2 * Copyright (c) 2011-2013 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/cdefs.h>
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/types.h>
44 RB_GENERATE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
48 hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2)
50 if (ip1->inum < ip2->inum)
52 if (ip1->inum > ip2->inum)
60 * HAMMER2 offers shared locks and exclusive locks on inodes.
62 * An inode's ip->chain pointer is resolved and stable while an inode is
63 * locked, and can be cleaned out at any time (become NULL) when an inode
66 * This function handles duplication races which can cause ip's cached
67 * chain to become stale.
69 * The underlying chain is also locked and returned.
71 * NOTE: We don't combine the inode/chain lock because putting away an
72 * inode would otherwise confuse multiple lock holders of the inode.
75 hammer2_inode_lock_ex(hammer2_inode_t *ip)
77 hammer2_chain_t *chain;
78 hammer2_chain_core_t *core;
80 hammer2_inode_ref(ip);
81 ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
86 if (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
87 spin_lock(&core->cst.spin);
88 while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
89 chain = TAILQ_NEXT(chain, core_entry);
90 hammer2_chain_ref(chain);
91 spin_unlock(&core->cst.spin);
92 hammer2_inode_repoint(ip, NULL, chain);
93 hammer2_chain_drop(chain);
95 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
96 if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
98 hammer2_chain_unlock(chain);
104 hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
107 * XXX this will catch parent directories too which we don't
111 hammer2_chain_unlock(chain);
112 ccms_thread_unlock(&ip->topo_cst);
113 hammer2_inode_drop(ip);
117 * NOTE: We don't combine the inode/chain lock because putting away an
118 * inode would otherwise confuse multiple lock holders of the inode.
120 * Shared locks are especially sensitive to having too many shared
121 * lock counts (from the same thread) on certain paths which might
122 * need to upgrade them. Only one count of a shared lock can be
126 hammer2_inode_lock_sh(hammer2_inode_t *ip)
128 hammer2_chain_t *chain;
130 hammer2_inode_ref(ip);
132 ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
135 KKASSERT(chain != NULL); /* for now */
136 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
137 HAMMER2_RESOLVE_SHARED);
140 * Resolve duplication races
142 if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
144 hammer2_chain_unlock(chain);
145 ccms_thread_unlock(&ip->topo_cst);
146 chain = hammer2_inode_lock_ex(ip);
147 hammer2_inode_unlock_ex(ip, chain);
153 hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
156 hammer2_chain_unlock(chain);
157 ccms_thread_unlock(&ip->topo_cst);
158 hammer2_inode_drop(ip);
162 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
164 return(ccms_thread_lock_temp_release(&ip->topo_cst));
168 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate)
170 ccms_thread_lock_temp_restore(&ip->topo_cst, ostate);
174 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
176 return(ccms_thread_lock_upgrade(&ip->topo_cst));
180 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate)
182 ccms_thread_lock_downgrade(&ip->topo_cst, ostate);
186 * Lookup an inode by inode number
189 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
194 spin_lock(&pmp->inum_spin);
195 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
197 hammer2_inode_ref(ip);
198 spin_unlock(&pmp->inum_spin);
206 * Adding a ref to an inode is only legal if the inode already has at least
210 hammer2_inode_ref(hammer2_inode_t *ip)
212 atomic_add_int(&ip->refs, 1);
216 * Drop an inode reference, freeing the inode when the last reference goes
220 hammer2_inode_drop(hammer2_inode_t *ip)
222 hammer2_pfsmount_t *pmp;
223 hammer2_inode_t *pip;
231 * Transition to zero, must interlock with
232 * the inode inumber lookup tree (if applicable).
234 * NOTE: The super-root inode has no pmp.
238 spin_lock(&pmp->inum_spin);
240 if (atomic_cmpset_int(&ip->refs, 1, 0)) {
241 KKASSERT(ip->topo_cst.count == 0);
242 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
243 atomic_clear_int(&ip->flags,
244 HAMMER2_INODE_ONRBTREE);
245 RB_REMOVE(hammer2_inode_tree,
246 &pmp->inum_tree, ip);
249 spin_unlock(&pmp->inum_spin);
256 * Cleaning out ip->chain isn't entirely
259 hammer2_inode_repoint(ip, NULL, NULL);
262 * We have to drop pip (if non-NULL) to
263 * dispose of our implied reference from
264 * ip->pip. We can simply loop on it.
267 KKASSERT((ip->flags &
268 HAMMER2_INODE_SROOT) == 0);
269 kfree(ip, pmp->minode);
270 atomic_add_long(&pmp->inmem_inodes, -1);
273 HAMMER2_INODE_SROOT);
274 kfree(ip, M_HAMMER2);
277 /* continue with pip (can be NULL) */
280 spin_unlock(&ip->pmp->inum_spin);
284 * Non zero transition
286 if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
293 * Get the vnode associated with the given inode, allocating the vnode if
294 * necessary. The vnode will be returned exclusively locked.
296 * The caller must lock the inode (shared or exclusive).
298 * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
302 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
304 hammer2_inode_data_t *ipdata;
305 hammer2_pfsmount_t *pmp;
310 KKASSERT(pmp != NULL);
312 ipdata = &ip->chain->data->ipdata;
316 * Attempt to reuse an existing vnode assignment. It is
317 * possible to race a reclaim so the vget() may fail. The
318 * inode must be unlocked during the vget() to avoid a
319 * deadlock against a reclaim.
324 * Inode must be unlocked during the vget() to avoid
325 * possible deadlocks, but leave the ip ref intact.
327 * vnode is held to prevent destruction during the
328 * vget(). The vget() can still fail if we lost
329 * a reclaim race on the vnode.
332 ostate = hammer2_inode_lock_temp_release(ip);
333 if (vget(vp, LK_EXCLUSIVE)) {
335 hammer2_inode_lock_temp_restore(ip, ostate);
338 hammer2_inode_lock_temp_restore(ip, ostate);
340 /* vp still locked and ref from vget */
342 kprintf("hammer2: igetv race %p/%p\n",
352 * No vnode exists, allocate a new vnode. Beware of
353 * allocation races. This function will return an
354 * exclusively locked and referenced vnode.
356 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
358 kprintf("hammer2: igetv getnewvnode failed %d\n",
365 * Lock the inode and check for an allocation race.
367 ostate = hammer2_inode_lock_upgrade(ip);
368 if (ip->vp != NULL) {
371 hammer2_inode_lock_downgrade(ip, ostate);
375 switch (ipdata->type) {
376 case HAMMER2_OBJTYPE_DIRECTORY:
379 case HAMMER2_OBJTYPE_REGFILE:
381 vinitvmio(vp, ipdata->size,
383 (int)ipdata->size & HAMMER2_LBUFMASK);
385 case HAMMER2_OBJTYPE_SOFTLINK:
387 * XXX for now we are using the generic file_read
388 * and file_write code so we need a buffer cache
392 vinitvmio(vp, ipdata->size,
394 (int)ipdata->size & HAMMER2_LBUFMASK);
396 case HAMMER2_OBJTYPE_CDEV:
399 case HAMMER2_OBJTYPE_BDEV:
400 vp->v_ops = &pmp->mp->mnt_vn_spec_ops;
401 if (ipdata->type != HAMMER2_OBJTYPE_CDEV)
403 addaliasu(vp, ipdata->rmajor, ipdata->rminor);
405 case HAMMER2_OBJTYPE_FIFO:
407 vp->v_ops = &pmp->mp->mnt_vn_fifo_ops;
410 panic("hammer2: unhandled objtype %d", ipdata->type);
414 if (ip == pmp->iroot)
415 vsetflags(vp, VROOT);
419 hammer2_inode_ref(ip); /* vp association */
420 hammer2_inode_lock_downgrade(ip, ostate);
425 * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
427 if (hammer2_debug & 0x0002) {
428 kprintf("igetv vp %p refs 0x%08x aux 0x%08x\n",
429 vp, vp->v_refcnt, vp->v_auxrefs);
435 * The passed-in chain must be locked and the returned inode will also be
436 * locked. This routine typically locates or allocates the inode, assigns
437 * ip->chain (adding a ref to chain if necessary), and returns the inode.
439 * The hammer2_inode structure regulates the interface between the high level
440 * kernel VNOPS API and the filesystem backend (the chains).
442 * WARNING! This routine sucks up the chain's lock (makes it part of the
443 * inode lock from the point of view of the inode lock API),
444 * so callers need to be careful.
446 * WARNING! The mount code is allowed to pass dip == NULL for iroot and
447 * is allowed to pass pmp == NULL and dip == NULL for sroot.
450 hammer2_inode_get(hammer2_pfsmount_t *pmp, hammer2_inode_t *dip,
451 hammer2_chain_t *chain)
453 hammer2_inode_t *nip;
455 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
458 * Interlocked lookup/ref of the inode. This code is only needed
459 * when looking up inodes with nlinks != 0 (TODO: optimize out
460 * otherwise and test for duplicates).
464 nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
467 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
468 if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { /* race */
469 ccms_thread_unlock(&nip->topo_cst);
470 hammer2_inode_drop(nip);
473 if (nip->chain != chain)
474 hammer2_inode_repoint(nip, NULL, chain);
477 * Consolidated nip/nip->chain is locked (chain locked
484 * We couldn't find the inode number, create a new inode.
487 nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
488 atomic_add_long(&pmp->inmem_inodes, 1);
489 hammer2_chain_memory_wakeup(pmp);
491 nip = kmalloc(sizeof(*nip), M_HAMMER2, M_WAITOK | M_ZERO);
492 nip->flags = HAMMER2_INODE_SROOT;
494 nip->inum = chain->data->ipdata.inum;
495 nip->size = chain->data->ipdata.size;
496 nip->mtime = chain->data->ipdata.mtime;
497 hammer2_inode_repoint(nip, NULL, chain);
498 nip->pip = dip; /* can be NULL */
500 hammer2_inode_ref(dip); /* ref dip for nip->pip */
505 * ref and lock on nip gives it state compatible to after a
506 * hammer2_inode_lock_ex() call.
509 ccms_cst_init(&nip->topo_cst, &nip->chain);
510 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
511 /* combination of thread lock and chain lock == inode lock */
514 * Attempt to add the inode. If it fails we raced another inode
515 * get. Undo all the work and try again.
518 spin_lock(&pmp->inum_spin);
519 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
520 spin_unlock(&pmp->inum_spin);
521 ccms_thread_unlock(&nip->topo_cst);
522 hammer2_inode_drop(nip);
525 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
526 spin_unlock(&pmp->inum_spin);
533 * Create a new inode in the specified directory using the vattr to
534 * figure out the type of inode.
536 * If no error occurs the new inode with its chain locked is returned in
537 * *nipp, otherwise an error is returned and *nipp is set to NULL.
539 * If vap and/or cred are NULL the related fields are not set and the
540 * inode type defaults to a directory. This is used when creating PFSs
541 * under the super-root, so the inode number is set to 1 in this case.
543 * dip is not locked on entry.
546 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
547 struct vattr *vap, struct ucred *cred,
548 const uint8_t *name, size_t name_len,
549 hammer2_chain_t **chainp, int *errorp)
551 hammer2_inode_data_t *dipdata;
552 hammer2_inode_data_t *nipdata;
553 hammer2_chain_t *chain;
554 hammer2_chain_t *parent;
555 hammer2_inode_t *nip;
556 hammer2_key_t key_dummy;
564 int cache_index = -1;
566 lhc = hammer2_dirhash(name, name_len);
570 * Locate the inode or indirect block to create the new
571 * entry in. At the same time check for key collisions
572 * and iterate until we don't get one.
574 * NOTE: hidden inodes do not have iterators.
577 parent = hammer2_inode_lock_ex(dip);
578 dipdata = &dip->chain->data->ipdata;
579 dip_uid = dipdata->uid;
580 dip_gid = dipdata->gid;
581 dip_mode = dipdata->mode;
582 dip_algo = dipdata->comp_algo;
586 chain = hammer2_chain_lookup(&parent, &key_dummy,
587 lhc, lhc, &cache_index, 0);
590 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
592 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
594 hammer2_chain_unlock(chain);
600 error = hammer2_chain_create(trans, &parent, &chain,
602 HAMMER2_BREF_TYPE_INODE,
603 HAMMER2_INODE_BYTES);
607 * Cleanup and handle retries.
609 if (error == EAGAIN) {
610 hammer2_chain_ref(parent);
611 hammer2_inode_unlock_ex(dip, parent);
612 hammer2_chain_wait(parent);
613 hammer2_chain_drop(parent);
616 hammer2_inode_unlock_ex(dip, parent);
619 KKASSERT(chain == NULL);
625 * Set up the new inode.
627 * NOTE: *_get() integrates chain's lock into the inode lock.
629 * NOTE: Only one new inode can currently be created per
630 * transaction. If the need arises we can adjust
631 * hammer2_trans_init() to allow more.
633 * NOTE: nipdata will have chain's blockset data.
635 chain->data->ipdata.inum = trans->sync_tid;
636 nip = hammer2_inode_get(dip->pmp, dip, chain);
637 nipdata = &chain->data->ipdata;
640 KKASSERT(trans->inodes_created == 0);
641 nipdata->type = hammer2_get_obj_type(vap->va_type);
642 nipdata->inum = trans->sync_tid;
643 ++trans->inodes_created;
645 switch (nipdata->type) {
646 case HAMMER2_OBJTYPE_CDEV:
647 case HAMMER2_OBJTYPE_BDEV:
648 nipdata->rmajor = vap->va_rmajor;
649 nipdata->rminor = vap->va_rminor;
655 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
659 /* Inherit parent's inode compression mode. */
660 nip->comp_heuristic = 0;
661 nipdata->comp_algo = dip_algo;
662 nipdata->version = HAMMER2_INODE_VERSION_ONE;
663 hammer2_update_time(&nipdata->ctime);
664 nipdata->mtime = nipdata->ctime;
666 nipdata->mode = vap->va_mode;
669 if (dip && dip->pmp) {
670 xuid = hammer2_to_unix_xid(&dip_uid);
671 xuid = vop_helper_create_uid(dip->pmp->mp,
677 /* super-root has no dip and/or pmp */
680 if (vap->va_vaflags & VA_UID_UUID_VALID)
681 nipdata->uid = vap->va_uid_uuid;
682 else if (vap->va_uid != (uid_t)VNOVAL)
683 hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
685 hammer2_guid_to_uuid(&nipdata->uid, xuid);
687 if (vap->va_vaflags & VA_GID_UUID_VALID)
688 nipdata->gid = vap->va_gid_uuid;
689 else if (vap->va_gid != (gid_t)VNOVAL)
690 hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
692 nipdata->gid = dip_gid;
696 * Regular files and softlinks allow a small amount of data to be
697 * directly embedded in the inode. This flag will be cleared if
698 * the size is extended past the embedded limit.
700 if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
701 nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
702 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
705 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
706 bcopy(name, nipdata->filename, name_len);
707 nipdata->name_key = lhc;
708 nipdata->name_len = name_len;
715 * chain may have been moved around by the create.
718 hammer2_chain_refactor(hammer2_chain_t **chainp)
720 hammer2_chain_t *chain = *chainp;
721 hammer2_chain_core_t *core;
724 while (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
725 spin_lock(&core->cst.spin);
726 chain = TAILQ_NEXT(chain, core_entry);
727 while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
728 chain = TAILQ_NEXT(chain, core_entry);
729 hammer2_chain_ref(chain);
730 spin_unlock(&core->cst.spin);
731 KKASSERT(chain->core == core);
733 hammer2_chain_unlock(*chainp);
734 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
735 HAMMER2_RESOLVE_NOREF); /* eat ref */
741 * ochain represents the target file inode. We need to move it to the
742 * specified common parent directory (dip) and rename it to a special
743 * invisible "0xINODENUMBER" filename.
745 * We use chain_duplicate and duplicate ochain at the new location,
746 * renaming it appropriately. We create a temporary chain and
747 * then delete it to placemark where the duplicate will go. Both of
748 * these use the inode number for (lhc) (the key), generating the
749 * invisible filename.
753 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
754 hammer2_inode_t *dip, int *errorp)
756 hammer2_inode_data_t *nipdata;
757 hammer2_chain_t *parent;
758 hammer2_chain_t *ochain;
759 hammer2_chain_t *nchain;
760 hammer2_chain_t *tmp;
761 hammer2_key_t key_dummy;
763 hammer2_blockref_t bref;
764 int cache_index = -1;
768 lhc = ochain->data->ipdata.inum;
769 KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
772 * Locate the inode or indirect block to create the new
773 * entry in. lhc represents the inode number so there is
774 * no collision iteration.
776 * There should be no key collisions with invisible inode keys.
778 * WARNING! Must use inode_lock_ex() on dip to handle a stale
782 parent = hammer2_inode_lock_ex(dip);
783 /*parent = hammer2_chain_lookup_init(dip->chain, 0);*/
784 nchain = hammer2_chain_lookup(&parent, &key_dummy,
785 lhc, lhc, &cache_index, 0);
787 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
788 nchain, parent, dip, dip->chain);
789 hammer2_chain_unlock(nchain);
798 * Create entry in common parent directory using the seek position
802 KKASSERT(nchain == NULL);
803 *errorp = hammer2_chain_create(trans, &parent, &nchain,
805 HAMMER2_BREF_TYPE_INODE,/* n/a */
806 HAMMER2_INODE_BYTES); /* n/a */
807 hammer2_chain_refactor(&ochain);
812 * Cleanup and handle retries.
814 if (*errorp == EAGAIN) {
815 hammer2_chain_ref(parent);
816 /* hammer2_chain_lookup_done(parent); */
817 hammer2_inode_unlock_ex(dip, parent);
818 hammer2_chain_wait(parent);
819 hammer2_chain_drop(parent);
824 * Handle the error case
827 KKASSERT(nchain == NULL);
828 hammer2_inode_unlock_ex(dip, parent);
829 /*hammer2_chain_lookup_done(parent);*/
834 * Use chain as a placeholder for (lhc), delete it and replace
835 * it with our duplication.
837 * Gain a second lock on ochain for the duplication function to
838 * unlock, maintain the caller's original lock across the call.
840 * This is a bit messy.
842 hammer2_chain_delete(trans, nchain, HAMMER2_DELETE_WILLDUP);
843 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
846 bref.key = lhc; /* invisible dir entry key */
848 hammer2_chain_duplicate(trans, &parent, &tmp, &bref, 0);
849 hammer2_inode_unlock_ex(dip, parent);
850 /*hammer2_chain_lookup_done(parent);*/
851 hammer2_chain_unlock(nchain); /* no longer needed */
854 * Now set chain to our duplicate and modify it appropriately.
856 * Directory entries are inodes but this is a hidden hardlink
857 * target. The name isn't used but to ease debugging give it
858 * a name after its inode number.
861 tmp = NULL; /* safety */
863 hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
864 nipdata = &nchain->data->ipdata;
865 ksnprintf(nipdata->filename, sizeof(nipdata->filename),
866 "0x%016jx", (intmax_t)nipdata->inum);
867 nipdata->name_len = strlen(nipdata->filename);
868 nipdata->name_key = lhc;
874 * Connect the target inode represented by (*chainp) to the media topology
875 * at (dip, name, len).
877 * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
878 * entry instead of connecting (*chainp).
880 * If hlink is FALSE this function uses chain_duplicate() to make a copy
881 * if (*chainp) in the directory entry. (*chainp) is likely to be deleted
882 * by the caller in this case (e.g. rename).
885 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
886 hammer2_inode_t *dip, hammer2_chain_t **chainp,
887 const uint8_t *name, size_t name_len)
889 hammer2_inode_data_t *ipdata;
890 hammer2_chain_t *nchain;
891 hammer2_chain_t *parent;
892 hammer2_chain_t *ochain;
893 hammer2_key_t key_dummy;
895 int cache_index = -1;
901 * Since ochain is either disconnected from the topology or represents
902 * a hardlink terminus which is always a parent of or equal to dip,
903 * we should be able to safely lock dip->chain for our setup.
905 * WARNING! Must use inode_lock_ex() on dip to handle a stale
908 parent = hammer2_inode_lock_ex(dip);
909 /*parent = hammer2_chain_lookup_init(dip->chain, 0);*/
911 lhc = hammer2_dirhash(name, name_len);
914 * Locate the inode or indirect block to create the new
915 * entry in. At the same time check for key collisions
916 * and iterate until we don't get one.
920 nchain = hammer2_chain_lookup(&parent, &key_dummy,
921 lhc, lhc, &cache_index, 0);
924 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
926 hammer2_chain_unlock(nchain);
934 * Hardlink pointer needed, create totally fresh
937 KKASSERT(nchain == NULL);
938 error = hammer2_chain_create(trans, &parent, &nchain,
940 HAMMER2_BREF_TYPE_INODE,
941 HAMMER2_INODE_BYTES);
942 hammer2_chain_refactor(&ochain);
945 * Reconnect the original chain and rename. Use
946 * chain_duplicate(). The caller will likely delete
947 * or has already deleted the original chain in
950 * NOTE: chain_duplicate() generates a new chain
951 * with CHAIN_DELETED cleared (ochain typically
952 * has it set from the file unlink).
956 hammer2_chain_duplicate(trans, NULL, &nchain, NULL, 0);
957 error = hammer2_chain_create(trans, &parent, &nchain,
959 HAMMER2_BREF_TYPE_INODE,
960 HAMMER2_INODE_BYTES);
967 KKASSERT(error != EAGAIN);
968 hammer2_inode_unlock_ex(dip, parent);
969 /*hammer2_chain_lookup_done(parent);*/
973 * nchain should be NULL on error, leave ochain (== *chainp) alone.
976 KKASSERT(nchain == NULL);
981 * Directory entries are inodes so if the name has changed we have
982 * to update the inode.
984 * When creating an OBJTYPE_HARDLINK entry remember to unlock the
985 * chain, the caller will access the hardlink via the actual hardlink
986 * target file and not the hardlink pointer entry, so we must still
989 if (hlink && hammer2_hardlink_enable >= 0) {
991 * Create the HARDLINK pointer. oip represents the hardlink
992 * target in this situation.
994 * We will return ochain (the hardlink target).
996 hammer2_chain_modify(trans, &nchain,
997 HAMMER2_MODIFY_ASSERTNOCOPY);
998 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
999 ipdata = &nchain->data->ipdata;
1000 bcopy(name, ipdata->filename, name_len);
1001 ipdata->name_key = lhc;
1002 ipdata->name_len = name_len;
1003 ipdata->target_type = ochain->data->ipdata.type;
1004 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1005 ipdata->inum = ochain->data->ipdata.inum;
1007 hammer2_chain_unlock(nchain);
1010 } else if (hlink && hammer2_hardlink_enable < 0) {
1012 * Create a snapshot (hardlink fake mode for debugging).
1013 * (ochain already flushed above so we can just copy the
1016 * Since this is a snapshot we return nchain in the fake
1019 hammer2_chain_modify(trans, &nchain,
1020 HAMMER2_MODIFY_ASSERTNOCOPY);
1021 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1022 ipdata = &nchain->data->ipdata;
1023 *ipdata = ochain->data->ipdata;
1024 bcopy(name, ipdata->filename, name_len);
1025 ipdata->name_key = lhc;
1026 ipdata->name_len = name_len;
1027 atomic_clear_int(&nchain->core->flags,
1028 HAMMER2_CORE_COUNTEDBREFS);
1029 kprintf("created fake hardlink %*.*s\n",
1030 (int)name_len, (int)name_len, name);
1033 * nchain is a duplicate of ochain at the new location.
1034 * We must fixup the name stored in oip. The bref key
1035 * has already been set up.
1037 hammer2_chain_modify(trans, &nchain,
1038 HAMMER2_MODIFY_ASSERTNOCOPY);
1039 ipdata = &nchain->data->ipdata;
1041 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1042 bcopy(name, ipdata->filename, name_len);
1043 ipdata->name_key = lhc;
1044 ipdata->name_len = name_len;
1049 * We are replacing ochain with nchain, unlock ochain. In the
1050 * case where ochain is left unchanged the code above sets
1051 * nchain to ochain and ochain to NULL, resulting in a NOP here.
1054 hammer2_chain_unlock(ochain);
1061 * Repoint ip->chain to nchain. Caller must hold the inode exclusively
1064 * ip->chain is set to nchain. The prior chain in ip->chain is dropped
1065 * and nchain is ref'd.
1068 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1069 hammer2_chain_t *nchain)
1071 hammer2_chain_t *ochain;
1072 hammer2_inode_t *opip;
1075 * Repoint ip->chain if requested.
1080 hammer2_chain_ref(nchain);
1082 hammer2_chain_drop(ochain);
1085 * Repoint ip->pip if requested (non-NULL pip).
1087 if (pip && ip->pip != pip) {
1089 hammer2_inode_ref(pip);
1092 hammer2_inode_drop(opip);
1097 * Unlink the file from the specified directory inode. The directory inode
1098 * does not need to be locked.
1100 * isdir determines whether a directory/non-directory check should be made.
1101 * No check is made if isdir is set to -1.
1103 * NOTE! The underlying file can still be active with open descriptors
1104 * or if the chain is being manually held (e.g. for rename).
1106 * The caller is responsible for fixing up ip->chain if e.g. a
1107 * rename occurs (see chain_duplicate()).
1110 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1111 const uint8_t *name, size_t name_len,
1112 int isdir, int *hlinkp)
1114 hammer2_inode_data_t *ipdata;
1115 hammer2_chain_t *parent;
1116 hammer2_chain_t *ochain;
1117 hammer2_chain_t *chain;
1118 hammer2_chain_t *dparent;
1119 hammer2_chain_t *dchain;
1120 hammer2_key_t key_dummy;
1121 hammer2_key_t key_next;
1124 int cache_index = -1;
1129 lhc = hammer2_dirhash(name, name_len);
1132 * Search for the filename in the directory
1136 parent = hammer2_inode_lock_ex(dip);
1137 chain = hammer2_chain_lookup(&parent, &key_next,
1138 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1141 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1142 name_len == chain->data->ipdata.name_len &&
1143 bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1146 chain = hammer2_chain_next(&parent, chain, &key_next,
1148 lhc + HAMMER2_DIRHASH_LOMASK,
1151 hammer2_inode_unlock_ex(dip, NULL); /* retain parent */
1154 * Not found or wrong type (isdir < 0 disables the type check).
1155 * If a hardlink pointer, type checks use the hardlink target.
1157 if (chain == NULL) {
1161 if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1164 type = chain->data->ipdata.target_type;
1167 if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1171 if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1177 * Hardlink must be resolved. We can't hold parent locked while we
1178 * do this or we could deadlock.
1180 * On success chain will be adjusted to point at the hardlink target
1181 * and ochain will point to the hardlink pointer in the original
1182 * directory. Otherwise chain remains pointing to the original.
1184 if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1185 hammer2_chain_unlock(parent);
1187 error = hammer2_hardlink_find(dip, &chain, &ochain);
1191 * If this is a directory the directory must be empty. However, if
1192 * isdir < 0 we are doing a rename and the directory does not have
1193 * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1194 * and the directory does not have to be empty.
1196 * NOTE: We check the full key range here which covers both visible
1197 * and invisible entries. Theoretically there should be no
1198 * invisible (hardlink target) entries if there are no visible
1201 if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1202 dparent = hammer2_chain_lookup_init(chain, 0);
1203 dchain = hammer2_chain_lookup(&dparent, &key_dummy,
1204 0, (hammer2_key_t)-1,
1206 HAMMER2_LOOKUP_NODATA);
1208 hammer2_chain_unlock(dchain);
1209 hammer2_chain_lookup_done(dparent);
1213 hammer2_chain_lookup_done(dparent);
1219 * Ok, we can now unlink the chain. We always decrement nlinks even
1220 * if the entry can be deleted in case someone has the file open and
1223 * The chain itself will no longer be in the on-media topology but
1224 * can still be flushed to the media (e.g. if an open descriptor
1225 * remains). When the last vnode/ip ref goes away the chain will
1226 * be marked unmodified, avoiding any further (now unnecesary) I/O.
1228 * A non-NULL ochain indicates a hardlink.
1232 * Delete the original hardlink pointer.
1234 * NOTE: parent from above is NULL when ochain != NULL
1235 * so we can reuse it.
1237 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1238 hammer2_chain_delete(trans, ochain, 0);
1239 hammer2_chain_unlock(ochain);
1242 * Then decrement nlinks on hardlink target, deleting
1243 * the target when nlinks drops to 0.
1245 hammer2_chain_modify(trans, &chain, 0);
1246 --chain->data->ipdata.nlinks;
1247 if (chain->data->ipdata.nlinks == 0)
1248 hammer2_chain_delete(trans, chain, 0);
1251 * Otherwise this was not a hardlink and we can just
1252 * remove the entry and decrement nlinks.
1254 * NOTE: *_get() integrates chain's lock into the inode lock.
1256 hammer2_chain_modify(trans, &chain, 0);
1257 ipdata = &chain->data->ipdata;
1259 hammer2_chain_delete(trans, chain, 0);
1265 hammer2_chain_unlock(chain);
1267 hammer2_chain_lookup_done(parent);
1269 hammer2_chain_drop(ochain);
1275 * Given an exclusively locked inode we consolidate its chain for hardlink
1276 * creation, adding (nlinks) to the file's link count and potentially
1277 * relocating the inode to a directory common to ip->pip and tdip.
1279 * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1280 * and returning a new locked chain.
1282 * NOTE! This function will also replace ip->chain.
1285 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1286 hammer2_chain_t **chainp,
1287 hammer2_inode_t *tdip, int nlinks)
1289 hammer2_inode_data_t *ipdata;
1290 hammer2_inode_t *fdip;
1291 hammer2_inode_t *cdip;
1292 hammer2_chain_t *chain;
1293 hammer2_chain_t *nchain;
1297 if (nlinks == 0 && /* no hardlink needed */
1298 (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1301 if (hammer2_hardlink_enable < 0) { /* fake hardlinks */
1305 if (hammer2_hardlink_enable == 0) { /* disallow hardlinks */
1306 hammer2_chain_unlock(chain);
1312 * cdip will be returned with a ref, but not locked.
1315 cdip = hammer2_inode_common_parent(fdip, tdip);
1318 * If no change in the hardlink's target directory is required and
1319 * this is already a hardlink target, all we need to do is adjust
1322 * XXX The common parent is a big wiggly due to duplication from
1323 * renames. Compare the core (RBTREE) pointer instead of the
1327 (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1329 hammer2_chain_modify(trans, &chain, 0);
1330 chain->data->ipdata.nlinks += nlinks;
1337 * We either have to move an existing hardlink target or we have
1338 * to create a fresh hardlink target.
1340 * Hardlink targets are hidden inodes in a parent directory common
1341 * to all directory entries referencing the hardlink.
1343 nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1347 * Bump nlinks on duplicated hidden inode, repoint
1350 hammer2_chain_modify(trans, &nchain, 0);
1351 nchain->data->ipdata.nlinks += nlinks;
1352 hammer2_inode_repoint(ip, cdip, nchain);
1355 * If the old chain is not a hardlink target then replace
1356 * it with a OBJTYPE_HARDLINK pointer.
1358 * If the old chain IS a hardlink target then delete it.
1360 if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1362 * Replace original non-hardlink that's been dup'd
1363 * with a special hardlink directory entry. We must
1364 * set the DIRECTDATA flag to prevent sub-chains
1365 * from trying to synchronize to the inode if the
1366 * file is extended afterwords.
1368 * DELDUP_RECORE causes the new chain to NOT inherit
1369 * the old chain's core (sub-tree). The new chain
1370 * will already be marked modified.
1372 /*hammer2_chain_modify(trans, &chain, 0);*/
1373 hammer2_chain_delete_duplicate(trans, &chain,
1374 HAMMER2_DELDUP_RECORE);
1375 ipdata = &chain->data->ipdata;
1376 ipdata->target_type = ipdata->type;
1377 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1385 bzero(&ipdata->uid, sizeof(ipdata->uid));
1386 bzero(&ipdata->gid, sizeof(ipdata->gid));
1387 ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1388 ipdata->cap_flags = 0;
1392 ipdata->iparent = 0; /* XXX */
1393 ipdata->pfs_type = 0;
1394 ipdata->pfs_inum = 0;
1395 bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1396 bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1397 ipdata->data_quota = 0;
1398 ipdata->data_count = 0;
1399 ipdata->inode_quota = 0;
1400 ipdata->inode_count = 0;
1401 ipdata->attr_tid = 0;
1402 ipdata->dirent_tid = 0;
1403 bzero(&ipdata->u, sizeof(ipdata->u));
1404 /* XXX transaction ids */
1406 hammer2_chain_delete(trans, chain, 0);
1410 * Return the new chain.
1412 hammer2_chain_unlock(chain);
1418 hammer2_chain_unlock(chain);
1423 * Cleanup, chain/nchain already dealt with.
1427 hammer2_inode_drop(cdip);
1433 * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1434 * inode while (*chainp) points to the resolved (hidden hardlink
1435 * target) inode. In this situation when nlinks is 1 we wish to
1436 * deconsolidate the hardlink, moving it back to the directory that now
1437 * represents the only remaining link.
1440 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1441 hammer2_inode_t *dip,
1442 hammer2_chain_t **chainp,
1443 hammer2_chain_t **ochainp)
1445 if (*ochainp == NULL)
1452 * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1453 * with an obj_type of HAMMER2_OBJTYPE_HARDLINK. This routine will gobble
1454 * the *chainp and return a new locked *chainp representing the file target
1455 * (the original *chainp will be unlocked).
1457 * When a match is found the chain representing the original HARDLINK
1458 * will be returned in *ochainp with a ref, but not locked.
1460 * When no match is found *chainp is set to NULL and EIO is returned.
1461 * (*ochainp) will still be set to the original chain with a ref but not
1465 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1466 hammer2_chain_t **ochainp)
1468 hammer2_chain_t *chain = *chainp;
1469 hammer2_chain_t *parent;
1470 hammer2_inode_t *ip;
1471 hammer2_inode_t *pip;
1472 hammer2_key_t key_dummy;
1474 int cache_index = -1;
1477 hammer2_inode_ref(pip); /* for loop */
1478 hammer2_chain_ref(chain); /* for (*ochainp) */
1482 * Locate the hardlink. pip is referenced and not locked,
1487 lhc = chain->data->ipdata.inum;
1488 hammer2_chain_unlock(chain);
1491 while ((ip = pip) != NULL) {
1492 parent = hammer2_inode_lock_ex(ip);
1493 hammer2_inode_drop(ip); /* loop */
1494 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1495 chain = hammer2_chain_lookup(&parent, &key_dummy,
1496 lhc, lhc, &cache_index, 0);
1497 hammer2_chain_lookup_done(parent); /* discard parent */
1500 pip = ip->pip; /* safe, ip held locked */
1502 hammer2_inode_ref(pip); /* loop */
1503 hammer2_inode_unlock_ex(ip, NULL);
1507 * chain is locked, ip is locked. Unlock ip, return the locked
1508 * chain. *ipp is already set w/a ref count and not locked.
1510 * (parent is already unlocked).
1513 hammer2_inode_unlock_ex(ip, NULL);
1516 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1517 /* already locked */
1525 * Find the directory common to both fdip and tdip, hold and return
1529 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1531 hammer2_inode_t *scan1;
1532 hammer2_inode_t *scan2;
1535 * We used to have a depth field but it complicated matters too
1536 * much for directory renames. So now its ugly. Check for
1537 * simple cases before giving up and doing it the expensive way.
1539 * XXX need a bottom-up topology stability lock
1541 if (fdip == tdip || fdip == tdip->pip) {
1542 hammer2_inode_ref(fdip);
1545 if (fdip->pip == tdip) {
1546 hammer2_inode_ref(tdip);
1553 for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1555 while (scan2->pmp == tdip->pmp) {
1556 if (scan1 == scan2) {
1557 hammer2_inode_ref(scan1);
1565 panic("hammer2_inode_common_parent: no common parent %p %p\n",
1572 * Synchronize the inode's frontend state with the chain state prior
1573 * to any explicit flush of the inode or any strategy write call.
1575 * Called with a locked inode.
1578 hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
1579 hammer2_chain_t **chainp)
1581 hammer2_inode_data_t *ipdata;
1582 hammer2_chain_t *parent;
1583 hammer2_chain_t *chain;
1584 hammer2_key_t lbase;
1585 hammer2_key_t key_next;
1588 ipdata = &ip->chain->data->ipdata;
1590 if (ip->flags & HAMMER2_INODE_MTIME) {
1591 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1592 atomic_clear_int(&ip->flags, HAMMER2_INODE_MTIME);
1593 ipdata->mtime = ip->mtime;
1595 if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size < ipdata->size) {
1596 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1597 ipdata->size = ip->size;
1598 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1601 * We must delete any chains beyond the EOF. The chain
1602 * straddling the EOF will be pending in the bioq.
1604 lbase = (ipdata->size + HAMMER2_PBUFMASK64) &
1605 ~HAMMER2_PBUFMASK64;
1606 parent = hammer2_chain_lookup_init(ip->chain, 0);
1607 chain = hammer2_chain_lookup(&parent, &key_next,
1608 lbase, (hammer2_key_t)-1,
1610 HAMMER2_LOOKUP_NODATA);
1613 * Degenerate embedded case, nothing to loop on
1615 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1616 hammer2_chain_unlock(chain);
1619 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
1620 hammer2_chain_delete(trans, chain, 0);
1622 chain = hammer2_chain_next(&parent, chain, &key_next,
1623 key_next, (hammer2_key_t)-1,
1625 HAMMER2_LOOKUP_NODATA);
1627 hammer2_chain_lookup_done(parent);
1629 if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size > ipdata->size) {
1630 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1631 ipdata->size = ip->size;
1632 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1635 * When resizing larger we may not have any direct-data
1638 if ((ipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
1639 ip->size > HAMMER2_EMBEDDED_BYTES) {
1640 ipdata->op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
1641 bzero(&ipdata->u.blockset, sizeof(ipdata->u.blockset));