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 spin_lock(&chain->core->cst.spin);
87 if (hammer2_chain_refactor_test(chain, 1)) {
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);
99 spin_unlock(&chain->core->cst.spin);
102 KKASSERT(chain != NULL); /* for now */
103 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
106 * Resolve duplication races
108 if (hammer2_chain_refactor_test(chain, 1)) {
109 hammer2_chain_unlock(chain);
116 hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
119 * XXX this will catch parent directories too which we don't
123 hammer2_chain_unlock(chain);
124 ccms_thread_unlock(&ip->topo_cst);
125 hammer2_inode_drop(ip);
129 * NOTE: We don't combine the inode/chain lock because putting away an
130 * inode would otherwise confuse multiple lock holders of the inode.
132 * Shared locks are especially sensitive to having too many shared
133 * lock counts (from the same thread) on certain paths which might
134 * need to upgrade them. Only one count of a shared lock can be
138 hammer2_inode_lock_sh(hammer2_inode_t *ip)
140 hammer2_chain_t *chain;
142 hammer2_inode_ref(ip);
144 ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
147 KKASSERT(chain != NULL); /* for now */
148 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
149 HAMMER2_RESOLVE_SHARED);
152 * Resolve duplication races
154 if (hammer2_chain_refactor_test(chain, 1)) {
155 hammer2_chain_unlock(chain);
156 ccms_thread_unlock(&ip->topo_cst);
157 chain = hammer2_inode_lock_ex(ip);
158 hammer2_inode_unlock_ex(ip, chain);
165 hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
168 hammer2_chain_unlock(chain);
169 ccms_thread_unlock(&ip->topo_cst);
170 hammer2_inode_drop(ip);
174 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
176 return(ccms_thread_lock_temp_release(&ip->topo_cst));
180 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate)
182 ccms_thread_lock_temp_restore(&ip->topo_cst, ostate);
186 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
188 return(ccms_thread_lock_upgrade(&ip->topo_cst));
192 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate)
194 ccms_thread_lock_downgrade(&ip->topo_cst, ostate);
198 * Lookup an inode by inode number
201 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
206 spin_lock(&pmp->inum_spin);
207 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
209 hammer2_inode_ref(ip);
210 spin_unlock(&pmp->inum_spin);
218 * Adding a ref to an inode is only legal if the inode already has at least
222 hammer2_inode_ref(hammer2_inode_t *ip)
224 atomic_add_int(&ip->refs, 1);
228 * Drop an inode reference, freeing the inode when the last reference goes
232 hammer2_inode_drop(hammer2_inode_t *ip)
234 hammer2_pfsmount_t *pmp;
235 hammer2_inode_t *pip;
243 * Transition to zero, must interlock with
244 * the inode inumber lookup tree (if applicable).
246 * NOTE: The super-root inode has no pmp.
250 spin_lock(&pmp->inum_spin);
252 if (atomic_cmpset_int(&ip->refs, 1, 0)) {
253 KKASSERT(ip->topo_cst.count == 0);
254 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
255 atomic_clear_int(&ip->flags,
256 HAMMER2_INODE_ONRBTREE);
257 RB_REMOVE(hammer2_inode_tree,
258 &pmp->inum_tree, ip);
261 spin_unlock(&pmp->inum_spin);
268 * Cleaning out ip->chain isn't entirely
271 hammer2_inode_repoint(ip, NULL, NULL);
274 * We have to drop pip (if non-NULL) to
275 * dispose of our implied reference from
276 * ip->pip. We can simply loop on it.
279 KKASSERT((ip->flags &
280 HAMMER2_INODE_SROOT) == 0);
281 kfree(ip, pmp->minode);
282 atomic_add_long(&pmp->inmem_inodes, -1);
285 HAMMER2_INODE_SROOT);
286 kfree(ip, M_HAMMER2);
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_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;
563 lhc = hammer2_dirhash(name, name_len);
567 * Locate the inode or indirect block to create the new
568 * entry in. At the same time check for key collisions
569 * and iterate until we don't get one.
571 * NOTE: hidden inodes do not have iterators.
574 parent = hammer2_inode_lock_ex(dip);
575 dipdata = &dip->chain->data->ipdata;
576 dip_uid = dipdata->uid;
577 dip_gid = dipdata->gid;
578 dip_mode = dipdata->mode;
582 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
585 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
587 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
589 hammer2_chain_unlock(chain);
595 error = hammer2_chain_create(trans, &parent, &chain,
597 HAMMER2_BREF_TYPE_INODE,
598 HAMMER2_INODE_BYTES);
602 * Cleanup and handle retries.
604 if (error == EAGAIN) {
605 hammer2_chain_ref(parent);
606 hammer2_inode_unlock_ex(dip, parent);
607 hammer2_chain_wait(parent);
608 hammer2_chain_drop(parent);
611 hammer2_inode_unlock_ex(dip, parent);
614 KKASSERT(chain == NULL);
620 * Set up the new inode.
622 * NOTE: *_get() integrates chain's lock into the inode lock.
624 * NOTE: Only one new inode can currently be created per
625 * transaction. If the need arises we can adjust
626 * hammer2_trans_init() to allow more.
628 * NOTE: nipdata will have chain's blockset data.
630 chain->data->ipdata.inum = trans->sync_tid;
631 nip = hammer2_inode_get(dip->pmp, dip, chain);
632 nipdata = &chain->data->ipdata;
635 KKASSERT(trans->inodes_created == 0);
636 nipdata->type = hammer2_get_obj_type(vap->va_type);
637 nipdata->inum = trans->sync_tid;
638 ++trans->inodes_created;
640 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
644 /* Inherit parent's inode compression mode. */
645 nip->comp_heuristic = 0;
646 nipdata->comp_algo = dipdata->comp_algo;
647 nipdata->version = HAMMER2_INODE_VERSION_ONE;
648 hammer2_update_time(&nipdata->ctime);
649 nipdata->mtime = nipdata->ctime;
651 nipdata->mode = vap->va_mode;
655 xuid = hammer2_to_unix_xid(&dip_uid);
656 xuid = vop_helper_create_uid(dip->pmp->mp,
664 if (vap->va_vaflags & VA_UID_UUID_VALID)
665 nipdata->uid = vap->va_uid_uuid;
666 else if (vap->va_uid != (uid_t)VNOVAL)
667 hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
669 hammer2_guid_to_uuid(&nipdata->uid, xuid);
671 if (vap->va_vaflags & VA_GID_UUID_VALID)
672 nipdata->gid = vap->va_gid_uuid;
673 else if (vap->va_gid != (gid_t)VNOVAL)
674 hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
676 nipdata->gid = dip_gid;
680 * Regular files and softlinks allow a small amount of data to be
681 * directly embedded in the inode. This flag will be cleared if
682 * the size is extended past the embedded limit.
684 if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
685 nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
686 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
689 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
690 bcopy(name, nipdata->filename, name_len);
691 nipdata->name_key = lhc;
692 nipdata->name_len = name_len;
699 * chain may have been moved around by the create.
703 hammer2_chain_refactor(hammer2_chain_t **chainp)
705 hammer2_chain_t *chain = *chainp;
706 hammer2_chain_core_t *core;
709 spin_lock(&core->cst.spin);
710 while (hammer2_chain_refactor_test(chain, 1)) {
711 chain = chain->next_parent;
712 while (hammer2_chain_refactor_test(chain, 1))
713 chain = chain->next_parent;
714 hammer2_chain_ref(chain);
715 spin_unlock(&core->cst.spin);
717 hammer2_chain_unlock(*chainp);
718 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
719 HAMMER2_RESOLVE_NOREF); /* eat ref */
721 spin_lock(&core->cst.spin);
723 spin_unlock(&core->cst.spin);
727 * ochain represents the target file inode. We need to move it to the
728 * specified common parent directory (dip) and rename it to a special
729 * invisible "0xINODENUMBER" filename.
731 * We use chain_duplicate and duplicate ochain at the new location,
732 * renaming it appropriately. We create a temporary chain and
733 * then delete it to placemark where the duplicate will go. Both of
734 * these use the inode number for (lhc) (the key), generating the
735 * invisible filename.
739 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
740 hammer2_inode_t *dip, int *errorp)
742 hammer2_inode_data_t *nipdata;
743 hammer2_chain_t *parent;
744 hammer2_chain_t *ochain;
745 hammer2_chain_t *nchain;
746 hammer2_chain_t *tmp;
748 hammer2_blockref_t bref;
752 lhc = ochain->data->ipdata.inum;
753 KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
756 * Locate the inode or indirect block to create the new
757 * entry in. lhc represents the inode number so there is
758 * no collision iteration.
760 * There should be no key collisions with invisible inode keys.
763 parent = hammer2_chain_lookup_init(dip->chain, 0);
764 nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
766 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
767 nchain, parent, dip, dip->chain);
768 hammer2_chain_unlock(nchain);
777 * Create entry in common parent directory using the seek position
781 KKASSERT(nchain == NULL);
782 *errorp = hammer2_chain_create(trans, &parent, &nchain,
784 HAMMER2_BREF_TYPE_INODE,/* n/a */
785 HAMMER2_INODE_BYTES); /* n/a */
786 hammer2_chain_refactor(&ochain);
791 * Cleanup and handle retries.
793 if (*errorp == EAGAIN) {
794 hammer2_chain_ref(parent);
795 hammer2_chain_lookup_done(parent);
796 hammer2_chain_wait(parent);
797 hammer2_chain_drop(parent);
802 * Handle the error case
805 KKASSERT(nchain == NULL);
806 hammer2_chain_lookup_done(parent);
811 * Use chain as a placeholder for (lhc), delete it and replace
812 * it with our duplication.
814 * Gain a second lock on ochain for the duplication function to
815 * unlock, maintain the caller's original lock across the call.
817 * This is a bit messy.
819 hammer2_chain_delete(trans, nchain, HAMMER2_DELETE_WILLDUP);
820 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
823 bref.key = lhc; /* invisible dir entry key */
825 hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
826 hammer2_chain_lookup_done(parent);
827 hammer2_chain_unlock(nchain); /* no longer needed */
830 * Now set chain to our duplicate and modify it appropriately.
832 * Directory entries are inodes but this is a hidden hardlink
833 * target. The name isn't used but to ease debugging give it
834 * a name after its inode number.
837 tmp = NULL; /* safety */
839 hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
840 nipdata = &nchain->data->ipdata;
841 ksnprintf(nipdata->filename, sizeof(nipdata->filename),
842 "0x%016jx", (intmax_t)nipdata->inum);
843 nipdata->name_len = strlen(nipdata->filename);
844 nipdata->name_key = lhc;
850 * Connect the target inode represented by (*chainp) to the media topology
851 * at (dip, name, len).
853 * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
854 * entry instead of connecting (*chainp).
856 * If hlink is FALSE this function uses chain_duplicate() to make a copy
857 * if (*chainp) in the directory entry. (*chainp) is likely to be deleted
858 * by the caller in this case (e.g. rename).
861 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
862 hammer2_inode_t *dip, hammer2_chain_t **chainp,
863 const uint8_t *name, size_t name_len)
865 hammer2_inode_data_t *ipdata;
866 hammer2_chain_t *nchain;
867 hammer2_chain_t *parent;
868 hammer2_chain_t *ochain;
875 * Since ochain is either disconnected from the topology or represents
876 * a hardlink terminus which is always a parent of or equal to dip,
877 * we should be able to safely lock dip->chain for our setup.
879 parent = hammer2_chain_lookup_init(dip->chain, 0);
881 lhc = hammer2_dirhash(name, name_len);
884 * Locate the inode or indirect block to create the new
885 * entry in. At the same time check for key collisions
886 * and iterate until we don't get one.
890 nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
893 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
895 hammer2_chain_unlock(nchain);
903 * Hardlink pointer needed, create totally fresh
906 KKASSERT(nchain == NULL);
907 error = hammer2_chain_create(trans, &parent, &nchain,
909 HAMMER2_BREF_TYPE_INODE,
910 HAMMER2_INODE_BYTES);
911 hammer2_chain_refactor(&ochain);
914 * Reconnect the original chain and rename. Use
915 * chain_duplicate(). The caller will likely delete
916 * or has already deleted the original chain in
919 * NOTE: chain_duplicate() generates a new chain
920 * with CHAIN_DELETED cleared (ochain typically
921 * has it set from the file unlink).
925 hammer2_chain_duplicate(trans, NULL, -1, &nchain, NULL);
926 error = hammer2_chain_create(trans, &parent, &nchain,
928 HAMMER2_BREF_TYPE_INODE,
929 HAMMER2_INODE_BYTES);
936 KKASSERT(error != EAGAIN);
937 hammer2_chain_lookup_done(parent);
941 * nchain should be NULL on error, leave ochain (== *chainp) alone.
944 KKASSERT(nchain == NULL);
949 * Directory entries are inodes so if the name has changed we have
950 * to update the inode.
952 * When creating an OBJTYPE_HARDLINK entry remember to unlock the
953 * chain, the caller will access the hardlink via the actual hardlink
954 * target file and not the hardlink pointer entry, so we must still
957 if (hlink && hammer2_hardlink_enable >= 0) {
959 * Create the HARDLINK pointer. oip represents the hardlink
960 * target in this situation.
962 * We will return ochain (the hardlink target).
964 hammer2_chain_modify(trans, &nchain,
965 HAMMER2_MODIFY_ASSERTNOCOPY);
966 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
967 ipdata = &nchain->data->ipdata;
968 atomic_set_int(&nchain->flags, HAMMER2_CHAIN_HARDLINK);
969 bcopy(name, ipdata->filename, name_len);
970 ipdata->name_key = lhc;
971 ipdata->name_len = name_len;
972 ipdata->target_type = ochain->data->ipdata.type;
973 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
974 ipdata->inum = ochain->data->ipdata.inum;
976 hammer2_chain_unlock(nchain);
979 } else if (hlink && hammer2_hardlink_enable < 0) {
981 * Create a snapshot (hardlink fake mode for debugging).
982 * (ochain already flushed above so we can just copy the
985 * Since this is a snapshot we return nchain in the fake
988 hammer2_chain_modify(trans, &nchain,
989 HAMMER2_MODIFY_ASSERTNOCOPY);
990 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
991 ipdata = &nchain->data->ipdata;
992 *ipdata = ochain->data->ipdata;
993 bcopy(name, ipdata->filename, name_len);
994 ipdata->name_key = lhc;
995 ipdata->name_len = name_len;
996 kprintf("created fake hardlink %*.*s\n",
997 (int)name_len, (int)name_len, name);
1000 * nchain is a duplicate of ochain at the new location.
1001 * We must fixup the name stored in oip. The bref key
1002 * has already been set up.
1004 hammer2_chain_modify(trans, &nchain,
1005 HAMMER2_MODIFY_ASSERTNOCOPY);
1006 ipdata = &nchain->data->ipdata;
1008 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1009 bcopy(name, ipdata->filename, name_len);
1010 ipdata->name_key = lhc;
1011 ipdata->name_len = name_len;
1016 * We are replacing ochain with nchain, unlock ochain. In the
1017 * case where ochain is left unchanged the code above sets
1018 * nchain to ochain and ochain to NULL, resulting in a NOP here.
1021 hammer2_chain_unlock(ochain);
1028 * Repoint ip->chain to nchain. Caller must hold the inode exclusively
1031 * ip->chain is set to nchain. The prior chain in ip->chain is dropped
1032 * and nchain is ref'd.
1035 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1036 hammer2_chain_t *nchain)
1038 hammer2_chain_t *ochain;
1039 hammer2_inode_t *opip;
1042 * Repoint ip->chain if requested.
1047 hammer2_chain_ref(nchain);
1049 hammer2_chain_drop(ochain);
1052 * Flag the chain for the refactor test
1054 if (nchain && nchain->data && nchain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK)
1055 atomic_set_int(&nchain->flags, HAMMER2_CHAIN_HARDLINK);
1058 * Repoint ip->pip if requested (non-NULL pip).
1060 if (pip && ip->pip != pip) {
1062 hammer2_inode_ref(pip);
1065 hammer2_inode_drop(opip);
1070 * Unlink the file from the specified directory inode. The directory inode
1071 * does not need to be locked.
1073 * isdir determines whether a directory/non-directory check should be made.
1074 * No check is made if isdir is set to -1.
1076 * NOTE! The underlying file can still be active with open descriptors
1077 * or if the chain is being manually held (e.g. for rename).
1079 * The caller is responsible for fixing up ip->chain if e.g. a
1080 * rename occurs (see chain_duplicate()).
1083 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1084 const uint8_t *name, size_t name_len,
1085 int isdir, int *hlinkp)
1087 hammer2_inode_data_t *ipdata;
1088 hammer2_chain_t *parent;
1089 hammer2_chain_t *ochain;
1090 hammer2_chain_t *chain;
1091 hammer2_chain_t *dparent;
1092 hammer2_chain_t *dchain;
1099 lhc = hammer2_dirhash(name, name_len);
1102 * Search for the filename in the directory
1106 parent = hammer2_inode_lock_ex(dip);
1107 chain = hammer2_chain_lookup(&parent,
1108 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1111 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1112 name_len == chain->data->ipdata.name_len &&
1113 bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1116 chain = hammer2_chain_next(&parent, chain,
1117 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1120 hammer2_inode_unlock_ex(dip, NULL); /* retain parent */
1123 * Not found or wrong type (isdir < 0 disables the type check).
1124 * If a hardlink pointer, type checks use the hardlink target.
1126 if (chain == NULL) {
1130 if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1133 type = chain->data->ipdata.target_type;
1136 if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1140 if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1146 * Hardlink must be resolved. We can't hold parent locked while we
1147 * do this or we could deadlock.
1149 * On success chain will be adjusted to point at the hardlink target
1150 * and ochain will point to the hardlink pointer in the original
1151 * directory. Otherwise chain remains pointing to the original.
1153 if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1154 hammer2_chain_unlock(parent);
1156 error = hammer2_hardlink_find(dip, &chain, &ochain);
1160 * If this is a directory the directory must be empty. However, if
1161 * isdir < 0 we are doing a rename and the directory does not have
1162 * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1163 * and the directory does not have to be empty.
1165 * NOTE: We check the full key range here which covers both visible
1166 * and invisible entries. Theoretically there should be no
1167 * invisible (hardlink target) entries if there are no visible
1170 if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1171 dparent = hammer2_chain_lookup_init(chain, 0);
1172 dchain = hammer2_chain_lookup(&dparent,
1173 0, (hammer2_key_t)-1,
1174 HAMMER2_LOOKUP_NODATA);
1176 hammer2_chain_unlock(dchain);
1177 hammer2_chain_lookup_done(dparent);
1181 hammer2_chain_lookup_done(dparent);
1187 * Ok, we can now unlink the chain. We always decrement nlinks even
1188 * if the entry can be deleted in case someone has the file open and
1191 * The chain itself will no longer be in the on-media topology but
1192 * can still be flushed to the media (e.g. if an open descriptor
1193 * remains). When the last vnode/ip ref goes away the chain will
1194 * be marked unmodified, avoiding any further (now unnecesary) I/O.
1196 * A non-NULL ochain indicates a hardlink.
1200 * Delete the original hardlink pointer.
1202 * NOTE: parent from above is NULL when ochain != NULL
1203 * so we can reuse it.
1205 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1206 hammer2_chain_delete(trans, ochain, 0);
1207 hammer2_chain_unlock(ochain);
1210 * Then decrement nlinks on hardlink target, deleting
1211 * the target when nlinks drops to 0.
1213 hammer2_chain_modify(trans, &chain, 0);
1214 --chain->data->ipdata.nlinks;
1215 if (chain->data->ipdata.nlinks == 0)
1216 hammer2_chain_delete(trans, chain, 0);
1219 * Otherwise this was not a hardlink and we can just
1220 * remove the entry and decrement nlinks.
1222 * NOTE: *_get() integrates chain's lock into the inode lock.
1224 hammer2_chain_modify(trans, &chain, 0);
1225 ipdata = &chain->data->ipdata;
1227 hammer2_chain_delete(trans, chain, 0);
1233 hammer2_chain_unlock(chain);
1235 hammer2_chain_lookup_done(parent);
1237 hammer2_chain_drop(ochain);
1243 * Given an exclusively locked inode we consolidate its chain for hardlink
1244 * creation, adding (nlinks) to the file's link count and potentially
1245 * relocating the inode to a directory common to ip->pip and tdip.
1247 * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1248 * and returning a new locked chain.
1250 * NOTE! This function will also replace ip->chain.
1253 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1254 hammer2_chain_t **chainp,
1255 hammer2_inode_t *tdip, int nlinks)
1257 hammer2_inode_data_t *ipdata;
1258 hammer2_inode_t *fdip;
1259 hammer2_inode_t *cdip;
1260 hammer2_chain_t *chain;
1261 hammer2_chain_t *nchain;
1265 if (nlinks == 0 && /* no hardlink needed */
1266 (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1269 if (hammer2_hardlink_enable < 0) { /* fake hardlinks */
1273 if (hammer2_hardlink_enable == 0) { /* disallow hardlinks */
1274 hammer2_chain_unlock(chain);
1280 * cdip will be returned with a ref, but not locked.
1283 cdip = hammer2_inode_common_parent(fdip, tdip);
1286 * If no change in the hardlink's target directory is required and
1287 * this is already a hardlink target, all we need to do is adjust
1290 * XXX The common parent is a big wiggly due to duplication from
1291 * renames. Compare the core (RBTREE) pointer instead of the
1295 (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1297 hammer2_chain_modify(trans, &chain, 0);
1298 chain->data->ipdata.nlinks += nlinks;
1305 * We either have to move an existing hardlink target or we have
1306 * to create a fresh hardlink target.
1308 * Hardlink targets are hidden inodes in a parent directory common
1309 * to all directory entries referencing the hardlink.
1311 nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1315 * Bump nlinks on duplicated hidden inode, repoint
1318 hammer2_chain_modify(trans, &nchain, 0);
1319 nchain->data->ipdata.nlinks += nlinks;
1320 hammer2_inode_repoint(ip, cdip, nchain);
1323 * If the old chain is not a hardlink target then replace
1324 * it with a OBJTYPE_HARDLINK pointer.
1326 * If the old chain IS a hardlink target then delete it.
1328 if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1330 * Replace original non-hardlink that's been dup'd
1331 * with a special hardlink directory entry. We must
1332 * set the DIRECTDATA flag to prevent sub-chains
1333 * from trying to synchronize to the inode if the
1334 * file is extended afterwords.
1336 hammer2_chain_modify(trans, &chain, 0);
1337 hammer2_chain_delete_duplicate(trans, &chain,
1338 HAMMER2_DELDUP_RECORE);
1339 atomic_set_int(&chain->flags, HAMMER2_CHAIN_HARDLINK);
1340 ipdata = &chain->data->ipdata;
1341 ipdata->target_type = ipdata->type;
1342 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1350 bzero(&ipdata->uid, sizeof(ipdata->uid));
1351 bzero(&ipdata->gid, sizeof(ipdata->gid));
1352 ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1353 ipdata->cap_flags = 0;
1357 ipdata->iparent = 0; /* XXX */
1358 ipdata->pfs_type = 0;
1359 ipdata->pfs_inum = 0;
1360 bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1361 bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1362 ipdata->data_quota = 0;
1363 ipdata->data_count = 0;
1364 ipdata->inode_quota = 0;
1365 ipdata->inode_count = 0;
1366 ipdata->attr_tid = 0;
1367 ipdata->dirent_tid = 0;
1368 bzero(&ipdata->u, sizeof(ipdata->u));
1369 /* XXX transaction ids */
1371 hammer2_chain_delete(trans, chain, 0);
1375 * Return the new chain.
1377 hammer2_chain_unlock(chain);
1383 hammer2_chain_unlock(chain);
1388 * Cleanup, chain/nchain already dealt with.
1392 hammer2_inode_drop(cdip);
1398 * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1399 * inode while (*chainp) points to the resolved (hidden hardlink
1400 * target) inode. In this situation when nlinks is 1 we wish to
1401 * deconsolidate the hardlink, moving it back to the directory that now
1402 * represents the only remaining link.
1405 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1406 hammer2_inode_t *dip,
1407 hammer2_chain_t **chainp,
1408 hammer2_chain_t **ochainp)
1410 if (*ochainp == NULL)
1417 * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1418 * with an obj_type of HAMMER2_OBJTYPE_HARDLINK. This routine will gobble
1419 * the *chainp and return a new locked *chainp representing the file target
1420 * (the original *chainp will be unlocked).
1422 * When a match is found the chain representing the original HARDLINK
1423 * will be returned in *ochainp with a ref, but not locked.
1425 * When no match is found *chainp is set to NULL and EIO is returned.
1426 * (*ochainp) will still be set to the original chain with a ref but not
1430 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1431 hammer2_chain_t **ochainp)
1433 hammer2_chain_t *chain = *chainp;
1434 hammer2_chain_t *parent;
1435 hammer2_inode_t *ip;
1436 hammer2_inode_t *pip;
1440 hammer2_inode_ref(pip); /* for loop */
1441 hammer2_chain_ref(chain); /* for (*ochainp) */
1445 * Locate the hardlink. pip is referenced and not locked,
1450 lhc = chain->data->ipdata.inum;
1451 hammer2_chain_unlock(chain);
1454 while ((ip = pip) != NULL) {
1455 parent = hammer2_inode_lock_ex(ip);
1456 hammer2_inode_drop(ip); /* loop */
1457 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1458 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
1459 hammer2_chain_lookup_done(parent); /* discard parent */
1462 pip = ip->pip; /* safe, ip held locked */
1464 hammer2_inode_ref(pip); /* loop */
1465 hammer2_inode_unlock_ex(ip, NULL);
1469 * chain is locked, ip is locked. Unlock ip, return the locked
1470 * chain. *ipp is already set w/a ref count and not locked.
1472 * (parent is already unlocked).
1475 hammer2_inode_unlock_ex(ip, NULL);
1478 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1479 /* already locked */
1487 * Find the directory common to both fdip and tdip, hold and return
1491 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1493 hammer2_inode_t *scan1;
1494 hammer2_inode_t *scan2;
1497 * We used to have a depth field but it complicated matters too
1498 * much for directory renames. So now its ugly. Check for
1499 * simple cases before giving up and doing it the expensive way.
1501 * XXX need a bottom-up topology stability lock
1503 if (fdip == tdip || fdip == tdip->pip) {
1504 hammer2_inode_ref(fdip);
1507 if (fdip->pip == tdip) {
1508 hammer2_inode_ref(tdip);
1515 for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1517 while (scan2->pmp == tdip->pmp) {
1518 if (scan1 == scan2) {
1519 hammer2_inode_ref(scan1);
1527 panic("hammer2_inode_common_parent: no common parent %p %p\n",
1534 * Synchronize the inode's frontend state with the chain state prior
1535 * to any explicit flush of the inode or any strategy write call.
1537 * Called with a locked inode.
1540 hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
1541 hammer2_chain_t **chainp)
1543 hammer2_inode_data_t *ipdata;
1544 hammer2_chain_t *parent;
1545 hammer2_chain_t *chain;
1546 hammer2_key_t lbase;
1548 ipdata = &ip->chain->data->ipdata;
1550 if (ip->flags & HAMMER2_INODE_MTIME) {
1551 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1552 atomic_clear_int(&ip->flags, HAMMER2_INODE_MTIME);
1553 ipdata->mtime = ip->mtime;
1555 if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size < ipdata->size) {
1556 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1557 ipdata->size = ip->size;
1558 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1561 * We must delete any chains beyond the EOF. The chain
1562 * straddling the EOF will be pending in the bioq.
1564 lbase = (ipdata->size + HAMMER2_PBUFMASK64) &
1565 ~HAMMER2_PBUFMASK64;
1566 parent = hammer2_chain_lookup_init(ip->chain, 0);
1567 chain = hammer2_chain_lookup(&parent,
1568 lbase, (hammer2_key_t)-1,
1569 HAMMER2_LOOKUP_NODATA);
1572 * Degenerate embedded case, nothing to loop on
1574 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1575 hammer2_chain_unlock(chain);
1578 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
1579 hammer2_chain_delete(trans, chain, 0);
1581 chain = hammer2_chain_next(&parent, chain,
1582 lbase, (hammer2_key_t)-1,
1583 HAMMER2_LOOKUP_NODATA);
1585 hammer2_chain_lookup_done(parent);
1587 if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size > ipdata->size) {
1588 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1589 ipdata->size = ip->size;
1590 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1593 * When resizing larger we may not have any direct-data
1596 if ((ipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
1597 ip->size > HAMMER2_EMBEDDED_BYTES) {
1598 ipdata->op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
1599 bzero(&ipdata->u.blockset, sizeof(ipdata->u.blockset));