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_mount_t *hmp;
233 hammer2_pfsmount_t *pmp;
234 hammer2_inode_t *pip;
242 * Transition to zero, must interlock with
243 * the inode inumber lookup tree (if applicable).
245 * NOTE: The super-root inode has no pmp.
249 spin_lock(&pmp->inum_spin);
251 if (atomic_cmpset_int(&ip->refs, 1, 0)) {
252 KKASSERT(ip->topo_cst.count == 0);
253 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
254 atomic_clear_int(&ip->flags,
255 HAMMER2_INODE_ONRBTREE);
256 RB_REMOVE(hammer2_inode_tree,
257 &pmp->inum_tree, ip);
260 spin_unlock(&pmp->inum_spin);
269 * Cleaning out ip->chain isn't entirely
272 hammer2_inode_repoint(ip, NULL, NULL);
275 * We have to drop pip (if non-NULL) to
276 * dispose of our implied reference from
277 * ip->pip. We can simply loop on it.
280 KKASSERT((ip->flags &
281 HAMMER2_INODE_SROOT) == 0);
282 kfree(ip, pmp->minode);
285 HAMMER2_INODE_SROOT);
286 kfree(ip, hmp->mchain);
289 /* continue with pip (can be NULL) */
292 spin_unlock(&ip->pmp->inum_spin);
296 * Non zero transition
298 if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
305 * Get the vnode associated with the given inode, allocating the vnode if
306 * necessary. The vnode will be returned exclusively locked.
308 * The caller must lock the inode (shared or exclusive).
310 * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
314 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
316 hammer2_inode_data_t *ipdata;
317 hammer2_pfsmount_t *pmp;
322 KKASSERT(pmp != NULL);
324 ipdata = &ip->chain->data->ipdata;
328 * Attempt to reuse an existing vnode assignment. It is
329 * possible to race a reclaim so the vget() may fail. The
330 * inode must be unlocked during the vget() to avoid a
331 * deadlock against a reclaim.
336 * Inode must be unlocked during the vget() to avoid
337 * possible deadlocks, but leave the ip ref intact.
339 * vnode is held to prevent destruction during the
340 * vget(). The vget() can still fail if we lost
341 * a reclaim race on the vnode.
343 vhold_interlocked(vp);
344 ostate = hammer2_inode_lock_temp_release(ip);
345 if (vget(vp, LK_EXCLUSIVE)) {
347 hammer2_inode_lock_temp_restore(ip, ostate);
350 hammer2_inode_lock_temp_restore(ip, ostate);
352 /* vp still locked and ref from vget */
354 kprintf("hammer2: igetv race %p/%p\n",
364 * No vnode exists, allocate a new vnode. Beware of
365 * allocation races. This function will return an
366 * exclusively locked and referenced vnode.
368 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
370 kprintf("hammer2: igetv getnewvnode failed %d\n",
377 * Lock the inode and check for an allocation race.
379 ostate = hammer2_inode_lock_upgrade(ip);
380 if (ip->vp != NULL) {
383 hammer2_inode_lock_downgrade(ip, ostate);
387 switch (ipdata->type) {
388 case HAMMER2_OBJTYPE_DIRECTORY:
391 case HAMMER2_OBJTYPE_REGFILE:
393 vinitvmio(vp, ipdata->size,
395 (int)ipdata->size & HAMMER2_LBUFMASK);
397 case HAMMER2_OBJTYPE_SOFTLINK:
399 * XXX for now we are using the generic file_read
400 * and file_write code so we need a buffer cache
404 vinitvmio(vp, ipdata->size,
406 (int)ipdata->size & HAMMER2_LBUFMASK);
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 %d aux %d\n",
429 vp, vp->v_sysref.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_mount_t *hmp, hammer2_pfsmount_t *pmp,
451 hammer2_inode_t *dip, 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);
489 nip = kmalloc(sizeof(*nip), hmp->mchain, M_WAITOK | M_ZERO);
490 nip->flags = HAMMER2_INODE_SROOT;
492 nip->inum = chain->data->ipdata.inum;
493 hammer2_inode_repoint(nip, NULL, chain);
494 nip->pip = dip; /* can be NULL */
496 hammer2_inode_ref(dip); /* ref dip for nip->pip */
502 * ref and lock on nip gives it state compatible to after a
503 * hammer2_inode_lock_ex() call.
506 ccms_cst_init(&nip->topo_cst, &nip->chain);
507 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
508 /* combination of thread lock and chain lock == inode lock */
511 * Attempt to add the inode. If it fails we raced another inode
512 * get. Undo all the work and try again.
515 spin_lock(&pmp->inum_spin);
516 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
517 spin_unlock(&pmp->inum_spin);
518 ccms_thread_unlock(&nip->topo_cst);
519 hammer2_inode_drop(nip);
522 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
523 spin_unlock(&pmp->inum_spin);
530 * Create a new inode in the specified directory using the vattr to
531 * figure out the type of inode.
533 * If no error occurs the new inode with its chain locked is returned in
534 * *nipp, otherwise an error is returned and *nipp is set to NULL.
536 * If vap and/or cred are NULL the related fields are not set and the
537 * inode type defaults to a directory. This is used when creating PFSs
538 * under the super-root, so the inode number is set to 1 in this case.
540 * dip is not locked on entry.
543 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
544 struct vattr *vap, struct ucred *cred,
545 const uint8_t *name, size_t name_len,
546 hammer2_chain_t **chainp, int *errorp)
548 hammer2_inode_data_t *dipdata;
549 hammer2_inode_data_t *nipdata;
550 hammer2_mount_t *hmp;
551 hammer2_chain_t *chain;
552 hammer2_chain_t *parent;
553 hammer2_inode_t *nip;
562 lhc = hammer2_dirhash(name, name_len);
566 * Locate the inode or indirect block to create the new
567 * entry in. At the same time check for key collisions
568 * and iterate until we don't get one.
570 * NOTE: hidden inodes do not have iterators.
573 parent = hammer2_inode_lock_ex(dip);
574 dipdata = &dip->chain->data->ipdata;
575 dip_uid = dipdata->uid;
576 dip_gid = dipdata->gid;
577 dip_mode = dipdata->mode;
581 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
584 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
586 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
588 hammer2_chain_unlock(chain);
593 error = hammer2_chain_create(trans, &parent, &chain,
595 HAMMER2_BREF_TYPE_INODE,
596 HAMMER2_INODE_BYTES);
600 * Cleanup and handle retries.
602 if (error == EAGAIN) {
603 hammer2_chain_ref(parent);
604 hammer2_inode_unlock_ex(dip, parent);
605 hammer2_chain_wait(parent);
606 hammer2_chain_drop(parent);
609 hammer2_inode_unlock_ex(dip, parent);
612 KKASSERT(chain == NULL);
618 * Set up the new inode.
620 * NOTE: *_get() integrates chain's lock into the inode lock.
622 * NOTE: Only one new inode can currently be created per
623 * transaction. If the need arises we can adjust
624 * hammer2_trans_init() to allow more.
626 chain->data->ipdata.inum = trans->sync_tid;
627 nip = hammer2_inode_get(dip->hmp, dip->pmp, dip, chain);
628 nipdata = &chain->data->ipdata;
631 KKASSERT(trans->inodes_created == 0);
632 nipdata->type = hammer2_get_obj_type(vap->va_type);
633 nipdata->inum = trans->sync_tid;
634 ++trans->inodes_created;
636 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
639 nipdata->version = HAMMER2_INODE_VERSION_ONE;
640 hammer2_update_time(&nipdata->ctime);
641 nipdata->mtime = nipdata->ctime;
643 nipdata->mode = vap->va_mode;
647 xuid = hammer2_to_unix_xid(&dip_uid);
648 xuid = vop_helper_create_uid(dip->pmp->mp,
656 if (vap->va_vaflags & VA_UID_UUID_VALID)
657 nipdata->uid = vap->va_uid_uuid;
658 else if (vap->va_uid != (uid_t)VNOVAL)
659 hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
661 hammer2_guid_to_uuid(&nipdata->uid, xuid);
663 if (vap->va_vaflags & VA_GID_UUID_VALID)
664 nipdata->gid = vap->va_gid_uuid;
665 else if (vap->va_gid != (gid_t)VNOVAL)
666 hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
668 nipdata->gid = dip_gid;
672 * Regular files and softlinks allow a small amount of data to be
673 * directly embedded in the inode. This flag will be cleared if
674 * the size is extended past the embedded limit.
676 if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
677 nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
678 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
681 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
682 bcopy(name, nipdata->filename, name_len);
683 nipdata->name_key = lhc;
684 nipdata->name_len = name_len;
691 * chain may have been moved around by the create.
695 hammer2_chain_refactor(hammer2_chain_t **chainp)
697 hammer2_chain_t *chain = *chainp;
698 hammer2_chain_core_t *core;
701 spin_lock(&core->cst.spin);
702 while (hammer2_chain_refactor_test(chain, 1)) {
703 chain = chain->next_parent;
704 while (hammer2_chain_refactor_test(chain, 1))
705 chain = chain->next_parent;
706 hammer2_chain_ref(chain);
707 spin_unlock(&core->cst.spin);
709 hammer2_chain_unlock(*chainp);
710 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
711 HAMMER2_RESOLVE_NOREF); /* eat ref */
713 spin_lock(&core->cst.spin);
715 spin_unlock(&core->cst.spin);
719 * ochain represents the target file inode. We need to move it to the
720 * specified common parent directory (dip) and rename it to a special
721 * invisible "0xINODENUMBER" filename.
723 * We use chain_duplicate and duplicate ochain at the new location,
724 * renaming it appropriately. We create a temporary chain and
725 * then delete it to placemark where the duplicate will go. Both of
726 * these use the inode number for (lhc) (the key), generating the
727 * invisible filename.
731 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
732 hammer2_inode_t *dip, int *errorp)
734 hammer2_inode_data_t *nipdata;
735 hammer2_mount_t *hmp;
736 hammer2_chain_t *parent;
737 hammer2_chain_t *ochain;
738 hammer2_chain_t *nchain;
739 hammer2_chain_t *tmp;
741 hammer2_blockref_t bref;
746 lhc = ochain->data->ipdata.inum;
747 KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
750 * Locate the inode or indirect block to create the new
751 * entry in. lhc represents the inode number so there is
752 * no collision iteration.
754 * There should be no key collisions with invisible inode keys.
757 parent = hammer2_chain_lookup_init(dip->chain, 0);
758 nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
760 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
761 nchain, parent, dip, dip->chain);
762 hammer2_chain_unlock(nchain);
771 * Create entry in common parent directory using the seek position
775 KKASSERT(nchain == NULL);
776 *errorp = hammer2_chain_create(trans, &parent, &nchain,
778 HAMMER2_BREF_TYPE_INODE,/* n/a */
779 HAMMER2_INODE_BYTES); /* n/a */
780 hammer2_chain_refactor(&ochain);
785 * Cleanup and handle retries.
787 if (*errorp == EAGAIN) {
788 hammer2_chain_ref(parent);
789 hammer2_chain_lookup_done(parent);
790 hammer2_chain_wait(parent);
791 hammer2_chain_drop(parent);
796 * Handle the error case
799 KKASSERT(nchain == NULL);
800 hammer2_chain_lookup_done(parent);
805 * Use chain as a placeholder for (lhc), delete it and replace
806 * it with our duplication.
808 * Gain a second lock on ochain for the duplication function to
809 * unlock, maintain the caller's original lock across the call.
811 * This is a bit messy.
813 hammer2_chain_delete(trans, nchain);
814 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
817 bref.key = lhc; /* invisible dir entry key */
819 hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
820 hammer2_chain_lookup_done(parent);
821 hammer2_chain_unlock(nchain); /* no longer needed */
824 * Now set chain to our duplicate and modify it appropriately.
826 * Directory entries are inodes but this is a hidden hardlink
827 * target. The name isn't used but to ease debugging give it
828 * a name after its inode number.
831 tmp = NULL; /* safety */
833 hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
834 nipdata = &nchain->data->ipdata;
835 ksnprintf(nipdata->filename, sizeof(nipdata->filename),
836 "0x%016jx", (intmax_t)nipdata->inum);
837 nipdata->name_len = strlen(nipdata->filename);
838 nipdata->name_key = lhc;
844 * Connect the target inode represented by (*chainp) to the media topology
845 * at (dip, name, len).
847 * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
848 * entry instead of connecting (*chainp).
850 * If hlink is FALSE this function uses chain_duplicate() to make a copy
851 * if (*chainp) in the directory entry. (*chainp) is likely to be deleted
852 * by the caller in this case (e.g. rename).
855 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
856 hammer2_inode_t *dip, hammer2_chain_t **chainp,
857 const uint8_t *name, size_t name_len)
859 hammer2_inode_data_t *ipdata;
860 hammer2_mount_t *hmp;
861 hammer2_chain_t *nchain;
862 hammer2_chain_t *parent;
863 hammer2_chain_t *ochain;
872 * Since ochain is either disconnected from the topology or represents
873 * a hardlink terminus which is always a parent of or equal to dip,
874 * we should be able to safely lock dip->chain for our setup.
876 parent = hammer2_chain_lookup_init(dip->chain, 0);
878 lhc = hammer2_dirhash(name, name_len);
881 * Locate the inode or indirect block to create the new
882 * entry in. At the same time check for key collisions
883 * and iterate until we don't get one.
887 nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
890 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
892 hammer2_chain_unlock(nchain);
900 * Hardlink pointer needed, create totally fresh
903 KKASSERT(nchain == NULL);
904 error = hammer2_chain_create(trans, &parent, &nchain,
906 HAMMER2_BREF_TYPE_INODE,
907 HAMMER2_INODE_BYTES);
908 hammer2_chain_refactor(&ochain);
911 * Reconnect the original chain and rename. Use
912 * chain_duplicate(). The caller will likely delete
913 * or has already deleted the original chain in
916 * NOTE: chain_duplicate() generates a new chain
917 * with CHAIN_DELETED cleared (ochain typically
918 * has it set from the file unlink).
922 hammer2_chain_duplicate(trans, NULL, -1, &nchain, NULL);
923 error = hammer2_chain_create(trans, &parent, &nchain,
925 HAMMER2_BREF_TYPE_INODE,
926 HAMMER2_INODE_BYTES);
933 KKASSERT(error != EAGAIN);
934 hammer2_chain_lookup_done(parent);
938 * nchain should be NULL on error, leave ochain (== *chainp) alone.
941 KKASSERT(nchain == NULL);
946 * Directory entries are inodes so if the name has changed we have
947 * to update the inode.
949 * When creating an OBJTYPE_HARDLINK entry remember to unlock the
950 * chain, the caller will access the hardlink via the actual hardlink
951 * target file and not the hardlink pointer entry, so we must still
954 if (hlink && hammer2_hardlink_enable >= 0) {
956 * Create the HARDLINK pointer. oip represents the hardlink
957 * target in this situation.
959 * We will return ochain (the hardlink target).
961 hammer2_chain_modify(trans, &nchain,
962 HAMMER2_MODIFY_ASSERTNOCOPY);
963 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
964 ipdata = &nchain->data->ipdata;
965 bcopy(name, ipdata->filename, name_len);
966 ipdata->name_key = lhc;
967 ipdata->name_len = name_len;
968 ipdata->target_type = ochain->data->ipdata.type;
969 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
970 ipdata->inum = ochain->data->ipdata.inum;
972 hammer2_chain_unlock(nchain);
975 } else if (hlink && hammer2_hardlink_enable < 0) {
977 * Create a snapshot (hardlink fake mode for debugging).
978 * (ochain already flushed above so we can just copy the
981 * Since this is a snapshot we return nchain in the fake
984 hammer2_chain_modify(trans, &nchain,
985 HAMMER2_MODIFY_ASSERTNOCOPY);
986 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
987 ipdata = &nchain->data->ipdata;
988 *ipdata = ochain->data->ipdata;
989 bcopy(name, ipdata->filename, name_len);
990 ipdata->name_key = lhc;
991 ipdata->name_len = name_len;
992 kprintf("created fake hardlink %*.*s\n",
993 (int)name_len, (int)name_len, name);
996 * nchain is a duplicate of ochain at the new location.
997 * We must fixup the name stored in oip. The bref key
998 * has already been set up.
1000 hammer2_chain_modify(trans, &nchain,
1001 HAMMER2_MODIFY_ASSERTNOCOPY);
1002 ipdata = &nchain->data->ipdata;
1004 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1005 bcopy(name, ipdata->filename, name_len);
1006 ipdata->name_key = lhc;
1007 ipdata->name_len = name_len;
1012 * We are replacing ochain with nchain, unlock ochain. In the
1013 * case where ochain is left unchanged the code above sets
1014 * nchain to ochain and ochain to NULL, resulting in a NOP here.
1017 hammer2_chain_unlock(ochain);
1024 * Repoint ip->chain to nchain. Caller must hold the inode exclusively
1027 * ip->chain is set to nchain. The prior chain in ip->chain is dropped
1028 * and nchain is ref'd.
1031 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1032 hammer2_chain_t *nchain)
1034 hammer2_chain_t *ochain;
1035 hammer2_inode_t *opip;
1038 * Repoint ip->chain if requested.
1043 hammer2_chain_ref(nchain);
1045 hammer2_chain_drop(ochain);
1048 * Repoint ip->pip if requested (non-NULL pip).
1050 if (pip && ip->pip != pip) {
1052 hammer2_inode_ref(pip);
1055 hammer2_inode_drop(opip);
1060 * Unlink the file from the specified directory inode. The directory inode
1061 * does not need to be locked.
1063 * isdir determines whether a directory/non-directory check should be made.
1064 * No check is made if isdir is set to -1.
1066 * NOTE! This function does not prevent the underlying file from still
1067 * being used if it has other refs (such as from an inode, or if it's
1068 * chain is manually held). However, the caller is responsible for
1069 * fixing up ip->chain if e.g. a rename occurs (see chain_duplicate()).
1072 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1073 const uint8_t *name, size_t name_len,
1074 int isdir, int *hlinkp)
1076 hammer2_inode_data_t *ipdata;
1077 hammer2_mount_t *hmp;
1078 hammer2_chain_t *parent;
1079 hammer2_chain_t *ochain;
1080 hammer2_chain_t *chain;
1081 hammer2_chain_t *dparent;
1082 hammer2_chain_t *dchain;
1090 lhc = hammer2_dirhash(name, name_len);
1093 * Search for the filename in the directory
1097 parent = hammer2_inode_lock_ex(dip);
1098 chain = hammer2_chain_lookup(&parent,
1099 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1102 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1103 name_len == chain->data->ipdata.name_len &&
1104 bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1107 chain = hammer2_chain_next(&parent, chain,
1108 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1111 hammer2_inode_unlock_ex(dip, NULL); /* retain parent */
1114 * Not found or wrong type (isdir < 0 disables the type check).
1115 * If a hardlink pointer, type checks use the hardlink target.
1117 if (chain == NULL) {
1121 if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1124 type = chain->data->ipdata.target_type;
1127 if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1131 if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1137 * Hardlink must be resolved. We can't hold parent locked while we
1138 * do this or we could deadlock.
1140 * On success chain will be adjusted to point at the hardlink target
1141 * and ochain will point to the hardlink pointer in the original
1142 * directory. Otherwise chain remains pointing to the original.
1144 if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1145 hammer2_chain_unlock(parent);
1147 error = hammer2_hardlink_find(dip, &chain, &ochain);
1151 * If this is a directory the directory must be empty. However, if
1152 * isdir < 0 we are doing a rename and the directory does not have
1153 * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1154 * and the directory does not have to be empty.
1156 * NOTE: We check the full key range here which covers both visible
1157 * and invisible entries. Theoretically there should be no
1158 * invisible (hardlink target) entries if there are no visible
1161 if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1162 dparent = hammer2_chain_lookup_init(chain, 0);
1163 dchain = hammer2_chain_lookup(&dparent,
1164 0, (hammer2_key_t)-1,
1165 HAMMER2_LOOKUP_NODATA);
1167 hammer2_chain_unlock(dchain);
1168 hammer2_chain_lookup_done(dparent);
1172 hammer2_chain_lookup_done(dparent);
1178 * Ok, we can now unlink the chain. We always decrement nlinks even
1179 * if the entry can be deleted in case someone has the file open and
1182 * The chain itself will no longer be in the on-media topology but
1183 * can still be flushed to the media (e.g. if an open descriptor
1184 * remains). When the last vnode/ip ref goes away the chain will
1185 * be marked unmodified, avoiding any further (now unnecesary) I/O.
1187 * A non-NULL ochain indicates a hardlink.
1191 * Delete the original hardlink pointer.
1193 * NOTE: parent from above is NULL when ochain != NULL
1194 * so we can reuse it.
1196 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1197 hammer2_chain_delete(trans, ochain);
1198 hammer2_chain_unlock(ochain);
1201 * Then decrement nlinks on hardlink target, deleting
1202 * the target when nlinks drops to 0.
1204 hammer2_chain_modify(trans, &chain, 0);
1205 --chain->data->ipdata.nlinks;
1206 if (chain->data->ipdata.nlinks == 0)
1207 hammer2_chain_delete(trans, chain);
1210 * Otherwise this was not a hardlink and we can just
1211 * remove the entry and decrement nlinks.
1213 * NOTE: *_get() integrates chain's lock into the inode lock.
1215 hammer2_chain_modify(trans, &chain, 0);
1216 ipdata = &chain->data->ipdata;
1218 hammer2_chain_delete(trans, chain);
1224 hammer2_chain_unlock(chain);
1226 hammer2_chain_lookup_done(parent);
1228 hammer2_chain_drop(ochain);
1234 * Given an exclusively locked inode we consolidate its chain for hardlink
1235 * creation, adding (nlinks) to the file's link count and potentially
1236 * relocating the inode to a directory common to ip->pip and tdip.
1238 * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1239 * and returning a new locked chain.
1241 * NOTE! This function will also replace ip->chain.
1244 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1245 hammer2_chain_t **chainp,
1246 hammer2_inode_t *tdip, int nlinks)
1248 hammer2_inode_data_t *ipdata;
1249 hammer2_mount_t *hmp;
1250 hammer2_inode_t *fdip;
1251 hammer2_inode_t *cdip;
1252 hammer2_chain_t *chain;
1253 hammer2_chain_t *nchain;
1259 if (nlinks == 0 && /* no hardlink needed */
1260 (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1263 if (hammer2_hardlink_enable < 0) { /* fake hardlinks */
1267 if (hammer2_hardlink_enable == 0) { /* disallow hardlinks */
1268 hammer2_chain_unlock(chain);
1274 * cdip will be returned with a ref, but not locked.
1277 cdip = hammer2_inode_common_parent(fdip, tdip);
1280 * If no change in the hardlink's target directory is required and
1281 * this is already a hardlink target, all we need to do is adjust
1284 * XXX The common parent is a big wiggly due to duplication from
1285 * renames. Compare the core (RBTREE) pointer instead of the
1289 (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1291 hammer2_chain_modify(trans, &chain, 0);
1292 chain->data->ipdata.nlinks += nlinks;
1299 * We either have to move an existing hardlink target or we have
1300 * to create a fresh hardlink target.
1302 * Hardlink targets are hidden inodes in a parent directory common
1303 * to all directory entries referencing the hardlink.
1305 nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1309 * Bump nlinks on duplicated hidden inode, repoint
1312 hammer2_chain_modify(trans, &nchain, 0);
1313 nchain->data->ipdata.nlinks += nlinks;
1314 hammer2_inode_repoint(ip, cdip, nchain);
1317 * If the old chain is not a hardlink target then replace
1318 * it with a OBJTYPE_HARDLINK pointer.
1320 * If the old chain IS a hardlink target then delete it.
1322 if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1324 * Replace original non-hardlink that's been dup'd
1325 * with a special hardlink directory entry. We must
1326 * set the DIRECTDATA flag to prevent sub-chains
1327 * from trying to synchronize to the inode if the
1328 * file is extended afterwords.
1330 hammer2_chain_modify(trans, &chain, 0);
1331 hammer2_chain_delete_duplicate(trans, &chain,
1332 HAMMER2_DELDUP_RECORE);
1333 ipdata = &chain->data->ipdata;
1334 ipdata->target_type = ipdata->type;
1335 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1343 bzero(&ipdata->uid, sizeof(ipdata->uid));
1344 bzero(&ipdata->gid, sizeof(ipdata->gid));
1345 ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1346 ipdata->cap_flags = 0;
1350 ipdata->iparent = 0; /* XXX */
1351 ipdata->pfs_type = 0;
1352 ipdata->pfs_inum = 0;
1353 bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1354 bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1355 ipdata->data_quota = 0;
1356 ipdata->data_count = 0;
1357 ipdata->inode_quota = 0;
1358 ipdata->inode_count = 0;
1359 ipdata->attr_tid = 0;
1360 ipdata->dirent_tid = 0;
1361 bzero(&ipdata->u, sizeof(ipdata->u));
1362 /* XXX transaction ids */
1364 hammer2_chain_delete(trans, chain);
1368 * Return the new chain.
1370 hammer2_chain_unlock(chain);
1376 hammer2_chain_unlock(chain);
1381 * Cleanup, chain/nchain already dealt with.
1385 hammer2_inode_drop(cdip);
1391 * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1392 * inode while (*chainp) points to the resolved (hidden hardlink
1393 * target) inode. In this situation when nlinks is 1 we wish to
1394 * deconsolidate the hardlink, moving it back to the directory that now
1395 * represents the only remaining link.
1398 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1399 hammer2_inode_t *dip,
1400 hammer2_chain_t **chainp,
1401 hammer2_chain_t **ochainp)
1403 if (*ochainp == NULL)
1410 * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1411 * with an obj_type of HAMMER2_OBJTYPE_HARDLINK. This routine will gobble
1412 * the *chainp and return a new locked *chainp representing the file target
1413 * (the original *chainp will be unlocked).
1415 * When a match is found the chain representing the original HARDLINK
1416 * will be returned in *ochainp with a ref, but not locked.
1418 * When no match is found *chainp is set to NULL and EIO is returned.
1419 * (*ochainp) will still be set to the original chain with a ref but not
1423 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1424 hammer2_chain_t **ochainp)
1426 hammer2_chain_t *chain = *chainp;
1427 hammer2_chain_t *parent;
1428 hammer2_inode_t *ip;
1429 hammer2_inode_t *pip;
1433 hammer2_inode_ref(pip); /* for loop */
1434 hammer2_chain_ref(chain); /* for (*ochainp) */
1438 * Locate the hardlink. pip is referenced and not locked,
1443 lhc = chain->data->ipdata.inum;
1444 hammer2_chain_unlock(chain);
1447 while ((ip = pip) != NULL) {
1448 parent = hammer2_inode_lock_ex(ip);
1449 hammer2_inode_drop(ip); /* loop */
1450 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1451 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
1452 hammer2_chain_lookup_done(parent); /* discard parent */
1455 pip = ip->pip; /* safe, ip held locked */
1457 hammer2_inode_ref(pip); /* loop */
1458 hammer2_inode_unlock_ex(ip, NULL);
1462 * chain is locked, ip is locked. Unlock ip, return the locked
1463 * chain. *ipp is already set w/a ref count and not locked.
1465 * (parent is already unlocked).
1468 hammer2_inode_unlock_ex(ip, NULL);
1471 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1472 /* already locked */
1480 * Find the directory common to both fdip and tdip, hold and return
1484 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1486 hammer2_inode_t *scan1;
1487 hammer2_inode_t *scan2;
1490 * We used to have a depth field but it complicated matters too
1491 * much for directory renames. So now its ugly. Check for
1492 * simple cases before giving up and doing it the expensive way.
1494 * XXX need a bottom-up topology stability lock
1496 if (fdip == tdip || fdip == tdip->pip) {
1497 hammer2_inode_ref(fdip);
1500 if (fdip->pip == tdip) {
1501 hammer2_inode_ref(tdip);
1508 for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1510 while (scan2->pmp == tdip->pmp) {
1511 if (scan1 == scan2) {
1512 hammer2_inode_ref(scan1);
1520 panic("hammer2_inode_common_parent: no common parent %p %p\n",