 if (dip->cache[3].node) {
542 hammer_cache_node(&ip->cache[1],
545 hammer_cache_node(&dip->cache[2], cursor.node);
549 * The file should not contain any data past the file size
550 * stored in the inode. Setting save_trunc_off to the
551 * file size instead of max reduces B-Tree lookup overheads
552 * on append by allowing the flusher to avoid checking for
555 ip->save_trunc_off = ip->ino_data.size;
558 * Locate and assign the pseudofs management structure to
561 if (dip && dip->obj_localization == ip->obj_localization) {
562 ip->pfsm = dip->pfsm;
563 hammer_ref(&ip->pfsm->lock);
565 ip->pfsm = hammer_load_pseudofs(trans,
566 ip->obj_localization,
568 *errorp = 0; /* ignore ENOENT */
573 * The inode is placed on the red-black tree and will be synced to
574 * the media when flushed or by the filesystem sync. If this races
575 * another instantiation/lookup the insertion will fail.
578 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
579 hammer_free_inode(ip);
580 hammer_done_cursor(&cursor);
583 ip->flags |= HAMMER_INODE_ONDISK;
585 if (ip->flags & HAMMER_INODE_RSV_INODES) {
586 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
590 hammer_free_inode(ip);
593 hammer_done_cursor(&cursor);
596 * NEWINODE is only set if the inode becomes dirty later,
597 * setting it here just leads to unnecessary stalls.
599 * trans->flags |= HAMMER_TRANSF_NEWINODE;
605 * Get a dummy inode to placemark a broken directory entry.
607 struct hammer_inode *
608 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
609 int64_t obj_id, hammer_tid_t asof, uint32_t localization,
610 int flags, int *errorp)
612 hammer_mount_t hmp = trans->hmp;
613 struct hammer_inode *ip;
616 * Determine if we already have an inode cached. If we do then
619 * If we find an inode with no vnode we have to mark the
620 * transaction such that hammer_inode_waitreclaims() is
621 * called later on to avoid building up an infinite number
622 * of inodes. Otherwise we can continue to * add new inodes
623 * faster then they can be disposed of, even with the tsleep
626 * If we find a non-fake inode we return an error. Only fake
627 * inodes can be returned by this routine.
631 ip = __hammer_find_inode(trans, obj_id, asof, localization);
633 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
637 hammer_ref(&ip->lock);
642 * Allocate a new inode structure and deal with races later.
644 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
645 ++hammer_count_inodes;
649 ip->obj_localization = localization;
651 ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
652 ip->cache[0].ip = ip;
653 ip->cache[1].ip = ip;
654 ip->cache[2].ip = ip;
655 ip->cache[3].ip = ip;
656 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
657 0x7FFFFFFFFFFFFFFFLL;
658 RB_INIT(&ip->rec_tree);
659 TAILQ_INIT(&ip->target_list);
660 hammer_ref(&ip->lock);
663 * Populate the dummy inode. Leave everything zero'd out.
665 * (ip->ino_leaf and ip->ino_data)
667 * Make the dummy inode a FIFO object which most copy programs
668 * will properly ignore.
670 ip->save_trunc_off = ip->ino_data.size;
671 ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
674 * Locate and assign the pseudofs management structure to
677 if (dip && dip->obj_localization == ip->obj_localization) {
678 ip->pfsm = dip->pfsm;
679 hammer_ref(&ip->pfsm->lock);
681 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
683 *errorp = 0; /* ignore ENOENT */
687 * The inode is placed on the red-black tree and will be synced to
688 * the media when flushed or by the filesystem sync. If this races
689 * another instantiation/lookup the insertion will fail.
691 * NOTE: Do not set HAMMER_INODE_ONDISK. The inode is a fake.
694 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
695 hammer_free_inode(ip);
699 if (ip->flags & HAMMER_INODE_RSV_INODES) {
700 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
703 hammer_free_inode(ip);
706 trans->flags |= HAMMER_TRANSF_NEWINODE;
711 * Return a referenced inode only if it is in our inode cache.
712 * Dummy inodes do not count.
714 struct hammer_inode *
715 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
716 hammer_tid_t asof, uint32_t localization)
718 struct hammer_inode *ip;
720 ip = __hammer_find_inode(trans, obj_id, asof, localization);
722 if (ip->flags & HAMMER_INODE_DUMMY)
725 hammer_ref(&ip->lock);
731 * Return a referenced inode only if it is in our inode cache.
732 * This function does not reference inode.
734 static struct hammer_inode *
735 __hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
736 hammer_tid_t asof, uint32_t localization)
738 hammer_mount_t hmp = trans->hmp;
739 struct hammer_inode_info iinfo;
740 struct hammer_inode *ip;
742 iinfo.obj_id = obj_id;
743 iinfo.obj_asof = asof;
744 iinfo.obj_localization = localization;
746 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
752 * Create a new filesystem object, returning the inode in *ipp. The
753 * returned inode will be referenced. The inode is created in-memory.
755 * If pfsm is non-NULL the caller wishes to create the root inode for
759 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
761 hammer_inode_t dip, const char *name, int namelen,
762 hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
774 * Disallow the creation of new inodes in directories which
775 * have been deleted. In HAMMER, this will cause a record
776 * syncing assertion later on in the flush code.
778 if (dip && dip->ino_data.nlinks == 0) {
786 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
787 ++hammer_count_inodes;
789 trans->flags |= HAMMER_TRANSF_NEWINODE;
792 KKASSERT(pfsm->localization != HAMMER_DEF_LOCALIZATION);
793 ip->obj_id = HAMMER_OBJID_ROOT;
794 ip->obj_localization = pfsm->localization;
796 KKASSERT(dip != NULL);
797 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
798 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
799 ip->obj_localization = dip->obj_localization;
802 KKASSERT(ip->obj_id != 0);
803 ip->obj_asof = hmp->asof;
805 ip->flush_state = HAMMER_FST_IDLE;
806 ip->flags = HAMMER_INODE_DDIRTY |
807 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
808 ip->cache[0].ip = ip;
809 ip->cache[1].ip = ip;
810 ip->cache[2].ip = ip;
811 ip->cache[3].ip = ip;
813 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
814 /* ip->save_trunc_off = 0; (already zero) */
815 RB_INIT(&ip->rec_tree);
816 TAILQ_INIT(&ip->target_list);
818 ip->ino_data.atime = trans->time;
819 ip->ino_data.mtime = trans->time;
820 ip->ino_data.size = 0;
821 ip->ino_data.nlinks = 0;
824 * A nohistory designator on the parent directory is inherited by
825 * the child. We will do this even for pseudo-fs creation... the
826 * sysad can turn it off.
829 ip->ino_data.uflags = dip->ino_data.uflags &
830 (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
833 ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
834 ip->ino_leaf.base.localization = ip->obj_localization |
835 HAMMER_LOCALIZE_INODE;
836 ip->ino_leaf.base.obj_id = ip->obj_id;
837 ip->ino_leaf.base.key = 0;
838 ip->ino_leaf.base.create_tid = 0;
839 ip->ino_leaf.base.delete_tid = 0;
840 ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
841 ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
843 ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
844 ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
845 ip->ino_data.mode = vap->va_mode;
846 ip->ino_data.ctime = trans->time;
849 * If we are running version 2 or greater directory entries are
850 * inode-localized instead of data-localized.
852 if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
853 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
854 ip->ino_data.cap_flags |=
855 HAMMER_INODE_CAP_DIR_LOCAL_INO;
858 if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
859 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
860 ip->ino_data.cap_flags |=
861 HAMMER_INODE_CAP_DIRHASH_ALG1;
866 * Setup the ".." pointer. This only needs to be done for directories
867 * but we do it for all objects as a recovery aid if dip exists.
868 * The inode is probably a PFS root if dip is NULL.
871 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
873 switch(ip->ino_leaf.base.obj_type) {
874 case HAMMER_OBJTYPE_CDEV:
875 case HAMMER_OBJTYPE_BDEV:
876 ip->ino_data.rmajor = vap->va_rmajor;
877 ip->ino_data.rminor = vap->va_rminor;
884 * Calculate default uid/gid and overwrite with information from
888 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
889 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
890 xuid, cred, &vap->va_mode);
894 ip->ino_data.mode = vap->va_mode;
896 if (vap->va_vaflags & VA_UID_UUID_VALID)
897 ip->ino_data.uid = vap->va_uid_uuid;
898 else if (vap->va_uid != (uid_t)VNOVAL)
899 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
901 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
903 if (vap->va_vaflags & VA_GID_UUID_VALID)
904 ip->ino_data.gid = vap->va_gid_uuid;
905 else if (vap->va_gid != (gid_t)VNOVAL)
906 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
908 ip->ino_data.gid = dip->ino_data.gid;
910 hammer_ref(&ip->lock);
914 hammer_ref(&pfsm->lock);
916 } else if (dip->obj_localization == ip->obj_localization) {
917 ip->pfsm = dip->pfsm;
918 hammer_ref(&ip->pfsm->lock);
921 ip->pfsm = hammer_load_pseudofs(trans,
922 ip->obj_localization,
924 error = 0; /* ignore ENOENT */
928 hammer_free_inode(ip);
930 } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
931 hpanic("duplicate obj_id %llx", (long long)ip->obj_id);
933 hammer_free_inode(ip);
940 * Final cleanup / freeing of an inode structure
943 hammer_free_inode(hammer_inode_t ip)
945 struct hammer_mount *hmp;
948 KKASSERT(hammer_oneref(&ip->lock));
949 hammer_uncache_node(&ip->cache[0]);
950 hammer_uncache_node(&ip->cache[1]);
951 hammer_uncache_node(&ip->cache[2]);
952 hammer_uncache_node(&ip->cache[3]);
953 hammer_inode_wakereclaims(ip);
955 hammer_clear_objid(ip);
956 --hammer_count_inodes;
959 hammer_rel_pseudofs(hmp, ip->pfsm);
962 kfree(ip, hmp->m_inodes);
966 * Retrieve pseudo-fs data. NULL will never be returned.
968 * If an error occurs *errorp will be set and a default template is returned,
969 * otherwise *errorp is set to 0. Typically when an error occurs it will
972 hammer_pseudofs_inmem_t
973 hammer_load_pseudofs(hammer_transaction_t trans,
974 uint32_t localization, int *errorp)
976 hammer_mount_t hmp = trans->hmp;
978 hammer_pseudofs_inmem_t pfsm;
979 struct hammer_cursor cursor;
983 pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
985 hammer_ref(&pfsm->lock);
991 * PFS records are associated with the root inode (not the PFS root
992 * inode, but the real root). Avoid an infinite recursion if loading
993 * the PFS for the real root.
996 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
998 HAMMER_DEF_LOCALIZATION, 0, errorp);
1003 pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
1004 pfsm->localization = localization;
1005 pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
1006 pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
1008 hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
1009 cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION |
1010 HAMMER_LOCALIZE_MISC;
1011 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1012 cursor.key_beg.create_tid = 0;
1013 cursor.key_beg.delete_tid = 0;
1014 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1015 cursor.key_beg.obj_type = 0;
1016 cursor.key_beg.key = localization;
1017 cursor.asof = HAMMER_MAX_TID;
1018 cursor.flags |= HAMMER_CURSOR_ASOF;
1021 *errorp = hammer_ip_lookup(&cursor);
1023 *errorp = hammer_btree_lookup(&cursor);
1025 *errorp = hammer_ip_resolve_data(&cursor);
1027 if (cursor.data->pfsd.mirror_flags &
1028 HAMMER_PFSD_DELETED) {
1031 bytes = cursor.leaf->data_len;
1032 if (bytes > sizeof(pfsm->pfsd))
1033 bytes = sizeof(pfsm->pfsd);
1034 bcopy(cursor.data, &pfsm->pfsd, bytes);
1038 hammer_done_cursor(&cursor);
1040 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1041 hammer_ref(&pfsm->lock);
1043 hammer_rel_inode(ip, 0);
1044 if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
1045 kfree(pfsm, hmp->m_misc);
1052 * Store pseudo-fs data. The backend will automatically delete any prior
1053 * on-disk pseudo-fs data but we have to delete in-memory versions.
1056 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
1058 struct hammer_cursor cursor;
1059 hammer_record_t record;
1064 * PFS records are associated with the root inode (not the PFS root
1065 * inode, but the real root).
1067 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1068 HAMMER_DEF_LOCALIZATION, 0, &error);
1070 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1071 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1072 cursor.key_beg.localization = ip->obj_localization |
1073 HAMMER_LOCALIZE_MISC;
1074 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1075 cursor.key_beg.create_tid = 0;
1076 cursor.key_beg.delete_tid = 0;
1077 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1078 cursor.key_beg.obj_type = 0;
1079 cursor.key_beg.key = pfsm->localization;
1080 cursor.asof = HAMMER_MAX_TID;
1081 cursor.flags |= HAMMER_CURSOR_ASOF;
1084 * Replace any in-memory version of the record.
1086 error = hammer_ip_lookup(&cursor);
1087 if (error == 0 && hammer_cursor_inmem(&cursor)) {
1088 record = cursor.iprec;
1089 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1090 KKASSERT(cursor.deadlk_rec == NULL);
1091 hammer_ref(&record->lock);
1092 cursor.deadlk_rec = record;
1095 record->flags |= HAMMER_RECF_DELETED_FE;
1101 * Allocate replacement general record. The backend flush will
1102 * delete any on-disk version of the record.
1104 if (error == 0 || error == ENOENT) {
1105 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1106 record->type = HAMMER_MEM_RECORD_GENERAL;
1108 record->leaf.base.localization = ip->obj_localization |
1109 HAMMER_LOCALIZE_MISC;
1110 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1111 record->leaf.base.key = pfsm->localization;
1112 record->leaf.data_len = sizeof(pfsm->pfsd);
1113 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1114 error = hammer_ip_add_record(trans, record);
1116 hammer_done_cursor(&cursor);
1117 if (error == EDEADLK)
1119 hammer_rel_inode(ip, 0);
1124 * Create a root directory for a PFS if one does not alredy exist.
1126 * The PFS root stands alone so we must also bump the nlinks count
1127 * to prevent it from being destroyed on release.
1130 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1131 hammer_pseudofs_inmem_t pfsm)
1137 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1138 pfsm->localization, 0, &error);
1143 error = hammer_create_inode(trans, &vap, cred,
1147 ++ip->ino_data.nlinks;
1148 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1152 hammer_rel_inode(ip, 0);
1157 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1158 * if we are unable to disassociate all the inodes.
1162 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1166 hammer_ref(&ip->lock);
1167 if (ip->vp && (ip->vp->v_flag & VPFSROOT)) {
1169 * The hammer pfs-upgrade directive itself might have the
1170 * root of the pfs open. Just allow it.
1175 * Don't allow any subdirectories or files to be open.
1177 if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1178 vclean_unlocked(ip->vp);
1179 if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1182 res = -1; /* stop, someone is using the inode */
1184 hammer_rel_inode(ip, 0);
1189 hammer_unload_pseudofs(hammer_transaction_t trans, uint32_t localization)
1194 for (try = res = 0; try < 4; ++try) {
1195 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1196 hammer_inode_pfs_cmp,
1197 hammer_unload_pseudofs_callback,
1199 if (res == 0 && try > 1)
1201 hammer_flusher_sync(trans->hmp);
1210 * Release a reference on a PFS
1213 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1215 hammer_rel(&pfsm->lock);
1216 if (hammer_norefs(&pfsm->lock)) {
1217 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1218 kfree(pfsm, hmp->m_misc);
1223 * Called by hammer_sync_inode().
1226 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1228 hammer_transaction_t trans = cursor->trans;
1229 hammer_record_t record;
1237 * If the inode has a presence on-disk then locate it and mark
1238 * it deleted, setting DELONDISK.
1240 * The record may or may not be physically deleted, depending on
1241 * the retention policy.
1243 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1244 HAMMER_INODE_ONDISK) {
1245 hammer_normalize_cursor(cursor);
1246 cursor->key_beg.localization = ip->obj_localization |
1247 HAMMER_LOCALIZE_INODE;
1248 cursor->key_beg.obj_id = ip->obj_id;
1249 cursor->key_beg.key = 0;
1250 cursor->key_beg.create_tid = 0;
1251 cursor->key_beg.delete_tid = 0;
1252 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1253 cursor->key_beg.obj_type = 0;
1254 cursor->asof = ip->obj_asof;
1255 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1256 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1257 cursor->flags |= HAMMER_CURSOR_BACKEND;
1259 error = hammer_btree_lookup(cursor);
1260 if (hammer_debug_inode)
1261 hdkprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1264 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1265 if (hammer_debug_inode)
1266 hdkprintf("error %d\n", error);
1268 ip->flags |= HAMMER_INODE_DELONDISK;
1271 hammer_cache_node(&ip->cache[0], cursor->node);
1273 if (error == EDEADLK) {
1274 hammer_done_cursor(cursor);
1275 error = hammer_init_cursor(trans, cursor,
1277 if (hammer_debug_inode)
1278 hdkprintf("IPDED %p %d\n", ip, error);
1285 * Ok, write out the initial record or a new record (after deleting
1286 * the old one), unless the DELETED flag is set. This routine will
1287 * clear DELONDISK if it writes out a record.
1289 * Update our inode statistics if this is the first application of
1290 * the inode on-disk.
1292 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1294 * Generate a record and write it to the media. We clean-up
1295 * the state before releasing so we do not have to set-up
1298 record = hammer_alloc_mem_record(ip, 0);
1299 record->type = HAMMER_MEM_RECORD_INODE;
1300 record->flush_state = HAMMER_FST_FLUSH;
1301 record->leaf = ip->sync_ino_leaf;
1302 record->leaf.base.create_tid = trans->tid;
1303 record->leaf.data_len = sizeof(ip->sync_ino_data);
1304 record->leaf.create_ts = trans->time32;
1305 record->data = (void *)&ip->sync_ino_data;
1306 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1309 * If this flag is set we cannot sync the new file size
1310 * because we haven't finished related truncations. The
1311 * inode will be flushed in another flush group to finish
1314 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1315 ip->sync_ino_data.size != ip->ino_data.size) {
1317 ip->sync_ino_data.size = ip->ino_data.size;
1323 error = hammer_ip_sync_record_cursor(cursor, record);
1324 if (hammer_debug_inode)
1325 hdkprintf("GENREC %p rec %08x %d\n",
1326 ip, record->flags, error);
1327 if (error != EDEADLK)
1329 hammer_done_cursor(cursor);
1330 error = hammer_init_cursor(trans, cursor,
1332 if (hammer_debug_inode)
1333 hdkprintf("GENREC reinit %d\n", error);
1339 * Note: The record was never on the inode's record tree
1340 * so just wave our hands importantly and destroy it.
1342 record->flags |= HAMMER_RECF_COMMITTED;
1343 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1344 record->flush_state = HAMMER_FST_IDLE;
1345 ++ip->rec_generation;
1346 hammer_rel_mem_record(record);
1352 if (hammer_debug_inode)
1353 hdkprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1354 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1355 HAMMER_INODE_SDIRTY |
1356 HAMMER_INODE_ATIME |
1357 HAMMER_INODE_MTIME);
1358 ip->flags &= ~HAMMER_INODE_DELONDISK;
1360 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1363 * Root volume count of inodes
1365 hammer_sync_lock_sh(trans);
1366 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1367 hammer_modify_volume_field(trans,
1370 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1371 hammer_modify_volume_done(trans->rootvol);
1372 ip->flags |= HAMMER_INODE_ONDISK;
1373 if (hammer_debug_inode)
1374 hdkprintf("NOWONDISK %p\n", ip);
1376 hammer_sync_unlock(trans);
1381 * If the inode has been destroyed, clean out any left-over flags
1382 * that may have been set by the frontend.
1384 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1385 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1386 HAMMER_INODE_SDIRTY |
1387 HAMMER_INODE_ATIME |
1388 HAMMER_INODE_MTIME);
1394 * Update only the itimes fields.
1396 * ATIME can be updated without generating any UNDO. MTIME is updated
1397 * with UNDO so it is guaranteed to be synchronized properly in case of
1400 * Neither field is included in the B-Tree leaf element's CRC, which is how
1401 * we can get away with updating ATIME the way we do.
1404 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1406 hammer_transaction_t trans = cursor->trans;
1410 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1411 HAMMER_INODE_ONDISK) {
1415 hammer_normalize_cursor(cursor);
1416 cursor->key_beg.localization = ip->obj_localization |
1417 HAMMER_LOCALIZE_INODE;
1418 cursor->key_beg.obj_id = ip->obj_id;
1419 cursor->key_beg.key = 0;
1420 cursor->key_beg.create_tid = 0;
1421 cursor->key_beg.delete_tid = 0;
1422 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1423 cursor->key_beg.obj_type = 0;
1424 cursor->asof = ip->obj_asof;
1425 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1426 cursor->flags |= HAMMER_CURSOR_ASOF;
1427 cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1428 cursor->flags |= HAMMER_CURSOR_GET_DATA;
1429 cursor->flags |= HAMMER_CURSOR_BACKEND;
1431 error = hammer_btree_lookup(cursor);
1433 hammer_cache_node(&ip->cache[0], cursor->node);
1434 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1436 * Updating MTIME requires an UNDO. Just cover
1437 * both atime and mtime.
1439 hammer_sync_lock_sh(trans);
1440 hammer_modify_buffer(trans, cursor->data_buffer,
1441 &cursor->data->inode.mtime,
1442 sizeof(cursor->data->inode.atime) +
1443 sizeof(cursor->data->inode.mtime));
1444 cursor->data->inode.atime = ip->sync_ino_data.atime;
1445 cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1446 hammer_modify_buffer_done(cursor->data_buffer);
1447 hammer_sync_unlock(trans);
1448 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1450 * Updating atime only can be done in-place with
1453 hammer_sync_lock_sh(trans);
1454 hammer_modify_buffer_noundo(trans, cursor->data_buffer);
1455 cursor->data->inode.atime = ip->sync_ino_data.atime;
1456 hammer_modify_buffer_done(cursor->data_buffer);
1457 hammer_sync_unlock(trans);
1459 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1461 if (error == EDEADLK) {
1462 hammer_done_cursor(cursor);
1463 error = hammer_init_cursor(trans, cursor, &ip->cache[0], ip);
1471 * Release a reference on an inode, flush as requested.
1473 * On the last reference we queue the inode to the flusher for its final
1477 hammer_rel_inode(struct hammer_inode *ip, int flush)
1480 * Handle disposition when dropping the last ref.
1483 if (hammer_oneref(&ip->lock)) {
1485 * Determine whether on-disk action is needed for
1486 * the inode's final disposition.
1488 KKASSERT(ip->vp == NULL);
1489 hammer_inode_unloadable_check(ip, 0);
1490 if (ip->flags & HAMMER_INODE_MODMASK) {
1491 hammer_flush_inode(ip, 0);
1492 } else if (hammer_oneref(&ip->lock)) {
1493 hammer_unload_inode(ip);
1498 hammer_flush_inode(ip, 0);
1501 * The inode still has multiple refs, try to drop
1504 KKASSERT(hammer_isactive(&ip->lock) >= 1);
1505 if (hammer_isactive(&ip->lock) > 1) {
1506 hammer_rel(&ip->lock);
1514 * Unload and destroy the specified inode. Must be called with one remaining
1515 * reference. The reference is disposed of.
1517 * The inode must be completely clean.
1520 hammer_unload_inode(struct hammer_inode *ip)
1522 hammer_mount_t hmp = ip->hmp;
1524 KASSERT(hammer_oneref(&ip->lock),
1525 ("hammer_unload_inode: %d refs", hammer_isactive(&ip->lock)));
1526 KKASSERT(ip->vp == NULL);
1527 KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1528 KKASSERT(ip->cursor_ip_refs == 0);
1529 KKASSERT(hammer_notlocked(&ip->lock));
1530 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1532 KKASSERT(RB_EMPTY(&ip->rec_tree));
1533 KKASSERT(TAILQ_EMPTY(&ip->target_list));
1535 if (ip->flags & HAMMER_INODE_RDIRTY) {
1536 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1537 ip->flags &= ~HAMMER_INODE_RDIRTY;
1539 RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1541 hammer_free_inode(ip);
1546 * Called during unmounting if a critical error occured. The in-memory
1547 * inode and all related structures are destroyed.
1549 * If a critical error did not occur the unmount code calls the standard
1550 * release and asserts that the inode is gone.
1553 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1555 hammer_record_t rec;
1558 * Get rid of the inodes in-memory records, regardless of their
1559 * state, and clear the mod-mask.
1561 while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1562 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1563 rec->target_ip = NULL;
1564 if (rec->flush_state == HAMMER_FST_SETUP)
1565 rec->flush_state = HAMMER_FST_IDLE;
1567 while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1568 if (rec->flush_state == HAMMER_FST_FLUSH)
1569 --rec->flush_group->refs;
1571 hammer_ref(&rec->lock);
1572 KKASSERT(hammer_oneref(&rec->lock));
1573 rec->flush_state = HAMMER_FST_IDLE;
1574 rec->flush_group = NULL;
1575 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1576 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1577 ++ip->rec_generation;
1578 hammer_rel_mem_record(rec);
1580 ip->flags &= ~HAMMER_INODE_MODMASK;
1581 ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1582 KKASSERT(ip->vp == NULL);
1585 * Remove the inode from any flush group, force it idle. FLUSH
1586 * and SETUP states have an inode ref.
1588 switch(ip->flush_state) {
1589 case HAMMER_FST_FLUSH:
1590 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1591 --ip->flush_group->refs;
1592 ip->flush_group = NULL;
1594 case HAMMER_FST_SETUP:
1595 hammer_rel(&ip->lock);
1596 ip->flush_state = HAMMER_FST_IDLE;
1598 case HAMMER_FST_IDLE:
1603 * There shouldn't be any associated vnode. The unload needs at
1604 * least one ref, if we do have a vp steal its ip ref.
1607 hdkprintf("Unexpected vnode association ip %p vp %p\n",
1609 ip->vp->v_data = NULL;
1612 hammer_ref(&ip->lock);
1614 hammer_unload_inode(ip);
1619 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1620 * the read-only flag for cached inodes.
1622 * This routine is called from a RB_SCAN().
1625 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1627 hammer_mount_t hmp = ip->hmp;
1629 if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1630 ip->flags |= HAMMER_INODE_RO;
1632 ip->flags &= ~HAMMER_INODE_RO;
1637 * A transaction has modified an inode, requiring updates as specified by
1640 * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1641 * and not including size changes due to write-append
1642 * (but other size changes are included).
1643 * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1645 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1646 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1647 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1648 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1651 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1654 * ronly of 0 or 2 does not trigger assertion.
1655 * 2 is a special error state
1657 KKASSERT(ip->hmp->ronly != 1 ||
1658 (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1659 HAMMER_INODE_SDIRTY |
1660 HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1661 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1662 if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1663 ip->flags |= HAMMER_INODE_RSV_INODES;
1664 ++ip->hmp->rsv_inodes;
1668 * Set the NEWINODE flag in the transaction if the inode
1669 * transitions to a dirty state. This is used to track
1670 * the load on the inode cache.
1673 (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1674 (flags & HAMMER_INODE_MODMASK)) {
1675 trans->flags |= HAMMER_TRANSF_NEWINODE;
1677 if (flags & HAMMER_INODE_MODMASK)
1678 hammer_inode_dirty(ip);
1683 * Attempt to quickly update the atime for a hammer inode. Return 0 on
1684 * success, -1 on failure.
1686 * We attempt to update the atime with only the ip lock and not the
1687 * whole filesystem lock in order to improve concurrency. We can only
1688 * do this safely if the ATIME flag is already pending on the inode.
1690 * This function is called via a vnops path (ip pointer is stable) without
1694 hammer_update_atime_quick(hammer_inode_t ip)
1699 if ((ip->flags & HAMMER_INODE_RO) ||
1700 (ip->hmp->mp->mnt_flag & MNT_NOATIME)) {
1702 * Silently indicate success on read-only mount/snap
1705 } else if (ip->flags & HAMMER_INODE_ATIME) {
1707 * Double check with inode lock held against backend. This
1708 * is only safe if all we need to do is update
1712 hammer_lock_ex(&ip->lock);
1713 if (ip->flags & HAMMER_INODE_ATIME) {
1714 ip->ino_data.atime =
1715 (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec;
1718 hammer_unlock(&ip->lock);
1724 * Request that an inode be flushed. This whole mess cannot block and may
1725 * recurse (if not synchronous). Once requested HAMMER will attempt to
1726 * actively flush the inode until the flush can be done.
1728 * The inode may already be flushing, or may be in a setup state. We can
1729 * place the inode in a flushing state if it is currently idle and flag it
1730 * to reflush if it is currently flushing.
1732 * Upon return if the inode could not be flushed due to a setup
1733 * dependancy, then it will be automatically flushed when the dependancy
1737 hammer_flush_inode(hammer_inode_t ip, int flags)
1740 hammer_flush_group_t flg;
1744 * fill_flush_group is the first flush group we may be able to
1745 * continue filling, it may be open or closed but it will always
1746 * be past the currently flushing (running) flg.
1748 * next_flush_group is the next open flush group.
1751 while ((flg = hmp->fill_flush_group) != NULL) {
1752 KKASSERT(flg->running == 0);
1753 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1754 flg->total_count <= hammer_autoflush) {
1757 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1758 hammer_flusher_async(ip->hmp, flg);
1761 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1762 flg->seq = hmp->flusher.next++;
1763 if (hmp->next_flush_group == NULL)
1764 hmp->next_flush_group = flg;
1765 if (hmp->fill_flush_group == NULL)
1766 hmp->fill_flush_group = flg;
1767 RB_INIT(&flg->flush_tree);
1768 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1772 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1773 * state we have to put it back into an IDLE state so we can
1774 * drop the extra ref.
1776 * If we have a parent dependancy we must still fall through
1779 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1780 if (ip->flush_state == HAMMER_FST_SETUP &&
1781 TAILQ_EMPTY(&ip->target_list)) {
1782 ip->flush_state = HAMMER_FST_IDLE;
1783 hammer_rel_inode(ip, 0);
1785 if (ip->flush_state == HAMMER_FST_IDLE)
1790 * Our flush action will depend on the current state.
1792 switch(ip->flush_state) {
1793 case HAMMER_FST_IDLE:
1795 * We have no dependancies and can flush immediately. Some
1796 * our children may not be flushable so we have to re-test
1797 * with that additional knowledge.
1799 hammer_flush_inode_core(ip, flg, flags);
1801 case HAMMER_FST_SETUP:
1803 * Recurse upwards through dependancies via target_list
1804 * and start their flusher actions going if possible.
1806 * 'good' is our connectivity. -1 means we have none and
1807 * can't flush, 0 means there weren't any dependancies, and
1808 * 1 means we have good connectivity.
1810 good = hammer_setup_parent_inodes(ip, 0, flg);
1814 * We can continue if good >= 0. Determine how
1815 * many records under our inode can be flushed (and
1818 hammer_flush_inode_core(ip, flg, flags);
1821 * Parent has no connectivity, tell it to flush
1822 * us as soon as it does.
1824 * The REFLUSH flag is also needed to trigger
1825 * dependancy wakeups.
1827 ip->flags |= HAMMER_INODE_CONN_DOWN |
1828 HAMMER_INODE_REFLUSH;
1829 if (flags & HAMMER_FLUSH_SIGNAL) {
1830 ip->flags |= HAMMER_INODE_RESIGNAL;
1831 hammer_flusher_async(ip->hmp, flg);
1835 case HAMMER_FST_FLUSH:
1837 * We are already flushing, flag the inode to reflush
1838 * if needed after it completes its current flush.
1840 * The REFLUSH flag is also needed to trigger
1841 * dependancy wakeups.
1843 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1844 ip->flags |= HAMMER_INODE_REFLUSH;
1845 if (flags & HAMMER_FLUSH_SIGNAL) {
1846 ip->flags |= HAMMER_INODE_RESIGNAL;
1847 hammer_flusher_async(ip->hmp, flg);
1854 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1855 * ip which reference our ip.
1857 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1858 * so for now do not ref/deref the structures. Note that if we use the
1859 * ref/rel code later, the rel CAN block.
1862 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1863 hammer_flush_group_t flg)
1865 hammer_record_t depend;
1870 * If we hit our recursion limit and we have parent dependencies
1871 * We cannot continue. Returning < 0 will cause us to be flagged
1872 * for reflush. Returning -2 cuts off additional dependency checks
1873 * because they are likely to also hit the depth limit.
1875 * We cannot return < 0 if there are no dependencies or there might
1876 * not be anything to wakeup (ip).
1878 if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1879 if (hammer_debug_general & 0x10000)
1880 hkrateprintf(&hammer_gen_krate,
1881 "Warning: depth limit reached on "
1882 "setup recursion, inode %p %016llx\n",
1883 ip, (long long)ip->obj_id);
1891 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1892 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1893 KKASSERT(depend->target_ip == ip);
1894 if (r < 0 && good == 0)
1900 * If we failed due to the recursion depth limit then stop
1910 * This helper function takes a record representing the dependancy between
1911 * the parent inode and child inode.
1913 * record = record in question (*rec in below)
1914 * record->ip = parent inode (*pip in below)
1915 * record->target_ip = child inode (*ip in below)
1917 * *pip--------------\
1920 * \ip /\\\\\ rbtree of recs from parent inode's view
1924 * \------*rec------target_ip------>*ip
1925 * ...target_entry<----...----->target_list<---...
1926 * list of recs from inode's view
1928 * We are asked to recurse upwards and convert the record from SETUP
1929 * to FLUSH if possible.
1931 * Return 1 if the record gives us connectivity
1933 * Return 0 if the record is not relevant
1935 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1938 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1939 hammer_flush_group_t flg)
1944 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1948 * If the record is already flushing, is it in our flush group?
1950 * If it is in our flush group but it is a general record or a
1951 * delete-on-disk, it does not improve our connectivity (return 0),
1952 * and if the target inode is not trying to destroy itself we can't
1953 * allow the operation yet anyway (the second return -1).
1955 if (record->flush_state == HAMMER_FST_FLUSH) {
1957 * If not in our flush group ask the parent to reflush
1958 * us as soon as possible.
1960 if (record->flush_group != flg) {
1961 pip->flags |= HAMMER_INODE_REFLUSH;
1962 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1967 * If in our flush group everything is already set up,
1968 * just return whether the record will improve our
1969 * visibility or not.
1971 if (record->type == HAMMER_MEM_RECORD_ADD)
1977 * It must be a setup record. Try to resolve the setup dependancies
1978 * by recursing upwards so we can place ip on the flush list.
1980 * Limit ourselves to 20 levels of recursion to avoid blowing out
1981 * the kernel stack. If we hit the recursion limit we can't flush
1982 * until the parent flushes. The parent will flush independantly
1983 * on its own and ultimately a deep recursion will be resolved.
1985 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1987 good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1990 * If good < 0 the parent has no connectivity and we cannot safely
1991 * flush the directory entry, which also means we can't flush our
1992 * ip. Flag us for downward recursion once the parent's
1993 * connectivity is resolved. Flag the parent for [re]flush or it
1994 * may not check for downward recursions.
1997 pip->flags |= HAMMER_INODE_REFLUSH;
1998 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2003 * We are go, place the parent inode in a flushing state so we can
2004 * place its record in a flushing state. Note that the parent
2005 * may already be flushing. The record must be in the same flush
2006 * group as the parent.
2008 if (pip->flush_state != HAMMER_FST_FLUSH)
2009 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
2010 KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
2013 * It is possible for a rename to create a loop in the recursion
2014 * and revisit a record. This will result in the record being
2015 * placed in a flush state unexpectedly. This check deals with
2018 if (record->flush_state == HAMMER_FST_FLUSH) {
2019 if (record->type == HAMMER_MEM_RECORD_ADD)
2024 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
2027 if (record->type == HAMMER_MEM_RECORD_DEL &&
2028 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
2030 * Regardless of flushing state we cannot sync this path if the
2031 * record represents a delete-on-disk but the target inode
2032 * is not ready to sync its own deletion.
2034 * XXX need to count effective nlinks to determine whether
2035 * the flush is ok, otherwise removing a hardlink will
2036 * just leave the DEL record to rot.
2038 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
2042 if (pip->flush_group == flg) {
2044 * Because we have not calculated nlinks yet we can just
2045 * set records to the flush state if the parent is in
2046 * the same flush group as we are.
2048 record->flush_state = HAMMER_FST_FLUSH;
2049 record->flush_group = flg;
2050 ++record->flush_group->refs;
2051 hammer_ref(&record->lock);
2054 * A general directory-add contributes to our visibility.
2056 * Otherwise it is probably a directory-delete or
2057 * delete-on-disk record and does not contribute to our
2058 * visbility (but we can still flush it).
2060 if (record->type == HAMMER_MEM_RECORD_ADD)
2065 * If the parent is not in our flush group we cannot
2066 * flush this record yet, there is no visibility.
2067 * We tell the parent to reflush and mark ourselves
2068 * so the parent knows it should flush us too.
2070 pip->flags |= HAMMER_INODE_REFLUSH;
2071 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2077 * This is the core routine placing an inode into the FST_FLUSH state.
2080 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
2082 hammer_mount_t hmp = ip->hmp;
2086 * Set flush state and prevent the flusher from cycling into
2087 * the next flush group. Do not place the ip on the list yet.
2088 * Inodes not in the idle state get an extra reference.
2090 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
2091 if (ip->flush_state == HAMMER_FST_IDLE)
2092 hammer_ref(&ip->lock);
2093 ip->flush_state = HAMMER_FST_FLUSH;
2094 ip->flush_group = flg;
2095 ++hmp->flusher.group_lock;
2096 ++hmp->count_iqueued;
2097 ++hammer_count_iqueued;
2099 hammer_redo_fifo_start_flush(ip);
2103 * We need to be able to vfsync/truncate from the backend.
2105 * XXX Any truncation from the backend will acquire the vnode
2108 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2109 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2110 ip->flags |= HAMMER_INODE_VHELD;
2116 * Figure out how many in-memory records we can actually flush
2117 * (not including inode meta-data, buffers, etc).
2119 KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2120 if (flags & HAMMER_FLUSH_RECURSION) {
2122 * If this is a upwards recursion we do not want to
2123 * recurse down again!
2127 } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2129 * No new records are added if we must complete a flush
2130 * from a previous cycle, but we do have to move the records
2131 * from the previous cycle to the current one.
2134 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2135 hammer_syncgrp_child_callback, NULL);
2141 * Normal flush, scan records and bring them into the flush.
2142 * Directory adds and deletes are usually skipped (they are
2143 * grouped with the related inode rather then with the
2146 * go_count can be negative, which means the scan aborted
2147 * due to the flush group being over-full and we should
2148 * flush what we have.
2150 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2151 hammer_setup_child_callback, NULL);
2155 * This is a more involved test that includes go_count. If we
2156 * can't flush, flag the inode and return. If go_count is 0 we
2157 * were are unable to flush any records in our rec_tree and
2158 * must ignore the XDIRTY flag.
2160 if (go_count == 0) {
2161 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2162 --hmp->count_iqueued;
2163 --hammer_count_iqueued;
2166 ip->flush_state = HAMMER_FST_SETUP;
2167 ip->flush_group = NULL;
2168 if (flags & HAMMER_FLUSH_SIGNAL) {
2169 ip->flags |= HAMMER_INODE_REFLUSH |
2170 HAMMER_INODE_RESIGNAL;
2172 ip->flags |= HAMMER_INODE_REFLUSH;
2175 if (ip->flags & HAMMER_INODE_VHELD) {
2176 ip->flags &= ~HAMMER_INODE_VHELD;
2182 * REFLUSH is needed to trigger dependancy wakeups
2183 * when an inode is in SETUP.
2185 ip->flags |= HAMMER_INODE_REFLUSH;
2186 if (--hmp->flusher.group_lock == 0)
2187 wakeup(&hmp->flusher.group_lock);
2193 * Snapshot the state of the inode for the backend flusher.
2195 * We continue to retain save_trunc_off even when all truncations
2196 * have been resolved as an optimization to determine if we can
2197 * skip the B-Tree lookup for overwrite deletions.
2199 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2200 * and stays in ip->flags. Once set, it stays set until the
2201 * inode is destroyed.
2203 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2204 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2205 ip->sync_trunc_off = ip->trunc_off;
2206 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2207 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2208 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2211 * The save_trunc_off used to cache whether the B-Tree
2212 * holds any records past that point is not used until
2213 * after the truncation has succeeded, so we can safely
2216 if (ip->save_trunc_off > ip->sync_trunc_off)
2217 ip->save_trunc_off = ip->sync_trunc_off;
2219 ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2220 ~HAMMER_INODE_TRUNCATED);
2221 ip->sync_ino_leaf = ip->ino_leaf;
2222 ip->sync_ino_data = ip->ino_data;
2223 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2226 * The flusher list inherits our inode and reference.
2228 KKASSERT(flg->running == 0);
2229 RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2230 if (--hmp->flusher.group_lock == 0)
2231 wakeup(&hmp->flusher.group_lock);
2234 * Auto-flush the group if it grows too large. Make sure the
2235 * inode reclaim wait pipeline continues to work.
2237 if (flg->total_count >= hammer_autoflush ||
2238 flg->total_count >= hammer_limit_reclaims / 4) {
2239 if (hmp->fill_flush_group == flg)
2240 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2241 hammer_flusher_async(hmp, flg);
2246 * Callback for scan of ip->rec_tree. Try to include each record in our
2247 * flush. ip->flush_group has been set but the inode has not yet been
2248 * moved into a flushing state.
2250 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2253 * We return 1 for any record placed or found in FST_FLUSH, which prevents
2254 * the caller from shortcutting the flush.
2257 hammer_setup_child_callback(hammer_record_t rec, void *data)
2259 hammer_flush_group_t flg;
2260 hammer_inode_t target_ip;
2265 * Records deleted or committed by the backend are ignored.
2266 * Note that the flush detects deleted frontend records at
2267 * multiple points to deal with races. This is just the first
2268 * line of defense. The only time HAMMER_RECF_DELETED_FE cannot
2269 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2270 * messes up link-count calculations.
2272 * NOTE: Don't get confused between record deletion and, say,
2273 * directory entry deletion. The deletion of a directory entry
2274 * which is on-media has nothing to do with the record deletion
2277 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2278 HAMMER_RECF_COMMITTED)) {
2279 if (rec->flush_state == HAMMER_FST_FLUSH) {
2280 KKASSERT(rec->flush_group == rec->ip->flush_group);
2289 * If the record is in an idle state it has no dependancies and
2293 flg = ip->flush_group;
2296 switch(rec->flush_state) {
2297 case HAMMER_FST_IDLE:
2299 * The record has no setup dependancy, we can flush it.
2301 KKASSERT(rec->target_ip == NULL);
2302 rec->flush_state = HAMMER_FST_FLUSH;
2303 rec->flush_group = flg;
2305 hammer_ref(&rec->lock);
2308 case HAMMER_FST_SETUP:
2310 * The record has a setup dependancy. These are typically
2311 * directory entry adds and deletes. Such entries will be
2312 * flushed when their inodes are flushed so we do not
2313 * usually have to add them to the flush here. However,
2314 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2315 * it is asking us to flush this record (and it).
2317 target_ip = rec->target_ip;
2318 KKASSERT(target_ip != NULL);
2319 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2322 * If the target IP is already flushing in our group
2323 * we could associate the record, but target_ip has
2324 * already synced ino_data to sync_ino_data and we
2325 * would also have to adjust nlinks. Plus there are
2326 * ordering issues for adds and deletes.
2328 * Reflush downward if this is an ADD, and upward if
2331 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2332 if (rec->type == HAMMER_MEM_RECORD_ADD)
2333 ip->flags |= HAMMER_INODE_REFLUSH;
2335 target_ip->flags |= HAMMER_INODE_REFLUSH;
2340 * Target IP is not yet flushing. This can get complex
2341 * because we have to be careful about the recursion.
2343 * Directories create an issue for us in that if a flush
2344 * of a directory is requested the expectation is to flush
2345 * any pending directory entries, but this will cause the
2346 * related inodes to recursively flush as well. We can't
2347 * really defer the operation so just get as many as we
2351 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2352 (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2354 * We aren't reclaiming and the target ip was not
2355 * previously prevented from flushing due to this
2356 * record dependancy. Do not flush this record.
2361 if (flg->total_count + flg->refs >
2362 ip->hmp->undo_rec_limit) {
2364 * Our flush group is over-full and we risk blowing
2365 * out the UNDO FIFO. Stop the scan, flush what we
2366 * have, then reflush the directory.
2368 * The directory may be forced through multiple
2369 * flush groups before it can be completely
2372 ip->flags |= HAMMER_INODE_RESIGNAL |
2373 HAMMER_INODE_REFLUSH;
2375 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2377 * If the target IP is not flushing we can force
2378 * it to flush, even if it is unable to write out
2379 * any of its own records we have at least one in
2380 * hand that we CAN deal with.
2382 rec->flush_state = HAMMER_FST_FLUSH;
2383 rec->flush_group = flg;
2385 hammer_ref(&rec->lock);
2386 hammer_flush_inode_core(target_ip, flg,
2387 HAMMER_FLUSH_RECURSION);
2391 * General or delete-on-disk record.
2393 * XXX this needs help. If a delete-on-disk we could
2394 * disconnect the target. If the target has its own
2395 * dependancies they really need to be flushed.
2399 rec->flush_state = HAMMER_FST_FLUSH;
2400 rec->flush_group = flg;
2402 hammer_ref(&rec->lock);
2403 hammer_flush_inode_core(target_ip, flg,
2404 HAMMER_FLUSH_RECURSION);
2408 case HAMMER_FST_FLUSH:
2410 * The record could be part of a previous flush group if the
2411 * inode is a directory (the record being a directory entry).
2412 * Once the flush group was closed a hammer_test_inode()
2413 * function can cause a new flush group to be setup, placing
2414 * the directory inode itself in a new flush group.
2416 * When associated with a previous flush group we count it
2417 * as if it were in our current flush group, since it will
2418 * effectively be flushed by the time we flush our current
2422 rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2423 rec->flush_group == flg);
2432 * This version just moves records already in a flush state to the new
2433 * flush group and that is it.
2436 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2438 hammer_inode_t ip = rec->ip;
2440 switch(rec->flush_state) {
2441 case HAMMER_FST_FLUSH:
2442 KKASSERT(rec->flush_group == ip->flush_group);
2452 * Wait for a previously queued flush to complete.
2454 * If a critical error occured we don't try to wait.
2457 hammer_wait_inode(hammer_inode_t ip)
2460 * The inode can be in a SETUP state in which case RESIGNAL
2461 * should be set. If RESIGNAL is not set then the previous
2462 * flush completed and a later operation placed the inode
2463 * in a passive setup state again, so we're done.
2465 * The inode can be in a FLUSH state in which case we
2466 * can just wait for completion.
2468 while (ip->flush_state == HAMMER_FST_FLUSH ||
2469 (ip->flush_state == HAMMER_FST_SETUP &&
2470 (ip->flags & HAMMER_INODE_RESIGNAL))) {
2472 * Don't try to flush on a critical error
2474 if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
2478 * If the inode was already being flushed its flg
2479 * may not have been queued to the backend. We
2480 * have to make sure it gets queued or we can wind
2481 * up blocked or deadlocked (particularly if we are
2482 * the vnlru thread).
2484 if (ip->flush_state == HAMMER_FST_FLUSH) {
2485 KKASSERT(ip->flush_group);
2486 if (ip->flush_group->closed == 0) {
2487 if (hammer_debug_inode) {
2488 hkprintf("debug: forcing "
2489 "async flush ip %016jx\n",
2490 (intmax_t)ip->obj_id);
2492 hammer_flusher_async(ip->hmp, ip->flush_group);
2493 continue; /* retest */
2498 * In a flush state with the flg queued to the backend
2499 * or in a setup state with RESIGNAL set, we can safely
2502 ip->flags |= HAMMER_INODE_FLUSHW;
2503 tsleep(&ip->flags, 0, "hmrwin", 0);
2508 * The inode may have been in a passive setup state,
2509 * call flush to make sure we get signaled.
2511 if (ip->flush_state == HAMMER_FST_SETUP)
2512 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2518 * Called by the backend code when a flush has been completed.
2519 * The inode has already been removed from the flush list.
2521 * A pipelined flush can occur, in which case we must re-enter the
2522 * inode on the list and re-copy its fields.
2525 hammer_flush_inode_done(hammer_inode_t ip, int error)
2530 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2535 * Auto-reflush if the backend could not completely flush
2536 * the inode. This fixes a case where a deferred buffer flush
2537 * could cause fsync to return early.
2539 if (ip->sync_flags & HAMMER_INODE_MODMASK)
2540 ip->flags |= HAMMER_INODE_REFLUSH;
2543 * Merge left-over flags back into the frontend and fix the state.
2544 * Incomplete truncations are retained by the backend.
2547 ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2548 ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2551 * The backend may have adjusted nlinks, so if the adjusted nlinks
2552 * does not match the fronttend set the frontend's DDIRTY flag again.
2554 if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2555 ip->flags |= HAMMER_INODE_DDIRTY;
2558 * Fix up the dirty buffer status.
2560 if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2561 ip->flags |= HAMMER_INODE_BUFS;
2563 hammer_redo_fifo_end_flush(ip);
2566 * Re-set the XDIRTY flag if some of the inode's in-memory records
2567 * could not be flushed.
2569 KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2570 (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2571 (!RB_EMPTY(&ip->rec_tree) &&
2572 (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2575 * Do not lose track of inodes which no longer have vnode
2576 * assocations, otherwise they may never get flushed again.
2578 * The reflush flag can be set superfluously, causing extra pain
2579 * for no reason. If the inode is no longer modified it no longer
2580 * needs to be flushed.
2582 if (ip->flags & HAMMER_INODE_MODMASK) {
2584 ip->flags |= HAMMER_INODE_REFLUSH;
2586 ip->flags &= ~HAMMER_INODE_REFLUSH;
2590 * The fs token is held but the inode lock is not held. Because this
2591 * is a backend flush it is possible that the vnode has no references
2592 * and cause a reclaim race inside vsetisdirty() if/when it blocks.
2594 * Therefore, we must lock the inode around this particular dirtying
2595 * operation. We don't have to around other dirtying operations
2596 * where the vnode is implicitly or explicitly held.
2598 if (ip->flags & HAMMER_INODE_MODMASK) {
2599 hammer_lock_ex(&ip->lock);
2600 hammer_inode_dirty(ip);
2601 hammer_unlock(&ip->lock);
2605 * Adjust the flush state.
2607 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2609 * We were unable to flush out all our records, leave the
2610 * inode in a flush state and in the current flush group.
2611 * The flush group will be re-run.
2613 * This occurs if the UNDO block gets too full or there is
2614 * too much dirty meta-data and allows the flusher to
2615 * finalize the UNDO block and then re-flush.
2617 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2621 * Remove from the flush_group
2623 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2624 ip->flush_group = NULL;
2628 * Clean up the vnode ref and tracking counts.
2630 if (ip->flags & HAMMER_INODE_VHELD) {
2631 ip->flags &= ~HAMMER_INODE_VHELD;
2635 --hmp->count_iqueued;
2636 --hammer_count_iqueued;
2639 * And adjust the state.
2641 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2642 ip->flush_state = HAMMER_FST_IDLE;
2645 ip->flush_state = HAMMER_FST_SETUP;
2650 * If the frontend is waiting for a flush to complete,
2653 if (ip->flags & HAMMER_INODE_FLUSHW) {
2654 ip->flags &= ~HAMMER_INODE_FLUSHW;
2659 * If the frontend made more changes and requested another
2660 * flush, then try to get it running.
2662 * Reflushes are aborted when the inode is errored out.
2664 if (ip->flags & HAMMER_INODE_REFLUSH) {
2665 ip->flags &= ~HAMMER_INODE_REFLUSH;
2666 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2667 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2668 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2670 hammer_flush_inode(ip, 0);
2676 * If we have no parent dependancies we can clear CONN_DOWN
2678 if (TAILQ_EMPTY(&ip->target_list))
2679 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2682 * If the inode is now clean drop the space reservation.
2684 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2685 (ip->flags & HAMMER_INODE_RSV_INODES)) {
2686 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2690 ip->flags &= ~HAMMER_INODE_SLAVEFLUSH;
2693 hammer_rel_inode(ip, 0);
2697 * Called from hammer_sync_inode() to synchronize in-memory records
2701 hammer_sync_record_callback(hammer_record_t record, void *data)
2703 hammer_cursor_t cursor = data;
2704 hammer_transaction_t trans = cursor->trans;
2705 hammer_mount_t hmp = trans->hmp;
2709 * Skip records that do not belong to the current flush.
2711 ++hammer_stats_record_iterations;
2712 if (record->flush_state != HAMMER_FST_FLUSH)
2715 if (record->flush_group != record->ip->flush_group) {
2716 hdkprintf("rec %p ip %p bad flush group %p %p\n",
2719 record->flush_group,
2720 record->ip->flush_group);
2721 if (hammer_debug_critical)
2725 KKASSERT(record->flush_group == record->ip->flush_group);
2728 * Interlock the record using the BE flag. Once BE is set the
2729 * frontend cannot change the state of FE.
2731 * NOTE: If FE is set prior to us setting BE we still sync the
2732 * record out, but the flush completion code converts it to
2733 * a delete-on-disk record instead of destroying it.
2735 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2736 record->flags |= HAMMER_RECF_INTERLOCK_BE;
2739 * The backend has already disposed of the record.
2741 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2747 * If the whole inode is being deleted and all on-disk records will
2748 * be deleted very soon, we can't sync any new records to disk
2749 * because they will be deleted in the same transaction they were
2750 * created in (delete_tid == create_tid), which will assert.
2752 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2753 * that we currently panic on.
2755 if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2756 switch(record->type) {
2757 case HAMMER_MEM_RECORD_DATA:
2759 * We don't have to do anything, if the record was
2760 * committed the space will have been accounted for
2764 case HAMMER_MEM_RECORD_GENERAL:
2766 * Set deleted-by-backend flag. Do not set the
2767 * backend committed flag, because we are throwing
2770 record->flags |= HAMMER_RECF_DELETED_BE;
2771 ++record->ip->rec_generation;
2774 case HAMMER_MEM_RECORD_ADD:
2775 hpanic("illegal add during inode deletion record %p",
2777 break; /* NOT REACHED */
2778 case HAMMER_MEM_RECORD_INODE:
2779 hpanic("attempt to sync inode record %p?", record);
2780 break; /* NOT REACHED */
2781 case HAMMER_MEM_RECORD_DEL:
2783 * Follow through and issue the on-disk deletion
2790 * If DELETED_FE is set special handling is needed for directory
2791 * entries. Dependant pieces related to the directory entry may
2792 * have already been synced to disk. If this occurs we have to
2793 * sync the directory entry and then change the in-memory record
2794 * from an ADD to a DELETE to cover the fact that it's been
2795 * deleted by the frontend.
2797 * A directory delete covering record (MEM_RECORD_DEL) can never
2798 * be deleted by the frontend.
2800 * Any other record type (aka DATA) can be deleted by the frontend.
2801 * XXX At the moment the flusher must skip it because there may
2802 * be another data record in the flush group for the same block,
2803 * meaning that some frontend data changes can leak into the backend's
2804 * synchronization point.
2806 if (record->flags & HAMMER_RECF_DELETED_FE) {
2807 if (record->type == HAMMER_MEM_RECORD_ADD) {
2809 * Convert a front-end deleted directory-add to
2810 * a directory-delete entry later.
2812 record->flags |= HAMMER_RECF_CONVERT_DELETE;
2815 * Dispose of the record (race case). Mark as
2816 * deleted by backend (and not committed).
2818 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2819 record->flags |= HAMMER_RECF_DELETED_BE;
2820 ++record->ip->rec_generation;
2827 * Assign the create_tid for new records. Deletions already
2828 * have the record's entire key properly set up.
2830 if (record->type != HAMMER_MEM_RECORD_DEL) {
2831 record->leaf.base.create_tid = trans->tid;
2832 record->leaf.create_ts = trans->time32;
2836 * This actually moves the record to the on-media B-Tree. We
2837 * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2838 * indicating that the related REDO_WRITE(s) have been committed.
2840 * During recovery any REDO_TERM's within the nominal recovery span
2841 * are ignored since the related meta-data is being undone, causing
2842 * any matching REDO_WRITEs to execute. The REDO_TERMs outside
2843 * the nominal recovery span will match against REDO_WRITEs and
2844 * prevent them from being executed (because the meta-data has
2845 * already been synchronized).
2847 if (record->flags & HAMMER_RECF_REDO) {
2848 KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2849 hammer_generate_redo(trans, record->ip,
2850 record->leaf.base.key -
2851 record->leaf.data_len,
2852 HAMMER_REDO_TERM_WRITE,
2854 record->leaf.data_len);
2858 error = hammer_ip_sync_record_cursor(cursor, record);
2859 if (error != EDEADLK)
2861 hammer_done_cursor(cursor);
2862 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2867 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2872 hammer_flush_record_done(record, error);
2875 * Do partial finalization if we have built up too many dirty
2876 * buffers. Otherwise a buffer cache deadlock can occur when
2877 * doing things like creating tens of thousands of tiny files.
2879 * We must release our cursor lock to avoid a 3-way deadlock
2880 * due to the exclusive sync lock the finalizer must get.
2882 * WARNING: See warnings in hammer_unlock_cursor() function.
2884 if (hammer_flusher_meta_limit(hmp) ||
2885 vm_page_count_severe()) {
2886 hammer_unlock_cursor(cursor);
2887 hammer_flusher_finalize(trans, 0);
2888 hammer_lock_cursor(cursor);
2894 * Backend function called by the flusher to sync an inode to media.
2897 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2899 struct hammer_cursor cursor;
2900 hammer_node_t tmp_node;
2901 hammer_record_t depend;
2902 hammer_record_t next;
2903 int error, tmp_error;
2906 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2909 error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2914 * Any directory records referencing this inode which are not in
2915 * our current flush group must adjust our nlink count for the
2916 * purposes of synchronizating to disk.
2918 * Records which are in our flush group can be unlinked from our
2919 * inode now, potentially allowing the inode to be physically
2922 * This cannot block.
2924 nlinks = ip->ino_data.nlinks;
2925 next = TAILQ_FIRST(&ip->target_list);
2926 while ((depend = next) != NULL) {
2927 next = TAILQ_NEXT(depend, target_entry);
2928 if (depend->flush_state == HAMMER_FST_FLUSH &&
2929 depend->flush_group == ip->flush_group) {
2931 * If this is an ADD that was deleted by the frontend
2932 * the frontend nlinks count will have already been
2933 * decremented, but the backend is going to sync its
2934 * directory entry and must account for it. The
2935 * record will be converted to a delete-on-disk when
2938 * If the ADD was not deleted by the frontend we
2939 * can remove the dependancy from our target_list.
2941 if (depend->flags & HAMMER_RECF_DELETED_FE) {
2944 TAILQ_REMOVE(&ip->target_list, depend,
2946 depend->target_ip = NULL;
2948 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2950 * Not part of our flush group and not deleted by
2951 * the front-end, adjust the link count synced to
2952 * the media (undo what the frontend did when it
2953 * queued the record).
2955 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2956 switch(depend->type) {
2957 case HAMMER_MEM_RECORD_ADD:
2960 case HAMMER_MEM_RECORD_DEL:
2970 * Set dirty if we had to modify the link count.
2972 if (ip->sync_ino_data.nlinks != nlinks) {
2973 KKASSERT((int64_t)nlinks >= 0);
2974 ip->sync_ino_data.nlinks = nlinks;
2975 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2979 * If there is a trunction queued destroy any data past the (aligned)
2980 * truncation point. Userland will have dealt with the buffer
2981 * containing the truncation point for us.
2983 * We don't flush pending frontend data buffers until after we've
2984 * dealt with the truncation.
2986 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2988 * Interlock trunc_off. The VOP front-end may continue to
2989 * make adjustments to it while we are blocked.
2992 off_t aligned_trunc_off;
2995 trunc_off = ip->sync_trunc_off;
2996 blkmask = hammer_blocksize(trunc_off) - 1;
2997 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
3000 * Delete any whole blocks on-media. The front-end has
3001 * already cleaned out any partial block and made it
3002 * pending. The front-end may have updated trunc_off
3003 * while we were blocked so we only use sync_trunc_off.
3005 * This operation can blow out the buffer cache, EWOULDBLOCK
3006 * means we were unable to complete the deletion. The
3007 * deletion will update sync_trunc_off in that case.
3009 error = hammer_ip_delete_range(&cursor, ip,
3011 0x7FFFFFFFFFFFFFFFLL, 2);
3012 if (error == EWOULDBLOCK) {
3013 ip->flags |= HAMMER_INODE_WOULDBLOCK;
3015 goto defer_buffer_flush;
3022 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
3024 * XXX we do this even if we did not previously generate
3025 * a REDO_TRUNC record. This operation may enclosed the
3026 * range for multiple prior truncation entries in the REDO
3029 if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
3030 (ip->flags & HAMMER_INODE_RDIRTY)) {
3031 hammer_generate_redo(trans, ip, aligned_trunc_off,
3032 HAMMER_REDO_TERM_TRUNC,
3037 * Clear the truncation flag on the backend after we have
3038 * completed the deletions. Backend data is now good again
3039 * (including new records we are about to sync, below).
3041 * Leave sync_trunc_off intact. As we write additional
3042 * records the backend will update sync_trunc_off. This
3043 * tells the backend whether it can skip the overwrite
3044 * test. This should work properly even when the backend
3045 * writes full blocks where the truncation point straddles
3046 * the block because the comparison is against the base
3047 * offset of the record.
3049 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3050 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
3056 * Now sync related records. These will typically be directory
3057 * entries, records tracking direct-writes, or delete-on-disk records.
3060 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
3061 hammer_sync_record_callback, &cursor);
3067 hammer_cache_node(&ip->cache[1], cursor.node);
3070 * Re-seek for inode update, assuming our cache hasn't been ripped
3071 * out from under us.
3074 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
3076 hammer_cursor_downgrade(&cursor);
3077 hammer_lock_sh(&tmp_node->lock);
3078 if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
3079 hammer_cursor_seek(&cursor, tmp_node, 0);
3080 hammer_unlock(&tmp_node->lock);
3081 hammer_rel_node(tmp_node);
3087 * If we are deleting the inode the frontend had better not have
3088 * any active references on elements making up the inode.
3090 * The call to hammer_ip_delete_clean() cleans up auxillary records
3091 * but not DB or DATA records. Those must have already been deleted
3092 * by the normal truncation mechanic.
3094 if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
3095 RB_EMPTY(&ip->rec_tree) &&
3096 (ip->sync_flags & HAMMER_INODE_DELETING) &&
3097 (ip->flags & HAMMER_INODE_DELETED) == 0) {
3100 error = hammer_ip_delete_clean(&cursor, ip, &count1);
3102 ip->flags |= HAMMER_INODE_DELETED;
3103 ip->sync_flags &= ~HAMMER_INODE_DELETING;
3104 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3105 KKASSERT(RB_EMPTY(&ip->rec_tree));
3108 * Set delete_tid in both the frontend and backend
3109 * copy of the inode record. The DELETED flag handles
3110 * this, do not set DDIRTY.
3112 ip->ino_leaf.base.delete_tid = trans->tid;
3113 ip->sync_ino_leaf.base.delete_tid = trans->tid;
3114 ip->ino_leaf.delete_ts = trans->time32;
3115 ip->sync_ino_leaf.delete_ts = trans->time32;
3119 * Adjust the inode count in the volume header
3121 hammer_sync_lock_sh(trans);
3122 if (ip->flags & HAMMER_INODE_ONDISK) {
3123 hammer_modify_volume_field(trans,
3126 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
3127 hammer_modify_volume_done(trans->rootvol);
3129 hammer_sync_unlock(trans);
3135 ip->sync_flags &= ~HAMMER_INODE_BUFS;
3139 * Now update the inode's on-disk inode-data and/or on-disk record.
3140 * DELETED and ONDISK are managed only in ip->flags.
3142 * In the case of a defered buffer flush we still update the on-disk
3143 * inode to satisfy visibility requirements if there happen to be
3144 * directory dependancies.
3146 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
3147 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3149 * If deleted and on-disk, don't set any additional flags.
3150 * the delete flag takes care of things.
3152 * Clear flags which may have been set by the frontend.
3154 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3155 HAMMER_INODE_SDIRTY |
3156 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3157 HAMMER_INODE_DELETING);
3159 case HAMMER_INODE_DELETED:
3161 * Take care of the case where a deleted inode was never
3162 * flushed to the disk in the first place.
3164 * Clear flags which may have been set by the frontend.
3166 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3167 HAMMER_INODE_SDIRTY |
3168 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3169 HAMMER_INODE_DELETING);
3170 while (RB_ROOT(&ip->rec_tree)) {
3171 hammer_record_t record = RB_ROOT(&ip->rec_tree);
3172 hammer_ref(&record->lock);
3173 KKASSERT(hammer_oneref(&record->lock));
3174 record->flags |= HAMMER_RECF_DELETED_BE;
3175 ++record->ip->rec_generation;
3176 hammer_rel_mem_record(record);
3179 case HAMMER_INODE_ONDISK:
3181 * If already on-disk, do not set any additional flags.
3186 * If not on-disk and not deleted, set DDIRTY to force
3187 * an initial record to be written.
3189 * Also set the create_tid in both the frontend and backend
3190 * copy of the inode record.
3192 ip->ino_leaf.base.create_tid = trans->tid;
3193 ip->ino_leaf.create_ts = trans->time32;
3194 ip->sync_ino_leaf.base.create_tid = trans->tid;
3195 ip->sync_ino_leaf.create_ts = trans->time32;
3196 ip->sync_flags |= HAMMER_INODE_DDIRTY;
3201 * If DDIRTY or SDIRTY is set, write out a new record.
3202 * If the inode is already on-disk the old record is marked as
3205 * If DELETED is set hammer_update_inode() will delete the existing
3206 * record without writing out a new one.
3208 if (ip->flags & HAMMER_INODE_DELETED) {
3209 error = hammer_update_inode(&cursor, ip);
3211 if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3212 (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3213 error = hammer_update_itimes(&cursor, ip);
3215 if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3216 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3217 error = hammer_update_inode(&cursor, ip);
3220 if (ip->flags & HAMMER_INODE_MODMASK)
3221 hammer_inode_dirty(ip);
3223 hammer_critical_error(ip->hmp, ip, error,
3224 "while syncing inode");
3226 hammer_done_cursor(&cursor);
3231 * This routine is called when the OS is no longer actively referencing
3232 * the inode (but might still be keeping it cached), or when releasing
3233 * the last reference to an inode.
3235 * At this point if the inode's nlinks count is zero we want to destroy
3236 * it, which may mean destroying it on-media too.
3239 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3244 * Set the DELETING flag when the link count drops to 0 and the
3245 * OS no longer has any opens on the inode.
3247 * The backend will clear DELETING (a mod flag) and set DELETED
3248 * (a state flag) when it is actually able to perform the
3251 * Don't reflag the deletion if the flusher is currently syncing
3252 * one that was already flagged. A previously set DELETING flag
3253 * may bounce around flags and sync_flags until the operation is
3256 * Do not attempt to modify a snapshot inode (one set to read-only).
3258 if (ip->ino_data.nlinks == 0 &&
3259 ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3260 ip->flags |= HAMMER_INODE_DELETING;
3261 ip->flags |= HAMMER_INODE_TRUNCATED;
3265 if (hammer_get_vnode(ip, &vp) != 0)
3273 nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0, 0);
3274 if (ip->flags & HAMMER_INODE_MODMASK)
3275 hammer_inode_dirty(ip);
3282 * After potentially resolving a dependancy the inode is tested
3283 * to determine whether it needs to be reflushed.
3286 hammer_test_inode(hammer_inode_t ip)
3288 if (ip->flags & HAMMER_INODE_REFLUSH) {
3289 ip->flags &= ~HAMMER_INODE_REFLUSH;
3290 hammer_ref(&ip->lock);
3291 if (ip->flags & HAMMER_INODE_RESIGNAL) {
3292 ip->flags &= ~HAMMER_INODE_RESIGNAL;
3293 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3295 hammer_flush_inode(ip, 0);
3297 hammer_rel_inode(ip, 0);
3302 * Clear the RECLAIM flag on an inode. This occurs when the inode is
3303 * reassociated with a vp or just before it gets freed.
3305 * Pipeline wakeups to threads blocked due to an excessive number of
3306 * detached inodes. This typically occurs when atime updates accumulate
3307 * while scanning a directory tree.
3310 hammer_inode_wakereclaims(hammer_inode_t ip)
3312 struct hammer_reclaim *reclaim;
3313 hammer_mount_t hmp = ip->hmp;
3315 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3318 --hammer_count_reclaims;
3319 --hmp->count_reclaims;
3320 ip->flags &= ~HAMMER_INODE_RECLAIM;
3322 if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3323 KKASSERT(reclaim->count > 0);
3324 if (--reclaim->count == 0) {
3325 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3332 * Setup our reclaim pipeline. We only let so many detached (and dirty)
3333 * inodes build up before we start blocking. This routine is called
3334 * if a new inode is created or an inode is loaded from media.
3336 * When we block we don't care *which* inode has finished reclaiming,
3337 * as long as one does.
3339 * The reclaim pipeline is primarily governed by the auto-flush which is
3340 * 1/4 hammer_limit_reclaims. We don't want to block if the count is
3341 * less than 1/2 hammer_limit_reclaims. From 1/2 to full count is
3342 * dynamically governed.
3345 hammer_inode_waitreclaims(hammer_transaction_t trans)
3347 hammer_mount_t hmp = trans->hmp;
3348 struct hammer_reclaim reclaim;
3352 * Track inode load, delay if the number of reclaiming inodes is
3353 * between 2/4 and 4/4 hammer_limit_reclaims, depending.
3355 if (curthread->td_proc) {
3356 struct hammer_inostats *stats;
3358 stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3361 if (stats->count > hammer_limit_reclaims / 2)
3362 stats->count = hammer_limit_reclaims / 2;
3363 lower_limit = hammer_limit_reclaims - stats->count;
3364 if (hammer_debug_general & 0x10000) {
3365 hdkprintf("pid %5d limit %d\n",
3366 (int)curthread->td_proc->p_pid, lower_limit);
3369 lower_limit = hammer_limit_reclaims * 3 / 4;
3371 if (hmp->count_reclaims >= lower_limit) {
3373 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3374 tsleep(&reclaim, 0, "hmrrcm", hz);
3375 if (reclaim.count > 0)
3376 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3381 * Keep track of reclaim statistics on a per-pid basis using a loose
3382 * 4-way set associative hash table. Collisions inherit the count of
3383 * the previous entry.
3385 * NOTE: We want to be careful here to limit the chain size. If the chain
3386 * size is too large a pid will spread its stats out over too many
3387 * entries under certain types of heavy filesystem activity and
3388 * wind up not delaying long enough.
3391 struct hammer_inostats *
3392 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3394 struct hammer_inostats *stats;
3397 static volatile int iterator; /* we don't care about MP races */
3400 * Chain up to 4 times to find our entry.
3402 for (chain = 0; chain < 4; ++chain) {
3403 stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3404 if (stats->pid == pid)
3409 * Replace one of the four chaining entries with our new entry.
3412 stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3413 HAMMER_INOSTATS_HMASK];
3420 if (stats->count && stats->ltick != ticks) {
3421 delta = ticks - stats->ltick;
3422 stats->ltick = ticks;
3423 if (delta <= 0 || delta > hz * 60)
3426 stats->count = stats->count * hz / (hz + delta);
3428 if (hammer_debug_general & 0x10000)
3429 hdkprintf("pid %5d stats %d\n", (int)pid, stats->count);
3436 * XXX not used, doesn't work very well due to the large batching nature
3439 * A larger then normal backlog of inodes is sitting in the flusher,
3440 * enforce a general slowdown to let it catch up. This routine is only
3441 * called on completion of a non-flusher-related transaction which
3442 * performed B-Tree node I/O.
3444 * It is possible for the flusher to stall in a continuous load.
3445 * blogbench -i1000 -o seems to do a good job generating this sort of load.
3446 * If the flusher is unable to catch up the inode count can bloat until
3447 * we run out of kvm.
3449 * This is a bit of a hack.
3452 hammer_inode_waithard(hammer_mount_t hmp)
3457 if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3458 if (hmp->count_reclaims < hammer_limit_reclaims / 2 &&
3459 hmp->count_iqueued < hmp->count_inodes / 20) {
3460 hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3464 if (hmp->count_reclaims < hammer_limit_reclaims ||
3465 hmp->count_iqueued < hmp->count_inodes / 10) {
3468 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3472 * Block for one flush cycle.
3474 hammer_flusher_wait_next(hmp);