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 * The underlying chain is also locked and returned.
68 * NOTE: We don't combine the inode/chain lock because putting away an
69 * inode would otherwise confuse multiple lock holders of the inode.
72 hammer2_inode_lock_ex(hammer2_inode_t *ip)
74 hammer2_chain_t *chain;
76 hammer2_inode_ref(ip);
77 ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
80 * ip->chain fixup. Certain duplications used to move inodes
81 * into indirect blocks (for example) can cause ip->chain to
86 if (hammer2_chain_refactor_test(chain, 1)) {
87 spin_lock(&chain->core->cst.spin);
88 while (hammer2_chain_refactor_test(chain, 1))
89 chain = chain->next_parent;
90 if (ip->chain != chain) {
91 hammer2_chain_ref(chain);
92 spin_unlock(&chain->core->cst.spin);
93 hammer2_inode_repoint(ip, NULL, chain);
94 hammer2_chain_drop(chain);
96 spin_unlock(&chain->core->cst.spin);
100 KKASSERT(chain != NULL); /* for now */
101 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
104 * Resolve duplication races
106 if (hammer2_chain_refactor_test(chain, 1)) {
107 hammer2_chain_unlock(chain);
114 hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
117 * XXX this will catch parent directories too which we don't
121 hammer2_chain_unlock(chain);
122 ccms_thread_unlock(&ip->topo_cst);
123 hammer2_inode_drop(ip);
127 * NOTE: We don't combine the inode/chain lock because putting away an
128 * inode would otherwise confuse multiple lock holders of the inode.
130 * Shared locks are especially sensitive to having too many shared
131 * lock counts (from the same thread) on certain paths which might
132 * need to upgrade them. Only one count of a shared lock can be
136 hammer2_inode_lock_sh(hammer2_inode_t *ip)
138 hammer2_chain_t *chain;
140 hammer2_inode_ref(ip);
142 ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
145 KKASSERT(chain != NULL); /* for now */
146 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
147 HAMMER2_RESOLVE_SHARED);
150 * Resolve duplication races
152 if (hammer2_chain_refactor_test(chain, 1)) {
153 hammer2_chain_unlock(chain);
154 ccms_thread_unlock(&ip->topo_cst);
155 chain = hammer2_inode_lock_ex(ip);
156 hammer2_inode_unlock_ex(ip, chain);
163 hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
166 hammer2_chain_unlock(chain);
167 ccms_thread_unlock(&ip->topo_cst);
168 hammer2_inode_drop(ip);
172 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
174 return(ccms_thread_lock_temp_release(&ip->topo_cst));
178 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate)
180 ccms_thread_lock_temp_restore(&ip->topo_cst, ostate);
184 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
186 return(ccms_thread_lock_upgrade(&ip->topo_cst));
190 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate)
192 ccms_thread_lock_downgrade(&ip->topo_cst, ostate);
196 * Lookup an inode by inode number
199 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
204 spin_lock(&pmp->inum_spin);
205 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
207 hammer2_inode_ref(ip);
208 spin_unlock(&pmp->inum_spin);
216 * Adding a ref to an inode is only legal if the inode already has at least
220 hammer2_inode_ref(hammer2_inode_t *ip)
222 atomic_add_int(&ip->refs, 1);
226 * Drop an inode reference, freeing the inode when the last reference goes
230 hammer2_inode_drop(hammer2_inode_t *ip)
232 hammer2_pfsmount_t *pmp;
233 hammer2_inode_t *pip;
241 * Transition to zero, must interlock with
242 * the inode inumber lookup tree (if applicable).
244 * NOTE: The super-root inode has no pmp.
248 spin_lock(&pmp->inum_spin);
250 if (atomic_cmpset_int(&ip->refs, 1, 0)) {
251 KKASSERT(ip->topo_cst.count == 0);
252 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
253 atomic_clear_int(&ip->flags,
254 HAMMER2_INODE_ONRBTREE);
255 RB_REMOVE(hammer2_inode_tree,
256 &pmp->inum_tree, ip);
259 spin_unlock(&pmp->inum_spin);
266 * Cleaning out ip->chain isn't entirely
269 hammer2_inode_repoint(ip, NULL, NULL);
272 * We have to drop pip (if non-NULL) to
273 * dispose of our implied reference from
274 * ip->pip. We can simply loop on it.
277 KKASSERT((ip->flags &
278 HAMMER2_INODE_SROOT) == 0);
279 kfree(ip, pmp->minode);
282 HAMMER2_INODE_SROOT);
283 kfree(ip, M_HAMMER2);
286 /* continue with pip (can be NULL) */
289 spin_unlock(&ip->pmp->inum_spin);
293 * Non zero transition
295 if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
302 * Get the vnode associated with the given inode, allocating the vnode if
303 * necessary. The vnode will be returned exclusively locked.
305 * The caller must lock the inode (shared or exclusive).
307 * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
311 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
313 hammer2_inode_data_t *ipdata;
314 hammer2_pfsmount_t *pmp;
319 KKASSERT(pmp != NULL);
321 ipdata = &ip->chain->data->ipdata;
325 * Attempt to reuse an existing vnode assignment. It is
326 * possible to race a reclaim so the vget() may fail. The
327 * inode must be unlocked during the vget() to avoid a
328 * deadlock against a reclaim.
333 * Inode must be unlocked during the vget() to avoid
334 * possible deadlocks, but leave the ip ref intact.
336 * vnode is held to prevent destruction during the
337 * vget(). The vget() can still fail if we lost
338 * a reclaim race on the vnode.
340 vhold_interlocked(vp);
341 ostate = hammer2_inode_lock_temp_release(ip);
342 if (vget(vp, LK_EXCLUSIVE)) {
344 hammer2_inode_lock_temp_restore(ip, ostate);
347 hammer2_inode_lock_temp_restore(ip, ostate);
349 /* vp still locked and ref from vget */
351 kprintf("hammer2: igetv race %p/%p\n",
361 * No vnode exists, allocate a new vnode. Beware of
362 * allocation races. This function will return an
363 * exclusively locked and referenced vnode.
365 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
367 kprintf("hammer2: igetv getnewvnode failed %d\n",
374 * Lock the inode and check for an allocation race.
376 ostate = hammer2_inode_lock_upgrade(ip);
377 if (ip->vp != NULL) {
380 hammer2_inode_lock_downgrade(ip, ostate);
384 switch (ipdata->type) {
385 case HAMMER2_OBJTYPE_DIRECTORY:
388 case HAMMER2_OBJTYPE_REGFILE:
390 vinitvmio(vp, ipdata->size,
392 (int)ipdata->size & HAMMER2_LBUFMASK);
394 case HAMMER2_OBJTYPE_SOFTLINK:
396 * XXX for now we are using the generic file_read
397 * and file_write code so we need a buffer cache
401 vinitvmio(vp, ipdata->size,
403 (int)ipdata->size & HAMMER2_LBUFMASK);
407 panic("hammer2: unhandled objtype %d", ipdata->type);
411 if (ip == pmp->iroot)
412 vsetflags(vp, VROOT);
416 hammer2_inode_ref(ip); /* vp association */
417 hammer2_inode_lock_downgrade(ip, ostate);
422 * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
424 if (hammer2_debug & 0x0002) {
425 kprintf("igetv vp %p refs %d aux %d\n",
426 vp, vp->v_sysref.refcnt, vp->v_auxrefs);
432 * The passed-in chain must be locked and the returned inode will also be
433 * locked. This routine typically locates or allocates the inode, assigns
434 * ip->chain (adding a ref to chain if necessary), and returns the inode.
436 * The hammer2_inode structure regulates the interface between the high level
437 * kernel VNOPS API and the filesystem backend (the chains).
439 * WARNING! This routine sucks up the chain's lock (makes it part of the
440 * inode lock from the point of view of the inode lock API),
441 * so callers need to be careful.
443 * WARNING! The mount code is allowed to pass dip == NULL for iroot and
444 * is allowed to pass pmp == NULL and dip == NULL for sroot.
447 hammer2_inode_get(hammer2_pfsmount_t *pmp, hammer2_inode_t *dip,
448 hammer2_chain_t *chain)
450 hammer2_inode_t *nip;
452 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
455 * Interlocked lookup/ref of the inode. This code is only needed
456 * when looking up inodes with nlinks != 0 (TODO: optimize out
457 * otherwise and test for duplicates).
461 nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
464 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
465 if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { /* race */
466 ccms_thread_unlock(&nip->topo_cst);
467 hammer2_inode_drop(nip);
470 if (nip->chain != chain)
471 hammer2_inode_repoint(nip, NULL, chain);
474 * Consolidated nip/nip->chain is locked (chain locked
481 * We couldn't find the inode number, create a new inode.
484 nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
486 nip = kmalloc(sizeof(*nip), M_HAMMER2, M_WAITOK | M_ZERO);
487 nip->flags = HAMMER2_INODE_SROOT;
489 nip->inum = chain->data->ipdata.inum;
490 hammer2_inode_repoint(nip, NULL, chain);
491 nip->pip = dip; /* can be NULL */
493 hammer2_inode_ref(dip); /* ref dip for nip->pip */
498 * ref and lock on nip gives it state compatible to after a
499 * hammer2_inode_lock_ex() call.
502 ccms_cst_init(&nip->topo_cst, &nip->chain);
503 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
504 /* combination of thread lock and chain lock == inode lock */
507 * Attempt to add the inode. If it fails we raced another inode
508 * get. Undo all the work and try again.
511 spin_lock(&pmp->inum_spin);
512 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
513 spin_unlock(&pmp->inum_spin);
514 ccms_thread_unlock(&nip->topo_cst);
515 hammer2_inode_drop(nip);
518 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
519 spin_unlock(&pmp->inum_spin);
526 * Create a new inode in the specified directory using the vattr to
527 * figure out the type of inode.
529 * If no error occurs the new inode with its chain locked is returned in
530 * *nipp, otherwise an error is returned and *nipp is set to NULL.
532 * If vap and/or cred are NULL the related fields are not set and the
533 * inode type defaults to a directory. This is used when creating PFSs
534 * under the super-root, so the inode number is set to 1 in this case.
536 * dip is not locked on entry.
539 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
540 struct vattr *vap, struct ucred *cred,
541 const uint8_t *name, size_t name_len,
542 hammer2_chain_t **chainp, int *errorp)
544 hammer2_inode_data_t *dipdata;
545 hammer2_inode_data_t *nipdata;
546 hammer2_chain_t *chain;
547 hammer2_chain_t *parent;
548 hammer2_inode_t *nip;
556 lhc = hammer2_dirhash(name, name_len);
560 * Locate the inode or indirect block to create the new
561 * entry in. At the same time check for key collisions
562 * and iterate until we don't get one.
564 * NOTE: hidden inodes do not have iterators.
567 parent = hammer2_inode_lock_ex(dip);
568 dipdata = &dip->chain->data->ipdata;
569 dip_uid = dipdata->uid;
570 dip_gid = dipdata->gid;
571 dip_mode = dipdata->mode;
575 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
578 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
580 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
582 hammer2_chain_unlock(chain);
587 error = hammer2_chain_create(trans, &parent, &chain,
589 HAMMER2_BREF_TYPE_INODE,
590 HAMMER2_INODE_BYTES);
594 * Cleanup and handle retries.
596 if (error == EAGAIN) {
597 hammer2_chain_ref(parent);
598 hammer2_inode_unlock_ex(dip, parent);
599 hammer2_chain_wait(parent);
600 hammer2_chain_drop(parent);
603 hammer2_inode_unlock_ex(dip, parent);
606 KKASSERT(chain == NULL);
612 * Set up the new inode.
614 * NOTE: *_get() integrates chain's lock into the inode lock.
616 * NOTE: Only one new inode can currently be created per
617 * transaction. If the need arises we can adjust
618 * hammer2_trans_init() to allow more.
620 chain->data->ipdata.inum = trans->sync_tid;
621 nip = hammer2_inode_get(dip->pmp, dip, chain);
622 nipdata = &chain->data->ipdata;
625 KKASSERT(trans->inodes_created == 0);
626 nipdata->type = hammer2_get_obj_type(vap->va_type);
627 nipdata->inum = trans->sync_tid;
628 ++trans->inodes_created;
630 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
633 nipdata->version = HAMMER2_INODE_VERSION_ONE;
634 hammer2_update_time(&nipdata->ctime);
635 nipdata->mtime = nipdata->ctime;
637 nipdata->mode = vap->va_mode;
641 xuid = hammer2_to_unix_xid(&dip_uid);
642 xuid = vop_helper_create_uid(dip->pmp->mp,
650 if (vap->va_vaflags & VA_UID_UUID_VALID)
651 nipdata->uid = vap->va_uid_uuid;
652 else if (vap->va_uid != (uid_t)VNOVAL)
653 hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
655 hammer2_guid_to_uuid(&nipdata->uid, xuid);
657 if (vap->va_vaflags & VA_GID_UUID_VALID)
658 nipdata->gid = vap->va_gid_uuid;
659 else if (vap->va_gid != (gid_t)VNOVAL)
660 hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
662 nipdata->gid = dip_gid;
666 * Regular files and softlinks allow a small amount of data to be
667 * directly embedded in the inode. This flag will be cleared if
668 * the size is extended past the embedded limit.
670 if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
671 nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
672 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
675 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
676 bcopy(name, nipdata->filename, name_len);
677 nipdata->name_key = lhc;
678 nipdata->name_len = name_len;
685 * chain may have been moved around by the create.
689 hammer2_chain_refactor(hammer2_chain_t **chainp)
691 hammer2_chain_t *chain = *chainp;
692 hammer2_chain_core_t *core;
695 spin_lock(&core->cst.spin);
696 while (hammer2_chain_refactor_test(chain, 1)) {
697 chain = chain->next_parent;
698 while (hammer2_chain_refactor_test(chain, 1))
699 chain = chain->next_parent;
700 hammer2_chain_ref(chain);
701 spin_unlock(&core->cst.spin);
703 hammer2_chain_unlock(*chainp);
704 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
705 HAMMER2_RESOLVE_NOREF); /* eat ref */
707 spin_lock(&core->cst.spin);
709 spin_unlock(&core->cst.spin);
713 * ochain represents the target file inode. We need to move it to the
714 * specified common parent directory (dip) and rename it to a special
715 * invisible "0xINODENUMBER" filename.
717 * We use chain_duplicate and duplicate ochain at the new location,
718 * renaming it appropriately. We create a temporary chain and
719 * then delete it to placemark where the duplicate will go. Both of
720 * these use the inode number for (lhc) (the key), generating the
721 * invisible filename.
725 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
726 hammer2_inode_t *dip, int *errorp)
728 hammer2_inode_data_t *nipdata;
729 hammer2_chain_t *parent;
730 hammer2_chain_t *ochain;
731 hammer2_chain_t *nchain;
732 hammer2_chain_t *tmp;
734 hammer2_blockref_t bref;
738 lhc = ochain->data->ipdata.inum;
739 KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
742 * Locate the inode or indirect block to create the new
743 * entry in. lhc represents the inode number so there is
744 * no collision iteration.
746 * There should be no key collisions with invisible inode keys.
749 parent = hammer2_chain_lookup_init(dip->chain, 0);
750 nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
752 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
753 nchain, parent, dip, dip->chain);
754 hammer2_chain_unlock(nchain);
763 * Create entry in common parent directory using the seek position
767 KKASSERT(nchain == NULL);
768 *errorp = hammer2_chain_create(trans, &parent, &nchain,
770 HAMMER2_BREF_TYPE_INODE,/* n/a */
771 HAMMER2_INODE_BYTES); /* n/a */
772 hammer2_chain_refactor(&ochain);
777 * Cleanup and handle retries.
779 if (*errorp == EAGAIN) {
780 hammer2_chain_ref(parent);
781 hammer2_chain_lookup_done(parent);
782 hammer2_chain_wait(parent);
783 hammer2_chain_drop(parent);
788 * Handle the error case
791 KKASSERT(nchain == NULL);
792 hammer2_chain_lookup_done(parent);
797 * Use chain as a placeholder for (lhc), delete it and replace
798 * it with our duplication.
800 * Gain a second lock on ochain for the duplication function to
801 * unlock, maintain the caller's original lock across the call.
803 * This is a bit messy.
805 hammer2_chain_delete(trans, nchain);
806 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
809 bref.key = lhc; /* invisible dir entry key */
811 hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
812 hammer2_chain_lookup_done(parent);
813 hammer2_chain_unlock(nchain); /* no longer needed */
816 * Now set chain to our duplicate and modify it appropriately.
818 * Directory entries are inodes but this is a hidden hardlink
819 * target. The name isn't used but to ease debugging give it
820 * a name after its inode number.
823 tmp = NULL; /* safety */
825 hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
826 nipdata = &nchain->data->ipdata;
827 ksnprintf(nipdata->filename, sizeof(nipdata->filename),
828 "0x%016jx", (intmax_t)nipdata->inum);
829 nipdata->name_len = strlen(nipdata->filename);
830 nipdata->name_key = lhc;
836 * Connect the target inode represented by (*chainp) to the media topology
837 * at (dip, name, len).
839 * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
840 * entry instead of connecting (*chainp).
842 * If hlink is FALSE this function uses chain_duplicate() to make a copy
843 * if (*chainp) in the directory entry. (*chainp) is likely to be deleted
844 * by the caller in this case (e.g. rename).
847 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
848 hammer2_inode_t *dip, hammer2_chain_t **chainp,
849 const uint8_t *name, size_t name_len)
851 hammer2_inode_data_t *ipdata;
852 hammer2_chain_t *nchain;
853 hammer2_chain_t *parent;
854 hammer2_chain_t *ochain;
861 * Since ochain is either disconnected from the topology or represents
862 * a hardlink terminus which is always a parent of or equal to dip,
863 * we should be able to safely lock dip->chain for our setup.
865 parent = hammer2_chain_lookup_init(dip->chain, 0);
867 lhc = hammer2_dirhash(name, name_len);
870 * Locate the inode or indirect block to create the new
871 * entry in. At the same time check for key collisions
872 * and iterate until we don't get one.
876 nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
879 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
881 hammer2_chain_unlock(nchain);
889 * Hardlink pointer needed, create totally fresh
892 KKASSERT(nchain == NULL);
893 error = hammer2_chain_create(trans, &parent, &nchain,
895 HAMMER2_BREF_TYPE_INODE,
896 HAMMER2_INODE_BYTES);
897 hammer2_chain_refactor(&ochain);
900 * Reconnect the original chain and rename. Use
901 * chain_duplicate(). The caller will likely delete
902 * or has already deleted the original chain in
905 * NOTE: chain_duplicate() generates a new chain
906 * with CHAIN_DELETED cleared (ochain typically
907 * has it set from the file unlink).
911 hammer2_chain_duplicate(trans, NULL, -1, &nchain, NULL);
912 error = hammer2_chain_create(trans, &parent, &nchain,
914 HAMMER2_BREF_TYPE_INODE,
915 HAMMER2_INODE_BYTES);
922 KKASSERT(error != EAGAIN);
923 hammer2_chain_lookup_done(parent);
927 * nchain should be NULL on error, leave ochain (== *chainp) alone.
930 KKASSERT(nchain == NULL);
935 * Directory entries are inodes so if the name has changed we have
936 * to update the inode.
938 * When creating an OBJTYPE_HARDLINK entry remember to unlock the
939 * chain, the caller will access the hardlink via the actual hardlink
940 * target file and not the hardlink pointer entry, so we must still
943 if (hlink && hammer2_hardlink_enable >= 0) {
945 * Create the HARDLINK pointer. oip represents the hardlink
946 * target in this situation.
948 * We will return ochain (the hardlink target).
950 hammer2_chain_modify(trans, &nchain,
951 HAMMER2_MODIFY_ASSERTNOCOPY);
952 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
953 ipdata = &nchain->data->ipdata;
954 bcopy(name, ipdata->filename, name_len);
955 ipdata->name_key = lhc;
956 ipdata->name_len = name_len;
957 ipdata->target_type = ochain->data->ipdata.type;
958 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
959 ipdata->inum = ochain->data->ipdata.inum;
961 hammer2_chain_unlock(nchain);
964 } else if (hlink && hammer2_hardlink_enable < 0) {
966 * Create a snapshot (hardlink fake mode for debugging).
967 * (ochain already flushed above so we can just copy the
970 * Since this is a snapshot we return nchain in the fake
973 hammer2_chain_modify(trans, &nchain,
974 HAMMER2_MODIFY_ASSERTNOCOPY);
975 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
976 ipdata = &nchain->data->ipdata;
977 *ipdata = ochain->data->ipdata;
978 bcopy(name, ipdata->filename, name_len);
979 ipdata->name_key = lhc;
980 ipdata->name_len = name_len;
981 kprintf("created fake hardlink %*.*s\n",
982 (int)name_len, (int)name_len, name);
985 * nchain is a duplicate of ochain at the new location.
986 * We must fixup the name stored in oip. The bref key
987 * has already been set up.
989 hammer2_chain_modify(trans, &nchain,
990 HAMMER2_MODIFY_ASSERTNOCOPY);
991 ipdata = &nchain->data->ipdata;
993 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
994 bcopy(name, ipdata->filename, name_len);
995 ipdata->name_key = lhc;
996 ipdata->name_len = name_len;
1001 * We are replacing ochain with nchain, unlock ochain. In the
1002 * case where ochain is left unchanged the code above sets
1003 * nchain to ochain and ochain to NULL, resulting in a NOP here.
1006 hammer2_chain_unlock(ochain);
1013 * Repoint ip->chain to nchain. Caller must hold the inode exclusively
1016 * ip->chain is set to nchain. The prior chain in ip->chain is dropped
1017 * and nchain is ref'd.
1020 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1021 hammer2_chain_t *nchain)
1023 hammer2_chain_t *ochain;
1024 hammer2_inode_t *opip;
1027 * Repoint ip->chain if requested.
1032 hammer2_chain_ref(nchain);
1034 hammer2_chain_drop(ochain);
1037 * Repoint ip->pip if requested (non-NULL pip).
1039 if (pip && ip->pip != pip) {
1041 hammer2_inode_ref(pip);
1044 hammer2_inode_drop(opip);
1049 * Unlink the file from the specified directory inode. The directory inode
1050 * does not need to be locked.
1052 * isdir determines whether a directory/non-directory check should be made.
1053 * No check is made if isdir is set to -1.
1055 * NOTE! This function does not prevent the underlying file from still
1056 * being used if it has other refs (such as from an inode, or if it's
1057 * chain is manually held). However, the caller is responsible for
1058 * fixing up ip->chain if e.g. a rename occurs (see chain_duplicate()).
1061 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1062 const uint8_t *name, size_t name_len,
1063 int isdir, int *hlinkp)
1065 hammer2_inode_data_t *ipdata;
1066 hammer2_chain_t *parent;
1067 hammer2_chain_t *ochain;
1068 hammer2_chain_t *chain;
1069 hammer2_chain_t *dparent;
1070 hammer2_chain_t *dchain;
1077 lhc = hammer2_dirhash(name, name_len);
1080 * Search for the filename in the directory
1084 parent = hammer2_inode_lock_ex(dip);
1085 chain = hammer2_chain_lookup(&parent,
1086 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1089 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1090 name_len == chain->data->ipdata.name_len &&
1091 bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1094 chain = hammer2_chain_next(&parent, chain,
1095 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1098 hammer2_inode_unlock_ex(dip, NULL); /* retain parent */
1101 * Not found or wrong type (isdir < 0 disables the type check).
1102 * If a hardlink pointer, type checks use the hardlink target.
1104 if (chain == NULL) {
1108 if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1111 type = chain->data->ipdata.target_type;
1114 if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1118 if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1124 * Hardlink must be resolved. We can't hold parent locked while we
1125 * do this or we could deadlock.
1127 * On success chain will be adjusted to point at the hardlink target
1128 * and ochain will point to the hardlink pointer in the original
1129 * directory. Otherwise chain remains pointing to the original.
1131 if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1132 hammer2_chain_unlock(parent);
1134 error = hammer2_hardlink_find(dip, &chain, &ochain);
1138 * If this is a directory the directory must be empty. However, if
1139 * isdir < 0 we are doing a rename and the directory does not have
1140 * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1141 * and the directory does not have to be empty.
1143 * NOTE: We check the full key range here which covers both visible
1144 * and invisible entries. Theoretically there should be no
1145 * invisible (hardlink target) entries if there are no visible
1148 if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1149 dparent = hammer2_chain_lookup_init(chain, 0);
1150 dchain = hammer2_chain_lookup(&dparent,
1151 0, (hammer2_key_t)-1,
1152 HAMMER2_LOOKUP_NODATA);
1154 hammer2_chain_unlock(dchain);
1155 hammer2_chain_lookup_done(dparent);
1159 hammer2_chain_lookup_done(dparent);
1165 * Ok, we can now unlink the chain. We always decrement nlinks even
1166 * if the entry can be deleted in case someone has the file open and
1169 * The chain itself will no longer be in the on-media topology but
1170 * can still be flushed to the media (e.g. if an open descriptor
1171 * remains). When the last vnode/ip ref goes away the chain will
1172 * be marked unmodified, avoiding any further (now unnecesary) I/O.
1174 * A non-NULL ochain indicates a hardlink.
1178 * Delete the original hardlink pointer.
1180 * NOTE: parent from above is NULL when ochain != NULL
1181 * so we can reuse it.
1183 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1184 hammer2_chain_delete(trans, ochain);
1185 hammer2_chain_unlock(ochain);
1188 * Then decrement nlinks on hardlink target, deleting
1189 * the target when nlinks drops to 0.
1191 hammer2_chain_modify(trans, &chain, 0);
1192 --chain->data->ipdata.nlinks;
1193 if (chain->data->ipdata.nlinks == 0)
1194 hammer2_chain_delete(trans, chain);
1197 * Otherwise this was not a hardlink and we can just
1198 * remove the entry and decrement nlinks.
1200 * NOTE: *_get() integrates chain's lock into the inode lock.
1202 hammer2_chain_modify(trans, &chain, 0);
1203 ipdata = &chain->data->ipdata;
1205 hammer2_chain_delete(trans, chain);
1211 hammer2_chain_unlock(chain);
1213 hammer2_chain_lookup_done(parent);
1215 hammer2_chain_drop(ochain);
1221 * Given an exclusively locked inode we consolidate its chain for hardlink
1222 * creation, adding (nlinks) to the file's link count and potentially
1223 * relocating the inode to a directory common to ip->pip and tdip.
1225 * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1226 * and returning a new locked chain.
1228 * NOTE! This function will also replace ip->chain.
1231 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1232 hammer2_chain_t **chainp,
1233 hammer2_inode_t *tdip, int nlinks)
1235 hammer2_inode_data_t *ipdata;
1236 hammer2_inode_t *fdip;
1237 hammer2_inode_t *cdip;
1238 hammer2_chain_t *chain;
1239 hammer2_chain_t *nchain;
1243 if (nlinks == 0 && /* no hardlink needed */
1244 (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1247 if (hammer2_hardlink_enable < 0) { /* fake hardlinks */
1251 if (hammer2_hardlink_enable == 0) { /* disallow hardlinks */
1252 hammer2_chain_unlock(chain);
1258 * cdip will be returned with a ref, but not locked.
1261 cdip = hammer2_inode_common_parent(fdip, tdip);
1264 * If no change in the hardlink's target directory is required and
1265 * this is already a hardlink target, all we need to do is adjust
1268 * XXX The common parent is a big wiggly due to duplication from
1269 * renames. Compare the core (RBTREE) pointer instead of the
1273 (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1275 hammer2_chain_modify(trans, &chain, 0);
1276 chain->data->ipdata.nlinks += nlinks;
1283 * We either have to move an existing hardlink target or we have
1284 * to create a fresh hardlink target.
1286 * Hardlink targets are hidden inodes in a parent directory common
1287 * to all directory entries referencing the hardlink.
1289 nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1293 * Bump nlinks on duplicated hidden inode, repoint
1296 hammer2_chain_modify(trans, &nchain, 0);
1297 nchain->data->ipdata.nlinks += nlinks;
1298 hammer2_inode_repoint(ip, cdip, nchain);
1301 * If the old chain is not a hardlink target then replace
1302 * it with a OBJTYPE_HARDLINK pointer.
1304 * If the old chain IS a hardlink target then delete it.
1306 if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1308 * Replace original non-hardlink that's been dup'd
1309 * with a special hardlink directory entry. We must
1310 * set the DIRECTDATA flag to prevent sub-chains
1311 * from trying to synchronize to the inode if the
1312 * file is extended afterwords.
1314 hammer2_chain_modify(trans, &chain, 0);
1315 hammer2_chain_delete_duplicate(trans, &chain,
1316 HAMMER2_DELDUP_RECORE);
1317 ipdata = &chain->data->ipdata;
1318 ipdata->target_type = ipdata->type;
1319 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1327 bzero(&ipdata->uid, sizeof(ipdata->uid));
1328 bzero(&ipdata->gid, sizeof(ipdata->gid));
1329 ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1330 ipdata->cap_flags = 0;
1334 ipdata->iparent = 0; /* XXX */
1335 ipdata->pfs_type = 0;
1336 ipdata->pfs_inum = 0;
1337 bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1338 bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1339 ipdata->data_quota = 0;
1340 ipdata->data_count = 0;
1341 ipdata->inode_quota = 0;
1342 ipdata->inode_count = 0;
1343 ipdata->attr_tid = 0;
1344 ipdata->dirent_tid = 0;
1345 bzero(&ipdata->u, sizeof(ipdata->u));
1346 /* XXX transaction ids */
1348 hammer2_chain_delete(trans, chain);
1352 * Return the new chain.
1354 hammer2_chain_unlock(chain);
1360 hammer2_chain_unlock(chain);
1365 * Cleanup, chain/nchain already dealt with.
1369 hammer2_inode_drop(cdip);
1375 * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1376 * inode while (*chainp) points to the resolved (hidden hardlink
1377 * target) inode. In this situation when nlinks is 1 we wish to
1378 * deconsolidate the hardlink, moving it back to the directory that now
1379 * represents the only remaining link.
1382 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1383 hammer2_inode_t *dip,
1384 hammer2_chain_t **chainp,
1385 hammer2_chain_t **ochainp)
1387 if (*ochainp == NULL)
1394 * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1395 * with an obj_type of HAMMER2_OBJTYPE_HARDLINK. This routine will gobble
1396 * the *chainp and return a new locked *chainp representing the file target
1397 * (the original *chainp will be unlocked).
1399 * When a match is found the chain representing the original HARDLINK
1400 * will be returned in *ochainp with a ref, but not locked.
1402 * When no match is found *chainp is set to NULL and EIO is returned.
1403 * (*ochainp) will still be set to the original chain with a ref but not
1407 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1408 hammer2_chain_t **ochainp)
1410 hammer2_chain_t *chain = *chainp;
1411 hammer2_chain_t *parent;
1412 hammer2_inode_t *ip;
1413 hammer2_inode_t *pip;
1417 hammer2_inode_ref(pip); /* for loop */
1418 hammer2_chain_ref(chain); /* for (*ochainp) */
1422 * Locate the hardlink. pip is referenced and not locked,
1427 lhc = chain->data->ipdata.inum;
1428 hammer2_chain_unlock(chain);
1431 while ((ip = pip) != NULL) {
1432 parent = hammer2_inode_lock_ex(ip);
1433 hammer2_inode_drop(ip); /* loop */
1434 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1435 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
1436 hammer2_chain_lookup_done(parent); /* discard parent */
1439 pip = ip->pip; /* safe, ip held locked */
1441 hammer2_inode_ref(pip); /* loop */
1442 hammer2_inode_unlock_ex(ip, NULL);
1446 * chain is locked, ip is locked. Unlock ip, return the locked
1447 * chain. *ipp is already set w/a ref count and not locked.
1449 * (parent is already unlocked).
1452 hammer2_inode_unlock_ex(ip, NULL);
1455 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1456 /* already locked */
1464 * Find the directory common to both fdip and tdip, hold and return
1468 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1470 hammer2_inode_t *scan1;
1471 hammer2_inode_t *scan2;
1474 * We used to have a depth field but it complicated matters too
1475 * much for directory renames. So now its ugly. Check for
1476 * simple cases before giving up and doing it the expensive way.
1478 * XXX need a bottom-up topology stability lock
1480 if (fdip == tdip || fdip == tdip->pip) {
1481 hammer2_inode_ref(fdip);
1484 if (fdip->pip == tdip) {
1485 hammer2_inode_ref(tdip);
1492 for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1494 while (scan2->pmp == tdip->pmp) {
1495 if (scan1 == scan2) {
1496 hammer2_inode_ref(scan1);
1504 panic("hammer2_inode_common_parent: no common parent %p %p\n",