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>
42 static int hammer_unload_inode(struct hammer_inode *ip);
43 static void hammer_free_inode(hammer_inode_t ip);
44 static void hammer_flush_inode_core(hammer_inode_t ip,
45 hammer_flush_group_t flg, int flags);
46 static int hammer_setup_child_callback(hammer_record_t rec, void *data);
48 static int hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
50 static int hammer_setup_parent_inodes(hammer_inode_t ip,
51 hammer_flush_group_t flg);
52 static int hammer_setup_parent_inodes_helper(hammer_record_t record,
53 hammer_flush_group_t flg);
54 static void hammer_inode_wakereclaims(hammer_inode_t ip, int dowake);
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 * RB-Tree support for inode structures / special LOOKUP_INFO
85 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
87 if (info->obj_localization < ip->obj_localization)
89 if (info->obj_localization > ip->obj_localization)
91 if (info->obj_id < ip->obj_id)
93 if (info->obj_id > ip->obj_id)
95 if (info->obj_asof < ip->obj_asof)
97 if (info->obj_asof > ip->obj_asof)
103 * Used by hammer_scan_inode_snapshots() to locate all of an object's
104 * snapshots. Note that the asof field is not tested, which we can get
105 * away with because it is the lowest-priority field.
108 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
110 hammer_inode_info_t info = data;
112 if (ip->obj_localization > info->obj_localization)
114 if (ip->obj_localization < info->obj_localization)
116 if (ip->obj_id > info->obj_id)
118 if (ip->obj_id < info->obj_id)
124 * Used by hammer_unload_pseudofs() to locate all inodes associated with
128 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
130 u_int32_t localization = *(u_int32_t *)data;
131 if (ip->obj_localization > localization)
133 if (ip->obj_localization < localization)
139 * RB-Tree support for pseudofs structures
142 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
144 if (p1->localization < p2->localization)
146 if (p1->localization > p2->localization)
152 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
153 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
154 hammer_inode_info_cmp, hammer_inode_info_t);
155 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
156 hammer_pfs_rb_compare, u_int32_t, localization);
159 * The kernel is not actively referencing this vnode but is still holding
162 * This is called from the frontend.
165 hammer_vop_inactive(struct vop_inactive_args *ap)
167 struct hammer_inode *ip = VTOI(ap->a_vp);
178 * If the inode no longer has visibility in the filesystem try to
179 * recycle it immediately, even if the inode is dirty. Recycling
180 * it quickly allows the system to reclaim buffer cache and VM
181 * resources which can matter a lot in a heavily loaded system.
183 * This can deadlock in vfsync() if we aren't careful.
185 * Do not queue the inode to the flusher if we still have visibility,
186 * otherwise namespace calls such as chmod will unnecessarily generate
187 * multiple inode updates.
189 hammer_inode_unloadable_check(ip, 0);
190 if (ip->ino_data.nlinks == 0) {
191 if (ip->flags & HAMMER_INODE_MODMASK)
192 hammer_flush_inode(ip, 0);
199 * Release the vnode association. This is typically (but not always)
200 * the last reference on the inode.
202 * Once the association is lost we are on our own with regards to
203 * flushing the inode.
206 hammer_vop_reclaim(struct vop_reclaim_args *ap)
208 struct hammer_inode *ip;
214 if ((ip = vp->v_data) != NULL) {
219 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
220 ++hammer_count_reclaiming;
221 ++hmp->inode_reclaims;
222 ip->flags |= HAMMER_INODE_RECLAIM;
224 hammer_rel_inode(ip, 1);
230 * Return a locked vnode for the specified inode. The inode must be
231 * referenced but NOT LOCKED on entry and will remain referenced on
234 * Called from the frontend.
237 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
247 if ((vp = ip->vp) == NULL) {
248 error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
251 hammer_lock_ex(&ip->lock);
252 if (ip->vp != NULL) {
253 hammer_unlock(&ip->lock);
258 hammer_ref(&ip->lock);
262 obj_type = ip->ino_data.obj_type;
263 vp->v_type = hammer_get_vnode_type(obj_type);
265 hammer_inode_wakereclaims(ip, 0);
267 switch(ip->ino_data.obj_type) {
268 case HAMMER_OBJTYPE_CDEV:
269 case HAMMER_OBJTYPE_BDEV:
270 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
271 addaliasu(vp, ip->ino_data.rmajor,
272 ip->ino_data.rminor);
274 case HAMMER_OBJTYPE_FIFO:
275 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
282 * Only mark as the root vnode if the ip is not
283 * historical, otherwise the VFS cache will get
284 * confused. The other half of the special handling
285 * is in hammer_vop_nlookupdotdot().
287 * Pseudo-filesystem roots can be accessed via
288 * non-root filesystem paths and setting VROOT may
289 * confuse the namecache. Set VPFSROOT instead.
291 if (ip->obj_id == HAMMER_OBJID_ROOT &&
292 ip->obj_asof == hmp->asof) {
293 if (ip->obj_localization == 0)
296 vp->v_flag |= VPFSROOT;
299 vp->v_data = (void *)ip;
300 /* vnode locked by getnewvnode() */
301 /* make related vnode dirty if inode dirty? */
302 hammer_unlock(&ip->lock);
303 if (vp->v_type == VREG)
304 vinitvmio(vp, ip->ino_data.size);
309 * loop if the vget fails (aka races), or if the vp
310 * no longer matches ip->vp.
312 if (vget(vp, LK_EXCLUSIVE) == 0) {
323 * Locate all copies of the inode for obj_id compatible with the specified
324 * asof, reference, and issue the related call-back. This routine is used
325 * for direct-io invalidation and does not create any new inodes.
328 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
329 int (*callback)(hammer_inode_t ip, void *data),
332 hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
333 hammer_inode_info_cmp_all_history,
338 * Acquire a HAMMER inode. The returned inode is not locked. These functions
339 * do not attach or detach the related vnode (use hammer_get_vnode() for
342 * The flags argument is only applied for newly created inodes, and only
343 * certain flags are inherited.
345 * Called from the frontend.
347 struct hammer_inode *
348 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
349 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
350 int flags, int *errorp)
352 hammer_mount_t hmp = trans->hmp;
353 struct hammer_inode_info iinfo;
354 struct hammer_cursor cursor;
355 struct hammer_inode *ip;
359 * Determine if we already have an inode cached. If we do then
362 * If we find an inode with no vnode we have to mark the
363 * transaction such that hammer_inode_waitreclaims() is
364 * called later on to avoid building up an infinite number
365 * of inodes. Otherwise we can continue to * add new inodes
366 * faster then they can be disposed of, even with the tsleep
369 iinfo.obj_id = obj_id;
370 iinfo.obj_asof = asof;
371 iinfo.obj_localization = localization;
373 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
377 trans->flags |= HAMMER_TRANSF_NEWINODE;
379 hammer_ref(&ip->lock);
385 * Allocate a new inode structure and deal with races later.
387 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
388 ++hammer_count_inodes;
391 ip->obj_asof = iinfo.obj_asof;
392 ip->obj_localization = localization;
394 ip->flags = flags & HAMMER_INODE_RO;
395 ip->cache[0].ip = ip;
396 ip->cache[1].ip = ip;
398 ip->flags |= HAMMER_INODE_RO;
399 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
400 0x7FFFFFFFFFFFFFFFLL;
401 RB_INIT(&ip->rec_tree);
402 TAILQ_INIT(&ip->target_list);
403 hammer_ref(&ip->lock);
406 * Locate the on-disk inode. If this is a PFS root we always
407 * access the current version of the root inode and (if it is not
408 * a master) always access information under it with a snapshot
412 hammer_init_cursor(trans, &cursor, (dip ? &dip->cache[0] : NULL), NULL);
413 cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
414 cursor.key_beg.obj_id = ip->obj_id;
415 cursor.key_beg.key = 0;
416 cursor.key_beg.create_tid = 0;
417 cursor.key_beg.delete_tid = 0;
418 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
419 cursor.key_beg.obj_type = 0;
421 cursor.asof = iinfo.obj_asof;
422 cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
425 *errorp = hammer_btree_lookup(&cursor);
426 if (*errorp == EDEADLK) {
427 hammer_done_cursor(&cursor);
432 * On success the B-Tree lookup will hold the appropriate
433 * buffer cache buffers and provide a pointer to the requested
434 * information. Copy the information to the in-memory inode
435 * and cache the B-Tree node to improve future operations.
438 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
439 ip->ino_data = cursor.data->inode;
442 * cache[0] tries to cache the location of the object inode.
443 * The assumption is that it is near the directory inode.
445 * cache[1] tries to cache the location of the object data.
446 * The assumption is that it is near the directory data.
448 hammer_cache_node(&ip->cache[0], cursor.node);
449 if (dip && dip->cache[1].node)
450 hammer_cache_node(&ip->cache[1], dip->cache[1].node);
453 * The file should not contain any data past the file size
454 * stored in the inode. Setting save_trunc_off to the
455 * file size instead of max reduces B-Tree lookup overheads
456 * on append by allowing the flusher to avoid checking for
459 ip->save_trunc_off = ip->ino_data.size;
462 * Locate and assign the pseudofs management structure to
465 if (dip && dip->obj_localization == ip->obj_localization) {
466 ip->pfsm = dip->pfsm;
467 hammer_ref(&ip->pfsm->lock);
469 ip->pfsm = hammer_load_pseudofs(trans,
470 ip->obj_localization,
472 *errorp = 0; /* ignore ENOENT */
477 * The inode is placed on the red-black tree and will be synced to
478 * the media when flushed or by the filesystem sync. If this races
479 * another instantiation/lookup the insertion will fail.
482 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
483 hammer_free_inode(ip);
484 hammer_done_cursor(&cursor);
487 ip->flags |= HAMMER_INODE_ONDISK;
489 if (ip->flags & HAMMER_INODE_RSV_INODES) {
490 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
494 hammer_free_inode(ip);
497 hammer_done_cursor(&cursor);
498 trans->flags |= HAMMER_TRANSF_NEWINODE;
503 * Create a new filesystem object, returning the inode in *ipp. The
504 * returned inode will be referenced. The inode is created in-memory.
506 * If pfsm is non-NULL the caller wishes to create the root inode for
510 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
511 struct ucred *cred, hammer_inode_t dip,
512 hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
521 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
522 ++hammer_count_inodes;
524 trans->flags |= HAMMER_TRANSF_NEWINODE;
527 KKASSERT(pfsm->localization != 0);
528 ip->obj_id = HAMMER_OBJID_ROOT;
529 ip->obj_localization = pfsm->localization;
531 KKASSERT(dip != NULL);
532 ip->obj_id = hammer_alloc_objid(hmp, dip);
533 ip->obj_localization = dip->obj_localization;
536 KKASSERT(ip->obj_id != 0);
537 ip->obj_asof = hmp->asof;
539 ip->flush_state = HAMMER_FST_IDLE;
540 ip->flags = HAMMER_INODE_DDIRTY |
541 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
542 ip->cache[0].ip = ip;
543 ip->cache[1].ip = ip;
545 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
546 /* ip->save_trunc_off = 0; (already zero) */
547 RB_INIT(&ip->rec_tree);
548 TAILQ_INIT(&ip->target_list);
550 ip->ino_data.atime = trans->time;
551 ip->ino_data.mtime = trans->time;
552 ip->ino_data.size = 0;
553 ip->ino_data.nlinks = 0;
556 * A nohistory designator on the parent directory is inherited by
557 * the child. We will do this even for pseudo-fs creation... the
558 * sysad can turn it off.
561 ip->ino_data.uflags = dip->ino_data.uflags &
562 (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
565 ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
566 ip->ino_leaf.base.localization = ip->obj_localization +
567 HAMMER_LOCALIZE_INODE;
568 ip->ino_leaf.base.obj_id = ip->obj_id;
569 ip->ino_leaf.base.key = 0;
570 ip->ino_leaf.base.create_tid = 0;
571 ip->ino_leaf.base.delete_tid = 0;
572 ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
573 ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
575 ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
576 ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
577 ip->ino_data.mode = vap->va_mode;
578 ip->ino_data.ctime = trans->time;
581 * If we are running version 2 or greater we use dirhash algorithm #1
582 * which is semi-sorted. Algorithm #0 was just a pure crc.
584 if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
585 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
586 ip->ino_data.cap_flags |= HAMMER_INODE_CAP_DIRHASH_ALG1;
591 * Setup the ".." pointer. This only needs to be done for directories
592 * but we do it for all objects as a recovery aid.
595 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
598 * The parent_obj_localization field only applies to pseudo-fs roots.
599 * XXX this is no longer applicable, PFSs are no longer directly
600 * tied into the parent's directory structure.
602 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
603 ip->obj_id == HAMMER_OBJID_ROOT) {
604 ip->ino_data.ext.obj.parent_obj_localization =
605 dip->obj_localization;
609 switch(ip->ino_leaf.base.obj_type) {
610 case HAMMER_OBJTYPE_CDEV:
611 case HAMMER_OBJTYPE_BDEV:
612 ip->ino_data.rmajor = vap->va_rmajor;
613 ip->ino_data.rminor = vap->va_rminor;
620 * Calculate default uid/gid and overwrite with information from
624 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
625 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
626 xuid, cred, &vap->va_mode);
630 ip->ino_data.mode = vap->va_mode;
632 if (vap->va_vaflags & VA_UID_UUID_VALID)
633 ip->ino_data.uid = vap->va_uid_uuid;
634 else if (vap->va_uid != (uid_t)VNOVAL)
635 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
637 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
639 if (vap->va_vaflags & VA_GID_UUID_VALID)
640 ip->ino_data.gid = vap->va_gid_uuid;
641 else if (vap->va_gid != (gid_t)VNOVAL)
642 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
644 ip->ino_data.gid = dip->ino_data.gid;
646 hammer_ref(&ip->lock);
650 hammer_ref(&pfsm->lock);
652 } else if (dip->obj_localization == ip->obj_localization) {
653 ip->pfsm = dip->pfsm;
654 hammer_ref(&ip->pfsm->lock);
657 ip->pfsm = hammer_load_pseudofs(trans,
658 ip->obj_localization,
660 error = 0; /* ignore ENOENT */
664 hammer_free_inode(ip);
666 } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
667 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
669 hammer_free_inode(ip);
676 * Final cleanup / freeing of an inode structure
679 hammer_free_inode(hammer_inode_t ip)
681 struct hammer_mount *hmp;
684 KKASSERT(ip->lock.refs == 1);
685 hammer_uncache_node(&ip->cache[0]);
686 hammer_uncache_node(&ip->cache[1]);
687 hammer_inode_wakereclaims(ip, 1);
689 hammer_clear_objid(ip);
690 --hammer_count_inodes;
693 hammer_rel_pseudofs(hmp, ip->pfsm);
696 kfree(ip, hmp->m_inodes);
701 * Retrieve pseudo-fs data. NULL will never be returned.
703 * If an error occurs *errorp will be set and a default template is returned,
704 * otherwise *errorp is set to 0. Typically when an error occurs it will
707 hammer_pseudofs_inmem_t
708 hammer_load_pseudofs(hammer_transaction_t trans,
709 u_int32_t localization, int *errorp)
711 hammer_mount_t hmp = trans->hmp;
713 hammer_pseudofs_inmem_t pfsm;
714 struct hammer_cursor cursor;
718 pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
720 hammer_ref(&pfsm->lock);
726 * PFS records are stored in the root inode (not the PFS root inode,
727 * but the real root). Avoid an infinite recursion if loading
728 * the PFS for the real root.
731 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
733 HAMMER_DEF_LOCALIZATION, 0, errorp);
738 pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
739 pfsm->localization = localization;
740 pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
741 pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
743 hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
744 cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
745 HAMMER_LOCALIZE_MISC;
746 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
747 cursor.key_beg.create_tid = 0;
748 cursor.key_beg.delete_tid = 0;
749 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
750 cursor.key_beg.obj_type = 0;
751 cursor.key_beg.key = localization;
752 cursor.asof = HAMMER_MAX_TID;
753 cursor.flags |= HAMMER_CURSOR_ASOF;
756 *errorp = hammer_ip_lookup(&cursor);
758 *errorp = hammer_btree_lookup(&cursor);
760 *errorp = hammer_ip_resolve_data(&cursor);
762 if (cursor.data->pfsd.mirror_flags &
763 HAMMER_PFSD_DELETED) {
766 bytes = cursor.leaf->data_len;
767 if (bytes > sizeof(pfsm->pfsd))
768 bytes = sizeof(pfsm->pfsd);
769 bcopy(cursor.data, &pfsm->pfsd, bytes);
773 hammer_done_cursor(&cursor);
775 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
776 hammer_ref(&pfsm->lock);
778 hammer_rel_inode(ip, 0);
779 if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
780 kfree(pfsm, hmp->m_misc);
787 * Store pseudo-fs data. The backend will automatically delete any prior
788 * on-disk pseudo-fs data but we have to delete in-memory versions.
791 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
793 struct hammer_cursor cursor;
794 hammer_record_t record;
798 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
799 HAMMER_DEF_LOCALIZATION, 0, &error);
801 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
802 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
803 cursor.key_beg.localization = ip->obj_localization +
804 HAMMER_LOCALIZE_MISC;
805 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
806 cursor.key_beg.create_tid = 0;
807 cursor.key_beg.delete_tid = 0;
808 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
809 cursor.key_beg.obj_type = 0;
810 cursor.key_beg.key = pfsm->localization;
811 cursor.asof = HAMMER_MAX_TID;
812 cursor.flags |= HAMMER_CURSOR_ASOF;
814 error = hammer_ip_lookup(&cursor);
815 if (error == 0 && hammer_cursor_inmem(&cursor)) {
816 record = cursor.iprec;
817 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
818 KKASSERT(cursor.deadlk_rec == NULL);
819 hammer_ref(&record->lock);
820 cursor.deadlk_rec = record;
823 record->flags |= HAMMER_RECF_DELETED_FE;
827 if (error == 0 || error == ENOENT) {
828 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
829 record->type = HAMMER_MEM_RECORD_GENERAL;
831 record->leaf.base.localization = ip->obj_localization +
832 HAMMER_LOCALIZE_MISC;
833 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
834 record->leaf.base.key = pfsm->localization;
835 record->leaf.data_len = sizeof(pfsm->pfsd);
836 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
837 error = hammer_ip_add_record(trans, record);
839 hammer_done_cursor(&cursor);
840 if (error == EDEADLK)
842 hammer_rel_inode(ip, 0);
847 * Create a root directory for a PFS if one does not alredy exist.
849 * The PFS root stands alone so we must also bump the nlinks count
850 * to prevent it from being destroyed on release.
853 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
854 hammer_pseudofs_inmem_t pfsm)
860 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
861 pfsm->localization, 0, &error);
866 error = hammer_create_inode(trans, &vap, cred, NULL, pfsm, &ip);
868 ++ip->ino_data.nlinks;
869 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
873 hammer_rel_inode(ip, 0);
878 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
879 * if we are unable to disassociate all the inodes.
883 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
887 hammer_ref(&ip->lock);
888 if (ip->lock.refs == 2 && ip->vp)
889 vclean_unlocked(ip->vp);
890 if (ip->lock.refs == 1 && ip->vp == NULL)
893 res = -1; /* stop, someone is using the inode */
894 hammer_rel_inode(ip, 0);
899 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
904 for (try = res = 0; try < 4; ++try) {
905 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
906 hammer_inode_pfs_cmp,
907 hammer_unload_pseudofs_callback,
909 if (res == 0 && try > 1)
911 hammer_flusher_sync(trans->hmp);
920 * Release a reference on a PFS
923 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
925 hammer_unref(&pfsm->lock);
926 if (pfsm->lock.refs == 0) {
927 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
928 kfree(pfsm, hmp->m_misc);
933 * Called by hammer_sync_inode().
936 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
938 hammer_transaction_t trans = cursor->trans;
939 hammer_record_t record;
947 * If the inode has a presence on-disk then locate it and mark
948 * it deleted, setting DELONDISK.
950 * The record may or may not be physically deleted, depending on
951 * the retention policy.
953 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
954 HAMMER_INODE_ONDISK) {
955 hammer_normalize_cursor(cursor);
956 cursor->key_beg.localization = ip->obj_localization +
957 HAMMER_LOCALIZE_INODE;
958 cursor->key_beg.obj_id = ip->obj_id;
959 cursor->key_beg.key = 0;
960 cursor->key_beg.create_tid = 0;
961 cursor->key_beg.delete_tid = 0;
962 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
963 cursor->key_beg.obj_type = 0;
964 cursor->asof = ip->obj_asof;
965 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
966 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
967 cursor->flags |= HAMMER_CURSOR_BACKEND;
969 error = hammer_btree_lookup(cursor);
970 if (hammer_debug_inode)
971 kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
974 error = hammer_ip_delete_record(cursor, ip, trans->tid);
975 if (hammer_debug_inode)
976 kprintf(" error %d\n", error);
978 ip->flags |= HAMMER_INODE_DELONDISK;
981 hammer_cache_node(&ip->cache[0], cursor->node);
983 if (error == EDEADLK) {
984 hammer_done_cursor(cursor);
985 error = hammer_init_cursor(trans, cursor,
987 if (hammer_debug_inode)
988 kprintf("IPDED %p %d\n", ip, error);
995 * Ok, write out the initial record or a new record (after deleting
996 * the old one), unless the DELETED flag is set. This routine will
997 * clear DELONDISK if it writes out a record.
999 * Update our inode statistics if this is the first application of
1000 * the inode on-disk.
1002 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1004 * Generate a record and write it to the media. We clean-up
1005 * the state before releasing so we do not have to set-up
1008 record = hammer_alloc_mem_record(ip, 0);
1009 record->type = HAMMER_MEM_RECORD_INODE;
1010 record->flush_state = HAMMER_FST_FLUSH;
1011 record->leaf = ip->sync_ino_leaf;
1012 record->leaf.base.create_tid = trans->tid;
1013 record->leaf.data_len = sizeof(ip->sync_ino_data);
1014 record->leaf.create_ts = trans->time32;
1015 record->data = (void *)&ip->sync_ino_data;
1016 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1019 * If this flag is set we cannot sync the new file size
1020 * because we haven't finished related truncations. The
1021 * inode will be flushed in another flush group to finish
1024 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1025 ip->sync_ino_data.size != ip->ino_data.size) {
1027 ip->sync_ino_data.size = ip->ino_data.size;
1033 error = hammer_ip_sync_record_cursor(cursor, record);
1034 if (hammer_debug_inode)
1035 kprintf("GENREC %p rec %08x %d\n",
1036 ip, record->flags, error);
1037 if (error != EDEADLK)
1039 hammer_done_cursor(cursor);
1040 error = hammer_init_cursor(trans, cursor,
1042 if (hammer_debug_inode)
1043 kprintf("GENREC reinit %d\n", error);
1049 * The record isn't managed by the inode's record tree,
1050 * destroy it whether we succeed or fail.
1052 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1053 record->flags |= HAMMER_RECF_DELETED_FE | HAMMER_RECF_COMMITTED;
1054 record->flush_state = HAMMER_FST_IDLE;
1055 hammer_rel_mem_record(record);
1061 if (hammer_debug_inode)
1062 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1063 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1064 HAMMER_INODE_ATIME |
1065 HAMMER_INODE_MTIME);
1066 ip->flags &= ~HAMMER_INODE_DELONDISK;
1068 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1071 * Root volume count of inodes
1073 hammer_sync_lock_sh(trans);
1074 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1075 hammer_modify_volume_field(trans,
1078 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1079 hammer_modify_volume_done(trans->rootvol);
1080 ip->flags |= HAMMER_INODE_ONDISK;
1081 if (hammer_debug_inode)
1082 kprintf("NOWONDISK %p\n", ip);
1084 hammer_sync_unlock(trans);
1089 * If the inode has been destroyed, clean out any left-over flags
1090 * that may have been set by the frontend.
1092 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1093 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1094 HAMMER_INODE_ATIME |
1095 HAMMER_INODE_MTIME);
1101 * Update only the itimes fields.
1103 * ATIME can be updated without generating any UNDO. MTIME is updated
1104 * with UNDO so it is guaranteed to be synchronized properly in case of
1107 * Neither field is included in the B-Tree leaf element's CRC, which is how
1108 * we can get away with updating ATIME the way we do.
1111 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1113 hammer_transaction_t trans = cursor->trans;
1117 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1118 HAMMER_INODE_ONDISK) {
1122 hammer_normalize_cursor(cursor);
1123 cursor->key_beg.localization = ip->obj_localization +
1124 HAMMER_LOCALIZE_INODE;
1125 cursor->key_beg.obj_id = ip->obj_id;
1126 cursor->key_beg.key = 0;
1127 cursor->key_beg.create_tid = 0;
1128 cursor->key_beg.delete_tid = 0;
1129 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1130 cursor->key_beg.obj_type = 0;
1131 cursor->asof = ip->obj_asof;
1132 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1133 cursor->flags |= HAMMER_CURSOR_ASOF;
1134 cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1135 cursor->flags |= HAMMER_CURSOR_GET_DATA;
1136 cursor->flags |= HAMMER_CURSOR_BACKEND;
1138 error = hammer_btree_lookup(cursor);
1140 hammer_cache_node(&ip->cache[0], cursor->node);
1141 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1143 * Updating MTIME requires an UNDO. Just cover
1144 * both atime and mtime.
1146 hammer_sync_lock_sh(trans);
1147 hammer_modify_buffer(trans, cursor->data_buffer,
1148 HAMMER_ITIMES_BASE(&cursor->data->inode),
1149 HAMMER_ITIMES_BYTES);
1150 cursor->data->inode.atime = ip->sync_ino_data.atime;
1151 cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1152 hammer_modify_buffer_done(cursor->data_buffer);
1153 hammer_sync_unlock(trans);
1154 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1156 * Updating atime only can be done in-place with
1159 hammer_sync_lock_sh(trans);
1160 hammer_modify_buffer(trans, cursor->data_buffer,
1162 cursor->data->inode.atime = ip->sync_ino_data.atime;
1163 hammer_modify_buffer_done(cursor->data_buffer);
1164 hammer_sync_unlock(trans);
1166 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1168 if (error == EDEADLK) {
1169 hammer_done_cursor(cursor);
1170 error = hammer_init_cursor(trans, cursor,
1179 * Release a reference on an inode, flush as requested.
1181 * On the last reference we queue the inode to the flusher for its final
1185 hammer_rel_inode(struct hammer_inode *ip, int flush)
1187 /*hammer_mount_t hmp = ip->hmp;*/
1190 * Handle disposition when dropping the last ref.
1193 if (ip->lock.refs == 1) {
1195 * Determine whether on-disk action is needed for
1196 * the inode's final disposition.
1198 KKASSERT(ip->vp == NULL);
1199 hammer_inode_unloadable_check(ip, 0);
1200 if (ip->flags & HAMMER_INODE_MODMASK) {
1201 hammer_flush_inode(ip, 0);
1202 } else if (ip->lock.refs == 1) {
1203 hammer_unload_inode(ip);
1208 hammer_flush_inode(ip, 0);
1211 * The inode still has multiple refs, try to drop
1214 KKASSERT(ip->lock.refs >= 1);
1215 if (ip->lock.refs > 1) {
1216 hammer_unref(&ip->lock);
1224 * Unload and destroy the specified inode. Must be called with one remaining
1225 * reference. The reference is disposed of.
1227 * The inode must be completely clean.
1230 hammer_unload_inode(struct hammer_inode *ip)
1232 hammer_mount_t hmp = ip->hmp;
1234 KASSERT(ip->lock.refs == 1,
1235 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1236 KKASSERT(ip->vp == NULL);
1237 KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1238 KKASSERT(ip->cursor_ip_refs == 0);
1239 KKASSERT(ip->lock.lockcount == 0);
1240 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1242 KKASSERT(RB_EMPTY(&ip->rec_tree));
1243 KKASSERT(TAILQ_EMPTY(&ip->target_list));
1245 RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1247 hammer_free_inode(ip);
1252 * Called during unmounting if a critical error occured. The in-memory
1253 * inode and all related structures are destroyed.
1255 * If a critical error did not occur the unmount code calls the standard
1256 * release and asserts that the inode is gone.
1259 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1261 hammer_record_t rec;
1264 * Get rid of the inodes in-memory records, regardless of their
1265 * state, and clear the mod-mask.
1267 while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1268 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1269 rec->target_ip = NULL;
1270 if (rec->flush_state == HAMMER_FST_SETUP)
1271 rec->flush_state = HAMMER_FST_IDLE;
1273 while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1274 if (rec->flush_state == HAMMER_FST_FLUSH)
1275 --rec->flush_group->refs;
1277 hammer_ref(&rec->lock);
1278 KKASSERT(rec->lock.refs == 1);
1279 rec->flush_state = HAMMER_FST_IDLE;
1280 rec->flush_group = NULL;
1281 rec->flags |= HAMMER_RECF_DELETED_FE;
1282 rec->flags |= HAMMER_RECF_DELETED_BE;
1283 hammer_rel_mem_record(rec);
1285 ip->flags &= ~HAMMER_INODE_MODMASK;
1286 ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1287 KKASSERT(ip->vp == NULL);
1290 * Remove the inode from any flush group, force it idle. FLUSH
1291 * and SETUP states have an inode ref.
1293 switch(ip->flush_state) {
1294 case HAMMER_FST_FLUSH:
1295 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1296 --ip->flush_group->refs;
1297 ip->flush_group = NULL;
1299 case HAMMER_FST_SETUP:
1300 hammer_unref(&ip->lock);
1301 ip->flush_state = HAMMER_FST_IDLE;
1303 case HAMMER_FST_IDLE:
1308 * There shouldn't be any associated vnode. The unload needs at
1309 * least one ref, if we do have a vp steal its ip ref.
1312 kprintf("hammer_destroy_inode_callback: Unexpected "
1313 "vnode association ip %p vp %p\n", ip, ip->vp);
1314 ip->vp->v_data = NULL;
1317 hammer_ref(&ip->lock);
1319 hammer_unload_inode(ip);
1324 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1325 * the read-only flag for cached inodes.
1327 * This routine is called from a RB_SCAN().
1330 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1332 hammer_mount_t hmp = ip->hmp;
1334 if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1335 ip->flags |= HAMMER_INODE_RO;
1337 ip->flags &= ~HAMMER_INODE_RO;
1342 * A transaction has modified an inode, requiring updates as specified by
1345 * HAMMER_INODE_DDIRTY: Inode data has been updated
1346 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1347 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1348 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1349 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1352 hammer_modify_inode(hammer_inode_t ip, int flags)
1355 * ronly of 0 or 2 does not trigger assertion.
1356 * 2 is a special error state
1358 KKASSERT(ip->hmp->ronly != 1 ||
1359 (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1360 HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1361 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1362 if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1363 ip->flags |= HAMMER_INODE_RSV_INODES;
1364 ++ip->hmp->rsv_inodes;
1371 * Request that an inode be flushed. This whole mess cannot block and may
1372 * recurse (if not synchronous). Once requested HAMMER will attempt to
1373 * actively flush the inode until the flush can be done.
1375 * The inode may already be flushing, or may be in a setup state. We can
1376 * place the inode in a flushing state if it is currently idle and flag it
1377 * to reflush if it is currently flushing.
1379 * Upon return if the inode could not be flushed due to a setup
1380 * dependancy, then it will be automatically flushed when the dependancy
1384 hammer_flush_inode(hammer_inode_t ip, int flags)
1387 hammer_flush_group_t flg;
1391 * next_flush_group is the first flush group we can place the inode
1392 * in. It may be NULL. If it becomes full we append a new flush
1393 * group and make that the next_flush_group.
1396 while ((flg = hmp->next_flush_group) != NULL) {
1397 KKASSERT(flg->running == 0);
1398 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1400 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1401 hammer_flusher_async(ip->hmp, flg);
1404 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1405 hmp->next_flush_group = flg;
1406 TAILQ_INIT(&flg->flush_list);
1407 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1411 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1412 * state we have to put it back into an IDLE state so we can
1413 * drop the extra ref.
1415 * If we have a parent dependancy we must still fall through
1418 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1419 if (ip->flush_state == HAMMER_FST_SETUP &&
1420 TAILQ_EMPTY(&ip->target_list)) {
1421 ip->flush_state = HAMMER_FST_IDLE;
1422 hammer_rel_inode(ip, 0);
1424 if (ip->flush_state == HAMMER_FST_IDLE)
1429 * Our flush action will depend on the current state.
1431 switch(ip->flush_state) {
1432 case HAMMER_FST_IDLE:
1434 * We have no dependancies and can flush immediately. Some
1435 * our children may not be flushable so we have to re-test
1436 * with that additional knowledge.
1438 hammer_flush_inode_core(ip, flg, flags);
1440 case HAMMER_FST_SETUP:
1442 * Recurse upwards through dependancies via target_list
1443 * and start their flusher actions going if possible.
1445 * 'good' is our connectivity. -1 means we have none and
1446 * can't flush, 0 means there weren't any dependancies, and
1447 * 1 means we have good connectivity.
1449 good = hammer_setup_parent_inodes(ip, flg);
1453 * We can continue if good >= 0. Determine how
1454 * many records under our inode can be flushed (and
1457 hammer_flush_inode_core(ip, flg, flags);
1460 * Parent has no connectivity, tell it to flush
1461 * us as soon as it does.
1463 * The REFLUSH flag is also needed to trigger
1464 * dependancy wakeups.
1466 ip->flags |= HAMMER_INODE_CONN_DOWN |
1467 HAMMER_INODE_REFLUSH;
1468 if (flags & HAMMER_FLUSH_SIGNAL) {
1469 ip->flags |= HAMMER_INODE_RESIGNAL;
1470 hammer_flusher_async(ip->hmp, flg);
1474 case HAMMER_FST_FLUSH:
1476 * We are already flushing, flag the inode to reflush
1477 * if needed after it completes its current flush.
1479 * The REFLUSH flag is also needed to trigger
1480 * dependancy wakeups.
1482 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1483 ip->flags |= HAMMER_INODE_REFLUSH;
1484 if (flags & HAMMER_FLUSH_SIGNAL) {
1485 ip->flags |= HAMMER_INODE_RESIGNAL;
1486 hammer_flusher_async(ip->hmp, flg);
1493 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1494 * ip which reference our ip.
1496 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1497 * so for now do not ref/deref the structures. Note that if we use the
1498 * ref/rel code later, the rel CAN block.
1501 hammer_setup_parent_inodes(hammer_inode_t ip, hammer_flush_group_t flg)
1503 hammer_record_t depend;
1508 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1509 r = hammer_setup_parent_inodes_helper(depend, flg);
1510 KKASSERT(depend->target_ip == ip);
1511 if (r < 0 && good == 0)
1520 * This helper function takes a record representing the dependancy between
1521 * the parent inode and child inode.
1523 * record->ip = parent inode
1524 * record->target_ip = child inode
1526 * We are asked to recurse upwards and convert the record from SETUP
1527 * to FLUSH if possible.
1529 * Return 1 if the record gives us connectivity
1531 * Return 0 if the record is not relevant
1533 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1536 hammer_setup_parent_inodes_helper(hammer_record_t record,
1537 hammer_flush_group_t flg)
1543 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1548 * If the record is already flushing, is it in our flush group?
1550 * If it is in our flush group but it is a general record or a
1551 * delete-on-disk, it does not improve our connectivity (return 0),
1552 * and if the target inode is not trying to destroy itself we can't
1553 * allow the operation yet anyway (the second return -1).
1555 if (record->flush_state == HAMMER_FST_FLUSH) {
1557 * If not in our flush group ask the parent to reflush
1558 * us as soon as possible.
1560 if (record->flush_group != flg) {
1561 pip->flags |= HAMMER_INODE_REFLUSH;
1562 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1567 * If in our flush group everything is already set up,
1568 * just return whether the record will improve our
1569 * visibility or not.
1571 if (record->type == HAMMER_MEM_RECORD_ADD)
1577 * It must be a setup record. Try to resolve the setup dependancies
1578 * by recursing upwards so we can place ip on the flush list.
1580 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1582 good = hammer_setup_parent_inodes(pip, flg);
1585 * If good < 0 the parent has no connectivity and we cannot safely
1586 * flush the directory entry, which also means we can't flush our
1587 * ip. Flag the parent and us for downward recursion once the
1588 * parent's connectivity is resolved.
1591 /* pip->flags |= HAMMER_INODE_CONN_DOWN; set by recursion */
1592 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1597 * We are go, place the parent inode in a flushing state so we can
1598 * place its record in a flushing state. Note that the parent
1599 * may already be flushing. The record must be in the same flush
1600 * group as the parent.
1602 if (pip->flush_state != HAMMER_FST_FLUSH)
1603 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1604 KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1605 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1608 if (record->type == HAMMER_MEM_RECORD_DEL &&
1609 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1611 * Regardless of flushing state we cannot sync this path if the
1612 * record represents a delete-on-disk but the target inode
1613 * is not ready to sync its own deletion.
1615 * XXX need to count effective nlinks to determine whether
1616 * the flush is ok, otherwise removing a hardlink will
1617 * just leave the DEL record to rot.
1619 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1623 if (pip->flush_group == flg) {
1625 * Because we have not calculated nlinks yet we can just
1626 * set records to the flush state if the parent is in
1627 * the same flush group as we are.
1629 record->flush_state = HAMMER_FST_FLUSH;
1630 record->flush_group = flg;
1631 ++record->flush_group->refs;
1632 hammer_ref(&record->lock);
1635 * A general directory-add contributes to our visibility.
1637 * Otherwise it is probably a directory-delete or
1638 * delete-on-disk record and does not contribute to our
1639 * visbility (but we can still flush it).
1641 if (record->type == HAMMER_MEM_RECORD_ADD)
1646 * If the parent is not in our flush group we cannot
1647 * flush this record yet, there is no visibility.
1648 * We tell the parent to reflush and mark ourselves
1649 * so the parent knows it should flush us too.
1651 pip->flags |= HAMMER_INODE_REFLUSH;
1652 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1658 * This is the core routine placing an inode into the FST_FLUSH state.
1661 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1666 * Set flush state and prevent the flusher from cycling into
1667 * the next flush group. Do not place the ip on the list yet.
1668 * Inodes not in the idle state get an extra reference.
1670 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1671 if (ip->flush_state == HAMMER_FST_IDLE)
1672 hammer_ref(&ip->lock);
1673 ip->flush_state = HAMMER_FST_FLUSH;
1674 ip->flush_group = flg;
1675 ++ip->hmp->flusher.group_lock;
1676 ++ip->hmp->count_iqueued;
1677 ++hammer_count_iqueued;
1681 * If the flush group reaches the autoflush limit we want to signal
1682 * the flusher. This is particularly important for remove()s.
1684 if (flg->total_count == hammer_autoflush)
1685 flags |= HAMMER_FLUSH_SIGNAL;
1688 * We need to be able to vfsync/truncate from the backend.
1690 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1691 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1692 ip->flags |= HAMMER_INODE_VHELD;
1697 * Figure out how many in-memory records we can actually flush
1698 * (not including inode meta-data, buffers, etc).
1700 KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1701 if (flags & HAMMER_FLUSH_RECURSION) {
1703 * If this is a upwards recursion we do not want to
1704 * recurse down again!
1708 } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1710 * No new records are added if we must complete a flush
1711 * from a previous cycle, but we do have to move the records
1712 * from the previous cycle to the current one.
1715 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1716 hammer_syncgrp_child_callback, NULL);
1722 * Normal flush, scan records and bring them into the flush.
1723 * Directory adds and deletes are usually skipped (they are
1724 * grouped with the related inode rather then with the
1727 * go_count can be negative, which means the scan aborted
1728 * due to the flush group being over-full and we should
1729 * flush what we have.
1731 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1732 hammer_setup_child_callback, NULL);
1736 * This is a more involved test that includes go_count. If we
1737 * can't flush, flag the inode and return. If go_count is 0 we
1738 * were are unable to flush any records in our rec_tree and
1739 * must ignore the XDIRTY flag.
1741 if (go_count == 0) {
1742 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1743 --ip->hmp->count_iqueued;
1744 --hammer_count_iqueued;
1747 ip->flush_state = HAMMER_FST_SETUP;
1748 ip->flush_group = NULL;
1749 if (ip->flags & HAMMER_INODE_VHELD) {
1750 ip->flags &= ~HAMMER_INODE_VHELD;
1755 * REFLUSH is needed to trigger dependancy wakeups
1756 * when an inode is in SETUP.
1758 ip->flags |= HAMMER_INODE_REFLUSH;
1759 if (flags & HAMMER_FLUSH_SIGNAL) {
1760 ip->flags |= HAMMER_INODE_RESIGNAL;
1761 hammer_flusher_async(ip->hmp, flg);
1763 if (--ip->hmp->flusher.group_lock == 0)
1764 wakeup(&ip->hmp->flusher.group_lock);
1770 * Snapshot the state of the inode for the backend flusher.
1772 * We continue to retain save_trunc_off even when all truncations
1773 * have been resolved as an optimization to determine if we can
1774 * skip the B-Tree lookup for overwrite deletions.
1776 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1777 * and stays in ip->flags. Once set, it stays set until the
1778 * inode is destroyed.
1780 if (ip->flags & HAMMER_INODE_TRUNCATED) {
1781 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1782 ip->sync_trunc_off = ip->trunc_off;
1783 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1784 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1785 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1788 * The save_trunc_off used to cache whether the B-Tree
1789 * holds any records past that point is not used until
1790 * after the truncation has succeeded, so we can safely
1793 if (ip->save_trunc_off > ip->sync_trunc_off)
1794 ip->save_trunc_off = ip->sync_trunc_off;
1796 ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1797 ~HAMMER_INODE_TRUNCATED);
1798 ip->sync_ino_leaf = ip->ino_leaf;
1799 ip->sync_ino_data = ip->ino_data;
1800 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1801 #ifdef DEBUG_TRUNCATE
1802 if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1803 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1807 * The flusher list inherits our inode and reference.
1809 KKASSERT(flg->running == 0);
1810 TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
1811 if (--ip->hmp->flusher.group_lock == 0)
1812 wakeup(&ip->hmp->flusher.group_lock);
1814 if (flags & HAMMER_FLUSH_SIGNAL) {
1815 hammer_flusher_async(ip->hmp, flg);
1820 * Callback for scan of ip->rec_tree. Try to include each record in our
1821 * flush. ip->flush_group has been set but the inode has not yet been
1822 * moved into a flushing state.
1824 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1827 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1828 * the caller from shortcutting the flush.
1831 hammer_setup_child_callback(hammer_record_t rec, void *data)
1833 hammer_flush_group_t flg;
1834 hammer_inode_t target_ip;
1839 * Deleted records are ignored. Note that the flush detects deleted
1840 * front-end records at multiple points to deal with races. This is
1841 * just the first line of defense. The only time DELETED_FE cannot
1842 * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1844 * Don't get confused between record deletion and, say, directory
1845 * entry deletion. The deletion of a directory entry that is on
1846 * the media has nothing to do with the record deletion flags.
1848 if (rec->flags & (HAMMER_RECF_DELETED_FE|HAMMER_RECF_DELETED_BE)) {
1849 if (rec->flush_state == HAMMER_FST_FLUSH) {
1850 KKASSERT(rec->flush_group == rec->ip->flush_group);
1859 * If the record is in an idle state it has no dependancies and
1863 flg = ip->flush_group;
1866 switch(rec->flush_state) {
1867 case HAMMER_FST_IDLE:
1869 * The record has no setup dependancy, we can flush it.
1871 KKASSERT(rec->target_ip == NULL);
1872 rec->flush_state = HAMMER_FST_FLUSH;
1873 rec->flush_group = flg;
1875 hammer_ref(&rec->lock);
1878 case HAMMER_FST_SETUP:
1880 * The record has a setup dependancy. These are typically
1881 * directory entry adds and deletes. Such entries will be
1882 * flushed when their inodes are flushed so we do not
1883 * usually have to add them to the flush here. However,
1884 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
1885 * it is asking us to flush this record (and it).
1887 target_ip = rec->target_ip;
1888 KKASSERT(target_ip != NULL);
1889 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
1892 * If the target IP is already flushing in our group
1893 * we could associate the record, but target_ip has
1894 * already synced ino_data to sync_ino_data and we
1895 * would also have to adjust nlinks. Plus there are
1896 * ordering issues for adds and deletes.
1898 * Reflush downward if this is an ADD, and upward if
1901 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
1902 if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
1903 ip->flags |= HAMMER_INODE_REFLUSH;
1905 target_ip->flags |= HAMMER_INODE_REFLUSH;
1910 * Target IP is not yet flushing. This can get complex
1911 * because we have to be careful about the recursion.
1913 * Directories create an issue for us in that if a flush
1914 * of a directory is requested the expectation is to flush
1915 * any pending directory entries, but this will cause the
1916 * related inodes to recursively flush as well. We can't
1917 * really defer the operation so just get as many as we
1921 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
1922 (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
1924 * We aren't reclaiming and the target ip was not
1925 * previously prevented from flushing due to this
1926 * record dependancy. Do not flush this record.
1931 if (flg->total_count + flg->refs >
1932 ip->hmp->undo_rec_limit) {
1934 * Our flush group is over-full and we risk blowing
1935 * out the UNDO FIFO. Stop the scan, flush what we
1936 * have, then reflush the directory.
1938 * The directory may be forced through multiple
1939 * flush groups before it can be completely
1942 ip->flags |= HAMMER_INODE_RESIGNAL |
1943 HAMMER_INODE_REFLUSH;
1945 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
1947 * If the target IP is not flushing we can force
1948 * it to flush, even if it is unable to write out
1949 * any of its own records we have at least one in
1950 * hand that we CAN deal with.
1952 rec->flush_state = HAMMER_FST_FLUSH;
1953 rec->flush_group = flg;
1955 hammer_ref(&rec->lock);
1956 hammer_flush_inode_core(target_ip, flg,
1957 HAMMER_FLUSH_RECURSION);
1961 * General or delete-on-disk record.
1963 * XXX this needs help. If a delete-on-disk we could
1964 * disconnect the target. If the target has its own
1965 * dependancies they really need to be flushed.
1969 rec->flush_state = HAMMER_FST_FLUSH;
1970 rec->flush_group = flg;
1972 hammer_ref(&rec->lock);
1973 hammer_flush_inode_core(target_ip, flg,
1974 HAMMER_FLUSH_RECURSION);
1978 case HAMMER_FST_FLUSH:
1980 * The flush_group should already match.
1982 KKASSERT(rec->flush_group == flg);
1991 * This version just moves records already in a flush state to the new
1992 * flush group and that is it.
1995 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
1997 hammer_inode_t ip = rec->ip;
1999 switch(rec->flush_state) {
2000 case HAMMER_FST_FLUSH:
2001 KKASSERT(rec->flush_group == ip->flush_group);
2011 * Wait for a previously queued flush to complete.
2013 * If a critical error occured we don't try to wait.
2016 hammer_wait_inode(hammer_inode_t ip)
2018 hammer_flush_group_t flg;
2021 if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2022 while (ip->flush_state != HAMMER_FST_IDLE &&
2023 (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2024 if (ip->flush_state == HAMMER_FST_SETUP)
2025 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2026 if (ip->flush_state != HAMMER_FST_IDLE) {
2027 ip->flags |= HAMMER_INODE_FLUSHW;
2028 tsleep(&ip->flags, 0, "hmrwin", 0);
2035 * Called by the backend code when a flush has been completed.
2036 * The inode has already been removed from the flush list.
2038 * A pipelined flush can occur, in which case we must re-enter the
2039 * inode on the list and re-copy its fields.
2042 hammer_flush_inode_done(hammer_inode_t ip, int error)
2047 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2052 * Merge left-over flags back into the frontend and fix the state.
2053 * Incomplete truncations are retained by the backend.
2056 ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2057 ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2060 * The backend may have adjusted nlinks, so if the adjusted nlinks
2061 * does not match the fronttend set the frontend's RDIRTY flag again.
2063 if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2064 ip->flags |= HAMMER_INODE_DDIRTY;
2067 * Fix up the dirty buffer status.
2069 if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2070 ip->flags |= HAMMER_INODE_BUFS;
2074 * Re-set the XDIRTY flag if some of the inode's in-memory records
2075 * could not be flushed.
2077 KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2078 (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2079 (!RB_EMPTY(&ip->rec_tree) &&
2080 (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2083 * Do not lose track of inodes which no longer have vnode
2084 * assocations, otherwise they may never get flushed again.
2086 * The reflush flag can be set superfluously, causing extra pain
2087 * for no reason. If the inode is no longer modified it no longer
2088 * needs to be flushed.
2090 if (ip->flags & HAMMER_INODE_MODMASK) {
2092 ip->flags |= HAMMER_INODE_REFLUSH;
2094 ip->flags &= ~HAMMER_INODE_REFLUSH;
2098 * Adjust the flush state.
2100 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2102 * We were unable to flush out all our records, leave the
2103 * inode in a flush state and in the current flush group.
2104 * The flush group will be re-run.
2106 * This occurs if the UNDO block gets too full or there is
2107 * too much dirty meta-data and allows the flusher to
2108 * finalize the UNDO block and then re-flush.
2110 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2114 * Remove from the flush_group
2116 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2117 ip->flush_group = NULL;
2120 * Clean up the vnode ref and tracking counts.
2122 if (ip->flags & HAMMER_INODE_VHELD) {
2123 ip->flags &= ~HAMMER_INODE_VHELD;
2126 --hmp->count_iqueued;
2127 --hammer_count_iqueued;
2130 * And adjust the state.
2132 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2133 ip->flush_state = HAMMER_FST_IDLE;
2136 ip->flush_state = HAMMER_FST_SETUP;
2141 * If the frontend is waiting for a flush to complete,
2144 if (ip->flags & HAMMER_INODE_FLUSHW) {
2145 ip->flags &= ~HAMMER_INODE_FLUSHW;
2150 * If the frontend made more changes and requested another
2151 * flush, then try to get it running.
2153 * Reflushes are aborted when the inode is errored out.
2155 if (ip->flags & HAMMER_INODE_REFLUSH) {
2156 ip->flags &= ~HAMMER_INODE_REFLUSH;
2157 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2158 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2159 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2161 hammer_flush_inode(ip, 0);
2167 * If we have no parent dependancies we can clear CONN_DOWN
2169 if (TAILQ_EMPTY(&ip->target_list))
2170 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2173 * If the inode is now clean drop the space reservation.
2175 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2176 (ip->flags & HAMMER_INODE_RSV_INODES)) {
2177 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2182 hammer_rel_inode(ip, 0);
2186 * Called from hammer_sync_inode() to synchronize in-memory records
2190 hammer_sync_record_callback(hammer_record_t record, void *data)
2192 hammer_cursor_t cursor = data;
2193 hammer_transaction_t trans = cursor->trans;
2194 hammer_mount_t hmp = trans->hmp;
2198 * Skip records that do not belong to the current flush.
2200 ++hammer_stats_record_iterations;
2201 if (record->flush_state != HAMMER_FST_FLUSH)
2205 if (record->flush_group != record->ip->flush_group) {
2206 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2211 KKASSERT(record->flush_group == record->ip->flush_group);
2214 * Interlock the record using the BE flag. Once BE is set the
2215 * frontend cannot change the state of FE.
2217 * NOTE: If FE is set prior to us setting BE we still sync the
2218 * record out, but the flush completion code converts it to
2219 * a delete-on-disk record instead of destroying it.
2221 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2222 record->flags |= HAMMER_RECF_INTERLOCK_BE;
2225 * The backend may have already disposed of the record.
2227 if (record->flags & HAMMER_RECF_DELETED_BE) {
2233 * If the whole inode is being deleting all on-disk records will
2234 * be deleted very soon, we can't sync any new records to disk
2235 * because they will be deleted in the same transaction they were
2236 * created in (delete_tid == create_tid), which will assert.
2238 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2239 * that we currently panic on.
2241 if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2242 switch(record->type) {
2243 case HAMMER_MEM_RECORD_DATA:
2245 * We don't have to do anything, if the record was
2246 * committed the space will have been accounted for
2250 case HAMMER_MEM_RECORD_GENERAL:
2251 record->flags |= HAMMER_RECF_DELETED_FE;
2252 record->flags |= HAMMER_RECF_DELETED_BE;
2255 case HAMMER_MEM_RECORD_ADD:
2256 panic("hammer_sync_record_callback: illegal add "
2257 "during inode deletion record %p", record);
2258 break; /* NOT REACHED */
2259 case HAMMER_MEM_RECORD_INODE:
2260 panic("hammer_sync_record_callback: attempt to "
2261 "sync inode record %p?", record);
2262 break; /* NOT REACHED */
2263 case HAMMER_MEM_RECORD_DEL:
2265 * Follow through and issue the on-disk deletion
2272 * If DELETED_FE is set special handling is needed for directory
2273 * entries. Dependant pieces related to the directory entry may
2274 * have already been synced to disk. If this occurs we have to
2275 * sync the directory entry and then change the in-memory record
2276 * from an ADD to a DELETE to cover the fact that it's been
2277 * deleted by the frontend.
2279 * A directory delete covering record (MEM_RECORD_DEL) can never
2280 * be deleted by the frontend.
2282 * Any other record type (aka DATA) can be deleted by the frontend.
2283 * XXX At the moment the flusher must skip it because there may
2284 * be another data record in the flush group for the same block,
2285 * meaning that some frontend data changes can leak into the backend's
2286 * synchronization point.
2288 if (record->flags & HAMMER_RECF_DELETED_FE) {
2289 if (record->type == HAMMER_MEM_RECORD_ADD) {
2290 record->flags |= HAMMER_RECF_CONVERT_DELETE;
2292 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2293 record->flags |= HAMMER_RECF_DELETED_BE;
2300 * Assign the create_tid for new records. Deletions already
2301 * have the record's entire key properly set up.
2303 if (record->type != HAMMER_MEM_RECORD_DEL)
2304 record->leaf.base.create_tid = trans->tid;
2305 record->leaf.create_ts = trans->time32;
2307 error = hammer_ip_sync_record_cursor(cursor, record);
2308 if (error != EDEADLK)
2310 hammer_done_cursor(cursor);
2311 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2316 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2321 hammer_flush_record_done(record, error);
2324 * Do partial finalization if we have built up too many dirty
2325 * buffers. Otherwise a buffer cache deadlock can occur when
2326 * doing things like creating tens of thousands of tiny files.
2328 * We must release our cursor lock to avoid a 3-way deadlock
2329 * due to the exclusive sync lock the finalizer must get.
2331 if (hammer_flusher_meta_limit(hmp)) {
2332 hammer_unlock_cursor(cursor, 0);
2333 hammer_flusher_finalize(trans, 0);
2334 hammer_lock_cursor(cursor, 0);
2341 * Backend function called by the flusher to sync an inode to media.
2344 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2346 struct hammer_cursor cursor;
2347 hammer_node_t tmp_node;
2348 hammer_record_t depend;
2349 hammer_record_t next;
2350 int error, tmp_error;
2353 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2356 error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2361 * Any directory records referencing this inode which are not in
2362 * our current flush group must adjust our nlink count for the
2363 * purposes of synchronization to disk.
2365 * Records which are in our flush group can be unlinked from our
2366 * inode now, potentially allowing the inode to be physically
2369 * This cannot block.
2371 nlinks = ip->ino_data.nlinks;
2372 next = TAILQ_FIRST(&ip->target_list);
2373 while ((depend = next) != NULL) {
2374 next = TAILQ_NEXT(depend, target_entry);
2375 if (depend->flush_state == HAMMER_FST_FLUSH &&
2376 depend->flush_group == ip->flush_group) {
2378 * If this is an ADD that was deleted by the frontend
2379 * the frontend nlinks count will have already been
2380 * decremented, but the backend is going to sync its
2381 * directory entry and must account for it. The
2382 * record will be converted to a delete-on-disk when
2385 * If the ADD was not deleted by the frontend we
2386 * can remove the dependancy from our target_list.
2388 if (depend->flags & HAMMER_RECF_DELETED_FE) {
2391 TAILQ_REMOVE(&ip->target_list, depend,
2393 depend->target_ip = NULL;
2395 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2397 * Not part of our flush group
2399 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2400 switch(depend->type) {
2401 case HAMMER_MEM_RECORD_ADD:
2404 case HAMMER_MEM_RECORD_DEL:
2414 * Set dirty if we had to modify the link count.
2416 if (ip->sync_ino_data.nlinks != nlinks) {
2417 KKASSERT((int64_t)nlinks >= 0);
2418 ip->sync_ino_data.nlinks = nlinks;
2419 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2423 * If there is a trunction queued destroy any data past the (aligned)
2424 * truncation point. Userland will have dealt with the buffer
2425 * containing the truncation point for us.
2427 * We don't flush pending frontend data buffers until after we've
2428 * dealt with the truncation.
2430 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2432 * Interlock trunc_off. The VOP front-end may continue to
2433 * make adjustments to it while we are blocked.
2436 off_t aligned_trunc_off;
2439 trunc_off = ip->sync_trunc_off;
2440 blkmask = hammer_blocksize(trunc_off) - 1;
2441 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2444 * Delete any whole blocks on-media. The front-end has
2445 * already cleaned out any partial block and made it
2446 * pending. The front-end may have updated trunc_off
2447 * while we were blocked so we only use sync_trunc_off.
2449 * This operation can blow out the buffer cache, EWOULDBLOCK
2450 * means we were unable to complete the deletion. The
2451 * deletion will update sync_trunc_off in that case.
2453 error = hammer_ip_delete_range(&cursor, ip,
2455 0x7FFFFFFFFFFFFFFFLL, 2);
2456 if (error == EWOULDBLOCK) {
2457 ip->flags |= HAMMER_INODE_WOULDBLOCK;
2459 goto defer_buffer_flush;
2466 * Clear the truncation flag on the backend after we have
2467 * complete the deletions. Backend data is now good again
2468 * (including new records we are about to sync, below).
2470 * Leave sync_trunc_off intact. As we write additional
2471 * records the backend will update sync_trunc_off. This
2472 * tells the backend whether it can skip the overwrite
2473 * test. This should work properly even when the backend
2474 * writes full blocks where the truncation point straddles
2475 * the block because the comparison is against the base
2476 * offset of the record.
2478 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2479 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2485 * Now sync related records. These will typically be directory
2486 * entries, records tracking direct-writes, or delete-on-disk records.
2489 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2490 hammer_sync_record_callback, &cursor);
2496 hammer_cache_node(&ip->cache[1], cursor.node);
2499 * Re-seek for inode update, assuming our cache hasn't been ripped
2500 * out from under us.
2503 tmp_node = hammer_ref_node_safe(ip->hmp, &ip->cache[0], &error);
2505 hammer_cursor_downgrade(&cursor);
2506 hammer_lock_sh(&tmp_node->lock);
2507 if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2508 hammer_cursor_seek(&cursor, tmp_node, 0);
2509 hammer_unlock(&tmp_node->lock);
2510 hammer_rel_node(tmp_node);
2516 * If we are deleting the inode the frontend had better not have
2517 * any active references on elements making up the inode.
2519 * The call to hammer_ip_delete_clean() cleans up auxillary records
2520 * but not DB or DATA records. Those must have already been deleted
2521 * by the normal truncation mechanic.
2523 if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2524 RB_EMPTY(&ip->rec_tree) &&
2525 (ip->sync_flags & HAMMER_INODE_DELETING) &&
2526 (ip->flags & HAMMER_INODE_DELETED) == 0) {
2529 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2531 ip->flags |= HAMMER_INODE_DELETED;
2532 ip->sync_flags &= ~HAMMER_INODE_DELETING;
2533 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2534 KKASSERT(RB_EMPTY(&ip->rec_tree));
2537 * Set delete_tid in both the frontend and backend
2538 * copy of the inode record. The DELETED flag handles
2539 * this, do not set RDIRTY.
2541 ip->ino_leaf.base.delete_tid = trans->tid;
2542 ip->sync_ino_leaf.base.delete_tid = trans->tid;
2543 ip->ino_leaf.delete_ts = trans->time32;
2544 ip->sync_ino_leaf.delete_ts = trans->time32;
2548 * Adjust the inode count in the volume header
2550 hammer_sync_lock_sh(trans);
2551 if (ip->flags & HAMMER_INODE_ONDISK) {
2552 hammer_modify_volume_field(trans,
2555 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2556 hammer_modify_volume_done(trans->rootvol);
2558 hammer_sync_unlock(trans);
2564 ip->sync_flags &= ~HAMMER_INODE_BUFS;
2568 * Now update the inode's on-disk inode-data and/or on-disk record.
2569 * DELETED and ONDISK are managed only in ip->flags.
2571 * In the case of a defered buffer flush we still update the on-disk
2572 * inode to satisfy visibility requirements if there happen to be
2573 * directory dependancies.
2575 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2576 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2578 * If deleted and on-disk, don't set any additional flags.
2579 * the delete flag takes care of things.
2581 * Clear flags which may have been set by the frontend.
2583 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2584 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2585 HAMMER_INODE_DELETING);
2587 case HAMMER_INODE_DELETED:
2589 * Take care of the case where a deleted inode was never
2590 * flushed to the disk in the first place.
2592 * Clear flags which may have been set by the frontend.
2594 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2595 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2596 HAMMER_INODE_DELETING);
2597 while (RB_ROOT(&ip->rec_tree)) {
2598 hammer_record_t record = RB_ROOT(&ip->rec_tree);
2599 hammer_ref(&record->lock);
2600 KKASSERT(record->lock.refs == 1);
2601 record->flags |= HAMMER_RECF_DELETED_FE;
2602 record->flags |= HAMMER_RECF_DELETED_BE;
2603 hammer_rel_mem_record(record);
2606 case HAMMER_INODE_ONDISK:
2608 * If already on-disk, do not set any additional flags.
2613 * If not on-disk and not deleted, set DDIRTY to force
2614 * an initial record to be written.
2616 * Also set the create_tid in both the frontend and backend
2617 * copy of the inode record.
2619 ip->ino_leaf.base.create_tid = trans->tid;
2620 ip->ino_leaf.create_ts = trans->time32;
2621 ip->sync_ino_leaf.base.create_tid = trans->tid;
2622 ip->sync_ino_leaf.create_ts = trans->time32;
2623 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2628 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2629 * is already on-disk the old record is marked as deleted.
2631 * If DELETED is set hammer_update_inode() will delete the existing
2632 * record without writing out a new one.
2634 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2636 if (ip->flags & HAMMER_INODE_DELETED) {
2637 error = hammer_update_inode(&cursor, ip);
2639 if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2640 (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2641 error = hammer_update_itimes(&cursor, ip);
2643 if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2644 error = hammer_update_inode(&cursor, ip);
2648 hammer_critical_error(ip->hmp, ip, error,
2649 "while syncing inode");
2651 hammer_done_cursor(&cursor);
2656 * This routine is called when the OS is no longer actively referencing
2657 * the inode (but might still be keeping it cached), or when releasing
2658 * the last reference to an inode.
2660 * At this point if the inode's nlinks count is zero we want to destroy
2661 * it, which may mean destroying it on-media too.
2664 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2669 * Set the DELETING flag when the link count drops to 0 and the
2670 * OS no longer has any opens on the inode.
2672 * The backend will clear DELETING (a mod flag) and set DELETED
2673 * (a state flag) when it is actually able to perform the
2676 * Don't reflag the deletion if the flusher is currently syncing
2677 * one that was already flagged. A previously set DELETING flag
2678 * may bounce around flags and sync_flags until the operation is
2681 if (ip->ino_data.nlinks == 0 &&
2682 ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2683 ip->flags |= HAMMER_INODE_DELETING;
2684 ip->flags |= HAMMER_INODE_TRUNCATED;
2688 if (hammer_get_vnode(ip, &vp) != 0)
2696 vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2697 vnode_pager_setsize(ip->vp, 0);
2706 * After potentially resolving a dependancy the inode is tested
2707 * to determine whether it needs to be reflushed.
2710 hammer_test_inode(hammer_inode_t ip)
2712 if (ip->flags & HAMMER_INODE_REFLUSH) {
2713 ip->flags &= ~HAMMER_INODE_REFLUSH;
2714 hammer_ref(&ip->lock);
2715 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2716 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2717 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2719 hammer_flush_inode(ip, 0);
2721 hammer_rel_inode(ip, 0);
2726 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2727 * reassociated with a vp or just before it gets freed.
2729 * Pipeline wakeups to threads blocked due to an excessive number of
2730 * detached inodes. The reclaim count generates a bit of negative
2734 hammer_inode_wakereclaims(hammer_inode_t ip, int dowake)
2736 struct hammer_reclaim *reclaim;
2737 hammer_mount_t hmp = ip->hmp;
2739 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2742 --hammer_count_reclaiming;
2743 --hmp->inode_reclaims;
2744 ip->flags &= ~HAMMER_INODE_RECLAIM;
2746 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT || dowake) {
2747 reclaim = TAILQ_FIRST(&hmp->reclaim_list);
2748 if (reclaim && reclaim->count > 0 && --reclaim->count == 0) {
2749 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2756 * Setup our reclaim pipeline. We only let so many detached (and dirty)
2757 * inodes build up before we start blocking.
2759 * When we block we don't care *which* inode has finished reclaiming,
2760 * as lone as one does. This is somewhat heuristical... we also put a
2761 * cap on how long we are willing to wait.
2764 hammer_inode_waitreclaims(hammer_mount_t hmp)
2766 struct hammer_reclaim reclaim;
2769 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT)
2771 delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2772 (HAMMER_RECLAIM_WAIT * 3) + 1;
2775 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
2776 tsleep(&reclaim, 0, "hmrrcm", delay);
2777 if (reclaim.count > 0)
2778 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2783 * A larger then normal backlog of inodes is sitting in the flusher,
2784 * enforce a general slowdown to let it catch up. This routine is only
2785 * called on completion of a non-flusher-related transaction which
2786 * performed B-Tree node I/O.
2788 * It is possible for the flusher to stall in a continuous load.
2789 * blogbench -i1000 -o seems to do a good job generating this sort of load.
2790 * If the flusher is unable to catch up the inode count can bloat until
2791 * we run out of kvm.
2793 * This is a bit of a hack.
2796 hammer_inode_waithard(hammer_mount_t hmp)
2801 if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
2802 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT / 2 &&
2803 hmp->count_iqueued < hmp->count_inodes / 20) {
2804 hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
2808 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT ||
2809 hmp->count_iqueued < hmp->count_inodes / 10) {
2812 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
2816 * Block for one flush cycle.
2818 hammer_flusher_wait_next(hmp);