2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.114 2008/09/24 00:53:51 dillon Exp $
38 #include <vm/vm_extern.h>
40 static int hammer_unload_inode(struct hammer_inode *ip);
41 static void hammer_free_inode(hammer_inode_t ip);
42 static void hammer_flush_inode_core(hammer_inode_t ip,
43 hammer_flush_group_t flg, int flags);
44 static int hammer_setup_child_callback(hammer_record_t rec, void *data);
46 static int hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
48 static int hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
49 hammer_flush_group_t flg);
50 static int hammer_setup_parent_inodes_helper(hammer_record_t record,
51 int depth, hammer_flush_group_t flg);
52 static void hammer_inode_wakereclaims(hammer_inode_t ip);
53 static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
57 extern struct hammer_inode *HammerTruncIp;
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 u_int32_t localization = *(u_int32_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, u_int32_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);
245 hammer_rel_inode(ip, 1);
246 lwkt_reltoken(&hmp->fs_token);
252 * Return a locked vnode for the specified inode. The inode must be
253 * referenced but NOT LOCKED on entry and will remain referenced on
256 * Called from the frontend.
259 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
269 if ((vp = ip->vp) == NULL) {
270 error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
273 hammer_lock_ex(&ip->lock);
274 if (ip->vp != NULL) {
275 hammer_unlock(&ip->lock);
281 hammer_ref(&ip->lock);
285 obj_type = ip->ino_data.obj_type;
286 vp->v_type = hammer_get_vnode_type(obj_type);
288 hammer_inode_wakereclaims(ip);
290 switch(ip->ino_data.obj_type) {
291 case HAMMER_OBJTYPE_CDEV:
292 case HAMMER_OBJTYPE_BDEV:
293 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
294 addaliasu(vp, ip->ino_data.rmajor,
295 ip->ino_data.rminor);
297 case HAMMER_OBJTYPE_FIFO:
298 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
300 case HAMMER_OBJTYPE_REGFILE:
307 * Only mark as the root vnode if the ip is not
308 * historical, otherwise the VFS cache will get
309 * confused. The other half of the special handling
310 * is in hammer_vop_nlookupdotdot().
312 * Pseudo-filesystem roots can be accessed via
313 * non-root filesystem paths and setting VROOT may
314 * confuse the namecache. Set VPFSROOT instead.
316 if (ip->obj_id == HAMMER_OBJID_ROOT &&
317 ip->obj_asof == hmp->asof) {
318 if (ip->obj_localization == 0)
319 vsetflags(vp, VROOT);
321 vsetflags(vp, VPFSROOT);
324 vp->v_data = (void *)ip;
325 /* vnode locked by getnewvnode() */
326 /* make related vnode dirty if inode dirty? */
327 hammer_unlock(&ip->lock);
328 if (vp->v_type == VREG) {
329 vinitvmio(vp, ip->ino_data.size,
330 hammer_blocksize(ip->ino_data.size),
331 hammer_blockoff(ip->ino_data.size));
337 * Interlock vnode clearing. This does not prevent the
338 * vnode from going into a reclaimed state but it does
339 * prevent it from being destroyed or reused so the vget()
340 * will properly fail.
342 hammer_lock_ex(&ip->lock);
343 if ((vp = ip->vp) == NULL) {
344 hammer_unlock(&ip->lock);
347 vhold_interlocked(vp);
348 hammer_unlock(&ip->lock);
351 * loop if the vget fails (aka races), or if the vp
352 * no longer matches ip->vp.
354 if (vget(vp, LK_EXCLUSIVE) == 0) {
368 * Locate all copies of the inode for obj_id compatible with the specified
369 * asof, reference, and issue the related call-back. This routine is used
370 * for direct-io invalidation and does not create any new inodes.
373 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
374 int (*callback)(hammer_inode_t ip, void *data),
377 hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
378 hammer_inode_info_cmp_all_history,
383 * Acquire a HAMMER inode. The returned inode is not locked. These functions
384 * do not attach or detach the related vnode (use hammer_get_vnode() for
387 * The flags argument is only applied for newly created inodes, and only
388 * certain flags are inherited.
390 * Called from the frontend.
392 struct hammer_inode *
393 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
394 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
395 int flags, int *errorp)
397 hammer_mount_t hmp = trans->hmp;
398 struct hammer_node_cache *cachep;
399 struct hammer_inode_info iinfo;
400 struct hammer_cursor cursor;
401 struct hammer_inode *ip;
405 * Determine if we already have an inode cached. If we do then
408 * If we find an inode with no vnode we have to mark the
409 * transaction such that hammer_inode_waitreclaims() is
410 * called later on to avoid building up an infinite number
411 * of inodes. Otherwise we can continue to * add new inodes
412 * faster then they can be disposed of, even with the tsleep
415 * If we find a dummy inode we return a failure so dounlink
416 * (which does another lookup) doesn't try to mess with the
417 * link count. hammer_vop_nresolve() uses hammer_get_dummy_inode()
418 * to ref dummy inodes.
420 iinfo.obj_id = obj_id;
421 iinfo.obj_asof = asof;
422 iinfo.obj_localization = localization;
424 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
426 if (ip->flags & HAMMER_INODE_DUMMY) {
430 hammer_ref(&ip->lock);
436 * Allocate a new inode structure and deal with races later.
438 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
439 ++hammer_count_inodes;
442 ip->obj_asof = iinfo.obj_asof;
443 ip->obj_localization = localization;
445 ip->flags = flags & HAMMER_INODE_RO;
446 ip->cache[0].ip = ip;
447 ip->cache[1].ip = ip;
448 ip->cache[2].ip = ip;
449 ip->cache[3].ip = ip;
451 ip->flags |= HAMMER_INODE_RO;
452 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
453 0x7FFFFFFFFFFFFFFFLL;
454 RB_INIT(&ip->rec_tree);
455 TAILQ_INIT(&ip->target_list);
456 hammer_ref(&ip->lock);
459 * Locate the on-disk inode. If this is a PFS root we always
460 * access the current version of the root inode and (if it is not
461 * a master) always access information under it with a snapshot
464 * We cache recent inode lookups in this directory in dip->cache[2].
465 * If we can't find it we assume the inode we are looking for is
466 * close to the directory inode.
471 if (dip->cache[2].node)
472 cachep = &dip->cache[2];
474 cachep = &dip->cache[0];
476 hammer_init_cursor(trans, &cursor, cachep, NULL);
477 cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
478 cursor.key_beg.obj_id = ip->obj_id;
479 cursor.key_beg.key = 0;
480 cursor.key_beg.create_tid = 0;
481 cursor.key_beg.delete_tid = 0;
482 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
483 cursor.key_beg.obj_type = 0;
485 cursor.asof = iinfo.obj_asof;
486 cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
489 *errorp = hammer_btree_lookup(&cursor);
490 if (*errorp == EDEADLK) {
491 hammer_done_cursor(&cursor);
496 * On success the B-Tree lookup will hold the appropriate
497 * buffer cache buffers and provide a pointer to the requested
498 * information. Copy the information to the in-memory inode
499 * and cache the B-Tree node to improve future operations.
502 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
503 ip->ino_data = cursor.data->inode;
506 * cache[0] tries to cache the location of the object inode.
507 * The assumption is that it is near the directory inode.
509 * cache[1] tries to cache the location of the object data.
510 * We might have something in the governing directory from
511 * scan optimizations (see the strategy code in
514 * We update dip->cache[2], if possible, with the location
515 * of the object inode for future directory shortcuts.
517 hammer_cache_node(&ip->cache[0], cursor.node);
519 if (dip->cache[3].node) {
520 hammer_cache_node(&ip->cache[1],
523 hammer_cache_node(&dip->cache[2], cursor.node);
527 * The file should not contain any data past the file size
528 * stored in the inode. Setting save_trunc_off to the
529 * file size instead of max reduces B-Tree lookup overheads
530 * on append by allowing the flusher to avoid checking for
533 ip->save_trunc_off = ip->ino_data.size;
536 * Locate and assign the pseudofs management structure to
539 if (dip && dip->obj_localization == ip->obj_localization) {
540 ip->pfsm = dip->pfsm;
541 hammer_ref(&ip->pfsm->lock);
543 ip->pfsm = hammer_load_pseudofs(trans,
544 ip->obj_localization,
546 *errorp = 0; /* ignore ENOENT */
551 * The inode is placed on the red-black tree and will be synced to
552 * the media when flushed or by the filesystem sync. If this races
553 * another instantiation/lookup the insertion will fail.
556 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
557 hammer_free_inode(ip);
558 hammer_done_cursor(&cursor);
561 ip->flags |= HAMMER_INODE_ONDISK;
563 if (ip->flags & HAMMER_INODE_RSV_INODES) {
564 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
568 hammer_free_inode(ip);
571 hammer_done_cursor(&cursor);
574 * NEWINODE is only set if the inode becomes dirty later,
575 * setting it here just leads to unnecessary stalls.
577 * trans->flags |= HAMMER_TRANSF_NEWINODE;
583 * Get a dummy inode to placemark a broken directory entry.
585 struct hammer_inode *
586 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
587 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
588 int flags, int *errorp)
590 hammer_mount_t hmp = trans->hmp;
591 struct hammer_inode_info iinfo;
592 struct hammer_inode *ip;
595 * Determine if we already have an inode cached. If we do then
598 * If we find an inode with no vnode we have to mark the
599 * transaction such that hammer_inode_waitreclaims() is
600 * called later on to avoid building up an infinite number
601 * of inodes. Otherwise we can continue to * add new inodes
602 * faster then they can be disposed of, even with the tsleep
605 * If we find a non-fake inode we return an error. Only fake
606 * inodes can be returned by this routine.
608 iinfo.obj_id = obj_id;
609 iinfo.obj_asof = asof;
610 iinfo.obj_localization = localization;
613 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
615 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
619 hammer_ref(&ip->lock);
624 * Allocate a new inode structure and deal with races later.
626 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
627 ++hammer_count_inodes;
630 ip->obj_asof = iinfo.obj_asof;
631 ip->obj_localization = localization;
633 ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
634 ip->cache[0].ip = ip;
635 ip->cache[1].ip = ip;
636 ip->cache[2].ip = ip;
637 ip->cache[3].ip = ip;
638 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
639 0x7FFFFFFFFFFFFFFFLL;
640 RB_INIT(&ip->rec_tree);
641 TAILQ_INIT(&ip->target_list);
642 hammer_ref(&ip->lock);
645 * Populate the dummy inode. Leave everything zero'd out.
647 * (ip->ino_leaf and ip->ino_data)
649 * Make the dummy inode a FIFO object which most copy programs
650 * will properly ignore.
652 ip->save_trunc_off = ip->ino_data.size;
653 ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
656 * Locate and assign the pseudofs management structure to
659 if (dip && dip->obj_localization == ip->obj_localization) {
660 ip->pfsm = dip->pfsm;
661 hammer_ref(&ip->pfsm->lock);
663 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
665 *errorp = 0; /* ignore ENOENT */
669 * The inode is placed on the red-black tree and will be synced to
670 * the media when flushed or by the filesystem sync. If this races
671 * another instantiation/lookup the insertion will fail.
673 * NOTE: Do not set HAMMER_INODE_ONDISK. The inode is a fake.
676 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
677 hammer_free_inode(ip);
681 if (ip->flags & HAMMER_INODE_RSV_INODES) {
682 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
685 hammer_free_inode(ip);
688 trans->flags |= HAMMER_TRANSF_NEWINODE;
693 * Return a referenced inode only if it is in our inode cache.
695 * Dummy inodes do not count.
697 struct hammer_inode *
698 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
699 hammer_tid_t asof, u_int32_t localization)
701 hammer_mount_t hmp = trans->hmp;
702 struct hammer_inode_info iinfo;
703 struct hammer_inode *ip;
705 iinfo.obj_id = obj_id;
706 iinfo.obj_asof = asof;
707 iinfo.obj_localization = localization;
709 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
711 if (ip->flags & HAMMER_INODE_DUMMY)
714 hammer_ref(&ip->lock);
720 * Create a new filesystem object, returning the inode in *ipp. The
721 * returned inode will be referenced. The inode is created in-memory.
723 * If pfsm is non-NULL the caller wishes to create the root inode for
727 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
729 hammer_inode_t dip, const char *name, int namelen,
730 hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
741 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
742 ++hammer_count_inodes;
744 trans->flags |= HAMMER_TRANSF_NEWINODE;
747 KKASSERT(pfsm->localization != 0);
748 ip->obj_id = HAMMER_OBJID_ROOT;
749 ip->obj_localization = pfsm->localization;
751 KKASSERT(dip != NULL);
752 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
753 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
754 ip->obj_localization = dip->obj_localization;
757 KKASSERT(ip->obj_id != 0);
758 ip->obj_asof = hmp->asof;
760 ip->flush_state = HAMMER_FST_IDLE;
761 ip->flags = HAMMER_INODE_DDIRTY |
762 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
763 ip->cache[0].ip = ip;
764 ip->cache[1].ip = ip;
765 ip->cache[2].ip = ip;
766 ip->cache[3].ip = ip;
768 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
769 /* ip->save_trunc_off = 0; (already zero) */
770 RB_INIT(&ip->rec_tree);
771 TAILQ_INIT(&ip->target_list);
773 ip->ino_data.atime = trans->time;
774 ip->ino_data.mtime = trans->time;
775 ip->ino_data.size = 0;
776 ip->ino_data.nlinks = 0;
779 * A nohistory designator on the parent directory is inherited by
780 * the child. We will do this even for pseudo-fs creation... the
781 * sysad can turn it off.
784 ip->ino_data.uflags = dip->ino_data.uflags &
785 (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
788 ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
789 ip->ino_leaf.base.localization = ip->obj_localization +
790 HAMMER_LOCALIZE_INODE;
791 ip->ino_leaf.base.obj_id = ip->obj_id;
792 ip->ino_leaf.base.key = 0;
793 ip->ino_leaf.base.create_tid = 0;
794 ip->ino_leaf.base.delete_tid = 0;
795 ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
796 ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
798 ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
799 ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
800 ip->ino_data.mode = vap->va_mode;
801 ip->ino_data.ctime = trans->time;
804 * If we are running version 2 or greater directory entries are
805 * inode-localized instead of data-localized.
807 if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
808 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
809 ip->ino_data.cap_flags |=
810 HAMMER_INODE_CAP_DIR_LOCAL_INO;
813 if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
814 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
815 ip->ino_data.cap_flags |=
816 HAMMER_INODE_CAP_DIRHASH_ALG1;
821 * Setup the ".." pointer. This only needs to be done for directories
822 * but we do it for all objects as a recovery aid.
825 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
828 * The parent_obj_localization field only applies to pseudo-fs roots.
829 * XXX this is no longer applicable, PFSs are no longer directly
830 * tied into the parent's directory structure.
832 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
833 ip->obj_id == HAMMER_OBJID_ROOT) {
834 ip->ino_data.ext.obj.parent_obj_localization =
835 dip->obj_localization;
839 switch(ip->ino_leaf.base.obj_type) {
840 case HAMMER_OBJTYPE_CDEV:
841 case HAMMER_OBJTYPE_BDEV:
842 ip->ino_data.rmajor = vap->va_rmajor;
843 ip->ino_data.rminor = vap->va_rminor;
850 * Calculate default uid/gid and overwrite with information from
854 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
855 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
856 xuid, cred, &vap->va_mode);
860 ip->ino_data.mode = vap->va_mode;
862 if (vap->va_vaflags & VA_UID_UUID_VALID)
863 ip->ino_data.uid = vap->va_uid_uuid;
864 else if (vap->va_uid != (uid_t)VNOVAL)
865 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
867 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
869 if (vap->va_vaflags & VA_GID_UUID_VALID)
870 ip->ino_data.gid = vap->va_gid_uuid;
871 else if (vap->va_gid != (gid_t)VNOVAL)
872 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
874 ip->ino_data.gid = dip->ino_data.gid;
876 hammer_ref(&ip->lock);
880 hammer_ref(&pfsm->lock);
882 } else if (dip->obj_localization == ip->obj_localization) {
883 ip->pfsm = dip->pfsm;
884 hammer_ref(&ip->pfsm->lock);
887 ip->pfsm = hammer_load_pseudofs(trans,
888 ip->obj_localization,
890 error = 0; /* ignore ENOENT */
894 hammer_free_inode(ip);
896 } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
897 panic("hammer_create_inode: duplicate obj_id %llx",
898 (long long)ip->obj_id);
900 hammer_free_inode(ip);
907 * Final cleanup / freeing of an inode structure
910 hammer_free_inode(hammer_inode_t ip)
912 struct hammer_mount *hmp;
915 KKASSERT(hammer_oneref(&ip->lock));
916 hammer_uncache_node(&ip->cache[0]);
917 hammer_uncache_node(&ip->cache[1]);
918 hammer_uncache_node(&ip->cache[2]);
919 hammer_uncache_node(&ip->cache[3]);
920 hammer_inode_wakereclaims(ip);
922 hammer_clear_objid(ip);
923 --hammer_count_inodes;
926 hammer_rel_pseudofs(hmp, ip->pfsm);
929 kfree(ip, hmp->m_inodes);
934 * Retrieve pseudo-fs data. NULL will never be returned.
936 * If an error occurs *errorp will be set and a default template is returned,
937 * otherwise *errorp is set to 0. Typically when an error occurs it will
940 hammer_pseudofs_inmem_t
941 hammer_load_pseudofs(hammer_transaction_t trans,
942 u_int32_t localization, int *errorp)
944 hammer_mount_t hmp = trans->hmp;
946 hammer_pseudofs_inmem_t pfsm;
947 struct hammer_cursor cursor;
951 pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
953 hammer_ref(&pfsm->lock);
959 * PFS records are stored in the root inode (not the PFS root inode,
960 * but the real root). Avoid an infinite recursion if loading
961 * the PFS for the real root.
964 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
966 HAMMER_DEF_LOCALIZATION, 0, errorp);
971 pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
972 pfsm->localization = localization;
973 pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
974 pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
976 hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
977 cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
978 HAMMER_LOCALIZE_MISC;
979 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
980 cursor.key_beg.create_tid = 0;
981 cursor.key_beg.delete_tid = 0;
982 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
983 cursor.key_beg.obj_type = 0;
984 cursor.key_beg.key = localization;
985 cursor.asof = HAMMER_MAX_TID;
986 cursor.flags |= HAMMER_CURSOR_ASOF;
989 *errorp = hammer_ip_lookup(&cursor);
991 *errorp = hammer_btree_lookup(&cursor);
993 *errorp = hammer_ip_resolve_data(&cursor);
995 if (cursor.data->pfsd.mirror_flags &
996 HAMMER_PFSD_DELETED) {
999 bytes = cursor.leaf->data_len;
1000 if (bytes > sizeof(pfsm->pfsd))
1001 bytes = sizeof(pfsm->pfsd);
1002 bcopy(cursor.data, &pfsm->pfsd, bytes);
1006 hammer_done_cursor(&cursor);
1008 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1009 hammer_ref(&pfsm->lock);
1011 hammer_rel_inode(ip, 0);
1012 if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
1013 kfree(pfsm, hmp->m_misc);
1020 * Store pseudo-fs data. The backend will automatically delete any prior
1021 * on-disk pseudo-fs data but we have to delete in-memory versions.
1024 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
1026 struct hammer_cursor cursor;
1027 hammer_record_t record;
1031 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1032 HAMMER_DEF_LOCALIZATION, 0, &error);
1034 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1035 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1036 cursor.key_beg.localization = ip->obj_localization +
1037 HAMMER_LOCALIZE_MISC;
1038 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1039 cursor.key_beg.create_tid = 0;
1040 cursor.key_beg.delete_tid = 0;
1041 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1042 cursor.key_beg.obj_type = 0;
1043 cursor.key_beg.key = pfsm->localization;
1044 cursor.asof = HAMMER_MAX_TID;
1045 cursor.flags |= HAMMER_CURSOR_ASOF;
1048 * Replace any in-memory version of the record.
1050 error = hammer_ip_lookup(&cursor);
1051 if (error == 0 && hammer_cursor_inmem(&cursor)) {
1052 record = cursor.iprec;
1053 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1054 KKASSERT(cursor.deadlk_rec == NULL);
1055 hammer_ref(&record->lock);
1056 cursor.deadlk_rec = record;
1059 record->flags |= HAMMER_RECF_DELETED_FE;
1065 * Allocate replacement general record. The backend flush will
1066 * delete any on-disk version of the record.
1068 if (error == 0 || error == ENOENT) {
1069 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1070 record->type = HAMMER_MEM_RECORD_GENERAL;
1072 record->leaf.base.localization = ip->obj_localization +
1073 HAMMER_LOCALIZE_MISC;
1074 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1075 record->leaf.base.key = pfsm->localization;
1076 record->leaf.data_len = sizeof(pfsm->pfsd);
1077 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1078 error = hammer_ip_add_record(trans, record);
1080 hammer_done_cursor(&cursor);
1081 if (error == EDEADLK)
1083 hammer_rel_inode(ip, 0);
1088 * Create a root directory for a PFS if one does not alredy exist.
1090 * The PFS root stands alone so we must also bump the nlinks count
1091 * to prevent it from being destroyed on release.
1094 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1095 hammer_pseudofs_inmem_t pfsm)
1101 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1102 pfsm->localization, 0, &error);
1107 error = hammer_create_inode(trans, &vap, cred,
1111 ++ip->ino_data.nlinks;
1112 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1116 hammer_rel_inode(ip, 0);
1121 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1122 * if we are unable to disassociate all the inodes.
1126 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1130 hammer_ref(&ip->lock);
1131 if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1132 vclean_unlocked(ip->vp);
1133 if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1136 res = -1; /* stop, someone is using the inode */
1137 hammer_rel_inode(ip, 0);
1142 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1147 for (try = res = 0; try < 4; ++try) {
1148 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1149 hammer_inode_pfs_cmp,
1150 hammer_unload_pseudofs_callback,
1152 if (res == 0 && try > 1)
1154 hammer_flusher_sync(trans->hmp);
1163 * Release a reference on a PFS
1166 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1168 hammer_rel(&pfsm->lock);
1169 if (hammer_norefs(&pfsm->lock)) {
1170 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1171 kfree(pfsm, hmp->m_misc);
1176 * Called by hammer_sync_inode().
1179 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1181 hammer_transaction_t trans = cursor->trans;
1182 hammer_record_t record;
1190 * If the inode has a presence on-disk then locate it and mark
1191 * it deleted, setting DELONDISK.
1193 * The record may or may not be physically deleted, depending on
1194 * the retention policy.
1196 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1197 HAMMER_INODE_ONDISK) {
1198 hammer_normalize_cursor(cursor);
1199 cursor->key_beg.localization = ip->obj_localization +
1200 HAMMER_LOCALIZE_INODE;
1201 cursor->key_beg.obj_id = ip->obj_id;
1202 cursor->key_beg.key = 0;
1203 cursor->key_beg.create_tid = 0;
1204 cursor->key_beg.delete_tid = 0;
1205 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1206 cursor->key_beg.obj_type = 0;
1207 cursor->asof = ip->obj_asof;
1208 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1209 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1210 cursor->flags |= HAMMER_CURSOR_BACKEND;
1212 error = hammer_btree_lookup(cursor);
1213 if (hammer_debug_inode)
1214 kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1217 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1218 if (hammer_debug_inode)
1219 kprintf(" error %d\n", error);
1221 ip->flags |= HAMMER_INODE_DELONDISK;
1224 hammer_cache_node(&ip->cache[0], cursor->node);
1226 if (error == EDEADLK) {
1227 hammer_done_cursor(cursor);
1228 error = hammer_init_cursor(trans, cursor,
1230 if (hammer_debug_inode)
1231 kprintf("IPDED %p %d\n", ip, error);
1238 * Ok, write out the initial record or a new record (after deleting
1239 * the old one), unless the DELETED flag is set. This routine will
1240 * clear DELONDISK if it writes out a record.
1242 * Update our inode statistics if this is the first application of
1243 * the inode on-disk.
1245 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1247 * Generate a record and write it to the media. We clean-up
1248 * the state before releasing so we do not have to set-up
1251 record = hammer_alloc_mem_record(ip, 0);
1252 record->type = HAMMER_MEM_RECORD_INODE;
1253 record->flush_state = HAMMER_FST_FLUSH;
1254 record->leaf = ip->sync_ino_leaf;
1255 record->leaf.base.create_tid = trans->tid;
1256 record->leaf.data_len = sizeof(ip->sync_ino_data);
1257 record->leaf.create_ts = trans->time32;
1258 record->data = (void *)&ip->sync_ino_data;
1259 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1262 * If this flag is set we cannot sync the new file size
1263 * because we haven't finished related truncations. The
1264 * inode will be flushed in another flush group to finish
1267 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1268 ip->sync_ino_data.size != ip->ino_data.size) {
1270 ip->sync_ino_data.size = ip->ino_data.size;
1276 error = hammer_ip_sync_record_cursor(cursor, record);
1277 if (hammer_debug_inode)
1278 kprintf("GENREC %p rec %08x %d\n",
1279 ip, record->flags, error);
1280 if (error != EDEADLK)
1282 hammer_done_cursor(cursor);
1283 error = hammer_init_cursor(trans, cursor,
1285 if (hammer_debug_inode)
1286 kprintf("GENREC reinit %d\n", error);
1292 * Note: The record was never on the inode's record tree
1293 * so just wave our hands importantly and destroy it.
1295 record->flags |= HAMMER_RECF_COMMITTED;
1296 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1297 record->flush_state = HAMMER_FST_IDLE;
1298 ++ip->rec_generation;
1299 hammer_rel_mem_record(record);
1305 if (hammer_debug_inode)
1306 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1307 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1308 HAMMER_INODE_SDIRTY |
1309 HAMMER_INODE_ATIME |
1310 HAMMER_INODE_MTIME);
1311 ip->flags &= ~HAMMER_INODE_DELONDISK;
1313 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1316 * Root volume count of inodes
1318 hammer_sync_lock_sh(trans);
1319 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1320 hammer_modify_volume_field(trans,
1323 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1324 hammer_modify_volume_done(trans->rootvol);
1325 ip->flags |= HAMMER_INODE_ONDISK;
1326 if (hammer_debug_inode)
1327 kprintf("NOWONDISK %p\n", ip);
1329 hammer_sync_unlock(trans);
1334 * If the inode has been destroyed, clean out any left-over flags
1335 * that may have been set by the frontend.
1337 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1338 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1339 HAMMER_INODE_SDIRTY |
1340 HAMMER_INODE_ATIME |
1341 HAMMER_INODE_MTIME);
1347 * Update only the itimes fields.
1349 * ATIME can be updated without generating any UNDO. MTIME is updated
1350 * with UNDO so it is guaranteed to be synchronized properly in case of
1353 * Neither field is included in the B-Tree leaf element's CRC, which is how
1354 * we can get away with updating ATIME the way we do.
1357 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1359 hammer_transaction_t trans = cursor->trans;
1363 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1364 HAMMER_INODE_ONDISK) {
1368 hammer_normalize_cursor(cursor);
1369 cursor->key_beg.localization = ip->obj_localization +
1370 HAMMER_LOCALIZE_INODE;
1371 cursor->key_beg.obj_id = ip->obj_id;
1372 cursor->key_beg.key = 0;
1373 cursor->key_beg.create_tid = 0;
1374 cursor->key_beg.delete_tid = 0;
1375 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1376 cursor->key_beg.obj_type = 0;
1377 cursor->asof = ip->obj_asof;
1378 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1379 cursor->flags |= HAMMER_CURSOR_ASOF;
1380 cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1381 cursor->flags |= HAMMER_CURSOR_GET_DATA;
1382 cursor->flags |= HAMMER_CURSOR_BACKEND;
1384 error = hammer_btree_lookup(cursor);
1386 hammer_cache_node(&ip->cache[0], cursor->node);
1387 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1389 * Updating MTIME requires an UNDO. Just cover
1390 * both atime and mtime.
1392 hammer_sync_lock_sh(trans);
1393 hammer_modify_buffer(trans, cursor->data_buffer,
1394 HAMMER_ITIMES_BASE(&cursor->data->inode),
1395 HAMMER_ITIMES_BYTES);
1396 cursor->data->inode.atime = ip->sync_ino_data.atime;
1397 cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1398 hammer_modify_buffer_done(cursor->data_buffer);
1399 hammer_sync_unlock(trans);
1400 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1402 * Updating atime only can be done in-place with
1405 hammer_sync_lock_sh(trans);
1406 hammer_modify_buffer(trans, cursor->data_buffer,
1408 cursor->data->inode.atime = ip->sync_ino_data.atime;
1409 hammer_modify_buffer_done(cursor->data_buffer);
1410 hammer_sync_unlock(trans);
1412 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1414 if (error == EDEADLK) {
1415 hammer_done_cursor(cursor);
1416 error = hammer_init_cursor(trans, cursor,
1425 * Release a reference on an inode, flush as requested.
1427 * On the last reference we queue the inode to the flusher for its final
1431 hammer_rel_inode(struct hammer_inode *ip, int flush)
1433 /*hammer_mount_t hmp = ip->hmp;*/
1436 * Handle disposition when dropping the last ref.
1439 if (hammer_oneref(&ip->lock)) {
1441 * Determine whether on-disk action is needed for
1442 * the inode's final disposition.
1444 KKASSERT(ip->vp == NULL);
1445 hammer_inode_unloadable_check(ip, 0);
1446 if (ip->flags & HAMMER_INODE_MODMASK) {
1447 hammer_flush_inode(ip, 0);
1448 } else if (hammer_oneref(&ip->lock)) {
1449 hammer_unload_inode(ip);
1454 hammer_flush_inode(ip, 0);
1457 * The inode still has multiple refs, try to drop
1460 KKASSERT(hammer_isactive(&ip->lock) >= 1);
1461 if (hammer_isactive(&ip->lock) > 1) {
1462 hammer_rel(&ip->lock);
1470 * Unload and destroy the specified inode. Must be called with one remaining
1471 * reference. The reference is disposed of.
1473 * The inode must be completely clean.
1476 hammer_unload_inode(struct hammer_inode *ip)
1478 hammer_mount_t hmp = ip->hmp;
1480 KASSERT(hammer_oneref(&ip->lock),
1481 ("hammer_unload_inode: %d refs\n", hammer_isactive(&ip->lock)));
1482 KKASSERT(ip->vp == NULL);
1483 KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1484 KKASSERT(ip->cursor_ip_refs == 0);
1485 KKASSERT(hammer_notlocked(&ip->lock));
1486 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1488 KKASSERT(RB_EMPTY(&ip->rec_tree));
1489 KKASSERT(TAILQ_EMPTY(&ip->target_list));
1491 if (ip->flags & HAMMER_INODE_RDIRTY) {
1492 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1493 ip->flags &= ~HAMMER_INODE_RDIRTY;
1495 RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1497 hammer_free_inode(ip);
1502 * Called during unmounting if a critical error occured. The in-memory
1503 * inode and all related structures are destroyed.
1505 * If a critical error did not occur the unmount code calls the standard
1506 * release and asserts that the inode is gone.
1509 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1511 hammer_record_t rec;
1514 * Get rid of the inodes in-memory records, regardless of their
1515 * state, and clear the mod-mask.
1517 while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1518 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1519 rec->target_ip = NULL;
1520 if (rec->flush_state == HAMMER_FST_SETUP)
1521 rec->flush_state = HAMMER_FST_IDLE;
1523 while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1524 if (rec->flush_state == HAMMER_FST_FLUSH)
1525 --rec->flush_group->refs;
1527 hammer_ref(&rec->lock);
1528 KKASSERT(hammer_oneref(&rec->lock));
1529 rec->flush_state = HAMMER_FST_IDLE;
1530 rec->flush_group = NULL;
1531 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1532 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1533 ++ip->rec_generation;
1534 hammer_rel_mem_record(rec);
1536 ip->flags &= ~HAMMER_INODE_MODMASK;
1537 ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1538 KKASSERT(ip->vp == NULL);
1541 * Remove the inode from any flush group, force it idle. FLUSH
1542 * and SETUP states have an inode ref.
1544 switch(ip->flush_state) {
1545 case HAMMER_FST_FLUSH:
1546 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1547 --ip->flush_group->refs;
1548 ip->flush_group = NULL;
1550 case HAMMER_FST_SETUP:
1551 hammer_rel(&ip->lock);
1552 ip->flush_state = HAMMER_FST_IDLE;
1554 case HAMMER_FST_IDLE:
1559 * There shouldn't be any associated vnode. The unload needs at
1560 * least one ref, if we do have a vp steal its ip ref.
1563 kprintf("hammer_destroy_inode_callback: Unexpected "
1564 "vnode association ip %p vp %p\n", ip, ip->vp);
1565 ip->vp->v_data = NULL;
1568 hammer_ref(&ip->lock);
1570 hammer_unload_inode(ip);
1575 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1576 * the read-only flag for cached inodes.
1578 * This routine is called from a RB_SCAN().
1581 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1583 hammer_mount_t hmp = ip->hmp;
1585 if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1586 ip->flags |= HAMMER_INODE_RO;
1588 ip->flags &= ~HAMMER_INODE_RO;
1593 * A transaction has modified an inode, requiring updates as specified by
1596 * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1597 * and not including size changes due to write-append
1598 * (but other size changes are included).
1599 * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1601 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1602 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1603 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1604 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1607 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1610 * ronly of 0 or 2 does not trigger assertion.
1611 * 2 is a special error state
1613 KKASSERT(ip->hmp->ronly != 1 ||
1614 (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1615 HAMMER_INODE_SDIRTY |
1616 HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1617 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1618 if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1619 ip->flags |= HAMMER_INODE_RSV_INODES;
1620 ++ip->hmp->rsv_inodes;
1624 * Set the NEWINODE flag in the transaction if the inode
1625 * transitions to a dirty state. This is used to track
1626 * the load on the inode cache.
1629 (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1630 (flags & HAMMER_INODE_MODMASK)) {
1631 trans->flags |= HAMMER_TRANSF_NEWINODE;
1638 * Attempt to quickly update the atime for a hammer inode. Return 0 on
1639 * success, -1 on failure.
1641 * We attempt to update the atime with only the ip lock and not the
1642 * whole filesystem lock in order to improve concurrency. We can only
1643 * do this safely if the ATIME flag is already pending on the inode.
1645 * This function is called via a vnops path (ip pointer is stable) without
1649 hammer_update_atime_quick(hammer_inode_t ip)
1654 if ((ip->flags & HAMMER_INODE_RO) ||
1655 (ip->hmp->mp->mnt_flag & MNT_NOATIME)) {
1657 * Silently indicate success on read-only mount/snap
1660 } else if (ip->flags & HAMMER_INODE_ATIME) {
1662 * Double check with inode lock held against backend. This
1663 * is only safe if all we need to do is update
1667 hammer_lock_ex(&ip->lock);
1668 if (ip->flags & HAMMER_INODE_ATIME) {
1669 ip->ino_data.atime =
1670 (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec;
1673 hammer_unlock(&ip->lock);
1679 * Request that an inode be flushed. This whole mess cannot block and may
1680 * recurse (if not synchronous). Once requested HAMMER will attempt to
1681 * actively flush the inode until the flush can be done.
1683 * The inode may already be flushing, or may be in a setup state. We can
1684 * place the inode in a flushing state if it is currently idle and flag it
1685 * to reflush if it is currently flushing.
1687 * Upon return if the inode could not be flushed due to a setup
1688 * dependancy, then it will be automatically flushed when the dependancy
1692 hammer_flush_inode(hammer_inode_t ip, int flags)
1695 hammer_flush_group_t flg;
1699 * fill_flush_group is the first flush group we may be able to
1700 * continue filling, it may be open or closed but it will always
1701 * be past the currently flushing (running) flg.
1703 * next_flush_group is the next open flush group.
1706 while ((flg = hmp->fill_flush_group) != NULL) {
1707 KKASSERT(flg->running == 0);
1708 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1709 flg->total_count <= hammer_autoflush) {
1712 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1713 hammer_flusher_async(ip->hmp, flg);
1716 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1717 flg->seq = hmp->flusher.next++;
1718 if (hmp->next_flush_group == NULL)
1719 hmp->next_flush_group = flg;
1720 if (hmp->fill_flush_group == NULL)
1721 hmp->fill_flush_group = flg;
1722 RB_INIT(&flg->flush_tree);
1723 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1727 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1728 * state we have to put it back into an IDLE state so we can
1729 * drop the extra ref.
1731 * If we have a parent dependancy we must still fall through
1734 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1735 if (ip->flush_state == HAMMER_FST_SETUP &&
1736 TAILQ_EMPTY(&ip->target_list)) {
1737 ip->flush_state = HAMMER_FST_IDLE;
1738 hammer_rel_inode(ip, 0);
1740 if (ip->flush_state == HAMMER_FST_IDLE)
1745 * Our flush action will depend on the current state.
1747 switch(ip->flush_state) {
1748 case HAMMER_FST_IDLE:
1750 * We have no dependancies and can flush immediately. Some
1751 * our children may not be flushable so we have to re-test
1752 * with that additional knowledge.
1754 hammer_flush_inode_core(ip, flg, flags);
1756 case HAMMER_FST_SETUP:
1758 * Recurse upwards through dependancies via target_list
1759 * and start their flusher actions going if possible.
1761 * 'good' is our connectivity. -1 means we have none and
1762 * can't flush, 0 means there weren't any dependancies, and
1763 * 1 means we have good connectivity.
1765 good = hammer_setup_parent_inodes(ip, 0, flg);
1769 * We can continue if good >= 0. Determine how
1770 * many records under our inode can be flushed (and
1773 hammer_flush_inode_core(ip, flg, flags);
1776 * Parent has no connectivity, tell it to flush
1777 * us as soon as it does.
1779 * The REFLUSH flag is also needed to trigger
1780 * dependancy wakeups.
1782 ip->flags |= HAMMER_INODE_CONN_DOWN |
1783 HAMMER_INODE_REFLUSH;
1784 if (flags & HAMMER_FLUSH_SIGNAL) {
1785 ip->flags |= HAMMER_INODE_RESIGNAL;
1786 hammer_flusher_async(ip->hmp, flg);
1790 case HAMMER_FST_FLUSH:
1792 * We are already flushing, flag the inode to reflush
1793 * if needed after it completes its current flush.
1795 * The REFLUSH flag is also needed to trigger
1796 * dependancy wakeups.
1798 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1799 ip->flags |= HAMMER_INODE_REFLUSH;
1800 if (flags & HAMMER_FLUSH_SIGNAL) {
1801 ip->flags |= HAMMER_INODE_RESIGNAL;
1802 hammer_flusher_async(ip->hmp, flg);
1809 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1810 * ip which reference our ip.
1812 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1813 * so for now do not ref/deref the structures. Note that if we use the
1814 * ref/rel code later, the rel CAN block.
1817 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1818 hammer_flush_group_t flg)
1820 hammer_record_t depend;
1825 * If we hit our recursion limit and we have parent dependencies
1826 * We cannot continue. Returning < 0 will cause us to be flagged
1827 * for reflush. Returning -2 cuts off additional dependency checks
1828 * because they are likely to also hit the depth limit.
1830 * We cannot return < 0 if there are no dependencies or there might
1831 * not be anything to wakeup (ip).
1833 if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1834 kprintf("HAMMER Warning: depth limit reached on "
1835 "setup recursion, inode %p %016llx\n",
1836 ip, (long long)ip->obj_id);
1844 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1845 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1846 KKASSERT(depend->target_ip == ip);
1847 if (r < 0 && good == 0)
1853 * If we failed due to the recursion depth limit then stop
1863 * This helper function takes a record representing the dependancy between
1864 * the parent inode and child inode.
1866 * record->ip = parent inode
1867 * record->target_ip = child inode
1869 * We are asked to recurse upwards and convert the record from SETUP
1870 * to FLUSH if possible.
1872 * Return 1 if the record gives us connectivity
1874 * Return 0 if the record is not relevant
1876 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1879 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1880 hammer_flush_group_t flg)
1886 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1891 * If the record is already flushing, is it in our flush group?
1893 * If it is in our flush group but it is a general record or a
1894 * delete-on-disk, it does not improve our connectivity (return 0),
1895 * and if the target inode is not trying to destroy itself we can't
1896 * allow the operation yet anyway (the second return -1).
1898 if (record->flush_state == HAMMER_FST_FLUSH) {
1900 * If not in our flush group ask the parent to reflush
1901 * us as soon as possible.
1903 if (record->flush_group != flg) {
1904 pip->flags |= HAMMER_INODE_REFLUSH;
1905 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1910 * If in our flush group everything is already set up,
1911 * just return whether the record will improve our
1912 * visibility or not.
1914 if (record->type == HAMMER_MEM_RECORD_ADD)
1920 * It must be a setup record. Try to resolve the setup dependancies
1921 * by recursing upwards so we can place ip on the flush list.
1923 * Limit ourselves to 20 levels of recursion to avoid blowing out
1924 * the kernel stack. If we hit the recursion limit we can't flush
1925 * until the parent flushes. The parent will flush independantly
1926 * on its own and ultimately a deep recursion will be resolved.
1928 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1930 good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1933 * If good < 0 the parent has no connectivity and we cannot safely
1934 * flush the directory entry, which also means we can't flush our
1935 * ip. Flag us for downward recursion once the parent's
1936 * connectivity is resolved. Flag the parent for [re]flush or it
1937 * may not check for downward recursions.
1940 pip->flags |= HAMMER_INODE_REFLUSH;
1941 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1946 * We are go, place the parent inode in a flushing state so we can
1947 * place its record in a flushing state. Note that the parent
1948 * may already be flushing. The record must be in the same flush
1949 * group as the parent.
1951 if (pip->flush_state != HAMMER_FST_FLUSH)
1952 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1953 KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1956 * It is possible for a rename to create a loop in the recursion
1957 * and revisit a record. This will result in the record being
1958 * placed in a flush state unexpectedly. This check deals with
1961 if (record->flush_state == HAMMER_FST_FLUSH) {
1962 if (record->type == HAMMER_MEM_RECORD_ADD)
1967 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1970 if (record->type == HAMMER_MEM_RECORD_DEL &&
1971 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1973 * Regardless of flushing state we cannot sync this path if the
1974 * record represents a delete-on-disk but the target inode
1975 * is not ready to sync its own deletion.
1977 * XXX need to count effective nlinks to determine whether
1978 * the flush is ok, otherwise removing a hardlink will
1979 * just leave the DEL record to rot.
1981 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1985 if (pip->flush_group == flg) {
1987 * Because we have not calculated nlinks yet we can just
1988 * set records to the flush state if the parent is in
1989 * the same flush group as we are.
1991 record->flush_state = HAMMER_FST_FLUSH;
1992 record->flush_group = flg;
1993 ++record->flush_group->refs;
1994 hammer_ref(&record->lock);
1997 * A general directory-add contributes to our visibility.
1999 * Otherwise it is probably a directory-delete or
2000 * delete-on-disk record and does not contribute to our
2001 * visbility (but we can still flush it).
2003 if (record->type == HAMMER_MEM_RECORD_ADD)
2008 * If the parent is not in our flush group we cannot
2009 * flush this record yet, there is no visibility.
2010 * We tell the parent to reflush and mark ourselves
2011 * so the parent knows it should flush us too.
2013 pip->flags |= HAMMER_INODE_REFLUSH;
2014 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2020 * This is the core routine placing an inode into the FST_FLUSH state.
2023 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
2025 hammer_mount_t hmp = ip->hmp;
2029 * Set flush state and prevent the flusher from cycling into
2030 * the next flush group. Do not place the ip on the list yet.
2031 * Inodes not in the idle state get an extra reference.
2033 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
2034 if (ip->flush_state == HAMMER_FST_IDLE)
2035 hammer_ref(&ip->lock);
2036 ip->flush_state = HAMMER_FST_FLUSH;
2037 ip->flush_group = flg;
2038 ++hmp->flusher.group_lock;
2039 ++hmp->count_iqueued;
2040 ++hammer_count_iqueued;
2042 hammer_redo_fifo_start_flush(ip);
2046 * We need to be able to vfsync/truncate from the backend.
2048 * XXX Any truncation from the backend will acquire the vnode
2051 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2052 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2053 ip->flags |= HAMMER_INODE_VHELD;
2059 * Figure out how many in-memory records we can actually flush
2060 * (not including inode meta-data, buffers, etc).
2062 KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2063 if (flags & HAMMER_FLUSH_RECURSION) {
2065 * If this is a upwards recursion we do not want to
2066 * recurse down again!
2070 } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2072 * No new records are added if we must complete a flush
2073 * from a previous cycle, but we do have to move the records
2074 * from the previous cycle to the current one.
2077 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2078 hammer_syncgrp_child_callback, NULL);
2084 * Normal flush, scan records and bring them into the flush.
2085 * Directory adds and deletes are usually skipped (they are
2086 * grouped with the related inode rather then with the
2089 * go_count can be negative, which means the scan aborted
2090 * due to the flush group being over-full and we should
2091 * flush what we have.
2093 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2094 hammer_setup_child_callback, NULL);
2098 * This is a more involved test that includes go_count. If we
2099 * can't flush, flag the inode and return. If go_count is 0 we
2100 * were are unable to flush any records in our rec_tree and
2101 * must ignore the XDIRTY flag.
2103 if (go_count == 0) {
2104 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2105 --hmp->count_iqueued;
2106 --hammer_count_iqueued;
2109 ip->flush_state = HAMMER_FST_SETUP;
2110 ip->flush_group = NULL;
2111 if (flags & HAMMER_FLUSH_SIGNAL) {
2112 ip->flags |= HAMMER_INODE_REFLUSH |
2113 HAMMER_INODE_RESIGNAL;
2115 ip->flags |= HAMMER_INODE_REFLUSH;
2118 if (ip->flags & HAMMER_INODE_VHELD) {
2119 ip->flags &= ~HAMMER_INODE_VHELD;
2125 * REFLUSH is needed to trigger dependancy wakeups
2126 * when an inode is in SETUP.
2128 ip->flags |= HAMMER_INODE_REFLUSH;
2129 if (--hmp->flusher.group_lock == 0)
2130 wakeup(&hmp->flusher.group_lock);
2136 * Snapshot the state of the inode for the backend flusher.
2138 * We continue to retain save_trunc_off even when all truncations
2139 * have been resolved as an optimization to determine if we can
2140 * skip the B-Tree lookup for overwrite deletions.
2142 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2143 * and stays in ip->flags. Once set, it stays set until the
2144 * inode is destroyed.
2146 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2147 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2148 ip->sync_trunc_off = ip->trunc_off;
2149 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2150 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2151 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2154 * The save_trunc_off used to cache whether the B-Tree
2155 * holds any records past that point is not used until
2156 * after the truncation has succeeded, so we can safely
2159 if (ip->save_trunc_off > ip->sync_trunc_off)
2160 ip->save_trunc_off = ip->sync_trunc_off;
2162 ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2163 ~HAMMER_INODE_TRUNCATED);
2164 ip->sync_ino_leaf = ip->ino_leaf;
2165 ip->sync_ino_data = ip->ino_data;
2166 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2167 #ifdef DEBUG_TRUNCATE
2168 if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2169 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2173 * The flusher list inherits our inode and reference.
2175 KKASSERT(flg->running == 0);
2176 RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2177 if (--hmp->flusher.group_lock == 0)
2178 wakeup(&hmp->flusher.group_lock);
2181 * Auto-flush the group if it grows too large. Make sure the
2182 * inode reclaim wait pipeline continues to work.
2184 if (flg->total_count >= hammer_autoflush ||
2185 flg->total_count >= hammer_limit_reclaims / 4) {
2186 if (hmp->fill_flush_group == flg)
2187 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2188 hammer_flusher_async(hmp, flg);
2193 * Callback for scan of ip->rec_tree. Try to include each record in our
2194 * flush. ip->flush_group has been set but the inode has not yet been
2195 * moved into a flushing state.
2197 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2200 * We return 1 for any record placed or found in FST_FLUSH, which prevents
2201 * the caller from shortcutting the flush.
2204 hammer_setup_child_callback(hammer_record_t rec, void *data)
2206 hammer_flush_group_t flg;
2207 hammer_inode_t target_ip;
2212 * Records deleted or committed by the backend are ignored.
2213 * Note that the flush detects deleted frontend records at
2214 * multiple points to deal with races. This is just the first
2215 * line of defense. The only time HAMMER_RECF_DELETED_FE cannot
2216 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2217 * messes up link-count calculations.
2219 * NOTE: Don't get confused between record deletion and, say,
2220 * directory entry deletion. The deletion of a directory entry
2221 * which is on-media has nothing to do with the record deletion
2224 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2225 HAMMER_RECF_COMMITTED)) {
2226 if (rec->flush_state == HAMMER_FST_FLUSH) {
2227 KKASSERT(rec->flush_group == rec->ip->flush_group);
2236 * If the record is in an idle state it has no dependancies and
2240 flg = ip->flush_group;
2243 switch(rec->flush_state) {
2244 case HAMMER_FST_IDLE:
2246 * The record has no setup dependancy, we can flush it.
2248 KKASSERT(rec->target_ip == NULL);
2249 rec->flush_state = HAMMER_FST_FLUSH;
2250 rec->flush_group = flg;
2252 hammer_ref(&rec->lock);
2255 case HAMMER_FST_SETUP:
2257 * The record has a setup dependancy. These are typically
2258 * directory entry adds and deletes. Such entries will be
2259 * flushed when their inodes are flushed so we do not
2260 * usually have to add them to the flush here. However,
2261 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2262 * it is asking us to flush this record (and it).
2264 target_ip = rec->target_ip;
2265 KKASSERT(target_ip != NULL);
2266 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2269 * If the target IP is already flushing in our group
2270 * we could associate the record, but target_ip has
2271 * already synced ino_data to sync_ino_data and we
2272 * would also have to adjust nlinks. Plus there are
2273 * ordering issues for adds and deletes.
2275 * Reflush downward if this is an ADD, and upward if
2278 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2279 if (rec->type == HAMMER_MEM_RECORD_ADD)
2280 ip->flags |= HAMMER_INODE_REFLUSH;
2282 target_ip->flags |= HAMMER_INODE_REFLUSH;
2287 * Target IP is not yet flushing. This can get complex
2288 * because we have to be careful about the recursion.
2290 * Directories create an issue for us in that if a flush
2291 * of a directory is requested the expectation is to flush
2292 * any pending directory entries, but this will cause the
2293 * related inodes to recursively flush as well. We can't
2294 * really defer the operation so just get as many as we
2298 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2299 (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2301 * We aren't reclaiming and the target ip was not
2302 * previously prevented from flushing due to this
2303 * record dependancy. Do not flush this record.
2308 if (flg->total_count + flg->refs >
2309 ip->hmp->undo_rec_limit) {
2311 * Our flush group is over-full and we risk blowing
2312 * out the UNDO FIFO. Stop the scan, flush what we
2313 * have, then reflush the directory.
2315 * The directory may be forced through multiple
2316 * flush groups before it can be completely
2319 ip->flags |= HAMMER_INODE_RESIGNAL |
2320 HAMMER_INODE_REFLUSH;
2322 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2324 * If the target IP is not flushing we can force
2325 * it to flush, even if it is unable to write out
2326 * any of its own records we have at least one in
2327 * hand that we CAN deal with.
2329 rec->flush_state = HAMMER_FST_FLUSH;
2330 rec->flush_group = flg;
2332 hammer_ref(&rec->lock);
2333 hammer_flush_inode_core(target_ip, flg,
2334 HAMMER_FLUSH_RECURSION);
2338 * General or delete-on-disk record.
2340 * XXX this needs help. If a delete-on-disk we could
2341 * disconnect the target. If the target has its own
2342 * dependancies they really need to be flushed.
2346 rec->flush_state = HAMMER_FST_FLUSH;
2347 rec->flush_group = flg;
2349 hammer_ref(&rec->lock);
2350 hammer_flush_inode_core(target_ip, flg,
2351 HAMMER_FLUSH_RECURSION);
2355 case HAMMER_FST_FLUSH:
2357 * The record could be part of a previous flush group if the
2358 * inode is a directory (the record being a directory entry).
2359 * Once the flush group was closed a hammer_test_inode()
2360 * function can cause a new flush group to be setup, placing
2361 * the directory inode itself in a new flush group.
2363 * When associated with a previous flush group we count it
2364 * as if it were in our current flush group, since it will
2365 * effectively be flushed by the time we flush our current
2369 rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2370 rec->flush_group == flg);
2379 * This version just moves records already in a flush state to the new
2380 * flush group and that is it.
2383 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2385 hammer_inode_t ip = rec->ip;
2387 switch(rec->flush_state) {
2388 case HAMMER_FST_FLUSH:
2389 KKASSERT(rec->flush_group == ip->flush_group);
2399 * Wait for a previously queued flush to complete.
2401 * If a critical error occured we don't try to wait.
2404 hammer_wait_inode(hammer_inode_t ip)
2406 hammer_flush_group_t flg;
2411 * The inode can be in a SETUP state in which case RESIGNAL
2412 * should be set. If RESIGNAL is not set then the previous
2413 * flush completed and a later operation placed the inode
2414 * in a passive setup state again, so we're done.
2416 * The inode can be in a FLUSH state in which case we
2417 * can just wait for completion.
2419 while (ip->flush_state == HAMMER_FST_FLUSH ||
2420 (ip->flush_state == HAMMER_FST_SETUP &&
2421 (ip->flags & HAMMER_INODE_RESIGNAL))) {
2423 * Don't try to flush on a critical error
2425 if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
2429 * If the inode was already being flushed its flg
2430 * may not have been queued to the backend. We
2431 * have to make sure it gets queued or we can wind
2432 * up blocked or deadlocked (particularly if we are
2433 * the vnlru thread).
2435 if (ip->flush_state == HAMMER_FST_FLUSH) {
2436 KKASSERT(ip->flush_group);
2437 if (ip->flush_group->closed == 0) {
2438 kprintf("hammer: debug: forcing async "
2439 "flush ip %016jx\n",
2440 (intmax_t)ip->obj_id);
2441 hammer_flusher_async(ip->hmp,
2443 continue; /* retest */
2448 * In a flush state with the flg queued to the backend
2449 * or in a setup state with RESIGNAL set, we can safely
2452 ip->flags |= HAMMER_INODE_FLUSHW;
2453 tsleep(&ip->flags, 0, "hmrwin", 0);
2458 * The inode may have been in a passive setup state,
2459 * call flush to make sure we get signaled.
2461 if (ip->flush_state == HAMMER_FST_SETUP)
2462 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2468 * Called by the backend code when a flush has been completed.
2469 * The inode has already been removed from the flush list.
2471 * A pipelined flush can occur, in which case we must re-enter the
2472 * inode on the list and re-copy its fields.
2475 hammer_flush_inode_done(hammer_inode_t ip, int error)
2480 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2485 * Auto-reflush if the backend could not completely flush
2486 * the inode. This fixes a case where a deferred buffer flush
2487 * could cause fsync to return early.
2489 if (ip->sync_flags & HAMMER_INODE_MODMASK)
2490 ip->flags |= HAMMER_INODE_REFLUSH;
2493 * Merge left-over flags back into the frontend and fix the state.
2494 * Incomplete truncations are retained by the backend.
2497 ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2498 ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2501 * The backend may have adjusted nlinks, so if the adjusted nlinks
2502 * does not match the fronttend set the frontend's DDIRTY flag again.
2504 if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2505 ip->flags |= HAMMER_INODE_DDIRTY;
2508 * Fix up the dirty buffer status.
2510 if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2511 ip->flags |= HAMMER_INODE_BUFS;
2513 hammer_redo_fifo_end_flush(ip);
2516 * Re-set the XDIRTY flag if some of the inode's in-memory records
2517 * could not be flushed.
2519 KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2520 (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2521 (!RB_EMPTY(&ip->rec_tree) &&
2522 (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2525 * Do not lose track of inodes which no longer have vnode
2526 * assocations, otherwise they may never get flushed again.
2528 * The reflush flag can be set superfluously, causing extra pain
2529 * for no reason. If the inode is no longer modified it no longer
2530 * needs to be flushed.
2532 if (ip->flags & HAMMER_INODE_MODMASK) {
2534 ip->flags |= HAMMER_INODE_REFLUSH;
2536 ip->flags &= ~HAMMER_INODE_REFLUSH;
2540 * Adjust the flush state.
2542 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2544 * We were unable to flush out all our records, leave the
2545 * inode in a flush state and in the current flush group.
2546 * The flush group will be re-run.
2548 * This occurs if the UNDO block gets too full or there is
2549 * too much dirty meta-data and allows the flusher to
2550 * finalize the UNDO block and then re-flush.
2552 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2556 * Remove from the flush_group
2558 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2559 ip->flush_group = NULL;
2563 * Clean up the vnode ref and tracking counts.
2565 if (ip->flags & HAMMER_INODE_VHELD) {
2566 ip->flags &= ~HAMMER_INODE_VHELD;
2570 --hmp->count_iqueued;
2571 --hammer_count_iqueued;
2574 * And adjust the state.
2576 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2577 ip->flush_state = HAMMER_FST_IDLE;
2580 ip->flush_state = HAMMER_FST_SETUP;
2585 * If the frontend is waiting for a flush to complete,
2588 if (ip->flags & HAMMER_INODE_FLUSHW) {
2589 ip->flags &= ~HAMMER_INODE_FLUSHW;
2594 * If the frontend made more changes and requested another
2595 * flush, then try to get it running.
2597 * Reflushes are aborted when the inode is errored out.
2599 if (ip->flags & HAMMER_INODE_REFLUSH) {
2600 ip->flags &= ~HAMMER_INODE_REFLUSH;
2601 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2602 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2603 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2605 hammer_flush_inode(ip, 0);
2611 * If we have no parent dependancies we can clear CONN_DOWN
2613 if (TAILQ_EMPTY(&ip->target_list))
2614 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2617 * If the inode is now clean drop the space reservation.
2619 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2620 (ip->flags & HAMMER_INODE_RSV_INODES)) {
2621 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2625 ip->flags &= ~HAMMER_INODE_SLAVEFLUSH;
2628 hammer_rel_inode(ip, 0);
2632 * Called from hammer_sync_inode() to synchronize in-memory records
2636 hammer_sync_record_callback(hammer_record_t record, void *data)
2638 hammer_cursor_t cursor = data;
2639 hammer_transaction_t trans = cursor->trans;
2640 hammer_mount_t hmp = trans->hmp;
2644 * Skip records that do not belong to the current flush.
2646 ++hammer_stats_record_iterations;
2647 if (record->flush_state != HAMMER_FST_FLUSH)
2651 if (record->flush_group != record->ip->flush_group) {
2652 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2653 if (hammer_debug_critical)
2658 KKASSERT(record->flush_group == record->ip->flush_group);
2661 * Interlock the record using the BE flag. Once BE is set the
2662 * frontend cannot change the state of FE.
2664 * NOTE: If FE is set prior to us setting BE we still sync the
2665 * record out, but the flush completion code converts it to
2666 * a delete-on-disk record instead of destroying it.
2668 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2669 record->flags |= HAMMER_RECF_INTERLOCK_BE;
2672 * The backend has already disposed of the record.
2674 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2680 * If the whole inode is being deleted and all on-disk records will
2681 * be deleted very soon, we can't sync any new records to disk
2682 * because they will be deleted in the same transaction they were
2683 * created in (delete_tid == create_tid), which will assert.
2685 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2686 * that we currently panic on.
2688 if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2689 switch(record->type) {
2690 case HAMMER_MEM_RECORD_DATA:
2692 * We don't have to do anything, if the record was
2693 * committed the space will have been accounted for
2697 case HAMMER_MEM_RECORD_GENERAL:
2699 * Set deleted-by-backend flag. Do not set the
2700 * backend committed flag, because we are throwing
2703 record->flags |= HAMMER_RECF_DELETED_BE;
2704 ++record->ip->rec_generation;
2707 case HAMMER_MEM_RECORD_ADD:
2708 panic("hammer_sync_record_callback: illegal add "
2709 "during inode deletion record %p", record);
2710 break; /* NOT REACHED */
2711 case HAMMER_MEM_RECORD_INODE:
2712 panic("hammer_sync_record_callback: attempt to "
2713 "sync inode record %p?", record);
2714 break; /* NOT REACHED */
2715 case HAMMER_MEM_RECORD_DEL:
2717 * Follow through and issue the on-disk deletion
2724 * If DELETED_FE is set special handling is needed for directory
2725 * entries. Dependant pieces related to the directory entry may
2726 * have already been synced to disk. If this occurs we have to
2727 * sync the directory entry and then change the in-memory record
2728 * from an ADD to a DELETE to cover the fact that it's been
2729 * deleted by the frontend.
2731 * A directory delete covering record (MEM_RECORD_DEL) can never
2732 * be deleted by the frontend.
2734 * Any other record type (aka DATA) can be deleted by the frontend.
2735 * XXX At the moment the flusher must skip it because there may
2736 * be another data record in the flush group for the same block,
2737 * meaning that some frontend data changes can leak into the backend's
2738 * synchronization point.
2740 if (record->flags & HAMMER_RECF_DELETED_FE) {
2741 if (record->type == HAMMER_MEM_RECORD_ADD) {
2743 * Convert a front-end deleted directory-add to
2744 * a directory-delete entry later.
2746 record->flags |= HAMMER_RECF_CONVERT_DELETE;
2749 * Dispose of the record (race case). Mark as
2750 * deleted by backend (and not committed).
2752 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2753 record->flags |= HAMMER_RECF_DELETED_BE;
2754 ++record->ip->rec_generation;
2761 * Assign the create_tid for new records. Deletions already
2762 * have the record's entire key properly set up.
2764 if (record->type != HAMMER_MEM_RECORD_DEL) {
2765 record->leaf.base.create_tid = trans->tid;
2766 record->leaf.create_ts = trans->time32;
2770 * This actually moves the record to the on-media B-Tree. We
2771 * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2772 * indicating that the related REDO_WRITE(s) have been committed.
2774 * During recovery any REDO_TERM's within the nominal recovery span
2775 * are ignored since the related meta-data is being undone, causing
2776 * any matching REDO_WRITEs to execute. The REDO_TERMs outside
2777 * the nominal recovery span will match against REDO_WRITEs and
2778 * prevent them from being executed (because the meta-data has
2779 * already been synchronized).
2781 if (record->flags & HAMMER_RECF_REDO) {
2782 KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2783 hammer_generate_redo(trans, record->ip,
2784 record->leaf.base.key -
2785 record->leaf.data_len,
2786 HAMMER_REDO_TERM_WRITE,
2788 record->leaf.data_len);
2792 error = hammer_ip_sync_record_cursor(cursor, record);
2793 if (error != EDEADLK)
2795 hammer_done_cursor(cursor);
2796 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2801 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2806 hammer_flush_record_done(record, error);
2809 * Do partial finalization if we have built up too many dirty
2810 * buffers. Otherwise a buffer cache deadlock can occur when
2811 * doing things like creating tens of thousands of tiny files.
2813 * We must release our cursor lock to avoid a 3-way deadlock
2814 * due to the exclusive sync lock the finalizer must get.
2816 * WARNING: See warnings in hammer_unlock_cursor() function.
2818 if (hammer_flusher_meta_limit(hmp) ||
2819 vm_page_count_severe()) {
2820 hammer_unlock_cursor(cursor);
2821 hammer_flusher_finalize(trans, 0);
2822 hammer_lock_cursor(cursor);
2828 * Backend function called by the flusher to sync an inode to media.
2831 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2833 struct hammer_cursor cursor;
2834 hammer_node_t tmp_node;
2835 hammer_record_t depend;
2836 hammer_record_t next;
2837 int error, tmp_error;
2840 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2843 error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2848 * Any directory records referencing this inode which are not in
2849 * our current flush group must adjust our nlink count for the
2850 * purposes of synchronizating to disk.
2852 * Records which are in our flush group can be unlinked from our
2853 * inode now, potentially allowing the inode to be physically
2856 * This cannot block.
2858 nlinks = ip->ino_data.nlinks;
2859 next = TAILQ_FIRST(&ip->target_list);
2860 while ((depend = next) != NULL) {
2861 next = TAILQ_NEXT(depend, target_entry);
2862 if (depend->flush_state == HAMMER_FST_FLUSH &&
2863 depend->flush_group == ip->flush_group) {
2865 * If this is an ADD that was deleted by the frontend
2866 * the frontend nlinks count will have already been
2867 * decremented, but the backend is going to sync its
2868 * directory entry and must account for it. The
2869 * record will be converted to a delete-on-disk when
2872 * If the ADD was not deleted by the frontend we
2873 * can remove the dependancy from our target_list.
2875 if (depend->flags & HAMMER_RECF_DELETED_FE) {
2878 TAILQ_REMOVE(&ip->target_list, depend,
2880 depend->target_ip = NULL;
2882 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2884 * Not part of our flush group and not deleted by
2885 * the front-end, adjust the link count synced to
2886 * the media (undo what the frontend did when it
2887 * queued the record).
2889 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2890 switch(depend->type) {
2891 case HAMMER_MEM_RECORD_ADD:
2894 case HAMMER_MEM_RECORD_DEL:
2904 * Set dirty if we had to modify the link count.
2906 if (ip->sync_ino_data.nlinks != nlinks) {
2907 KKASSERT((int64_t)nlinks >= 0);
2908 ip->sync_ino_data.nlinks = nlinks;
2909 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2913 * If there is a trunction queued destroy any data past the (aligned)
2914 * truncation point. Userland will have dealt with the buffer
2915 * containing the truncation point for us.
2917 * We don't flush pending frontend data buffers until after we've
2918 * dealt with the truncation.
2920 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2922 * Interlock trunc_off. The VOP front-end may continue to
2923 * make adjustments to it while we are blocked.
2926 off_t aligned_trunc_off;
2929 trunc_off = ip->sync_trunc_off;
2930 blkmask = hammer_blocksize(trunc_off) - 1;
2931 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2934 * Delete any whole blocks on-media. The front-end has
2935 * already cleaned out any partial block and made it
2936 * pending. The front-end may have updated trunc_off
2937 * while we were blocked so we only use sync_trunc_off.
2939 * This operation can blow out the buffer cache, EWOULDBLOCK
2940 * means we were unable to complete the deletion. The
2941 * deletion will update sync_trunc_off in that case.
2943 error = hammer_ip_delete_range(&cursor, ip,
2945 0x7FFFFFFFFFFFFFFFLL, 2);
2946 if (error == EWOULDBLOCK) {
2947 ip->flags |= HAMMER_INODE_WOULDBLOCK;
2949 goto defer_buffer_flush;
2956 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
2958 * XXX we do this even if we did not previously generate
2959 * a REDO_TRUNC record. This operation may enclosed the
2960 * range for multiple prior truncation entries in the REDO
2963 if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
2964 (ip->flags & HAMMER_INODE_RDIRTY)) {
2965 hammer_generate_redo(trans, ip, aligned_trunc_off,
2966 HAMMER_REDO_TERM_TRUNC,
2971 * Clear the truncation flag on the backend after we have
2972 * completed the deletions. Backend data is now good again
2973 * (including new records we are about to sync, below).
2975 * Leave sync_trunc_off intact. As we write additional
2976 * records the backend will update sync_trunc_off. This
2977 * tells the backend whether it can skip the overwrite
2978 * test. This should work properly even when the backend
2979 * writes full blocks where the truncation point straddles
2980 * the block because the comparison is against the base
2981 * offset of the record.
2983 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2984 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2990 * Now sync related records. These will typically be directory
2991 * entries, records tracking direct-writes, or delete-on-disk records.
2994 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2995 hammer_sync_record_callback, &cursor);
3001 hammer_cache_node(&ip->cache[1], cursor.node);
3004 * Re-seek for inode update, assuming our cache hasn't been ripped
3005 * out from under us.
3008 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
3010 hammer_cursor_downgrade(&cursor);
3011 hammer_lock_sh(&tmp_node->lock);
3012 if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
3013 hammer_cursor_seek(&cursor, tmp_node, 0);
3014 hammer_unlock(&tmp_node->lock);
3015 hammer_rel_node(tmp_node);
3021 * If we are deleting the inode the frontend had better not have
3022 * any active references on elements making up the inode.
3024 * The call to hammer_ip_delete_clean() cleans up auxillary records
3025 * but not DB or DATA records. Those must have already been deleted
3026 * by the normal truncation mechanic.
3028 if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
3029 RB_EMPTY(&ip->rec_tree) &&
3030 (ip->sync_flags & HAMMER_INODE_DELETING) &&
3031 (ip->flags & HAMMER_INODE_DELETED) == 0) {
3034 error = hammer_ip_delete_clean(&cursor, ip, &count1);
3036 ip->flags |= HAMMER_INODE_DELETED;
3037 ip->sync_flags &= ~HAMMER_INODE_DELETING;
3038 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3039 KKASSERT(RB_EMPTY(&ip->rec_tree));
3042 * Set delete_tid in both the frontend and backend
3043 * copy of the inode record. The DELETED flag handles
3044 * this, do not set DDIRTY.
3046 ip->ino_leaf.base.delete_tid = trans->tid;
3047 ip->sync_ino_leaf.base.delete_tid = trans->tid;
3048 ip->ino_leaf.delete_ts = trans->time32;
3049 ip->sync_ino_leaf.delete_ts = trans->time32;
3053 * Adjust the inode count in the volume header
3055 hammer_sync_lock_sh(trans);
3056 if (ip->flags & HAMMER_INODE_ONDISK) {
3057 hammer_modify_volume_field(trans,
3060 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
3061 hammer_modify_volume_done(trans->rootvol);
3063 hammer_sync_unlock(trans);
3069 ip->sync_flags &= ~HAMMER_INODE_BUFS;
3073 * Now update the inode's on-disk inode-data and/or on-disk record.
3074 * DELETED and ONDISK are managed only in ip->flags.
3076 * In the case of a defered buffer flush we still update the on-disk
3077 * inode to satisfy visibility requirements if there happen to be
3078 * directory dependancies.
3080 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
3081 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3083 * If deleted and on-disk, don't set any additional flags.
3084 * the delete flag takes care of things.
3086 * Clear flags which may have been set by the frontend.
3088 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3089 HAMMER_INODE_SDIRTY |
3090 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3091 HAMMER_INODE_DELETING);
3093 case HAMMER_INODE_DELETED:
3095 * Take care of the case where a deleted inode was never
3096 * flushed to the disk in the first place.
3098 * Clear flags which may have been set by the frontend.
3100 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3101 HAMMER_INODE_SDIRTY |
3102 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3103 HAMMER_INODE_DELETING);
3104 while (RB_ROOT(&ip->rec_tree)) {
3105 hammer_record_t record = RB_ROOT(&ip->rec_tree);
3106 hammer_ref(&record->lock);
3107 KKASSERT(hammer_oneref(&record->lock));
3108 record->flags |= HAMMER_RECF_DELETED_BE;
3109 ++record->ip->rec_generation;
3110 hammer_rel_mem_record(record);
3113 case HAMMER_INODE_ONDISK:
3115 * If already on-disk, do not set any additional flags.
3120 * If not on-disk and not deleted, set DDIRTY to force
3121 * an initial record to be written.
3123 * Also set the create_tid in both the frontend and backend
3124 * copy of the inode record.
3126 ip->ino_leaf.base.create_tid = trans->tid;
3127 ip->ino_leaf.create_ts = trans->time32;
3128 ip->sync_ino_leaf.base.create_tid = trans->tid;
3129 ip->sync_ino_leaf.create_ts = trans->time32;
3130 ip->sync_flags |= HAMMER_INODE_DDIRTY;
3135 * If DDIRTY or SDIRTY is set, write out a new record.
3136 * If the inode is already on-disk the old record is marked as
3139 * If DELETED is set hammer_update_inode() will delete the existing
3140 * record without writing out a new one.
3142 * If *ONLY* the ITIMES flag is set we can update the record in-place.
3144 if (ip->flags & HAMMER_INODE_DELETED) {
3145 error = hammer_update_inode(&cursor, ip);
3147 if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3148 (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3149 error = hammer_update_itimes(&cursor, ip);
3151 if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3152 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3153 error = hammer_update_inode(&cursor, ip);
3157 hammer_critical_error(ip->hmp, ip, error,
3158 "while syncing inode");
3160 hammer_done_cursor(&cursor);
3165 * This routine is called when the OS is no longer actively referencing
3166 * the inode (but might still be keeping it cached), or when releasing
3167 * the last reference to an inode.
3169 * At this point if the inode's nlinks count is zero we want to destroy
3170 * it, which may mean destroying it on-media too.
3173 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3178 * Set the DELETING flag when the link count drops to 0 and the
3179 * OS no longer has any opens on the inode.
3181 * The backend will clear DELETING (a mod flag) and set DELETED
3182 * (a state flag) when it is actually able to perform the
3185 * Don't reflag the deletion if the flusher is currently syncing
3186 * one that was already flagged. A previously set DELETING flag
3187 * may bounce around flags and sync_flags until the operation is
3190 * Do not attempt to modify a snapshot inode (one set to read-only).
3192 if (ip->ino_data.nlinks == 0 &&
3193 ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3194 ip->flags |= HAMMER_INODE_DELETING;
3195 ip->flags |= HAMMER_INODE_TRUNCATED;
3199 if (hammer_get_vnode(ip, &vp) != 0)
3207 nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0);
3214 * After potentially resolving a dependancy the inode is tested
3215 * to determine whether it needs to be reflushed.
3218 hammer_test_inode(hammer_inode_t ip)
3220 if (ip->flags & HAMMER_INODE_REFLUSH) {
3221 ip->flags &= ~HAMMER_INODE_REFLUSH;
3222 hammer_ref(&ip->lock);
3223 if (ip->flags & HAMMER_INODE_RESIGNAL) {
3224 ip->flags &= ~HAMMER_INODE_RESIGNAL;
3225 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3227 hammer_flush_inode(ip, 0);
3229 hammer_rel_inode(ip, 0);
3234 * Clear the RECLAIM flag on an inode. This occurs when the inode is
3235 * reassociated with a vp or just before it gets freed.
3237 * Pipeline wakeups to threads blocked due to an excessive number of
3238 * detached inodes. This typically occurs when atime updates accumulate
3239 * while scanning a directory tree.
3242 hammer_inode_wakereclaims(hammer_inode_t ip)
3244 struct hammer_reclaim *reclaim;
3245 hammer_mount_t hmp = ip->hmp;
3247 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3250 --hammer_count_reclaims;
3251 --hmp->count_reclaims;
3252 ip->flags &= ~HAMMER_INODE_RECLAIM;
3254 if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3255 KKASSERT(reclaim->count > 0);
3256 if (--reclaim->count == 0) {
3257 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3264 * Setup our reclaim pipeline. We only let so many detached (and dirty)
3265 * inodes build up before we start blocking. This routine is called
3266 * if a new inode is created or an inode is loaded from media.
3268 * When we block we don't care *which* inode has finished reclaiming,
3269 * as long as one does.
3271 * The reclaim pipeline is primarily governed by the auto-flush which is
3272 * 1/4 hammer_limit_reclaims. We don't want to block if the count is
3273 * less than 1/2 hammer_limit_reclaims. From 1/2 to full count is
3274 * dynamically governed.
3277 hammer_inode_waitreclaims(hammer_transaction_t trans)
3279 hammer_mount_t hmp = trans->hmp;
3280 struct hammer_reclaim reclaim;
3284 * Track inode load, delay if the number of reclaiming inodes is
3285 * between 2/4 and 4/4 hammer_limit_reclaims, depending.
3287 if (curthread->td_proc) {
3288 struct hammer_inostats *stats;
3290 stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3293 if (stats->count > hammer_limit_reclaims / 2)
3294 stats->count = hammer_limit_reclaims / 2;
3295 lower_limit = hammer_limit_reclaims - stats->count;
3296 if (hammer_debug_general & 0x10000) {
3297 kprintf("pid %5d limit %d\n",
3298 (int)curthread->td_proc->p_pid, lower_limit);
3301 lower_limit = hammer_limit_reclaims * 3 / 4;
3303 if (hmp->count_reclaims >= lower_limit) {
3305 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3306 tsleep(&reclaim, 0, "hmrrcm", hz);
3307 if (reclaim.count > 0)
3308 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3313 * Keep track of reclaim statistics on a per-pid basis using a loose
3314 * 4-way set associative hash table. Collisions inherit the count of
3315 * the previous entry.
3317 * NOTE: We want to be careful here to limit the chain size. If the chain
3318 * size is too large a pid will spread its stats out over too many
3319 * entries under certain types of heavy filesystem activity and
3320 * wind up not delaying long enough.
3323 struct hammer_inostats *
3324 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3326 struct hammer_inostats *stats;
3329 static volatile int iterator; /* we don't care about MP races */
3332 * Chain up to 4 times to find our entry.
3334 for (chain = 0; chain < 4; ++chain) {
3335 stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3336 if (stats->pid == pid)
3341 * Replace one of the four chaining entries with our new entry.
3344 stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3345 HAMMER_INOSTATS_HMASK];
3352 if (stats->count && stats->ltick != ticks) {
3353 delta = ticks - stats->ltick;
3354 stats->ltick = ticks;
3355 if (delta <= 0 || delta > hz * 60)
3358 stats->count = stats->count * hz / (hz + delta);
3360 if (hammer_debug_general & 0x10000)
3361 kprintf("pid %5d stats %d\n", (int)pid, stats->count);
3368 * XXX not used, doesn't work very well due to the large batching nature
3371 * A larger then normal backlog of inodes is sitting in the flusher,
3372 * enforce a general slowdown to let it catch up. This routine is only
3373 * called on completion of a non-flusher-related transaction which
3374 * performed B-Tree node I/O.
3376 * It is possible for the flusher to stall in a continuous load.
3377 * blogbench -i1000 -o seems to do a good job generating this sort of load.
3378 * If the flusher is unable to catch up the inode count can bloat until
3379 * we run out of kvm.
3381 * This is a bit of a hack.
3384 hammer_inode_waithard(hammer_mount_t hmp)
3389 if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3390 if (hmp->count_reclaims < hammer_limit_reclaims / 2 &&
3391 hmp->count_iqueued < hmp->count_inodes / 20) {
3392 hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3396 if (hmp->count_reclaims < hammer_limit_reclaims ||
3397 hmp->count_iqueued < hmp->count_inodes / 10) {
3400 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3404 * Block for one flush cycle.
3406 hammer_flusher_wait_next(hmp);